The seasonal dynamics of Arctic surface
hydrology in permafrost environments
Anna Maria A. Trofaier
St Edmund’s College
Scott Polar Research Institute
University of Cambridge
This dissertation is submitted for the degree of
Doctor of Philosophy
2014

Declaration
This dissertation is the result of my own work and includes nothing which is the
outcome of work done in collaboration except where specifically indicated in the
text and Acknowledgments. It has not been submitted in whole or in part for a
degree at any other university. This thesis is 179 pages long, including text, illustra-
tions, tables and references.
Anna Maria Trofaier
01 April, 2014

To my late father
who was so proud of my academic ambitions but was denied the possibility to ever see me graduate.
And to my husband Tim
who believed in me and supported me every step of the way.

Acknowledgments
First of all, I would like to thank my supervisor Gareth Rees for his enthusiasm when I pro-
posed to undertake this PhD project, as well as his support throughout my time at the Scott
Polar Research Institute (SPRI).
At the beginning of my studies I was privileged to secure a DOC-fFORTE [Women in
Research and Technology] Fellowship of the Austrian Academy of Sciences (ÖAW). I would
like to thank the ÖAW, without its support this PhD research would not have been possible.
Furthermore, St Edmund’s College and the Geography Department have both generously sup-
ported me financially on several conference trips, as did SPRI.
My last year has been tumultuous on a personal level and I would like to say a big thank
you to my tutor at college, Anna Gannon, who was incredibly supportive of both me and Tim.
I would also like to acknowledge the University of Cambridge’s Lundgren Award and the St
Edmund’s College hardship fund.
My husband Tim, who followed me to Cambridge, not only is he the most amazing and
courageous person I know, but he also proofread the entire thesis. Thank you so much for
taking the time!
During the course of my PhD I met many amazing and inspirational people: My conference
buddies, my SPRI colleagues and all the friends I met at the University Centre in Svalbard
(UNIS) - one of the best experiences I have had in my life. Thank you for everyone’s friendship.
Meeting people who love their work is both encouraging and inspirational. It keeps reminding
me of what a great field we work in.
Thank you to all the PhD students and Post-docs at SPRI: Katja, Evan and TJ. Joe, Craig and
Christine. Ruth and Steve. Allen - for generally being a star. Ali - for being an amazing role
model. Thank you also to my colleagues in Vienna: Michael Hornácˇek who was a great help
with all things Java related, and Stefan Schlaffer who was my IDL helpdesk.
There are many more people I would like to thank for their friendship and support during
my time at Cambridge, too many to list. However, most importantly I would like to thank
Annett Bartsch at the Vienna University of Technology. As my MSc supervisor she introduced
me to microwave remote sensing. She is my mentor and my friend, and I am grateful she took
me under her wing.
Lastly, I would like to thank Professor Johannes Ortner and Professor Alan Heavens from
my Astrophysics days. I would not have been able to carry out this PhD without their encour-
v
agement and support.
Abstract
Climate-induced landscape evolution is resulting in changes to biogeochemical and hydrologi-
cal cycling. In the Arctic and sub-Arctic permafrost zones, rising air temperatures are warming,
and in some regions even thawing, the frozen ground. Permafrost is a carbon sink. The thermal
state of the ground therefore has important implications on carbon exchange with the atmo-
sphere. Permafrost thaw mobilises previously sequestered carbon stocks, potentially turning
these high latitude regions into a net carbon source.
Borehole temperature and active layer depth measurements are the traditional means for
monitoring permafrost, however these point measurements cannot easily be extrapolated to
the landscape-scale; Earth Observation (EO) data may be used for such purposes. It is widely
recognised that changes in the thermal state of permafrost may be associated with longterm
changes in surface hydrology. As the ground shifts from a frozen to a thawed state, Arctic
lakes display changes in surface extent. Therefore, it has become common practice to explore
lake dynamics, using these as indicators of permafrost change; dynamics being the keyword.
Surface hydrology is a dynamic process. Discharge studies in the Arctic and sub-Arctic regions
are associated with flashy hydrographs. Currently, however, remote sensing of permafrost lake
change is done on the scale of decades without explicitly taking seasonality and rapid hydrolog-
ical phenology into consideration. To examine the seasonal changes in Arctic surface hydrology
on the landscape scale high temporal resolution data are necessary. Synthetic aperture radar
instruments are exemplary for such a task.
The PhD research focuses on establishing operational techniques for mapping open surface
water using synthetic aperture radar data, investigating straightforward raster classification
methods and exploring their feasibility by undertaking map accuracy and sensitivity studies
(chapter 3). The results are then used to justify error propagation when developing an auto-
mated procedure that creates temporal composites of water body extent. These temporal water
body classifications are the main EO product used to identify and image seasonal surface water
change in Arctic permafrost environments (chapter 4). Furthermore, a terrain-based hydrolog-
ical study is undertaken to explore the context of the detected changes and possible links to
relief and stream channel network (chapter 5).
The aim of this PhD is to demonstrate a new method of dynamic monitoring using the Euro-
pean Space Agency’s Envisat Advanced Synthetic Aperture Radar, recommending its incorpo-
ration in longterm lake change studies. Technical feasibility is explored, the inherent trade-off
vii
between spatial and temporal resolution discussed. An automated surface water change de-
tection algorithm is developed and its applicability to monitoring spring floods is assessed;
noting possible modifications to the drainage system given present-day land-use and land-
cover changes that are taking place in the study area, the hydrocarbon-rich Yamalo-Nenets
Autonomous District in the North of West Siberia (chapter 6). The key significance of this
research is to improve the current knowledge of Arctic lake change by including spring flood
events and seasonality in the equation. Therefore, it is strongly believed that this research is of
benefit to the entire permafrost community.
Contents
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xx
List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiii
List of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxv
1 Introduction 1
1.1 Rationale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Research aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Structure of thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Author contributions to peer reviewed papers . . . . . . . . . . . . . . . . . . . . 6
2 Permafrost and Arctic surface hydrology 7
2.1 Warming permafrost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Active layer dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Subsurface water flow and taliks . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Arctic water bodies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.4 The West Siberian Lowland . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 RS of the Arctic tundra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.1 Optical Imagery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Radar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.3 Arctic lakes - The winter perspective . . . . . . . . . . . . . . . . . . . . . . 20
2.2.4 Snow detection and freeze/thaw cycles . . . . . . . . . . . . . . . . . . . . 21
2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Mapping inundation 25
3.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.1 Envisat ASAR WS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.2 ASAR WS data pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1.3 ALSO PRISM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.4 TerraSAR-X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.5 Data availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Water body detection method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
ix
3.2.1 Open surface water, a surface scatterer . . . . . . . . . . . . . . . . . . . . 34
3.2.2 The effects of aquatic, waterborne and marginal vegetation . . . . . . . . . 38
3.2.3 The effects of wind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 Case 1: Local scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4 Feasibility studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.4.1 Optimum threshold sensitivity test . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.2 Data fusion I: Optical/Radar (C-band) accuracy test . . . . . . . . . . . . . 46
3.4.3 Data fusion II: Radar (C-band/X-band) accuracy test . . . . . . . . . . . . 50
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.6 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4 Monitoring seasonal inundation 61
4.1 Case 1: Local Scale, cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 Spatial analysis of changes in water body extent . . . . . . . . . . . . . . . . . . . 62
4.2.1 Temporal composite water body classification . . . . . . . . . . . . . . . . 62
4.2.2 Temporal composite quality control . . . . . . . . . . . . . . . . . . . . . . 63
4.2.3 Water body dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.4 Wind biases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Water body statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3.1 TerraSAR-X and ASAR WS: Limnicities . . . . . . . . . . . . . . . . . . . . 71
4.3.2 TerraSAR-X and ASAR WS: Size-frequency comparison analysis . . . . . 71
4.3.3 ASAR WS: Early, mid and late season size-frequency distribution . . . . . 76
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.1 Change detection analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.2 Lake size-abundance distributions . . . . . . . . . . . . . . . . . . . . . . . 82
4.5 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5 Modelling stream channels 87
5.1 Case 1: Local scale, cont. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.1 Digital Elevation Models - types and features . . . . . . . . . . . . . . . . . 89
5.2.2 ASTER GDEM2 and ancillary materials . . . . . . . . . . . . . . . . . . . . 89
5.3 ASTER GDEM2 validation study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3.1 GDEM2 and map spot-heights distribution test . . . . . . . . . . . . . . . 91
5.3.2 GDEM2 and map spot-heights direct comparison . . . . . . . . . . . . . . 93
5.3.3 GDEM2 and DUE Permafrost DEM comparison . . . . . . . . . . . . . . . 95
5.3.4 Discussion: GDEM2 validation . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.4 GIS-based surface flow modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.4.1 Flow accumulation and channel network analysis . . . . . . . . . . . . . . 101
5.4.2 Topographic Wetness Index analysis . . . . . . . . . . . . . . . . . . . . . . 107
5.4.3 Discussion: Surface water modelling . . . . . . . . . . . . . . . . . . . . . . 115
5.5 Empirical lake basin analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.5.1 Background and Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.6 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6 Development of SWDA 127
6.1 Case 2: Regional scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.1.1 Subregion 1 - Taz river basin near Mangazeya (Ìàíãàçeß) . . . . . . . . . 130
6.1.2 Subregion 11a - Yamal peninsula . . . . . . . . . . . . . . . . . . . . . . . . 130
6.1.3 Subregion 11b - Gydan peninsula . . . . . . . . . . . . . . . . . . . . . . . 131
6.2 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.2.1 ALANIS Methane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.2.2 QuikSCAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.3 Development of the Surface Water Dynamics Algorithm . . . . . . . . . . . . . . 135
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.5.1 SWDA quality control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.5.2 Assessment of floods detected by SWDA . . . . . . . . . . . . . . . . . . . 141
6.5.3 The time-slice conundrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.6 Summary and conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7 Conclusions 147
7.1 Research synthesis and achieved objectives . . . . . . . . . . . . . . . . . . . . . . 148
7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.2.1 A dynamic thresholding algorithm . . . . . . . . . . . . . . . . . . . . . . . 151
7.2.2 Case 3: Pan-Siberian scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7.2.3 Field work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
7.3 Significance of this research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Epilogue 156

List of Figures
2.1 Section of the ground showing the relationship between permafrost, active layer
and taliks, taken from French (2007), courtesy of Ferrians et al. (1969) and the
United States Geological Survey 2.1(a). Schematic cross-section of a thermokarst
lake 2.1(b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Spot medallions on uplands (left panel) and tussock tundra landscape (right
panel). Photographs taken by A.M. Trofaier on 2 and 4 July 2012. . . . . . . . . . 14
2.3 Photographs of wetland tundra: A transect for Greening of the Arctic project
spectrometer measurements with dwarf birch (a), alas with thermokarst lakes
and Polar willow on the slopes of the depression, tussock tundra on the uplands
(b, top panel), lake with grasses along its shoreline and bordering willow thickets
in the background (b, bottom panel), a type of bog moss growing in an inundated
area (c) and water sedges along a shoreline of a lake (d). Photos taken by A.M.
Trofaier during the period of 1 - 3 July 2012. . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Timeline of SAR sensors. The y-axis shows decreasing resolution. Diagram
taken from Pope et al. (2014). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5 Graphs illustrating the effect of a phase change from water to ice on its dielectric
constant (left panel) and the backscattering coefficient (right panel) as presented
by Han and Lee (2013), who carried out a C-band radar experiment. . . . . . . . 19
3.1 Flow chart of the pre-processing chain undertaken at the Vienna University of
Technology (Sabel 2007) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 ASAR WS (HH polarisation) Level 1b coverage over the period December 2004
- February 2009 adapted from Sabel et al. (2011) . . . . . . . . . . . . . . . . . . . 32
3.3 Number of ASAR WS (HH polarisation) scenes covering the Arctic available for
the summer months July and August in 2007 and 2008, as presented in Bartsch
et al. (2012). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Sketch of a signal on a flat, unknown background of magnitude B, adapted from
Sivia and Skilling (2006). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
xiii
3.5 Bimodal backscatter distribution for a tundra region (top panel) and the corre-
sponding ASAR WS image from 20 August 2007 (bottom panel), as also shown
in (Bartsch et al. 2012). µdm is the mean of the dominant mode and σdm is its
standard deviation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.6 Timeseries of Envisat ASAR WS data over a tundra landscape, showing the
loss of the bimodal backscatter distribution due to increased open water sur-
face roughness on a windy day (25.07.2007) (Bartsch et al. 2012). The grey level
stretch of the backscatter distribution shown in the top panel goes from zero
(white) to -21 dB (black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.7 Local Scale study areas. ROI 1, ROI ALOS and ROI Terra partially overlap and
were used for the feasibility studies. ROI 2 exhibited strong wind biases. . . . . . 43
3.8 Optimum threshold for the 12 lakes (each depicted by a different colour). A
spline was fitted to the daily threshold mean. The standard deviation of the
entire threshold set was found to be 1.22 dB. . . . . . . . . . . . . . . . . . . . . . 46
3.9 Normalised ASAR WS image from 01 July 2007 (top panel). Particles identified
by sensitivity study (red) for a 1 dB bin threshold range from -10 to -15 dB (bot-
tom, left to right). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.10 An unsupervised PRISM classification is used as a land mask. The surroundings
are masked out with the value zero, revealing only the backscatter values of the
ASAR WS image that should be associated with surface water. . . . . . . . . . . . 49
3.11 Bimodal backscatter distribution of Gamma and mean filtered TerraSAR-X scene
from 13 August 2008. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.12 Spatial sub-scene of TerraSAR-X image a) after applying a Gamma filter to the
radar image, b) classification of Gamma filtered image using a -17 dB thresh-
old, c) after applying a 9 × 9 median filter to the Gamma filtered image and d)
classification of median filtered image using the manually determined optimum
threshold of -16.76 dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.13 Bimodal backscatter distribution of ASAR WS scene from 13 August 2008. . . . . 56
3.14 From left to right: TerraSAR-X classification (t = -16.76 dB), overlain by ASAR
WS classification (t = -14 dB) (top panel); overlain by ASAR WS classification (t
= -13 dB) and (t = -12 dB) (bottom panel). . . . . . . . . . . . . . . . . . . . . . . . 58
4.1 Pixel counts for each classification period in each year as a function of Envisat
ASAR WS measurements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.2 Cumulative change in pixel count due to water pixel loss and gain for ROI 1, as
presented in Trofaier et al. (2012). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Maps of pixel loss (drainage) and pixel gain (expansion) for classifications I,II
and III of region ROI 1 in 2007, 2008 and 2009. . . . . . . . . . . . . . . . . . . . . 68
4.4 Number of ASAR WS measurements and their associated pixel count for ROI 2
in 2008, as presented in Trofaier et al. (2012). . . . . . . . . . . . . . . . . . . . . . 69
4.5 A base 10 log-log plot of the lake frequency (dL, the number of lakes in a given
size interval) as a function of lake area for TerraSAR-X (blue) and ASAR WS
(red). Linear regression lines are shown for TerraSAR-X and ASAR WS also in
blue (left panel) and red (right panel), respectively. . . . . . . . . . . . . . . . . . . 74
4.6 The total lake area within each base 10 log width lake size interval for TerraSAR-
X (blue) and ASAR WS (red) is shown in panel (a). The difference in total water
body extent per lake size interval between these two sensors is depicted by the
black line in panel (b). The number of water bodies in each lake size interval is
given in panel (c), where the difference in the total number of water bodies, Ndiff,
is 1433. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.7 Histograms of water body extent for the early (A), mid (B and C) and late (D)
season periods in 2007 (a, b, c, d), 2008 (e, f, g, h) and 2009 (i, j, k, l). . . . . . . . . 77
4.8 Size distributions of water body extent for the early (A), mid (B and C) and late
(D) season periods in 2007 (a, b, c, d), 2008 (e, f, g, h) and 2009 (i, j, k, l). . . . . . . 78
4.9 Vegetated lake basin with arctophila emerging from the lake surface. Photo-
graph taken on 03 July 2012 by AM Trofaier. . . . . . . . . . . . . . . . . . . . . . . 81
4.10 Arctophila emerging from the surfaces of two lakes. Photographs taken on 01
July 2012 by AM Trofaier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.1 Example of a spike anomaly. Surface plot showing unrealistically high elevation
values (left panel) associated with low stereo coverage (right panel). . . . . . . . 90
5.2 The probability density functions for both (a) ASTER GDEM2 and (b) the Soviet
map spot-height data sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 The empirical cumulative distribution functions of both ASTER GDEM2 and the
map spot-height data sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4 Boxplot of the ASTER GDEM2 and Soviet map elevation sample data. The
means are not significantly different, although on average the GDEM2 sample
data are slightly higher than the map sample data. The outlier affecting the bias
is also clearly visible as the individual point. . . . . . . . . . . . . . . . . . . . . . 95
5.5 ASTER GDEM2 points as a function of map spot-heights. The red error bar in-
dicates the confidence level derived from the standard deviation (σ) of the linear
regression fit. The green symbols are those values classified as extreme differ-
ence values, deviating from the linear fit by ±6σ. The black line is the 1:1 corre-
spondence line. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.6 The difference in GDEM2 and the map spot-heights (i.e. the sample error, e(x))
versus the elevation taken from GDEM2. Linear regression fit given by dashed
blue line. Low R2 value was inspected by carrying out a one-tailed t-test showing
that the slope of the regression line is significantly different from zero. . . . . . . 97
5.7 a) D-DEM cover of the Yamal peninsula, the extent of the study area is illustrated
by the green frame, and b) D-DEM, c) GDEM2 and d) difference DEM of the
study area (WGS84 Lat/Lon). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.8 (a) GDEM2 and D-DEM (WGS84 UTM zone 42N) illustrating the outline of ele-
vation profiles in green (A - A’) and blue (B - B’) and (b) the elevation profile A
- A’ and (c) the elevation profile B - B’ for both GDEM2 and D-DEM in grey and
black, respectively. Colour stretch according to previous figure 5.7. . . . . . . . . 99
5.9 TerraSAR-X water body classification with outlines of the Mordy-Yakha and
Se-Yakha catchments (left panel) and legend (top right panel). Yamalo-Nenets
Autonomous District borders according to Stolbovoi and McCallum (2002) with
Vaskiny Dachi area extent indicator (bottom right panel). . . . . . . . . . . . . . . 104
5.10 Derived stream channel network, with associated Strahler segment link order.
The ASTER GDEM2 is used as the background layer. Anomalously high pixel
values were not removed; it is known from the literature that the maximum
elevation above sea level in this area is just under 60 m a.s.l. (Leibman et al.
2012). 0.07% of the pixels were higher than 60 m a.s.l., resulting in the depicted
colour stretch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.11 The frequency of stream segments for each drainage basin . . . . . . . . . . . . . 106
5.12 Slope gradient of the ASTER GDEM2. . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.13 Map of TWI for the Se-Yakha and Mordy-Yakha basins (top panel), with spring
floods depicted for 2007, 2008 and 2009 from left to right (bottom panel). . . . . . 112
5.14 Histograms of normalised TWI for the Se-Yakha basin (a), the Mordy-Yakha
basin (b) and both catchment areas combined (c). . . . . . . . . . . . . . . . . . . . 113
5.15 The empirical distributions functions of the TWI for the Se-Yakha (blue) and
Mordy-Yakha (green) basins. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.16 Mean TWI as a function of slope gradient (a), as well as drainage density for
each stream channel order also as a function of slope gradient (b) for the Se-
Yakha (black) and Mordy-Yakha (green) basins. Spline fit for drainage density
vs slope. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.17 Active layer detachment slide (left panel) and landslide body (right panel) in the
Vaskiny Dachi study area. Willow shrubs are good indicators of landslide ac-
tivity, as tall thickets of these willows are found on old landslide bottoms due
to greater nutrient availability (Walker et al. 2011). Approximate landslide body
and willow thicket outlined with dashed and solid black lines, respectively. Pho-
tographs taken by A.M. Trofaier on 4 July 2012. . . . . . . . . . . . . . . . . . . . . 117
5.18 SPOT image of the Vaskiny Dachi area, overlain by 2007 difference map I. . . . . 120
5.19 Changes in mean backscatter values over the spring/summer season 2007 for
selected water bodies (left panel) and a surrounding bufferzone (right panel). . . 121
5.20 SPOT image zoomed to the Bovanenkovo gas field, overlain by 2007 difference
map I (top panel) and by the modelled stream channel network (top right panel).
False colour Quickbird-2 scene of the Bovanenkovo gas field compound from
Forbes and Kumpula (2009) (bottom panel). . . . . . . . . . . . . . . . . . . . . . . 123
6.1 Permafrost extent in Eurasia. Distribution of continuous and discontinuous per-
mafrost according to Brown et al. (1997). The black frame indicates the extent
of figure 6.2. The borders of Russia’s administrative regions are outlined in
grey (Stolbovoi and McCallum 2002); the Yamalo-Nenets Autonomous District
is highlighted by a line-fill. The magenta frames represent the ALANIS methane
local wetlands product coverage across Eurasia with their associated identifica-
tion numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.2 Map of the Yamalo-Nenets Autonomous District of the Russian Federation (NW
Siberia) showing the location of the three study areas: Subregion 1 (Taz), subre-
gion 11a (Yamal), subregion 11b (Gydan). Permafrost extent according to Brown
et al. (1997). The borders of Russia’s administrative regions are outlined in grey
(Stolbovoi and McCallum 2002). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.3 Boxplot of QuikSCAT data for each study area. The median is indicated by the
thick black line (left panel). MODIS image (courtesy of NASA Earth Observa-
tory) of the Yamal peninsula on 29 June 2007. Frozen lakes on both the Yamal
and Gydan peninsulas are clearly visible (right panel). . . . . . . . . . . . . . . . 134
6.4 Lake in a partially drained basin on the Yamal peninsula (left panel) and lake
with arctophila emerging from the water surface and growing at its margins
(right panel). Photos taken by A.M. Trofaier on 1 and 3 July 2012, respectively. . 135
6.5 Histogram of changes in water area in subregion 1. Frequency given on a base
10 logarithmic scale. A strip chart of each change is given just above the x-axis. . 138
6.6 (a) Comparison of classifications of water body extent for subregion 1 for June
and July 2007. (b) SWDA classification for July-August 2007. (c) Radar timeseries
of disappearing open water body June-July 2007. . . . . . . . . . . . . . . . . . . . 139
6.7 Minimum (top), maximum (middle) and mean (bottom) number of Envisat ASAR
WS acquisitions used to classify water bodies for each 10-day composite over the
summer period July - August for each study area. . . . . . . . . . . . . . . . . . . 140
6.8 Surface hydrology classification of study area 11a. Seasonal changes in water ex-
tent for July-August 2007 are shown in red. The Obskaya-Bovanenkovo Railway
was partially traced from GoogleEarth (gaps were left where the spatial reso-
lution did not allow the identification of the railway tracks). In the north, the
southern part of the Sannikov (2012) region is outlined. . . . . . . . . . . . . . . . 141
6.9 Map of the YNAO (black frame) overlain by the region investigated by Smith
et al. (2005). Red squares indicate a disappearing lake, one of which is within
subregion 1 in the discontinuous zone. According to the figure presented in
Smith et al. (2005) the location of this disappearing lake seems to coincide with
the seasonal lake shown in figure 6.6c. Magenta frames are the ALANIS methane
subregions. Permafrost extent uses the same colour scheme as in figure 6.1. . . . 144
List of Tables
2.1 The dielectric constant of ice and water for C-band sensors. . . . . . . . . . . . . . 20
3.1 Characteristics of the ASAR WS processing grid (Sabel et al. 2011). . . . . . . . . 29
3.2 Geographical coordinates for ROI 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Geographical coordinates for ROI ALOS. . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Geographical coordinates for ROI Terra. . . . . . . . . . . . . . . . . . . . . . . . . 44
3.5 ASAR WS images used for sensitivity study. . . . . . . . . . . . . . . . . . . . . . 46
3.6 Confusion matrix comparing ASAR WS ISODATA classification with PRISM ref-
erence image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.7 Confusion matrix comparing ASAR WS threshold classification with PRISM ref-
erence image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.8 Confusion matrix comparing ASAR WS threshold classification (t = -14 dB) with
TerraSAR-X threshold classification as reference image. . . . . . . . . . . . . . . . 55
3.9 Confusion matrix comparing ASAR WS threshold classification (t = -13 dB) with
TerraSAR-X threshold classification as reference image. . . . . . . . . . . . . . . . 55
3.10 Confusion matrix comparing ASAR WS threshold classification (t = -12 dB) with
TerraSAR-X threshold classification as reference image. . . . . . . . . . . . . . . . 56
4.1 Classifications and respective time periods for each year. . . . . . . . . . . . . . . 63
4.2 Main statistical parameters of ASAR WS and TerraSAR-X water body data from
13 August 2008. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Main statistical parameters of ASAR WS and TerraSAR-X water body data from
13 August 2008. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.1 DEM sample statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Se-Yakha stream channel statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.3 Mordy-Yakha stream channel statistics . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4 TWI statistics at the location of the flooded areas identified from the 2007 differ-
ence map I. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.1 Main parameters of study areas and their lakes in 2007 . . . . . . . . . . . . . . . 137
xix
6.2 Number of Envisat ASAR WS acquisitions used to classify water bodies over the
first time period 1 - 10 July 2007 (SWDA class 1). . . . . . . . . . . . . . . . . . . . 139
List of Abbreviations
ALANIS Atmosphere-Land Interaction Study
ALD Active Layer Depth
ALOS Advanced Land Observing Satellite
ASAR Advanced Synthetic Aperture Radar
ASTER Advanced Spaceborne Thermal Emission and Reflection Radiometer
AWI Alfred Wegener Institute for Polar and Marine Research
BP Before Present
CALM Circumpolar Active Layer Monitoring
DEM Digital Elevation Model
DLR German Aerospace Centre
DN Digital Numbers
DUE Data User Element
DZAA Depth of Zero Annual Amplitude
ECDF Empirical Cumulative Distribution Function
ERS European Remote Sensing (satellite)
ESA European Space Agency
ETM+ (Landsat) Enhanced Thematic Mapper Plus
GDEM2 Global Digital Elevation Model version 2
GIS Geographic Information System
GLWD Global Lakes and Wetlands Database
GM Global Mode
GtC Gigatons of carbon
GWP Global Warming Potential
IDL Interactive Data Language
IPCC Intergovernmental Panel on Climate Change
ISODATA Iterative Self-Organising Data Analysis Technique Algorithm
JAXA Japan Aerospace Exploration Agency
K-S Kolmogorov-Smirnov (test)
LDD Local Drainage Direction
LST Land Surface Temperature
MAAT Mean Annual Air Temperature
MAGST Mean Annual Ground Surface Temperature
MMU Minimum Mapping Unit
MODIS Moderate-Resolution Imaging Spectroradiometer
MSS (Landsat) Multispectral Scanner
NAN Not A Number
NASA National Aeronautics and Space Administration
NDVI Normalised Difference Vegetation Index
NEST Next ESA SAR Toolbox
PAGE21 Changing Permafrost in the Arctic and its Global Effects in the 21st Century
PALSAR Phased Array type L-band Synthetic Aperture Radar
PNPT Poluy-Nadym-Pur-Taz
PRISM Panchromatic Remote-sensing Instrument for Stereo Mapping
QuikSCAT Quick Scatterometer
RMSE Root Mean Square Error
ROI Region Of Interest
RS Remote Sensing
SAR Synthetic Aperture Radar
SPOT Satellite Pour l’Observation de la Terre
SRTM Shuttle Radar Topography Mission
STSE Support To Science Element
SWE Snow Water Equivalent
TDR Time-Domain Reflectometer
TM (Landsat) Thematic Mapper
TTOP Temperature at the Top Of Permafrost
TWI Topographic Wetness Index
UTM Universal Transverse Mercator
WS Wide Swath
WSL West Siberian Lowland
YNAO Yamalo-Nenets Autonomous Okrug (District)

List of Variables
z Active layer depth
κ Cohen’s kappa coefficient
′r Dielectric constant
e(x) Error in sample elevation values of two datasets
λ Electromagnetic wavelength
hGDEM2(x) Elevation values for GDEM2 data
hmap(x) Elevation values for Soviet map data
l Latitude
θi Local incidence angles
N Number of observations
k Pareto cut-off value
γ Pareto shape parameter
z˙(τ) Rate of thawing front descent through the active layer (where τ is time)
A Radar scattering area
β0 Radar brightness
σ0 Radar backscatter coefficient
R2 Regression coefficient
t Threshold value for radar water body delineation
ϑ(z) Volumetric fraction of soil moisture (as a function of depth)

Chapter 1
Introduction
The Arctic circle (66° 33’ 44” N) is often depicted as the border between frozen and unfrozen
lands. This notion is not entirely true. Vast areas of the North, either side of the Arctic cir-
cle, are underlain by perennially frozen ground known as permafrost. In fact, approximately
24% of the Northern hemisphere is permafrost territory (United Nations Environment Pro-
gramme 2007). Permafrost is found at high latitudes as well as in mountain regions at high
altitudes. Permafrost landscapes are fragile. They are sensitive to changes in the climate and
therefore, given the modern-day trend in air temperatures, the warming of the atmosphere is
warming the frozen ground which in turn is contributing to an actively evolving landscape
(Intergovernmental Panel on Climate Change 2013).
The rising air temperatures are altering the geothermal regime of the ground. The warming,
and consequently possible thawing, of the frozen ground is associated with the melting of
ground ice within. This is of course a simplified statement but extensive research across the
Earth’s permafrost environments has shown a correlation between these rising air temperatures
and permafrost warming. Across Northern Alaska increasing permafrost temperatures (by up
to 2°C to 3°C at borehole depths of 20 m) have been measured since the 1980s (Forster et al.
2007). There is of course regional variability in these warming trends, and Romanovsky et al.
(2010b) discuss how these may be attributed to different meteorological conditions, such as
snow fall and its regional distribution. Iijima et al. (2010) observed rapid warming of both the
upper layers of the frozen ground in Eastern Siberia, discussing its warming due to early snow
cover, greater snow depth and therefore more spring melt waters. These factors are not to be
underestimated - spring meltwaters play a key role in the seasonal dynamics of Arctic surface
hydrology. Across Russia, similarly, borehole temperatures increased over the last 30 years
by a magnitude of 0.5°C to 2°C at the depth of zero annual amplitude DZAA (Romanovsky
et al. 2010a), the depth at which there is no measurable seasonal fluctuation in temperature
(Riseborough 1990).
In mountainous areas of the Arctic changes in meteorological conditions have led to ex-
treme avalanche events within the last few years (Eckerstorfer and Christiansen 2012) and it
can be expected that further slope instabilities arising from warming permafrost conditions
1
Chapter 1. Introduction
will lead to increased geohazards, such as avalanches, land and glacier slides (e.g. Kääb (2008)
and Blikra and Christiansen (2014)). In Northern regions, permafrost degradation is leading
to deterioration of infrastructure with implications on cold climate engineering (e.g. Stephani
et al. (2014)).
However, crucially, permafrost decay is mobilising old carbon stocks as the ground shifts
from its frozen to unfrozen state (Schuur et al. 2009, McGuire et al. 2009). These old carbon
stocks amount to over 1000 Gigatons of carbon (GtC). In fact, Schuur et al. (2008) estimate soil
carbon content of the Northern hemisphere to amount to 1672 petagrams of carbon (PgC), of
which 1024 PgC are considered vulnerable to permafrost thaw and its associated mobilisation,
lying within the topmost 3 m layers of ground. Current efforts are centred on identifying the
processes involved in, quantifying and modelling the removal of these C pools (MacDougall
et al. 2012). Flessa et al. (2008) find that both the distribution of permafrost and the thick-
ness of the active layer control the strong methane fluxes that are typically measured in tundra
wetland ecosystems. In particular, Arctic thermokarst lakes are associated with high carbon
emissions both in the form of carbon dioxide and methane. The latter being the more domi-
nant gas (e.g. Zimov et al. (1997), Walter et al. (2006, 2007) and Edwards et al. (2009)), while at
the same time also being the more potent gas compared to CO2, given it has a Global Warming
Potential (GWP) of 84 and 28 over a time horizon of 20 and 100 years respectively (Intergov-
ernmental Panel on Climate Change 2013). Models have shown that increased atmospheric
carbon exchange is leading to positive permafrost-carbon feedbacks with global importance.
Schneider von Deimling et al. (2012) predict that by the beginning of the 24th century 50% of
the carbon stocks stored in the upper soil layers (at 3 m depth) would potentially have been
released to the atmosphere. This amounts to approximately 600 GtC. Furthermore, it is yet
unclear what effect changes in surface hydrology have on the soil removal processes of carbon.
It has been shown by Lupascu et al. (2013) that, despite higher ecosystem respiration, wetter
conditions are associated with an increase in CO2 sink due to their dampening effect on the
removal of old carbon stocks, while on the other hand Myers-Smith et al. (2007) show methane
emissions increase with wetness conditions. This highlights the need to enhance the current
understanding of all processes, including surface hydrology, associated with changes to the
global carbon budget and the potential implications on the climate system.
Permafrost studies, once mostly observed in the context of cold climate engineering, are
moving increasingly into the focus of scientific research due to the role permafrost plays in
the climate system. Surface hydrology, and shifts associated with a warming climate, are of
significance to permafrost research. Assessing novel methods for monitoring surface hydrol-
ogy are therefore crucial to the understanding of complex links driving the warming Arctic
environment.
2
Chapter 1. Introduction Rationale
1.1 Rationale
As a result of the uncertainties associated with surface hydrological change in permafrost en-
vironments, the need to explicitly monitor changes in surface waters becomes prevalent. Satel-
lite and Earth Observation (EO) methods are established techniques for detecting large-scale
(landscape-scale) changes. With respect to water body monitoring, satellite images are used
to monitor the changes in permafrost thaw lake extent over the time scale of decades. While
this type of research is becoming ever more common practice and has developed into a field of
ongoing research interest, the focus has been on classifying lakes and comparing their extent
on an image to image basis. Change detection analyses are a common tool for determining
trends in geomorphological processes, such as thermokarst - a process associated with ground-
ice melt (see chapter 2). However, a crucial aspect of water body change has been neglected
entirely in these studies - until now. To date, no attempts have been made to identify and quan-
tify seasonal surface water changes by imaging these frequently over the course of the spring
season. The term rapid hydrological phenology is coined here, to emphasise the extreme seasonal-
ity of surface hydrology in the Arctic associated with flood events and flashy hydrographs; and
to allude to the fact that inter-annual comparison of satellite images needs to be done within
the same reference period, using ancillary meteorological and hydrological data as well as the
proposed method of frequent SAR flood dynamics imaging. Although the aforementioned
studies on thaw lake change use optical data taken within the summer months to establish the
changes in water body extent, cloud cover and satellite re-visit periods limit the availability of
images. Often, images early in the spring season are compared to those later in the summer
without taking hydrological activity into consideration. In a region where spring floods present
an important component of the hydrological year the demand for monitoring the dynamics of
seasonal inundation at the landscape scale is palpable. It is, therefore, imperative to rectify this
current absence.
The assessment of satellite data that may be used to identify the extent and runoff pat-
terns of spring floods and the development of a monitoring mechanism are crucial to modern
day understanding of permafrost surface hydrology. The free access to extensive archives of
satellite data through the European Space Agencey (ESA) as well as the National Aeronau-
tics and Space Administration (NASA) has become vital. Most importantly, access to frequent
synthetic aperture radar (SAR) timeseries taken by ESA’s Envisat advanced synthetic aperture
radar (ASAR) sensor enables a seasonal remote sensing analysis to be undertaken.
1.2 Research aim
Keeping the above rationale and its climate context in mind, the aim of this research is to im-
prove the understanding of Arctic spring flood events by establishing a new method to monitor
their spatial patterns periodically. The importance of monitoring intra-annual changes in Arc-
tic surface water extent is illustrated by developing a suitable monitoring scheme that takes
3
Research aim Chapter 1. Introduction
advantage of high periodic mapping of water bodies using active microwave data. Given the
high data volumes that are created by frequent SAR data, a straightforward image to image
analysis is not a plausible option. Therefore, the first goal of this PhD is to develop a more ef-
fective and computationally efficient method. Initially, the key outcome will be the production
of temporal composite classifications and the development of a new approach to change detec-
tion analysis using these temporal composites. Hence, the motivation of this PhD research has
to be seen, not merely in the context of permafrost surface hydrology, but also within its EO
reference frame, as a new approach to change detection analysis.
To support the analysis of permafrost lake change in the warming Arctic, this PhD research
suggests a method for optimising such an analysis by including information on seasonally in-
undated areas; their extent, their timings and patterns of their sequential summer retreat. Ini-
tially, this is done for a small area on the Yamal peninsula, and is followed up a regional analysis
of the Yamalo-Nenets Autonomous Okrug (YNAO). The developed method may, however, be
applied to any geographical location, and further research on this is planned for the future.
The aims of this research may therefore be listed as:
• To establish a method for automatically classifying water bodies using C-band ASAR WS
data, by first analysing different classification techniques and then choosing an optimum
solution that will allow computationally efficient automation of such a routine.
• To investigate how different classification techniques affect the derived water body prod-
uct and hence, to analyse the uncertainties in the established method through accuracy
and sensitivity tests.
• To develop an automated procedure that will enable efficient and effective processing
and application of large SAR data volumes, by creating temporal composites that will
provide information on water body extent and its dynamics over the spring/summer
season, investigating different methods of change detection analysis.
• To quantify changes in water body extent and to discuss their relevance to the distribution
of freshwater ecosystems, and scrutinise possible local impacts on the drainage system
due to human activities that may be related to flood events.
• To identify areas prone to wetting and discuss these in the light of ongoing land-cover
and land-use changes in the Arctic permafrost regions of the North of West Siberia.
4
Chapter 1. Introduction Structure of thesis
With ESA’s upcoming Sentinel-1 mission providing SAR data comparable to those from En-
visat, continuity is secured and longer timeseries analysis of frequent SAR data will ensure sea-
sonal change of Arctic water bodies to be considered on a inter-annual timespan, establishing
their relevance to longterm analyses of lake change. This PhD research topic reaches beyond
the permafrost community, with significance to climate modellers interested in wetland and
water body dynamics and their implications on carbon exchange with the atmosphere.
1.3 Structure of thesis
This thesis comprises seven chapters that are devised to illustrate the applicability of active
microwave remote sensing for detecting seasonal water body change. The chapters roughly
follow the chronological order of the work that was undertaken, although chapter 6 was in
fact done prior to chapter 5, having undergone peer review and been published in the journal
Environmental Research Letters in November 2013. It is anticipated that the content of chapter
5 will be published after the submission of this thesis. This chapter has introduced the topic
of permafrost research in its climate context. Chapter 2 gives a clear literature review of per-
mafrost research and advances in remote sensing of permafrost. The emphasis here is on the
current body of permafrost remote sensing studies that primarily deals with VNIR data. Per-
mafrost studies involving microwave remote sensing are also discussed, although their appli-
cation with respect to permafrost research is presently smaller than the aforementioned VNIR
investigations. Microwave remote sensing is particularly useful in the fields of snowcover, soil
moisture and freeze/thaw dynamics. These are also discussed in chapter 2, providing a clear
overview which will be referred at a later stage, as these are relevant methods that need to
be integrated in an analysis of seasonal inundation. Chapter 3 takes the reader through the
data used for mapping water body extent, the post-processing methods that were applied to
these data by colleagues, the active microwave mapping method that was further developed
and refined during this PhD research and map accuracy and sensitivity studies that are essen-
tial to error analysis of EO data. Chapter 4 elaborates on the monitoring procedure that was
devised, explaining the temporal composite approach and quality control thereof. Moreover,
the chapter discusses the first of the two water body change detection techniques applied to
demonstrate seasonal inundation, which is identified, mapped and quantified in the context
of previous water body distribution analyses. Chapter 5 provides a detailed analysis of the
effects relief have on surface water flow by undertaking terrain-based hydrological modelling,
embedded in a Geographic Information System (GIS). Further terrain-based techniques such
as the topographic wetness index (TWI) are investigated to model the flow of surface waters
and the accumulation thereof, comparing these to detected changes in surface inundation. Fi-
nally, chapter 6 discusses a second change detection technique, which is automated and may
be applied to assessing spring flood retreat. The final chapter reiterates the concluding results
of this research and suggests further research studies that would be beneficial to enhancing the
understanding surface hydrological change has on permafrost environments.
5
Author contributions to peer reviewed papers Chapter 1. Introduction
1.4 Author contributions to peer reviewed papers
While the paper Bartsch et al. (2012) was developed at the beginning of my PhD and only
included a short study of my own on fitting a double-Gaussian to the radar backscatter dis-
tribution in order to determine a classification threshold (see section 3.2.1), the other two peer
reviewed papers Trofaier et al. (2012) and Trofaier et al. (2013a) were conceived by myself. In
both cases I developed the concept, carried out the analysis and wrote the paper.
Co-author contribution for Trofaier et al. (2012):
W.G. Rees and A. Bartsch supervised this study, commenting on early drafts of the manuscript
and discussed the results with me. W.G. Rees in particular helped me refine the sensitivity
approach used for error analysis. D. Sabel developed the TU Wien in-house pre-processing
chain for the Envisat ASAR WS data and S. Schlaffer further developed updates to these pre-
processing algorithms.
Co-author contribution for Trofaier et al. (2013a):
A. Bartsch was the main collaborator for this study. She provided the ALANIS Methane
data and discussed the proposed algorithm concept with me. W.G. Rees provided native En-
glish editing and M.O. Leibman commented on the manuscript enabling me to set the study
into its appropriate geographical context by including her local knowledge of Yamal’s per-
mafrost environment.
The design of these studies was my own and the analysis was done under guidance of my
supervisors. Therefore the published materials and chapters presented here are almost entirely
based on my own work.
6
Chapter 2
Permafrost and Arctic surface
hydrology
2.1 Warming permafrost in the 21st century
Landscape evolution is a natural process. Periglacial environments have undergone modifica-
tions throughout the Quaternary. Nonetheless, the current trend of increasing air temperatures
are raising concerns about the implications on permafrost landscapes in Arctic and sub-Arctic
regions (Trofaier et al. 2012).
Permafrost is defined to be ground (rock, soil, sediment) that remains below 0 °C for more
than two consecutive years (French 2007). In fact, permafrost does not necessarily need to
contain water - either frozen or liquid (depending on salinity some water may not freeze at 0
°C) - ground ice, however, is typically found within permafrost. Therefore, perennially frozen
ground is vulnerable to atmospheric warming, as this warming promotes ground-ice melt,
which in turn leads to soil strength instabilities (Murton 2009).
Ground temperatures are a function of air temperatures (amongst other parameters). Local
conditions such as vegetation, topography, snow cover and soil substrate also influence the
temperature regime of the ground. Hence, climate and permafrost have a more complicated
relationship than is often portrayed. The mean annual air temperature (MAAT) is usually a
few degrees cooler than the mean annual ground surface temperature (MAGST) due to the
insulating properties of the winter snow cover known as the nival effect (vegetation has the
opposite effect but it is less dominant) (Smith and Riseborough 2002). The MAGST is therefore
greater than the temperature at the top of the permafrost (TTOP). The TTOP is always below 0
°C. Nonetheless, air temperatures do play an important role in permafrost warming. The net
effect of the local conditions including the so-called surface offset (due to vegetation and winter
snow cover) and of the air temperatures will influence (a) whether permafrost is present and
(b) whether over time a warming trend may be established.
During recent decades, a warming trend, measured with thermistors in deep boreholes,
has been observed in the upper permafrost layers (Romanovsky et al. 2010b). This warming
7
Warming permafrost Chapter 2. Permafrost and Arctic surface hydrology
affects geomorphological, hydrological and ecological processes associated with permafrost
conditions. One such process is known as thermokarst. When permafrost thaws, ground ice
ablation may result in topographic slumping. This is thermokarst subsidence. Thermokarst is
therefore a geomorphological process related to the thawing of frozen ground in ground ice
rich regions. Melt waters accumulate in the resulting depressions and form thermokarst lakes
or ponds, depending on their respective sizes. Tundra thaw ponds are of limited size; their
diameter often being below 2 km (French 2007). As this is a topic central to the PhD research,
the nature of these thermokarst lakes is discussed in more detail below in section 2.1.3.
2.1.1 Active layer dynamics
The active layer is the topmost layer of ground that undergoes seasonal freezing and thawing.
Site variables, such as vegetation patterns, soil moisture and soil type and texture influence the
ground thermal regime and affect the thaw depth of the active layer (Everett et al. 1981). The
seasonal dynamics of the soil temperature regime, also known as the annual freeze-thaw cycle,
will result in a seasonally frozen soil layer that overlies the perennially frozen ground. The
active layer plays a key role in the periglacial soil-water-plant system as it is the main unit to
which any interplay between ground, water and vegetation is constrained (Tyrtikov 1964).
The thickness of the permafrost and hence the location of the permafrost table is dependent
on the energy balance at the surface and the thermal properties of the ground components. It,
therefore, is determined by the internal heat gain and the heat loss from the surface (French
2007). Large thermal conductivities of the soil can ensure relatively thick active layers (Duguay
et al. 2005).
When considering active layer dynamics it is also important to mention the effects vege-
tation has on the seasonally frozen ground. Vegetation plays a key role in determining the
thickness of the active layer or, as it will be referred to henceforth, active layer depth (ALD). As
mentioned above, the insulating properties of vegetation cause a thermal offset, allowing for
ground temperatures to remain below the summer air temperatures. Low soil temperatures
result in low seasonal thaw depth, z. The active layer is often found to be thinnest beneath
vegetated areas and thickest beneath bare rock soils (French 2007). Furthermore, vegetation
contributes to changes in the ground thermal regime through soil moisture. Therefore, the
importance of soil water content with respect to freeze/thaw cycles, ALD and the variabil-
ity thereof in a changing climate is not to be underestimated (Anisimov et al. 2002). Changes
in vegetation patterns are often related to changes in soil moisture, which in turn lead to an
increased active layer and possible permafrost decay (Ström and Christensen 2007).
In general, any type of vegetation canopy or cover will help conserve the soil moisture, as
this cover decreases evaporation from the ground itself, and therefore it influences the ground
thermal regime. During the day, vegetation obstructs the evaporation of soil water, whereas
during the night, the condensation from the air on the vegetation is also a soil moisture con-
tributor. Furthermore, vegetation’s ability to retain precipitation also influences the ground
8
Chapter 2. Permafrost and Arctic surface hydrology Warming permafrost
thermal regime, as precipitation leads to ground thawing (Tyrtikov 1964).
Soil moisture and subsurface water flow are distinct features of active layer dynamics and
indeed soil moisture influences the rate at which the thawing front descends through the active
layer during the thaw season, z˙(τ). This rate has been described by Woo et al. (2004) to be
inversely proportional to the volumetric fraction of soil moisture as a function of depth, ϑ(z).
2.1.2 Subsurface water flow and taliks
Latent heat exchange between the atmosphere and the ground occurs through convective pro-
cesses at the interface. Furthermore, surface and subsurface waters influence the energy bal-
ance of the ground. These waters control the ground thermal regime through thermodynamic
interactions with the ground substrate (Duguay et al. 2005). As a potential energy source, sur-
face and subsurface waters may lead to unfrozen areas within permafrost known as taliks.
These taliks come in three main forms: (1) supra, (2) intra or (3) subpermafrost taliks, depend-
ing on where they are located in relation to the permafrost. Suprapermafrost are those taliks
that lie in between the active layer and the permafrost table. Intrapermafrost taliks lie within
and subpermafrost below the permafrost. Further, another categorisation can be made between
open taliks (unfrozen areas that reach from supra- to subpermafrost) and closed taliks, which
are mainly found beneath water bodies (French 2007) (see figure 2.1 for schematic).
In the discontinuous permafrost zone, as the name suggests, the permafrost is thin and
many taliks exist. This zone spans over the most southerly part of the Arctic, where mean
annual air temperatures are higher and the permafrost is hence more delicate (Anderson 2004).
In particular, suprapermafrost taliks are mainly associated with the discontinuous permafrost
zone were the seasonal frost does not penetrate through the ground entirely and hence does
not reach the permafrost layer (French 2007).
In the continuous permafrost zone taliks are found beneath rivers and lakes (Weise 1983).
Water bodies promote the development of taliks below their beds due to their ability to store
heat. They have high volumetric heat capacities and are hence associated with a disturbance
in the geothermal regime leading to disequilibria in the permafrost conditions (French 2007).
The size of a talik below a water body will depend on a number of factors such as the depth
and areal extent of the overlying water body, the water’s temperature, the ice thickness and
snow cover in the winter, as well as the type and composition of the bottom sediments and
the degree of their compaction (French 2007). Burn (2005) discuss the sensitivity of talik extent
and geometry to the thermal regime of lake bottoms, and the importance of water being a heat
storage.
The area of interest to this PhD research,the North of West Siberia, is described in chapter 3
(section 3.3) in more detail. In the context of taliks and subsurface flow, Rivkin (1995) find that
shallow lakes, common to this area, have depths of approximately 1.6 - 3 m. Those lakes that
are shallower than 1.6 m are mostly frozen to the bottom in the winter and the bed sediments
undergo seasonal thawing in the summer months, whereas lakes that are deeper than 1.6 m are
9
Warming permafrost Chapter 2. Permafrost and Arctic surface hydrology
associated with taliks. These findings concur with study results from Canada where large deep
lakes are predominantly associated with higher lake-bottom temperatures and hence the pres-
ence of taliks (Burn 2002), although it has been shown that the size of the lake has more impact
on permafrost and talik size than the lake-bottom temperature itself (Burn 2005), implying that
surface water extent promotes permafrost degradation.
Permafrost is often described as an impermeable frozen layer through which water can-
not percolate. However, rather than identifying it as an aquiclude both seasonal and peren-
nial frozen ground are to be seen as aquitards (van Everdingen 1990). Contrary to aquicludes
these do allow for groundwater transfer, albeit at a highly retarded rate (Smithson et al. 2008).
Nonetheless, the hydraulic conductivity of the ground may decrease due to segregated ice
lenses that may be present, resulting in an effectively impermeable layer (van Everdingen
1990). Therefore, the importance of taliks in relation to permafrost hydrology is that ground
water movements are nearly entirely restricted to these unfrozen areas within the permafrost
(French 2007). The vulnerability of Arctic permafrost environments to rising air temperatures
is modifying ecosystems (Jorgenson et al. 2013) and contributing to shifts in the hydrological
system (Alessa et al. 2008). Active layers are deepening and taliks are most likely growing
under changing climatic conditions (Zhou and Huang 2004, West and Plug 2008). Recent ob-
servation in increased base flow have been attributed to active layer deepening in a warming
climate, reiterating the active layer’s significance as a suprapermafrost aquifer and its poten-
tial for accommodating increases in net groundwater flux related to permafrost thaw (Quinton
et al. 2009, Walvoord et al. 2012).
(a) Section depicting the relationship between permafrost
and unfrozen ground.
(b) Cross-section of a thermokarst lake with talik
(thaw bubble) below.
Figure 2.1: Section of the ground showing the relationship between permafrost, active layer
and taliks, taken from French (2007), courtesy of Ferrians et al. (1969) and the United States
Geological Survey 2.1(a). Schematic cross-section of a thermokarst lake 2.1(b).
10
Chapter 2. Permafrost and Arctic surface hydrology Warming permafrost
2.1.3 Arctic water bodies
2.1.3.1 Arctic lakes
According to Lehner and Döll (2004) nearly a quarter of the world’s inland freshwaters lie at
high latitudes. Taking only water bodies greater than 10 ha at latitudes greater than 45.5° N into
consideration, approximately 414 400 km2 of these northern territories are covered by lakes (N
∼ 148300) (Smith et al. 2007). These water bodies may be of various geomorphological origin,
periglacial, glacial or even volcanic. However, many of these lakes are landscape units con-
trolled by permafrost conditions (Grosse et al. 2013). Lake abundance is an important concept,
since being able to quantify these freshwater ecosystems plays a key role in up-scaling aquatic
rates and processes that are of global importance. In particular, these inland freshwater systems
play a key role in the global carbon cycle (Dean and Gorham 1998, Cole et al. 2007).
Thermokarst lakes are particularly common Arctic landscape features. Their development
is related to a disturbance in the ground thermal regime and permafrost degradation. These
terrain disturbances lead to ground subsidence in which melt waters accumulate to create lakes
and ponds (Burn and Smith 1990). During the early Holocene alasses or alas-systems formed.
These are thermokarst depressions in which lakes developed and evolved (Morgenstern et al.
2013). The Arctic and sub-Arctic regions are extensively covered by these alas-systems. In
particular the vast tundra lowlands are strewn with an abundance of thermokarst lakes and
ponds.
In North America, the term thermokarst lakes is usually attributed to lakes lying within per-
mafrost regions in general. This potentially arises from remote sensing studies that can monitor
lakes on the landscape scale without explicitly distinguishing their origins. However, it is im-
portant to note that terminology in Russia is slightly different. Here, thermokarst lakes are
not just considered those that originated through past permafrost degradation but rather those
where permafrost is proven to be an actively influencing factor. There is a fine line between
these two definitions but it must be kept in mind when discussing the literature of two differ-
ent cultures. As this PhD research is almost exclusively a remote sensing project, thermokarst
lakes as well as any other open water surfaces, will be referred to as water bodies, in the context
of EO, to avoid any cultural mis-understandings.
Lakes in permafrost regions are subject to seasonal variations due to snowmelt and fluc-
tuations in precipitation, both of which can lead to changes in lake extent. The interaction of
water bodies with permafrost is crucial to permafrost hydrology. Water has a high heat ca-
pacity, which results in further ground thawing. Hence surface water itself and changes in its
extent affect permafrost processes, and these in turn affect lake extent. This reciprocity has
lead to the method of using thermokarst lake studies as an EO means to identify permafrost
change over the past 50 years (e.g. Yoshikawa and Hinzman (2003), Smith et al. (2005) and
Grosse et al. (2005)). Lake change, in combination with meteorological data, is used as a type of
proxy for permafrost decay, establishing trends in the climate dependencies of permafrost and
visualising its change in a warming climate.
11
Warming permafrost Chapter 2. Permafrost and Arctic surface hydrology
2.1.3.2 Arctic streams, rivers and floodplains
The Arctic accommodates large river ecosystems (e.g. Yenisei, Lena and Ob in Siberia and
the Mackenzie River in Canada) that flow across wetland tundra through extensive floodplain
deltas, eventually discharging in to the Arctic Ocean. Wetland tundra is strewn with innu-
merable streams whose interactions with thaw lakes are not to be neglected. The relationship
between Arctic streams and their surrounding watershed is sensitive. It was discussed above
how permafrost and thin active layers restrict water infiltration. Some stream water does how-
ever infiltrate the bed in the summer, often expanding the unfrozen zone where stream water
exchange between the stream and its bed may take place. This zone is known as the hyporheic
zone and its extent is also controlled by seasonal thaw depths (McKnight et al. 2008). Small
streams may freeze up completely in the winter, grounding the ice cover. Often only partial
freezing of the stream occurs and ice-jams may arise. All water above the jam effectively goes
into storage with associated floods occurring upstream (Jeffries et al. 2005).
As the extent of the streams’ hyporheic zones are constrained by the permafrost table, ex-
tensive spring melt waters will not easily infiltrate the ground. Summer precipitation and snow
and ice melt (including ice-jam failures) lead to sudden and extreme hydrological responses by
these streams. Therefore, permafrost catchments are often associated with flashy hydrographs
(Lyons and Finlay 2008). Rain and melt water increase stream activity, as its movements into
frozen ground are highly restricted (McKnight et al. 2008). This leads to extensive seasonal
inundation of floodplains and in general surface water flow on wetlands (Woo 1990). The
floodplains are also strewn with a myriad of lakes. These floodplain lakes are of thermokarst
origin, a result of patchy ground thawing and subsidence, and concentrated melt water filling
of these depressions become the lake basins (McKnight et al. 2008). With respect to the im-
pacts of surface water flow on permafrost, fluvial processes lead to thermal erosion of frozen
ground. This is due to the thermal effects of both heat conduction and convection of running
water (Murton 2009). Fluvial-thermal erosion promotes slope instability (French 2007). Rapid
drainage of thaw lakes has in the past been attributed to this erosion process known as ’tapping’
(Mackay 1988), although this drainage mechanism is not the only mode by which thaw lakes
may drain. Stream-flood-lake interactions are a prominent feature of Arctic surface hydrology
and therefore a key concept, central to this PhD research.
2.1.3.3 Arctic wetlands
The combination of water, vegetation and soil substrate is crucial to wetland ecosystems. These
wetland characteristics are the key features that are used in EO analyses. Saturated soils in the
Arctic are most commonly found in the Low Arctic where precipitation rates are high and mois-
ture from precipitation as well as from snow melt is retained during the summer months in the
seasonally frozen ground, the active layer. It is therefore quite common that wetlands co-exist
along side periglacial lakes, which have different and multifold geomorphological origins to
wetlands (Pienitz et al. 2008). Wetlands play an important role in the global carbon budget.
They are sources of carbon dioxide (CO2) and methane (CH4) emissions and hence have impli-
12
Chapter 2. Permafrost and Arctic surface hydrology Warming permafrost
cations on our climate. Arctic and sub-Arctic wetlands are to be considered especially fragile
ecosystems due to their location in the climate sensitive tundra and boreal biomes (Aber et al.
2012). Rising air temperatures are resulting in landscape evolution; changes in vegetation and
warming of permafrost associated with complex links to biogeochemical cycling (Callaghan
et al. 2005).
The highly dynamic surface hydrology of these high-latitudinal wetlands and surface wa-
ters may be a significant factor contributing to potential climate feedbacks through its effects
on, and reciprocal interactions with, the frozen ground below (Goswami et al. 2011). The signif-
icance of freshwater ecosystems regarding carbon exchange with the atmosphere are not to be
underestimated (Downing et al. 2006). Furthermore, understanding the seasonal variability of
these surface water bodies and their inherent processes is of utmost importance. To determine
the environmental and climatic impacts of lakes it is necessary to know their abundance and
extent (Grosse et al. 2008). The percentage ratio of surface water cover to land cover, known as
the limnicity, gives us some indication of this abundance (Likens 2009).
2.1.4 The West Siberian Lowland -
A large wetland ecosystem underlain by permafrost
This PhD research focuses on a large wetland environment in the West of North Siberia, known
as the West Siberian Lowland (WSL). The study areas of this PhD research will be discussed in
more detail later on, however, here a general description of this wetland environment is given.
The WSL’s boundaries are in the East the Ural mountain range, in the North the Kara Sea, in the
West the Yenisey river and in the South the steppes of Kazakhstan. The sheer size is such that it
encompasses approximately 16% of the Russian Federation, and it is the country’s lowest and
flattest landscape (Solomeshch 2005).
This is a peatland ecosystem, where approximately 16-25% of the landscape contains deep
peat layers. Peatland and mires are wetland ecosystems in which significant peat accumulation
takes place (Aber et al. 2012). Two types of mires are found in this region: (1) homogeneous
and (2) polygonal mires. The former are known to be found in floodplains and in thermokarst
depressions and alasses around lakes, whereas the latter are found in well-drained floodplains
where mineral soil gives rise to forest and meadow vegetation (Solomeshch 2005). Homoge-
neous mires are found in the North of the WSL, especially on the Yamal peninsula, which is the
main study area of this research. Homogeneous mires are found in the wet areas of the Yamal
peninsula, in the floodplains and depressions while the upland areas are covered by tussock
tundra vegetation, exhibiting typical periglacial behaviour such as bare spot medallions also
known as soil polygons (Bliss et al. 1981). Photographs of tussock tundra and spot medallions
on the Yamal peninsula are shown in figure 2.2. A high presence of Betula nana (dwarf birch),
Salix polaris (Polar willow) and Sphagnum balticum (Baltic bog moss) is found in these homoge-
neous mires. Other plants such as Carex aquatilis (water sedge), Carex disperma (softleaf sedge)
and Sphagnum squarrosum (spiky bog moss) mark this northern peatland, allowing it to be per-
13
RS of the Arctic tundra Chapter 2. Permafrost and Arctic surface hydrology
Figure 2.2: Spot medallions on uplands (left panel) and tussock tundra landscape (right panel).
Photographs taken by A.M. Trofaier on 2 and 4 July 2012.
ceived as a green landscape, albeit with less colourful flora than may be found in the southern
sub-Arctic areas (Solomeshch 2005). Figure 2.3 shows photos taken in the field in the summer
of 2012, illustrating the typical landscape and its vegetation cover.
2.2 Satellite remote sensing of the Arctic wetland tundra landscape
As pointed out by Smith and Riseborough (2002), the current understanding of permafrost
extent and its conditions is restricted by insufficient field measurements of the thermal state
of the ground. These insufficiencies are due to the fact that permafrost underlies landscapes
which are both remote and vast, in which regular sampling of active layer depth and borehole
temperature measurements are both difficult and expensive. This is where (satellite) remote
sensing comes into its own. Satellite data have been used in many studies to monitor lakes
over the vast tundra regions. However, due to their limited size, tundra ponds are not cap-
tured in satellite-based global land cover datasets (Frey and Smith 2007, Bartsch et al. 2008).
High spatial resolution data are required. Such data are available from optical sensors. The
availability of acquisitions depends on daylight and cloud conditions. Aerial photographs are
most suitable, but these are often limited due the logistics involved in acquisition. Satellite data
provide an alternative means for more extensive archived material. Landsat data have been
available since the 1970s, and multispectral instruments—the Thematic Mapper (TM) and, in
later missions, the Enhanced Thematic Mapper (ETM)—have frequently been used to capture
land cover changes in thermokarst regions (e.g. Frohn et al. (2005), Grosse et al. (2006), Ulrich
et al. (2009), Arp et al. (2011)). Other optical data are available for even longer periods back in
time. Corona data, which date back to the 1960s, and the more recent IKONOS data have been
of interest for permafrost studies (Yoshikawa and Hinzman 2003, Grosse et al. 2005).
Less extensive archives for active microwave data are available. However, these sensors,
14
Chapter 2. Permafrost and Arctic surface hydrology RS of the Arctic tundra
(a) (b)
(c) (d)
Figure 2.3: Photographs of wetland tundra: A transect for Greening of the Arctic project spec-
trometer measurements with dwarf birch (a), alas with thermokarst lakes and Polar willow on
the slopes of the depression, tussock tundra on the uplands (b, top panel), lake with grasses
along its shoreline and bordering willow thickets in the background (b, bottom panel), a type
of bog moss growing in an inundated area (c) and water sedges along a shoreline of a lake (d).
Photos taken by A.M. Trofaier during the period of 1 - 3 July 2012.
15
RS of the Arctic tundra Chapter 2. Permafrost and Arctic surface hydrology
Figure 2.4: Timeline of SAR sensors. The y-axis shows decreasing resolution. Diagram taken
from Pope et al. (2014).
such as synthetic aperture radar (SAR), are independent of cloud cover and illumination and
have proved to be a useful method for providing observations and monitoring features of the
hydrosphere (Bartsch et al. 2009). The timeline of SAR sensors with their respective spatial
resolutions are given in figure 2.4. Due to the less extensive archive and its often lower spatial
resolution, SAR has in the past been more readily shunned and the feasibility of using these
radar sensors for monitoring thermokarst lakes is a central point of this PhD research and is
explored and discussed in chapter 3 (section 3.4).
In general, representative data for the phenomenon of lake change are restricted to late
summer acquisitions, since most lakes are often still frozen at the beginning of summer. Lake
drainage and shrinkage have been reported based on optical satellite data from Alaska (Yoshikawa
and Hinzman 2003), Siberia (Smith et al. 2005), and Canada (Plug et al. 2008, Labrecque et al.
2009). However, both decreasing and increasing lake surface extent are reported, depend-
ing on geographical location (Grosse et al. 2011). In addition to lake change, remote sensing
techniques are a well-established method for monitoring wetland dynamics. Information on
seasonal and periodic inundation patterns can easily be retrieved from such techniques. Mi-
crowave satellites (both active and passive) provide valuable information on wetland hydrol-
ogy (Prigent 2001, Prigent et al. 2007, Schroeder et al. 2010, Watts et al. 2012). Active microwave
sensors may be used to map open water surfaces (Bartsch et al. 2009) and monitor their dynam-
ics (Bartsch et al. 2007a, 2012). Investigating wetland dynamics is a multi-scale problem; both
spatial and temporal aspects need to be considered. Furthermore, the inherent trade-off be-
tween spatial and temporal resolution of remotely sensed data is cause for concern. Careful
16
Chapter 2. Permafrost and Arctic surface hydrology RS of the Arctic tundra
considerations on which sensors provide the optimum resolution for the investigation of wet-
land processes are essential. The advantages of Earth Observation products are undeniable,
however, their suitability as a long-term monitoring method are influenced by data availability
and satellite lifetime (Bartsch et al. 2012). The limiting factors to long-term monitoring are a
matter of continuity of comparable and self-consistent data.
2.2.1 Optical Imagery
Optical sensors are often used in analysing thermokarst related processes. Yoshikawa and
Hinzman (2003) used aerial photographs as well as imagery from the IKONOS satellite to mon-
itor decadal changes in lake surface extent, while Hinkel et al. (2005) used Landsat 7 ETM+ im-
agery to identify thaw lakes and drained thaw lake basins and to further undertake a statistical
analysis of their orientation, shape and eccentricity.
In general, multispectral imagery is most applied in mapping vegetation and land cover
changes. With respect to vegetation, its potential is in identifying changes in plant production,
plant type and cover and phenological growth attributes, as well as the establishment of prox-
ies for biomass or chlorophyll content (Stow et al. 2004). Commonly, spectral vegetation indices
are used to infer trends. Spectral indices are calculated by mathematically combining at least
two spectral bands. The combination is chosen so that the spectral indices will more distinctly
represent biophysical parameters. Common vegetation indices are for example leaf-area index
(LAI) and normalised difference vegetation index (NDVI). The radiometric data captured in
the multispetral imagery is used to derive these vegetation indices. Most vegetation indices
make use of the so-called red-edge feature. This feature is represented by a distinct increase
in reflectance at a wavelength, λ, of about 700 nm, which is characteristic of green, chlorophyll
vegetation (Jones and Vaughan 2010). Studies that make use of vegetation indices in the Arc-
tic include those by Huemmrich et al. (2010) and Stow et al. (2004). The former examined the
effects a changing climate might have on NDVI (an increase in NDVI typically representing an
increase in green vegetation). The latter also made use of a NDVI time series, this study how-
ever, investigated its applicability to monitoring rapidly changing greenness trends in the Arc-
tic. Crossman et al. (2012) use the NDVI to identify areas of glacial fed groundwater upwelling
in a sub-Arctic permafrost environment in Alaska; these zones of groundwater upwelling are
known to be densely vegetated patches found on glacial floodplains that are otherwise unpro-
ductive areas, sparsely populated with vegetation.
In the North of West Siberia, which is the study area this PhD focuses on (see chapter 3, sec-
tion 3.3), NDVI techniques have been used to correlate warming trends to vegetation patterns.
Walker et al. (2009b) analyse NDVI trends (between 1982 and 2007) on the Yamal peninsula,
comparing these to a summer warmth index, derived from Land Surface Temperature (LST)
measurements (Raynolds et al. 2008), and sea-ice concentration (Bhatt et al. 2007). Their re-
sults show higher NDVI values are correlated to warmer LSTs and periods of lower sea-ice
concentrations. Although these results are consistent with other coastal area studies (e.g. Jia
17
RS of the Arctic tundra Chapter 2. Permafrost and Arctic surface hydrology
et al. (2003), Bunn et al. (2007), Walker et al. (2009a), Macias-Fauria et al. (2012)), the correlation
of NDVI to sea-ice concentration on Yamal was, however, found to be less significant than in
North America.
Another method for monitoring ecosystems is the technique of land cover classification.
Land cover classification methods can be used for monitoring anthropogenic disturbances to
the sensitive Arctic environment (e.g. Rees et al. (2003)), in particular in the North of West
Siberia, where extensive development projects of the gas industry are leading to significant
land-use changes (e.g. Kumpula et al. (2010, 2011, 2012) and Khitun et al. (2001)). Not only
land-use changes, but natural land cover changes may be monitored using satellite imagery
classification techniques. These methods are also applicable to vegetation mapping, where
multispectral imagery is classified through combining spectral information with ancillary data
(Jones and Vaughan 2010). Land cover classification of the Arctic tundra has multiple appli-
cations within change analysis studies. Ulrich et al. (2009) used field spectrometry data in
combination with a Landsat-7 ETM+ land cover classification to arrive at a base dataset that
would allow the quantification of long-term changes in the Lena Delta. Schneider et al. (2009)
similarly produced a Landsat 7 ETM+ classification of the Lena Delta. This study proposed
to classify land cover types based on their methane emission potential. The ancillary data for
this classification was provided by measurements of methane emissions, which were used to
determine the final land cover classes by merging those classes that after the initial supervised
classification procedure were found to have the same methane emission potential. The aim of
this study was to temporally upscale the methane emission data with the aid of the remotely-
sensed data. EO products are used for both upscaling the aggregated processes of landscape
units (such as regional lake methane emissions) (Downing 2009), as well as downscaling satel-
lite images with relatively large resolution cells to finer-scale subpixel land-cover classifications
(Muster et al. 2012).
Multispectral imagery is a powerful tool, that in combination with radar data enables issues
concerning the trade-off between spatial and temporal resolution to be addressed. This is of
significance for monitoring intra- as well as inter-annual dynamical processes at the appropri-
ate spatial resolution (Gala and Melesse 2012), especially in terms of hydrological monitoring.
Brown and Young (2006) emphasis the importance of monitoring the seasonal hydrodynamics
of Arctic wetland ecosystems, which currently remain undetected. This is a key concept of,
and the reasoning for undertaking, this PhD research. In order to compensate for the trade-off
between temporal and spatial resolution, a combination of data from multiple satellite sensors
is needed (McCabe et al. 2008).
2.2.2 Radar
The application of active microwave sensors, such as synthetic aperture radar (SAR), has par-
ticular advantages when investigating Arctic permafrost environments. Firstly, atmospheric
attenuation of microwave radiation is minimal. Microwaves are capable of cloud cover pene-
18
Chapter 2. Permafrost and Arctic surface hydrology RS of the Arctic tundra
Figure 2.5: Graphs illustrating the effect of a phase change from water to ice on its dielectric
constant (left panel) and the backscattering coefficient (right panel) as presented by Han and
Lee (2013), who carried out a C-band radar experiment.
tration; an important ability for monitoring the Arctic regions, as images can be taken through-
out the entire year, irrespective of weather conditions. Secondly, the dielectric properties of
water are such that microwaves can differentiate between a frozen and a thawing landscape.
Frozen water binds its molecules into a crystalline lattice resulting in a reduction of its dielectric
constant (see table 2.1). Figure 2.5 shows this decrease in dielectric constant as water undergoes
a phase change becoming ice. The right panel of the figure also shows the effect a phase change
has on the radar backscatter coefficient, as demonstrated by Han and Lee (2013) in a C-band
radar experiment. Therefore, due to the nature of water, terrain characteristics (such as vege-
tation, soil and snow) display large differences in their dielectric properties as the landscape
changes from a frozen to a thawed state (Kimball et al. 2004).
As the landscape thaws, and the melt waters accumulate, and infiltrate the ground, the ac-
tive layer will retain greater amounts of soil moisture in the early summer than it does later
on in the season. The greater the moisture levels the lower the ability of the radar sensor to
penetrate the ground. Needless to say, many features influence the radar signal (incidence
angle, surface roughness, canopy cover and moisture). In order to separate the soil moisture
signal from these noise features, techniques for isolating and removing these features need to
be developed (Schmugge et al. 2002). Nonetheless, empirical approaches to extract soil mois-
ture from barren soil surfaces which hence remain uninfluenced by vegetation, have also been
established. Through least-squares linear regression, it has been found by Ulaby and Batlivala
(1976) that a near linear relationship between soil moisture conditions and radar backscatter
exists. The model the authors propose identifies a relationship between the backscatter coef-
ficient due to surface scattering and the field capacity (the amount of moisture retained in the
soil after the excess has been removed (Hendriks 2010)).
Soil moisture retrieval algorithms often make use of both passive and active microwave
sensors, in order to account for biases introduced by vegetation (Entekhabi et al. 2010). Wag-
ner et al. (2003) use active microwave (scatterometer) data to derive a multi-year global soil
moisture dataset over the time period 1992 - 2000. SAR is an imaging radar system, that has
successfully been applied to wetland monitoring (Bartsch et al. 2007a, 2008). SAR is in par-
19
RS of the Arctic tundra Chapter 2. Permafrost and Arctic surface hydrology
Table 2.1: The dielectric constant of ice and water for C-band sensors.
Dielectric constant C-band
′ice 3.2
′water 69
ticular effective with respect to mapping inundation (seasonal and permanent). This is due
to SAR’s ability to distinguish water bodies from its surrounding land cover, once straightfor-
ward classification techniques have been applied. The response of the microwave signal due
to water bodies results in low backscatter values for the associated water pixels (in the digital
greyscale image, these low backscatter values are represented by low grey level values). The
low backscatter values are due to specular reflection of the incident radar beam off calm water
surfaces away from the sensor (Bartsch et al. 2008). However, SAR also has the potential to map
soil moisture. The above mentioned relationship between soil moisture and radar backscatter
is used for this purpose. However, surface roughness and vegetation introduce ambiguity into
the radar signal. These effects may be eliminated by subtraction of a reference image, however,
mapping soil moisture using SAR data does need further developing before it can become op-
erational (Wagner et al. 2007, 2008).
There are further applications of SAR with regard to permafrost studies, in particular re-
lated to thermokarst lakes. For example, SAR is used as a means for determining ice-on and
ice-off dates of lakes (Duguay et al. 2003), for the differentiation between grounded-ice (i.e. the
entire lake is frozen) and floating-ice covers (Arp et al. 2011) (as described in more detail below
in section 2.2.3), as well as a method for determining lake depth and lake bathymetry (Jeffries
et al. 1996, Kozlenko and Jeffries 2000).
2.2.3 Arctic lakes - The winter perspective
As air temperatures start to cool at the onset of the winter season, so does the water of lakes.
At 4 °C, water reaches its maximum density, heat loss occurs at the water surface and the dense
waters sink. As the lake continues to cool, a more stable, less dense water layer forms at the
surface of the lake which then may reach the freezing point, allowing ice to form at the surface
(Jeffries et al. 2005). Depending on lake size and depth, winter ice covers are either floating or
grounded. The latter occur for shallow lakes; these water bodies are completely frozen during
the winter season. The difference between floating ice and grounded ice is distinguishable with
active microwave sensors. The dielectric contrast of water between its liquid and solid phase is
so great that floating ice produces very high backscatter values, whereas grounded ice’s low di-
electric contrast to the sediments of lake bed, to which it is frozen, produce very low backscatter
values. The radar beam is attenuated and eventually absorbed by the soil. Therefore, synthetic
20
Chapter 2. Permafrost and Arctic surface hydrology RS of the Arctic tundra
aperture radar exploits the dielectric properties of water to provide information on lake ice.
This is in stark contrast to VIR and TIR sensors, which can only provide information on the
snow covers that lies upon the frozen lakes’ surfaces.
C-band sensors (λ ∼ 5.6 cm) are traditionally used for monitoring lake ice covers (e.g.
Jeffries et al. (1996), Duguay et al. (2003), Arp et al. (2012)). This is due to the substantial
difference between the dielectric constants, ′r, of ice and water. However, a recent study by
Engram et al. (2013) showed that L-band (λ ∼ 23.6 cm) SAR data may also be used to delineate
floating ice from grounded ice, although the results of the study showed that C-band data still
remain preferential over L-band data.
The significance of mentioning these lake ice cover studies here is to highlight the possibil-
ities SAR can provide for monitoring freshwater ecosystems. Furthermore, in the winter snow
will accumulate on these ice covers and contribute to meltwaters of a thawing landscape and
hence floods in spring. This aspect has not been investigated during the course of this PhD, but
it is noteworthy to mention that winter precipitation and the accumulation thereof in the form
of snow is the main factor contributing to high spring discharge in rivers and the spring floods
detected and explored in this PhD research.
2.2.4 Snow detection and freeze/thaw cycles
The two main influencing factors that snow has on permafrost are related to (1) the energy bal-
ance and (2) the hydrological input. Snow has a key role in altering radiative fluxes, providing
an insulating layer that causes the nival offset which in turn allows ground temperatures to be
higher than air temperatures in cold harsh winters. The heat budget of a snow pack has been
described in detail by Nakao et al. (1998) and will not be elaborated on here. Suffice to say that
snow also acts as a storage of thermal energy, where sensible heat fluxes due to temperature
gradients between the atmosphere and the snow pack and latent heat fluxes due to conden-
sation and sublimation of water vapour, impact on the heat budget of the snow pack. Snow,
therefore, is a heat exchanger, storing and transferring energy between its interfaces with both
the ground and the atmosphere. Snow also influences permafrost by being a source of surface
water, its melt waters accumulating in thermokarst depressions and floodplains, infiltrating
the active layer and increasing the gravimetric soil moisture content of the soil which in turn
influences the descent of the thawing front as mentioned in section 2.1.1. Indeed, this twofold
influence of snow on permafrost is further magnified by the fact that the rate at which the
thawing front descends through the active layer is inversely proportional to the soil moisture
content (z˙(τ) ∝ ϑ(z)−1). The more melt waters percolate through the active layer, the higher
the gravimetric soil moisture content will be and the slower the rate of thaw descent through
the active layer. In fact, Christiansen et al. (2010) showed that for study sites in the Nordic
area, where snow cover was found to be thin during their time of investigation, the surfaces
were dry and the water content of the soil was low. Here, ALD showed a clear correlation to
the local summer air temperatures. Furthermore, Seppälä (2004) emphasises the importance of
21
Summary Chapter 2. Permafrost and Arctic surface hydrology
snow re-distribution through wind on thermokarst activity. The wind-blown accumulation of
snow in thermokarst depressions may further insulate and warm the ground and associated
deepening of the active layer is to be expected in spring.
The distribution (and re-distribution) of snow is therefore an important feature to be consid-
ered in the context of permafrost surface hydrology, and therefor the development of snow dis-
tribution detection methods is an important part of cryospheric remote sensing studies. Moni-
toring snow may be done with both multispectral as well as radar sensors. The former allows
the identification of the reflectance of snow, its surface temperature, the snow water equivalent
(SWE), as well as snow grain size (e.g. Solberg and Andersen (1994), Salminen et al. (2009)
and Kuittinen (1988)). In addition, both active and passive microwave systems may be used to
monitor snow. SAR data have been used to differentiate snow from bare ground (Guneriussen
1997) and scatterometer data have been used to measure the distribution as well as determine
snow melt dates (Nghiem and Tsai 2001). These thaw events are intrinsically associated with
diurnal freeze/thaw cycles which effectively may be used to derive snow-off dates and there-
fore provide a method for deriving the length of the growing season (Bartsch et al. 2007b).
Indeed, in chapter 6 (section 6.2.2), such data are made use of to determine te snow-off season.
Other applications of scatterometer data concern snow conditions. Bartsch et al. (2010) go so far
as to use scatterometer data to evaluate the impacts different snow conditions have on reindeer
herding on the Yamal peninsula.
2.3 Summary
• Substantial changes in permafrost extent are expected to arise, predicteded by modern-
day trends in global warming. Active layer deepening will not only result in the removal
of old carbon stocks, but also change surface and subsurface hydrology. The relationship
of wetlands, open surface waters, subsurface water flow and soil moisture content on
permafrost needs closer inspection.
• The West Siberian Lowland is a peatland environment that encloses approximately 16%
of the territory of the Russian Federation. On the Yamal peninsula, homogeneous mires
dominate the landscape. They are found in alas systems and on wet floodplains. The
significance of this peatland ecosystem with respect to carbon accounting stresses the
need for continuous monitoring of wetland dynamics, such as seasonal inundation.
• Satellite remote sensing may be applied to monitoring water body dynamics on the landscape-
scale. Both VNIR and microwave (active and passive) data may be used for such a proce-
dure. The choice of sensor is process dependent. In terms of geomorphological processes,
which are often thought to be related to long timescales, VNIR data are used for studies
of lake change over decades. However, rapid hydrologial phenology need also be con-
sidered. Flash events related to highly dynamic hydrological processes, such as spring
22
Chapter 2. Permafrost and Arctic surface hydrology Summary
discharge, are not properly taken into consideration. Such processes need frequent mon-
itoring, which can only be achieved at a cost of spatial resolution. Radar sensors provide
periodic data for which changes in surface water body extent may be derived over the
summer period.
• Combining different satellite sensors to analyse data on snow (distribution and melt),
water body extent (frequent and longterm) and field data (ALD and permafrost borehole
temperatures) should provide a clearer picture of the significance of surface hydrology,
and any modifications thereof, in cold climate regions.
23

Chapter 3
Mapping inundated areas with active
microwave sensors
The core of this chapter explores the development and implementation of a lake mapping pro-
cedure using high temporal resolution active microwave data, which will be used in chapter 4
for analysing seasonal flooding and drainage. This calls for a discourse on the applicability of
radar sensors to identifying open surface water bodies, the need for frequent monitoring using
periodic data such as Envisat ASAR WS imagery and the feasibility of using ASAR WS given
its coarse spatial resolution.
After giving a detailed description of the data that are used this chapter puts forward and
discusses the methodology by which active microwave data are classified. This is a general
explanation of a specific radar water body classification method known as density-slicing or
thresholding, which has been implemented throughout this dissertation. The method of thresh-
olding for water body mapping follows the technique first put forward by Bartsch et al. (2008).
This method has since been further investigated (Trofaier 2010, Bartsch et al. 2012) and devel-
oped (Trofaier et al. 2012) and is presented here. It is common practice in science to provide
appropriate error analysis which of course also holds true in the field of remote sensing. Error
analysis in the form of multiple feasibility studies is therefore presented here in support of the
radar thresholding method. Three different feasibility studies were undertaken. The study ar-
eas differ for each study but are partially overlapping. It is hence proposed to call this the ‘local
scale’. Study area sizes depend on the data that are used for each feasibility study. However,
all of these study areas are located on the Yamal peninsula within an area of approximately 250
km2, a few kilometres South of the Bovanenkovo gas field (70° 29’ N, 68° E). This study area
is named Vaskiny Dachi, after a research camp (70° 17’ N, 68° 54’ E) that was first established
in 1988 as an engineering support to the construction project of the Obskaya-Bovanenkovo
railroad.
This chapter, therefore, focuses on local scale analysis on the Yamal peninsula. Section 3.4.1
was published in the proceedings of the Tenth International Conference on Permafrost and
where sentences have been used verbose these will be identified with the appropriate refer-
25
Data Chapter 3. Mapping inundation
ence Trofaier et al. (2012). Section 3.4.2 is a study that was presented in a talk at the Twelfth
Circumpolar Remote Sensing Symposium in Levi, Finland and was also part of a poster at the
ESA Living Planet Symposium in Edinburgh. Subsequently, it was submitted to the conference
proceedings (see Trofaier et al. (2013b)). The study in section 3.4.3 is yet to be disseminated
to the scientific community. Section 3.2.1 introduces the classification method and examines a
double-Gaussian fit of the bimodal backscatter distribution presented in Bartsch et al. (2012).
As the second author to this publication, the work on the double-Gaussian is my own contri-
bution and it seems fitting to elaborate on it here. Here follows the discussion of water body
mapping and associated feasibility studies.
3.1 Data
The data come from spaceborne sensors of varying spatial and temporal resolution. Primar-
ily, this research is founded on the application of Envisat ASAR WS data, its characteristics
enabling the frequent monitoring of surface inundation (as discussed below in section 3.1.1).
Additional data are presented here (ALOS PRISM and TerraSAR-X). These were used for the
feasibility studies as validation for handling ASAR WS data in order to conduct this research.
3.1.1 Envisat Advanced Synthetic Aperture Radar operating in Wide Swath mode
The European Space Agency’s Envisat (Environmental Satellite) Advanced Synthetic Aperture
Radar (ASAR) instrument operating in wide swath (WS) mode constitutes the main source of
data for this research. Envisat was launched in 2002 and its mission ended in 2012 after an
unexpected loss in communications with the satellite. ASAR is a C-band sensor with a centre
frequency of 5.331 GHz (λ = 5.6 cm). ASAR data were collected from 2003 onwards at several
modes, the main mode being Global Mode (GM) with a spatial resolution of 1 km. In WS a
high number of images are available for certain areas in the Arctic, although not at regular
intervals as the data were only acquirable on request and therefore, archived material is lim-
ited. WS - like GM - is a ScanSAR mode. ScanSAR is a technique that uses sub-swaths that are
each imaged at different pulse repetition frequencies (the rates at which a sequence of radar
pulses are transmitted) and then combined to produce a much wider swath coverage (Wood-
house 2006). Both GM and WS have a swath width of 405 km, divided into 5 sub-swaths at
their respective resolutions (Desnos et al. 2000). The spatial resolution of ASAR WS is 120 m
for all sub-swaths (Closa et al. 2003). However, the approximate nominal spatial resolution of
the WS mode product of 150 m is often listed in the literature (e.g. ESA (2007) and Santoro
et al. (2011)) as an undersampled pixel spacing of 75 m is used. This is rather coarse com-
pared to the high-resolution optical sensors such as Quickbird-2, SPOT and Landsat ETM+.
Their panchromatic bands each have a spatial resolution of 0.6 m, 1.5 m and 15 m, respectively.
Nonetheless, due to the inherent trade-off between spatial and temporal resolution, ASAR WS
allows improved mapping over larger regions and the majority of tundra surface hydrology
26
Chapter 3. Mapping inundation Data
is captured (Bartsch et al. 2008). In WS mode, data are available either in HH (Horizontally
transmitted - Horizontally received) or VV (Vertically transmitted - Vertically received) polar-
isations. Throughout this research project only HH polarised ASAR WS data were used. Due
to the specific water body classification method that was implemented, radar polarisation does
not play an important role. However, it is noted that if Envisat ASAR WS had been able to
produce cross-polarised data (HV or VH), then radar coss-polarisation would have also been
explored. A brief discussion on the differences in polarisation will be broached in section 3.2
where an explanation of the water body classification method is given. Since ASAR WS is the
main data source for this PhD research, the pre-processing steps are explained in more detail
below in the next section.
3.1.2 ASAR WS data pre-processing
The first few stages within the processing chain are usually referred to as pre-processing steps.
Pre-processing of ASAR WS data is computationally extensive and expensive. The job requires
an efficient setup and computer power beyond that of a single desktop. Years of development
(and investments) have achieved this setup at the Vienna University of Technology’s Depart-
ment of Geodesy and Geoinformation (formally known as the Institute of Photogrammetry and
Remote Sensing). They have a rolling-archive of ASAR Level 1b data (both WS and GM) which
any member of the European scientific community may download from ESA free of charge. In-
house as well as commercial software are used to produce an automatic pre-processing chain
whose output is a radar database. Here follows an account of the pre-processing steps that are
involved in producing an easily accessible and computationally efficient ASAR WS database.
A flow chart of the pre-processing steps is given in figure 3.1.
The pre-processing chain is divided into calibration and georeferencing. The former en-
compasses the conversion of the raw digital numbers, as measured by the sensors, into their
representative physical quantities. The latter comprises the assignment of the appropriate spa-
tial coordinates to each pixel. Signal calibration is the process of extracting any systematic
errors that arise due to radiometric and geometric properties of the data (Rees 2006). Standard
SAR data, such as Envisat ASAR WS, are ellipsoid geocoded products; they take the Earth’s
curvature into account but are yet to be rectified with respect to the side looking geometry of
the SAR system and the effects of local topography. Both radiometric and geometric accuracies
are terrain dependent (Loew and Mauser 2007) and need to be corrected appropriately in the
pre-processing stage. The pre-processing steps of the ASAR WS data are as follows:
i. Radiometric calibration. The detected active microwave signal describes amplitude and
phase properties at certain polarisations of the radar reflectivity (or backscatter coeffi-
cient) of the targeted scene (Oliver and Quegan 1998). These properties need to be de-
coded first, before this signal can then be converted into a radar image. Imaging radar
sensors measure a property called radar brightness, β0 (Woodhouse 2006). Sensor internal
calibrators relate the digital numbers (DNs) of each pixel to a calibrated radar brightness
27
Data Chapter 3. Mapping inundation
Figure 3.1: Flow chart of the pre-processing chain undertaken at the Vienna University of Tech-
nology (Sabel 2007)
value. In flat terrain, β0 expresses the backscatter coefficient σ0 [dB] via the scattering
area, A [m2]
σ0 =
β0
A
= β0 × sin θi (3.1)
where θi is the local incidence angle and A may be approximated as A~(sin θi)−1 as this
accounts for changes in scattering area due to near range to far range region changes. As
discussed in section 5.19, the backscatter coefficient is defined by scattering mechanisms.
Different incidence angles alter these scattering mechanisms; as do changes in the local
scattering area (Loew and Mauser 2007). These need to be resolved in the radiometric
calibration stage. In addition, terrain induced geometric distortions need to be corrected.
ii. Geometric correction. It is important to stress that SAR systems are side-looking sensors,
from which a radar intrinsic geometric distortion arises, the slant-range distortion. Areas
that slope upwards seem to be narrower than areas of equivalent size that slope down-
wards. This is known as foreshortening (Woodhouse 2006). Further geometric distor-
tions arise due to topography. The geometric correction procedure uses a digital eleva-
tion model (DEM) to compensate for any distortions that arise due to the local terrain.
Changes in relief (slope and aspect) result in a geometric bias of the incident radar beam,
which modifies the backscatter intensities (Loew et al. 2006). These distortions are known
as layover and shadowing and also result from the radar’s slant-angle. The former mis-
28
Chapter 3. Mapping inundation Data
Table 3.1: Characteristics of the ASAR WS processing grid (Sabel et al. 2011).
Datum Map projection Sampling distance Origin
WGS84 Plate Carrée 2.4 arc-seconds 180° W 90° S
interprets a scatterer at an elevated terrain height to be located at a reduced across-track
coordinate. This is due to the radar using the pulse delay time to determine across-track
coordinates of targets (Rees 2006) and the elevated scatterer appearing to be closer as it
has a reduced delay time. The latter distortion is a radar viewshed issue; certain areas
of the surface remain hidden to the radar sensor and are said to lie in the radar shadow.
While layover may be rectified by a suitable DEM, radar shadow can only be reprojected
to the appropriate geographic area unless data from different look directions can be ac-
quired - this, however, is constrained by the satellite’s orbital path and only similar view
directions are possible at the poles (Woodhouse 2006). An image is said to be geocoded
only once the geometric distortions have been removed and each pixel location is geo-
referenced to its geographic location. Once this has been achieved, the image matrix will
consist of a raster with the two-dimensional cell coordinates azimuth and range (Oliver
and Quegan 1998). Cartesian map coordinates are calculated by applying a transfer func-
tion to this geometrically corrected image matrix (Sabel 2007).
iii. Resampling. Each image is resampled to a fixed grid (sample distance 2.4 arc-seconds,
which corresponds to a pixel spacing of 75 m at the equator) and stored in a database.
The characteristics of the processing grid are given in table 3.1. The 150 m resolution
WS data are oversampled to a pixel spacing of 75 m. Bilinear interpolation is used to
assign the new set of pixel values (Sabel 2007). This WS grid, consists of 0.1° × 0.1°
gridboxes, resulting in 3600 columns and 1800 rows globally. The data of each gridbox
is stored as a binary file in time series stacks within the database. Appropriate meta-
files on local incidence angles, date and time of data acquisitions, the nature of the orbit
(ascending/descending) and log-files containing a list of the processed images are also
stored in this database (Sabel et al. 2011). The datasets are therefore resampled to pre-
defined tiles and stored as layer stacks.
iv. Normalisation. The radar backscatter coefficient, σ0, is a function of incidence angle, θi.
Therefore, the scattering field is defined by the surface roughness and the incidence angle
(Woodhouse 2006). This angular dependency of the backscatter coefficient needs to be
normalised in order to be able to compare and analyse measurements from images taken
at different incidence angles, as well as those taken within the same image at different
range positions (Sabel 2007). The chosen angle of normalisation is θ = 30°. Hence, a
29
Data Chapter 3. Mapping inundation
backscatter normalisation model, described by equation 3.2, is applied to the data.
σ0(30) = σ0(θi)− α(θi − 30) (3.2)
α is the so-called slope; this parameter determines the dependency of the backscatter
coefficient on the local incidence angle in units of decibel per degree. The retrieval of
both σ0(30) and α is possible with sufficiently large time-series data. The normalisation
angle, θ = 30°, was chosen to represent the mean incidence angle range, where incidence
angles of the ASAR instrument are within the range 17° to 42° (Sabel et al. 2012).
3.1.3 ALOS Panchromatic Remote-sensing Instrument for Stereo Mapping
The Advanced Land Observing Satellite (ALOS) was a Japan Aerospace Exploration Agency
(JAXA) mission, launched into a sun-synchronous orbit, with a revisit period of 46 days, in
2006. The satellite completed its operations in spring 2011, after an expected 5 year lifetime.
Its payload contained three sensors: The Advanced Visible and Near Infrared Radiometer
type 2 (AVNIR-2), the Phased Array type L-band Synthetic Aperture Radar (PALSAR) and
the Panchromatic Remote-sensing Instrument for Stereo Mapping (PRISM) (JAXA 2011). The
PRISM sensor is a high resolution panchromatic (λ = 0.52 − 0.77 µm) radiometer. PRISM’s
viewing was restricted to latitudes, l, −82° < l < 82°. The sensor had three individual optical
systems (forward, backward, nadir) (Igarashi 2001). The nadir radiometer consisted of 6 CCD-
arrays (nominal overlap of adjacent arrays was 32 pixels) comprising 4992 pixels. The forward
and backward radiometers were both made up by 8 CCD-arrays that contained 4928 pixels
(Schneider et al. 2008). Images taken at nadir have a spatial resolution of 2.5 m. The stereo
triplet optics were conceived in order to be used for DEM production through the application
of photogrammetric techniques.
ALOS PRISM triplet images were acquired on 09 July 2007 over the Yamal peninsula (Vaskiny
Dachi area ROI ALOS) and were orthorectified. Their projection is given in UTM 42N (Datum
WGS84). These PRISM images are used for the study presented in section 3.4.2 as validation
data for evaluating the accuracy of the radar lake classification method described in section 3.2.
The ALOS PRISM data have been made available for this PhD through the JAXA ALOS
PI agreement P090 with the Vienna University of Technology, Department of Geodesy and
Geoinformation, Research Groups Remote Sensing and Photogrammetry.
3.1.4 TerraSAR-X
TerraSAR-X is a German radar satellite that has been in operation for over 5 years, exceeding
its nominal lifetime. Due to its success, it has been granted an extended lifetime of up to 3
years. This SAR satellite operates in four modes: High Resolution SpotLight (1 m spatial res-
olution), SpotLight (2 m spatial resolution), StripMode (3 m spatial resolution) and ScanSAR
(18.5 m spatial resolution). As its name implies, TerraSAR-X is a X-band instrument with its
30
Chapter 3. Mapping inundation Data
wavelength centred on λ = 3.1 cm and it may be used in any of the three single, dual or quad
polarisations. TerraSAR-X follows a near-polar, sun-synchronous orbit with an 11 day revisit
cycle (Mather and Koch 2011).
TerraSAR-X data were acquired on 13 August 2008 over the Yamal peninsula (Vaskiny
Dachi area ROI Terra). The data have been reprojected into UTM 42N (Datum WGS84) and
are used in section 3.4.3 as validation data for a map accuracy assessment, as well as in sec-
tion 4.3 for investigating limnicity and the lake number distribution of the ROI Terra area. The
raw TerraSAR-X data were provided under a user agreement with the German Aerospace Cen-
tre (DLR). Pre-processing of the TerraSAR-X data was undertaken by colleagues at the Vienna
University of Technology, Department of Geodesy and Geoinformation, Research Groups Re-
mote Sensing and Photogrammetry, who are the principle investigators under the DLR user
agreement.
3.1.5 Data availability
The advantages of Remote Sensing products are undeniable, however, their suitability as a
long-term monitoring method are influenced by data availability and satellite lifetime (Bartsch
et al. 2012). Long-term monitoring requires continuous, comparable and self-consistent data.
The Envisat satellite was launched into orbit in spring 2002, transmitting data for over ten-
years, until, in 2012, communications with the satellite unexpectedly broke off. The instru-
ment’s lifetime was slightly shortened by this malfunction; it had been expected to remain
operating until the end of 2013. Fortunately, the wheels had already been set in motion for a
continuity mission, the Sentinel missions. There will be 6 Sentinel satellites in total. Sentinel-1A
is due to be launched at the end of 2013, followed by its twin, Sentinel-1B, in 2014. Both satel-
lites will be, as ASAR was before them, C-band synthetic aperture radar sensors. Continuity
missions are the foundations for long-term EOs, ensuring that comparable data are available
beyond a satellite mission’s lifetime.
Furthermore, data availability is dependent on temporal resolution and signal quality. This
is an issue that will vary with sensor type. A common problem for multispectral instruments
is the effect atmospheric propagation has on the detected signal. For visible-infrared (VIR)
sensors, the received signal is not the reflectance of the solar radiation (which is the quantity
of interest) but it is the radiance attenuated by the atmosphere. It is, therefore, important that
atmospheric correction procedures are applied to these images before any further analysis can
take place. In addition, clouds present a VIR obstruction; contaminated pixels that need to be
removed by a so-called cloud mask. This is a crucial issue for Arctic research, where cloud-free
days are all but minimal, and it is an opportunity for active microwave sensors to come into
their own. Microwaves belong to the long-wavelength regime of the electromagnetic spectrum
(mm - m). Not only do active microwave sensors provide their own illumination (hence the
name ‘active’), allowing them to monitor throughout day and night, but, importantly, they can
also penetrate through clouds, which increases data availability.
31
Data Chapter 3. Mapping inundation
Figure 3.2: ASAR WS (HH polarisation) Level 1b coverage over the period December 2004 -
February 2009 adapted from Sabel et al. (2011)
Availability of comparable data is absolutely crucial to timeseries analysis. This is where
Envisat ASAR WS demonstrates its value as a data source over the Yamal peninsula. Archived
material is abundant. The global level 1b coverage from 2004 to 2009 is given in figure 3.2.
The number of ASAR WS scenes covering the Arctic available for the summer months July and
August in 2007 and 2008 was presented in the paper Bartsch et al. (2012) and is reproduced
in figure 3.3. Furthermore, ASAR WS’ data are high temporal resolution data, and near-daily
images may be acquired. The temporal resolution of the ALOS PRISM and TerraSAR-X sensors
are far coarser than ASAR WS, as their acquisition time intervals are much longer. However,
these data are used during this PhD research as ancillary data for comparative studies. In these
studies, individual ASAR WS images is compared to ALOS PRISM and TerraSAR-X images
with the respective time stamp. The important point is that these reference data were acquired
on the same day as the ASAR WS data and may hence be used as a means for classification
accuracy assessment.
The choice of satellite instrument will obviously also depend on access to data in terms of
their commercial cost - European Space Agency data are free of charge to the scientific com-
munity, as are other data (e.g. Landsat), however, some data are only available through li-
censed agreements, others must be bought. As mentioned above, both ancillary data (ALOS
PRISM and TerraSAR-X) were acquired by Vienna University of Technology Principle Inves-
tigator agreements with the Japanese and German aerospace agencies, respectively. Envisat
ASAR WS was downloaded from ESA free of charge.
32
Chapter 3. Mapping inundation Data
Figure 3.3: Number of ASAR WS (HH polarisation) scenes covering the Arctic available for the
summer months July and August in 2007 and 2008, as presented in Bartsch et al. (2012).
33
Water body detection method Chapter 3. Mapping inundation
3.2 Water body detection method
Having detailed the data sources that are used for this research, the choice of classification
method needs to be considered. As established, Envisat ASAR WS is the primary data source,
the classification method is outlined for active microwave data. This section explores the
method by which water bodies are differentiated from their surroundings in radar imagery
and hence actually classified as water.
As mentioned in section 3.1.2 the backscatter coefficient is a function of incidence angle,
σ0 = f(θi). This was a simplified statement used to explain the need for normalisation of the
backscatter coefficient. This concept of backscattering will be revisited here, to support the
choice of water body classification technique, explaining the detection method of open surface
water bodies using active microwave sensors.
Scattering of electromagnetic radiation is a complex phenomenon. Radar remote sensing
distinguishes between surface scattering and volume scattering. The former accounts for the
scattering of the incident wave at the surface boundary (e.g. water, soil (except very dry soils
such as desert sands) and wet snow); the latter describes the scattering of microwaves off inho-
mogeneous media (e. g. vegetation canopies, snow flakes in a dry snow pack and air/gas bub-
bles in ice). Open surface waters, for obvious reasons, play an important role for analysing sea-
sonal lake dynamics with active microwaves, as do the freeze/thaw cycle and wet snow. These
concepts will be revisited in section 6.2.2 of chapter 6, where an analysis of the Yamalo-Nenets
Autonomous District takes advantage of scatterometer data to determine snow-off dates.
3.2.1 Open surface water, a surface scatterer
Apart from penetration depth related to the sensor’s wavelength and dielectric attenuation,
surface scattering models consider the surface roughness. Surface roughness may be consid-
ered as a deviation in height from the average elevation of a horizontal surface. The degree
of roughness will determine whether a surface may be considered smooth or rough, which in
microwave remote sensing is determined by the Fraunhofer criterion given by
hsmooth <
λ
32 cos θ
(3.3)
Within this criterion surface scattering is considered to be specular reflection, that is radi-
ation with incident angle θi and azimuthal angle φi will be scattered in the direction θr = θi
and φr = φi − pi only (Rees 2001). Indeed, this simple scattering model may be applied to
surface water, such as lakes and rivers. The physical principles of surface scattering enable the
application of a straightforward surface water body classification method, known as density
slicing or thresholding. According to the laws of reflection set out above, the radar beam is
reflected away from the sensor when it is incident on smooth open surface water bodies. This
results in low backscatter values. The entire SAR image will be composed of areas of low grey
34
Chapter 3. Mapping inundation Water body detection method
level values (water) as well as higher, normally distributed grey level values (surroundings).
The image’s histogram will follow a distinct bimodal backscatter distribution. When the wa-
ter fraction of a region of interest is below 50%, which is archetypal (Bartsch et al. 2012), the
bimodal backscatter distribution’s dominant mode results from diffuse scattering of the sur-
roundings. Applying a threshold value to the bimodal backscatter distribution will delineate
water bodies from their surroundings.
Given this distinct bimodal backscatter distribution, a simple classification technique known
as density slicing may be applied to delineate these water bodies from their surroundings. The
method uses threshold values (in this case only one threshold value that distinguishes the wa-
ter class from its surroundings) to classify image data. Density slicing is of course a form of
image enhancement; it allows certain features to be made more visible, however it does so with-
out performing any modifications to pixel information (Campbell 1996). Bartsch et al. (2008)
discuss the application of density slicing of open surface water bodies in tundra and taiga re-
gions. They use ASAR WS data to map water bodies within a large transect of Siberia, from the
Taymir peninsula to Lake Baikal. They validate their data with multiple datasets at different
scales (aerial photographs [<1:10 000], topographic maps [1:200 000], MODIS land cover clas-
sifications [500 m spatial resolution] and the Global Lakes and Wetlands Database [1:1 000 000
- 1:3 000 000] devised by Lehner and Döll (2004)). Comparison to aerial photographs resulted
in a kappa coefficient of κ = 0.82 and an overall classification accuracy of 95%; the kappa co-
efficient, κ, being a measure of inter-rater agreement, incorporating an adjustment for random
distribution and allocation (Cohen 1960) - this topic is discussed in section 3.4.2, where a map
accuracy study is presented. The level of confidence for water body classification of ASAR WS
data, as demonstrated by Bartsch et al. (2008), is a positive outcome and provides an adequate
base for further studies on mapping water body extent using the density slicing classification
method and, moreover, monitoring water body dynamics using these data. Nonetheless, the
theme of threshold-classification needs to be revisited, as it cannot, without justification, be
taken as a given methodology.
It is a common problem in data analysis to extract a signal despite the presence of back-
ground noise. In this case, the signal extraction (i.e. water body classification) may seem
straightforward, but appearances are deceiving. It is true that water bodies are associated
with the secondary peak of the bimodal backscatter distribution, while the surroundings may
be associated with the dominant peak (see figure 3.5). A threshold between these two modes
should allow the differentiation between water bodies and their surroundings. However, the
crucial aspect of the thresholding procedure is the choice of threshold - which is a less obvious
concept. The decisive question for density slicing is therefore: What is the optimum threshold?
This is an open question. Nonetheless, determining this optimum threshold, which will cap-
ture all the required information on water without introducing excessive noise into this water
signal, is essential.
The methods by which one might go about choosing the optimum threshold value include
choosing the saddle point value (trough value) or - as a more sophisticated approach - fitting a
35
Water body detection method Chapter 3. Mapping inundation
function to the distribution and using statistical parameterisation to extract a meaningful value.
Returning to basics and setting aside the a priori knowledge of the backscatter distribution
being bimodal, it is reasonable to expect the number of counts Nx, of a set of data {x} to be
proportional to the sum of the water signal and the background (Sivia and Skilling 2006). A
signal sitting on a flat background of amplitude B is given in figure 3.4. This is a statement on
noise, where the background is noise. In the given case of the bimodal backscatter distribution
of the active microwave data, the background is actually not noise as such. Landgrebe and
Malaret (1986) state noise in remote sensing systems may be associated with variations in the
spectral response which are non-information bearing. Noise may also arise due to atmospheric
effects, sensor specific effects such as thermal noise (known as preamplifier noise processes),
as well as quantisation noise related to the conversion of the digital signal (Landgrebe and
Malaret 1986). However, here, the set of data {xBG} associated with the surroundings (i.e.
related to the diffuse backscatter coefficients) are defined to be the background. This approach
is atypical but is meant to enable a clearer understanding of the attempt to delineate water
from the surrounding land class. Those data related to water are the signal. Considering the
argument, and later it will be shown that this is a reasonable assumption, that the backscatter
distribution follows a normal (i.e. Gaussian) distribution, then the expected number of counts
are given by
N = A exp
(−(xw − µ)2
2σ2
)
+B (3.4)
whereA is the amplitude of the signal, xw is the backscatter coefficient associated with water, µ
is the mean and σ is the standard deviation of the Gaussian function (Sivia and Skilling 2006).
However, as stated above, the background is not noise but is associated with diffuse scattering
and hence also follows a distinct distribution, so that equation 3.4 may be rewritten as
N = A exp
(−(xw − µ)2
2σ2
)
+ f(x)surroundings (3.5)
where fsurroundings is the backscatter distribution of the surroundings.
Using the a priori knowledge that the backscatter distribution is bimodal and assuming
fsurroundings also follows a Gaussian distribution then a double-Gaussian function may be fitted
to the backscatter data in order to determine the parameters that will enable the choice of a sta-
tistically meaningful threshold value. This parameter estimation model is presented by Bartsch
et al. (2012) who use a double-Gaussian fit to determine a threshold value. An example from
the summer 2007 is given in figure 3.5. As noted in section 3.1.2, radar backscatter needs to be
normalised so that images taken at different incidence angles may be compared, and also so
that different range positions within the image are comparable. Therefore, density slicing has
to be applied to normalised radar images in order to arrive at a purposeful threshold.
The double-Gaussian model is a program that outputs two standard deviations, one for
each mode (as the double-Gaussian function is a superposition of two Gaussian functions with
different means (Lewellen and Yoh 1993)). The standard deviation of the dominant mode,
36
Chapter 3. Mapping inundation Water body detection method
Figure 3.4: Sketch of a signal on a flat, unknown background of magnitude B, adapted from
Sivia and Skilling (2006).
σdominant, is used for determining the threshold value. The cut-off value of the dominant mode’s
trail is taken at approximately 3 standard deviations difference to the mean of the dominant
mode. At a 99% confidence level, this is generous in terms of the proportion of the surround-
ings’ signal that is rejected to avoid misclassifications, while still retaining enough of the water
body signal. In the given example from 20 August 2007 (see figure 3.5) µdominant − 3σdominant =
−14.4 dB.
For efficient analysis of ASAR WS’s near-daily images, a batch process needs to be devel-
oped. Hence the classification method is an automatic procedure. Each ASAR WS image will
however have similar, but clearly distinct optimum thresholds. Computer efficiency comes
at the cost of classifying each image according to its own, image specific, optimum threshold
value. The choice of this threshold value has fallen on -14 dB, after analysis of several images
and applying the µdominant−3σdominant threshold condition. A method first presented in Bartsch
et al. (2012). Furthermore, a sensitivity analysis of this choice is presented in section 3.4.1 where
the results of an increase in threshold value are examined, to support the selection of a -14 dB
threshold value.
It is noted here, that the double-Gaussian model is but one model that may be used as a
means of parameter estimation. Other studies have used alternative approaches, such as Mat-
gen et al. (2011) who describe the secondary mode of the backscatter distribution as following
a Gamma probability density function (pdf) with two parameters k and θ:
f(σ0/k, θ) =
(σ0 − σ01)k−1
θkΓ(k)
· exp
(
−(σ
0 − σ01)
θ
)
(3.6)
where σ01 is the minimum backscatter value within the image. Matgen et al. (2011) apply a
37
Water body detection method Chapter 3. Mapping inundation
non-linear curve-fitting algorithm to estimate the parameters k and θ. The Gamma distribution
mode is given by
σ0m = (k − 1) · θ + σ01, when k ≥ 1 (3.7)
Rearranging for theta enables a simplified fitting procedure. This method requires some tweak-
ing, as the gamma probability density function is only valid for positive values and therefore
the backscatter values need to be manipulated accordingly. Furthermore, this parameter es-
timation method only considers the secondary mode. Matgen et al. (2011) use an iterative
algorithm that progressively increases the threshold value, each time computing the difference
between the actual distribution (the backscatter histogram) and the Gamma-distribution, until
the cumulated difference is below 1% of the cumulated probability (the integral of the Gamma
pdf). Furthermore, the threshold value is then determined using a region growing approach,
allowing all water pixels associated with the seed region to be included in the final water body
classification. This adaptive approach is extremely work-intensive and complex. For our pur-
pose, it is excessively so, as the automated batch procedure is not of an adaptive nature; region
growing would have to be applied to each image individually. The choice of parameter esti-
mation method was therefore the double-Gaussian fit, which was applied to several images
before deciding on the threshold value of -14 dB. As noted earlier, a sensitivity analysis of this
threshold value is presented below in section 3.4.1.
3.2.2 The effects of aquatic, waterborne and marginal vegetation
A crucial aspect of water body classification by means of density slicing is that the method can
only determine open surface water bodies. Vegetation on the banks and at the borders of lakes,
as well as aquatic, waterborne vegetation alter the backscattering mechanism - specular reflec-
tion is superseded by double bounce reflection between the water and the vegetation (i.e. a
form of volume scattering). This results in the loss of a distinct radiometric water signal associ-
ated with the secondary mode of the backscatter distribution. The importance and magnitude
of this change in scattering mechanism is dependent on vegetation properties and restricted by
sensor characteristics (Horritt et al. 2003). Diffuse scattering due to double bounce is specific
to ’short’ wavelength radar systems. Other sensors (e.g. ALOS PALSAR), operating at longer
wavelength regimes (L-band, ∼1 - 2 GHz), have a greater penetration depth, so that flooded
areas below vegetation may also be distinguished (Evans et al. 2010). However, Envisat ASAR
is a C-band instrument that will experience diffuse backscattering due to emerging vegetation.
Small incidence angles may minimise the C-band limitations in penetration, as the path that
the incident and reflected beams have to travel is curtailed (Töyrä et al. 2001). The incidence
angle of θi = 30° is thought to be sufficiently small.
Nonetheless, the effects of vegetation on water body classification are not trivial and need to
be taken into account, especially when examining the dynamics of water bodies over the course
of the spring/summer period, as this period is also the growing season. Changes in water body
extent may indeed be related to vegetation phenology, possibly arising in misinterpretation.
38
Chapter 3. Mapping inundation Water body detection method
3.2.3 The effects of wind
Strong winds are common to high latitudinal (and altitudinal) areas. They are a major geomor-
phological agent that play a significant role in the redistribution of precipitation in the form of
snow (Seppälä 2004). The distribution of snow is important as surface runoff from the snow
pack is a crucial contribution to seasonal inundation and spring flood dynamics. As such, ex-
amining snow and its distribution is essential to an analysis of seasonal water body dynamics.
However, the influence of wind on snow redistribution in the north of West Siberia will be
discussed at a later stage. First, the general effects that wind has on active microwave moni-
toring methods are to be examined. Radar has long been an established system for studies of
wind velocities over water surfaces. In fact wind-scatterometers, which are active microwave
sensors that, unlike SAR, make detailed measurements of a target’s radar cross-section at the
cost of spatial resolution, are specifically designed for applications to be able to calculate ocean
surface wind vectors. This is done by exploiting the wave ripples that strong winds create on
the ocean surface (Woodhouse 2006). The concept of a scatterometer will be revisited in chapter
6 where data from such an instrument are used to investigate snow conditions. However, with
respect to this research, which takes advantage of a SAR imaging system and implements a
density-slicing classification technique, it is important to note that wave ripples on the surface
of open water bodies are a prominent example of surface roughening by wind. As mentioned
in section 3.2.1, once the Fraunhofer criterion is exceeded the surface is thought to be rough.
The non-coherent surface scattering component is dominant. Surface roughness is proportional
to this non-coherent component. The less the Fraunhofer criterion is fulfilled (i.e. the higher
hsmooth), the rougher the surface. The backscatter distribution ceases to be bimodal and the
separability of open water surfaces from the surrounding land cover class is lost (see figure
3.6).
In effect, this problem may be remedied by a large dataset, as data taken on wind affected
days are purely unused data and as long as there are other data within the time period of
classification, the issue is minimised. High data availability will ensure that change detection of
water body dynamics are not misinterpreted. Accounting for wind bias is therefore particularly
important in an analysis of seasonal surface water extent dynamics and is further discussed
below in section 4.2.4, after addressing the method of change detection.
39
Water body detection method Chapter 3. Mapping inundation
0
2.0e+03
4.0e+03
6.0e+03
8.0e+03
1.0e+04
1.2e+04
1.4e+04
1.6e+04
1.8e+04
-18 -16 -14 -12 -10 -8 -6 -4
Fr
eq
ue
nc
y
Backscatter value [dB]
µdm - 3σdm= -14.4 dB
µdm= -9.41 dB
Data
Gaussian fit
Figure 3.5: Bimodal backscatter distribution for a tundra region (top panel) and the correspond-
ing ASAR WS image from 20 August 2007 (bottom panel), as also shown in (Bartsch et al. 2012).
µdm is the mean of the dominant mode and σdm is its standard deviation.
40
Chapter 3. Mapping inundation Water body detection method
Figure 3.6: Timeseries of Envisat ASAR WS data over a tundra landscape, showing the loss of
the bimodal backscatter distribution due to increased open water surface roughness on a windy
day (25.07.2007) (Bartsch et al. 2012). The grey level stretch of the backscatter distribution
shown in the top panel goes from zero (white) to -21 dB (black).
41
Case 1: Local scale Chapter 3. Mapping inundation
Table 3.2: Geographical coordinates for ROI 1.
Corner points Latitude (N) Longitude (E)
Upper left 70° 30’ 2’’ 67° 53’ 59’’
Lower right 69° 48’ 5’’ 68° 59’ 57’’
3.3 Case 1: Local scale - Vaskiny Dachi, Yamal peninsula
The local scale focuses on three partially overlapping study areas on the Yamal peninsula.
These have been labelled regions of interest ROI 1, ROI ALOS and ROI Terra. The largest
of these three areas covers approximately 8 000 km2. Its coordinates are given in table 3.2. This
area is referred to as ROI 1, the label having been introduced in Trofaier et al. (2012). ROI 1 is
one of two regions that had been investigated in the Trofaier et al. (2012) study - the second re-
gion is referred to as ROI 2. The following section consists of three feasibility studies. Despite
each of these being addressed as local scale studies, not all the studies presented here cover
the same area, or indeed the entire ROI 1 area. The location of each local scale study area are
indicated in figure 3.7, where ROI 2 is also shown for completeness.
ROI 1 is the largest area of investigation. Its exact borders were chosen after the first tests
were run on the water body classification method of the ASAR WS data (which was first run
over the entire Yamal peninsula). Furthermore, ROI 1 encompasses the location of a Russian
research camp known as Vaskiny Dachi, which lies in the East of the study area (~70° 17’ N, 68°
54’ E). Active layer measurements and permafrost borehole temperatures at this site are contin-
uous and records date back to the 1990s (Leibman et al. 2012). More detail on the characteristics
of the Vaskiny Dachi region will be presented in chapter 4 (section 4.1), in which water body
change detection and seasonal limnicity changes are considered, as it is deemed more helpful
to discuss these in their geographical and geological context.
Some of the feasibility studies presented here use ancillary data that do not cover the entire
extent of ROI 1. They do, however, lie within the borders of ROI 1, which is why it was decided
to refer to these study areas as Vaskiny Dachi as well. The corner coordinates of each study area
are given in tables 3.2, 3.3 and 3.4 for ROI 1, ROI ALOS and ROI Terra, respectively.
3.4 Feasibility studies
It was stated above that the density-slicing classification procedure depends on the chosen
threshold of -14 dB. Previous studies have explored the threshold approach itself and found
it to produce satisfactory results (Bartsch et al. 2007a, 2008). As discussed in section 3.2.1, the
choice of threshold is not a straightforward matter, however, this was followed up with the
reasoning for the choice in the -14 dB threshold value. Nonetheless, testing the feasibility of
42
Chapter 3. Mapping inundation Feasibility studies
Figure 3.7: Local Scale study areas. ROI 1, ROI ALOS and ROI Terra partially overlap and were
used for the feasibility studies. ROI 2 exhibited strong wind biases.
43
Feasibility studies Chapter 3. Mapping inundation
Table 3.3: Geographical coordinates for ROI ALOS.
Corner points Latitude (N) Longitude (E)
Upper left 70° 16’ 28’’ 67° 45’ 52’’
Lower left 70° 07’ 36’’ 67° 46’ 24’’
Lower right 70° 07’ 45’’ 68° 10’ 20’’
Upper right 70° 16’ 36’’ 68° 10’ 01’’
Table 3.4: Geographical coordinates for ROI Terra.
Corner points Latitude (N) Longitude (E)
Upper left 70° 36’ 36’’ 67° 46’ 38’’
Lower left 70° 09’ 08’’ 68° 09’ 11’’
Lower right 70° 13’ 30’’ 68° 57’ 43’’
Upper right 70° 41’ 03’’ 68° 36’ 08’’
the density-slicing classification method as well as the chosen threshold value is essential. Un-
certainties in a classification (or any EO product for that matter) need to be made explicit. This
is especially important to EO products that may be accessed by the modelling community as
input and/or validation datasets (Comber et al. 2012). This section therefore deals with accu-
racy testing and error analysis. Two approaches are presented here. Firstly, for the purpose of
examining the chosen threshold, the density-slicing classification technique is assumed to be
an incontestable method. This is a sensitivity test that examines the optimum threshold value,
i.e. the threshold value that resolves all the water bodies without introducing excess noise in
the form of backscatter pixels from the land class. Secondly, map accuracy tests are undertaken
that compare the density-slicing technique to the unsupervised classification technique known
as ISODATA. ISODATA is a clustering algorithm that uses statistical methods to quantify pixel
similarities and hence classify the image into a set of classes, their number having been spec-
ified in advance, however without any preconception of class attributes and labelling (Rees
2001). The ISODATA algorithm depends on the statistical separability of the clusters. An a pos-
teriori approach allows the identification of each class with its associated label (Richards 1993).
Map accuracy testing is done for two separate studies, each comparing different sensors and
classification methods.
3.4.1 Optimum threshold sensitivity test
The first feasibility study is undertaken for the entire ROI 1 area (coordinates given in table
3.2). This study is an optimum threshold sensitivity study and is set up to examine how much
44
Chapter 3. Mapping inundation Feasibility studies
the classified water bodies increase in size by increasing the threshold value, while at the same
time trying to minimise the signal from the surroundings. In effect, if the signal from the
surrounding land class is treated as ‘noise’ (as claimed in section 3.2.1) then a signal-to-noise
problem arises. The ‘noise’ from the land class needs to be minimised, in order to maximise
the signal from the water bodies. This is done by choosing an optimum threshold value, as
determined by the sensitivity study.
The sensitivity test is performed on a set of randomly chosen lakes. This procedure was
done in the ImageJ software environment (Abramoff et al. 2004). The optimum threshold
for five ASAR WS images (see table 3.5) was determined manually by examining a range of
threshold values, t. The software identifies so-called particles (polygons with associated size
attributes). The optimum threshold of a set of 12 randomly selected lakes, each associated with
a particle, was explored. The software was then used to calculate the area and perimeter of each
lake for each threshold value. The error in the threshold value was investigated by analysing
the change in mean lake diameter and area as a function of threshold value.
dA = p¯ · dw |t1→t2 (3.8)
where dA is the change in area of the lake, p¯ =
1
2
(p1 + p2) is the average perimeter and dw is
the width between p1 and p2 after a change in threshold value from t1 to t2. Rearranging and
integrating gives us this perimeter width. The quantity of interest is this perimeter width, as it
represents the change in mean lake diameter.
dw =
∫ A2
A1
dA
p¯
(3.9)
The manually determined optimum threshold for each lake in each image showed that
this optimum threshold varied significantly from lake to lake and was also image dependent.
The average optimum threshold value of the entire set is -12.34 dB. However, the daily values
range from -15 to -9.8 dB. Figure 3.8 demonstrates this large spread. Each of the twelve lakes
is depicted in a different colour. A spline fit to the daily optimum threshold value averages is
given in grey. The standard deviation of the entire set of optimum threshold values was found
to be 1.22 dB. It was therefore decided to proceed with a threshold sensitivity analysis in steps
of 1 dB.
Six threshold values were examined (t = -10 to -15 dB, 1 dB interval). The perimeter width
of each lake for each threshold step within this range was calculated according to equation
3.9. Figure 3.9 shows the area of investigation for an ASAR WS image taken on 01 July 2007
given in UTM zone 42N (WGS 84) projection, as well as the particles that are identified for each
threshold value.
The sensitivity analysis of the threshold conditions showed that an increase in threshold
by 1 dB results in an average change in mean lake diameter of approximately 21.84 m and a
change in mean lake area of approximately 194.71 m2. The sensitivity test showed that the
45
Feasibility studies Chapter 3. Mapping inundation
Figure 3.8: Optimum threshold for the 12 lakes (each depicted by a different colour). A spline
was fitted to the daily threshold mean. The standard deviation of the entire threshold set was
found to be 1.22 dB.
Table 3.5: ASAR WS images used for sensitivity study.
Image Date: DD MM YYYY
1 01 07 2007
2 05 07 2007
3 06 07 2007
4 09 07 2007
5 17 07 2007
mean optimum threshold value is −12.34 dB, however, this mean value may not be the best
estimate of the image-specific optimum threshold, given the large spread of threshold values
established for each lake. For normally distributed threshold values, a σ error at a 68% con-
fidence level, would mean the optimum threshold value is −12.34 ± 1.22 dB. The choice of a
threshold value that is at the lower end of the optimum threshold range is related to the overall
loss in separability of the lakes from the surrounding land class per image. At a threshold of
approximately −10 dB, most of the selected lakes merge with neighbouring lakes. In addition,
the increase in noise from pixels that are classified as water but cannot be identified as lakes,
due to spatial resolution constraints, justifies the reasoning for setting the threshold value at
−14 dB (rounding down from −13.56 dB).
3.4.2 Data fusion I: Optical/Radar (C-band) accuracy test
The map accuracy study presented here uses optical imagery from the Japan Aerospace Explo-
ration Agency’s ALOS Panchromatic Remote-sensing Instrument for Stereo Mapping (PRSIM)
sensor as reference data to validate two different classification methods of the radar data. Both
the density-slicing method, as detailed in section 3.2.1, as well as an unsupervised classifi-
cation technique are examined. The aim of this study is to examine the map accuracy of the
thresholding technique and justify its applicability by relating the accuracy to that of a different
46
Chapter 3. Mapping inundation Feasibility studies
Figure 3.9: Normalised ASAR WS image from 01 July 2007 (top panel). Particles identified by
sensitivity study (red) for a 1 dB bin threshold range from -10 to -15 dB (bottom, left to right).
47
Feasibility studies Chapter 3. Mapping inundation
classification technique.
The ALOS PRISM images (nadir, forwards and backwards) used for this study were ac-
quired within the same time frame as an ASAR WS image. The spatial frame had to be reduced
from ROI 1 to a subset that contains the PRISM data. The coordinates of the spatial extent of
this accuracy test site are given in table 3.3. The corresponding date of acquisition is 9 July 2007.
In the absence of ground validation data it is of utmost importance that the validation dataset
(in this case the PRISM imagery) temporally coincides with the data that is being mapped.
The present study examines the accuracy of the thresholding technique and compares it to a
unsupervised ISODATA classification of the radar data, using the PRISM data on two accounts:
i. For creating an ‘inverse water mask’, i.e. a land mask.
ii. As validation data of the classification results.
The PRISM image was classified in the ERDAS IMAGINE software environment, using un-
supervised ISODATA classification. This classification algorithm first groups pixels into clus-
ters in the pixel-value space, and then optimises these clusters by migrating pixels from one
cluster to another in an iterative procedure according to the cluster means (Richards and Jia
2006). Initially 10 cluster classes (as well as a maximum of 10 iterations and 95% convergence
- the algorithm itself does not nominally converge (Rees 2013)) were specified resulting in 4
distinct classes associated with surface water. The remaining classes were merged into a mask-
ing class. Therefore, this mask is referred to as a land mask (see figure 3.10). An ISODATA
classification was then run over the masked ASAR WS image (using the aforementioned spec-
ifications again), clustering the backscatter values that - according to the PRISM classification
- are associated with surface inundation. Both lake classifications (ASAR threshold and ASAR
ISODATA) were vectorised and the lake perimeters inspected. The accuracy of each was deter-
mined by analysing a confusion matrix, using the optical PRISM image as validation; with no
field data available, this is an opportune and plausible validation method as the time frames
of PRISM and ASAR are identical. 200 random points were selected for both confusion ma-
trix analyses. In both cases, the ASAR data are the classified data and the PRISM data are the
reference data.
Two confusion matrices were constructed for both ASAR ISODATA and threshold classi-
fications. In both cases the user’s and producer’s accuracies were calculated. The numbers
are given in tables 3.6 and 3.7, respectively. The overall map accuracy of the ASAR ISODATA
classification is 91%. The ASAR thresholding classification is slightly worse with an overall
map accuracy of 89%. The Kappa coefficients, which incorporate an adjustment for random
distribution and allocation were also calculated according to Cohen (1960)’s definition
κ =
N
∑
k
xkk −
∑
k
xk+x+k
N2 −∑
k
xk+x+k
(3.10)
48
Chapter 3. Mapping inundation Feasibility studies
Figure 3.10: An unsupervised PRISM classification is used as a land mask. The surroundings
are masked out with the value zero, revealing only the backscatter values of the ASAR WS
image that should be associated with surface water.
49
Feasibility studies Chapter 3. Mapping inundation
Table 3.6: Confusion matrix comparing ASAR WS ISODATA classification with PRISM refer-
ence image.
Reference
Classified Water No water Total User Overall
Water 19 1 20 0.95
No water 17 163 180 0.906
Total 36 164 200
Producer 0.528 0.99 0.91
where xij is the confusion matrix with i rows and j columns,
∑
i
xii is the sum over the matrix
diagonal, xi+ =
∑
j
xij is the sum over all columns for row i, x+i =
∑
i
xij is the sum over all
rows for column j and N is the total number of tested pixels (Richards 1993). Monserud and
Leemans (1992) state that the Kappa statistic is an ideal method for rank ordering across mul-
tiple maps as well as across various classes within each map. They provide a table of Kappa
values and their associated degree of agreement. Rank ordering amongst several classes is ir-
relevant to this research, as only two classes (water and no water) exist. However, the Kappa
statistic may be used to compare the degree of agreement between observed data and reference
data taking chance agreement into consideration. The Kappa statistic is a normalised quan-
tity; when Kappa is equal to one there is perfect agreement, when Kappa is equal to zero the
map accuracy has effectively resulted by chance. The Kappa coefficients determined for the
present study are κISO = 0.63 and κthreshold = 0.60 for the ISODATA and threshold classifica-
tion, respectively. These two Kappa values fall within the category of good agreement accord-
ing to Monserud and Leemans (1992). By using the classified ALOS imagery as a land mask,
a more controlled ISODATA clustering of the backscatter coefficient could be achieved, hence
the ASAR ISODATA classification is improved. The thresholding technique on its own pro-
duces a Kappa coefficient of 0.60 with an overall map accuracy of 89%, which is a sufficiently
acceptable result.
This map accuracy study has shown that the density-slicing technique is a good method for
classifying radar data and that the threshold of -14 dB applied to ASAR WS data results in a
classification with satisfactory map accuracy.
3.4.3 Data fusion II: Radar (C-band/X-band) accuracy test
Having established that density-slicing is an adequate method for creating binary water/land
classifications from SAR data, a second map accuracy test is explored. This time, data from
two SAR sensors are compared, Envisat ASAR WS and TerraSAR-X. These sensors not only
differ in spatial resolution, but also vary spectrally (the sensor details have been discussed
50
Chapter 3. Mapping inundation Feasibility studies
Table 3.7: Confusion matrix comparing ASAR WS threshold classification with PRISM refer-
ence image.
Reference
Classified Water No water Total User Overall
Water 14 0 14 -
No water 22 164 186 0.88
Total 36 164 200
Producer 0.389 - 0.89
in section 3.1). ASAR is a C-band sensor, while TerraSAR-X’s wavelength regime lies in the
X-band. Being an X-band instrument, TerraSAR-X penetrates vegetation even less than the
ASAR C-band sensor (due to its shorter wavelength). However, at a spatial resolution of 3
m in StripMode, TerraSAR-X delineates open surface water to a much higher degree of detail
than ASAR WS. Hence, a second map accuracy test is undertaken, using classified TerraSAR-X
data for reference. The spatial extent of this study area covers the entire TerraSAR-X scene. Its
geographic coordinates are given in table 3.4.
First, the TerraSAR-X data needs to be classified. A scene, covering the Vaskiny Dachi
site, was georeferenced with the free NEST (Next ESA SAR Toolbox) software. This scene was
then post-processed in the ENVI/IDL software environment by initially converting the digital
numbers (DNs) in the amplitude image from linear to decibel (σ0 [dB] = 10 log10 σ0 [-]) be-
fore then applying a speckle filter. A statistical model is used to smooth speckle. The model
assumes that the speckle noise is statistically independent between pixels (Lee 1981). An adap-
tive filter, a Gamma filter, was applied using a 3 × 3 window size. Contrary to the multi-look
method, adaptive filters are able to preserve the texture of heterogeneous areas whilst sup-
pressing speckle in homogeneous areas. The Gamma filter is a Maximum A Posteriori filter, and
hence uses Bayesian statistics to reduce the influence of speckle. The filter works under the
assumption that the backscatter signal is Gamma distributed and as stated by Shi and Fung
(1994), the filter is given by
Rˆ(t0) =

I¯(t0) for CI(t0) < Cu
(α−L−1)I¯(t0)+
√
I¯2(t0)(α−L−1)2+4αLI¯(t0)
2α for Cu ≤ CI(t0) ≤ Cmax
I(t0) for CI(t0) > Cmax
(3.11)
where I is the observed image intensity (with variance σ2I and mean I¯) which is the product of
terrain reflectivity R(t) and speckle noise u(t) (with variance σ2u and mean u¯), these being func-
tions of spatial coordinates t = (x, y). t0 are the pixels that are to be filtered, L are the number of
looks and CI = σII¯ , Cu =
σu
u¯ and Cmax(t0) =
√
2Cu are variation coefficients. Senthilnath et al.
51
Feasibility studies Chapter 3. Mapping inundation
(2013) have found that the Gamma filter prevails over other speckle removal filters (such as Lee
and Frost) with respect to texture preservation, retaining edges and features in heterogeneous
regions.
Having applied a Gamma filter to the TerraSAR-X scene, the backscatter distribution was
examined in order to determine the optimum threshold for these X-band data. The backscatter
distribution and a double-Gaussian fit are shown in figure 3.11. Following the methodology
for determining a threshold value, detailed in section 3.2.1, a double-Gaussian was fitted to
the TerraSAR-X backscatter distribution to estimate a final set of parameters. The mean of the
dominant mode is at µdominant = −11.59 dB, σdominant = 3.07 dB. At a 95% confidence level,
the threshold is µdominant − 2σdominant = −17.73 dB. However, since the TerraSAR-X data are
provided in a single scene rather than a timeseries batch (as is the case for ASAR WS), a more
rigorous evaluation of a scene specific optimum threshold may be undertaken manually. It was
found, even at a more conservative threshold of -17 dB, that not only too much of the water
signal is eradicated but additional noise from the land class is still present. Therefore, a second
filter - a median filter - was applied. The median filter is a form of geometric enhancement; it is
an image smoothing technique that maintains edges by allocating the median brightness value
to each centre pixel of a filter region (Burger and Burge 2008). Often, it is useful to apply the
median filter twice consecutively, the second time with a larger kernel window. Iterative filter-
ing methods are used to further improve image quality (Forouzan and Araabi 2003). However,
as a Gamma filter had already been applied to the data, a single-iteration median filter was
used, and a 9 × 9 kernel window was chosen in order to increase the smoothing effects. After
applying this median filter, and the image having been deemed sufficiently smooth, avoiding
excess noise, the optimum threshold was found to be tTerraSAR-X = - 16.76 dB. Figure 3.12 depicts
the effects of this filter procedure and the results of the subsequent threshold classification.
The TerraSAR-X classification is then used as reference data for a map accuracy test of the
Envisat ASAR WS data. The map accuracy of the TerraSAR-X classification itself is not in-
vestigated, however, as long as no geo-coding issues have arisen, the effectiveness of these
reference data lies with the higher spatial resolution properties of the TerraSAR-X data versus
those of ASAR WS. The objective of this map accuracy test is to evaluate the applicability of the
density-slicing technique and the chosen threshold value to the ASAR WS data.
An ASAR WS image taken on the 13 August 2008, corresponding to the same time frame
as the TerraSAR-X data’s, was classified. First, since it is the threshold value used for batch
processing, the -14 dB threshold value was used to classify the data. As described in section
3.4.2, 200 random points were generated and a confusion matrix established. This matrix is
given in table 3.8. The overall map accuracy is determined to be 86%, similar to (3% worse
than) the result of the ASAR WS -14 dB threshold classification given in section 3.4.2. However,
closer inspection of the backscatter distribution, reveals that this ASAR WS scene’s backscatter
distribution is shifted and actually requires a different threshold value for optimisation. The
trough value is at -12.59 dB, while µdominant − 2σdominant = -12.39 dB.
Having established that -14 dB is not the optimum threshold for the ASAR WS scene from
52
Chapter 3. Mapping inundation Feasibility studies
 0
 1e+06
 2e+06
 3e+06
 4e+06
 5e+06
 6e+06
 7e+06
-35 -30 -25 -20 -15 -10 -5  0  5  10
Fr
eq
ue
nc
y
Backscatter [dB]
TerraSAR-X
double-Gaussian fit
Figure 3.11: Bimodal backscatter distribution of Gamma and mean filtered TerraSAR-X scene
from 13 August 2008.
53
Feasibility studies Chapter 3. Mapping inundation
Figure 3.12: Spatial sub-scene of TerraSAR-X image a) after applying a Gamma filter to the
radar image, b) classification of Gamma filtered image using a -17 dB threshold, c) after apply-
ing a 9 × 9 median filter to the Gamma filtered image and d) classification of median filtered
image using the manually determined optimum threshold of -16.76 dB.
54
Chapter 3. Mapping inundation Feasibility studies
Table 3.8: Confusion matrix comparing ASAR WS threshold classification (t = -14 dB) with
TerraSAR-X threshold classification as reference image.
Reference
Classified Water No water Total User Overall
Water 14 0 14 -
No water 28 158 186 0.85
Total 42 158 200
Producer 0.33 - 0.86
Table 3.9: Confusion matrix comparing ASAR WS threshold classification (t = -13 dB) with
TerraSAR-X threshold classification as reference image.
Reference
Classified Water No water Total User Overall
Water 22 0 22 -
No water 20 158 178 0.89
Total 42 158 200
Producer 0.52 - 0.9
13 August 2008, map accuracy tests of multiple ASAR WS water body threshold classifications
were carried out, again proceeding in 1 dB interval steps. Two further threshold values were
examined, t = -13 dB and t = -12 dB, expecting the latter to be the optimum choice as established
by the 95% confidence level above. Confusion matrices were constructed following the same
procedure as before. These are given in tables 3.9 and 3.10.
Unsurprisingly, given the a priori knowledge of the mean, as well as the standard deviation
derived from the double-Gaussian fit, the map accuracy increases with decreasing threshold.
Starting off with an overall accuracy of 86% for t = -14 dB, the map accuracy then jumps to
an improved 90% for t = -13 dB, before leveling off at 94% for t = -12 dB. This improvement
is exemplified in figure 3.14, where the TerraSAR-X data is overlain by the ASAR WS data.
Calculating the Kappa coefficients to adjust for random allocations of the sampling points, it is
found that each threshold classification has a Kappa coefficient of κ-14dB = 0.441, κ-13dB = 0.635,
κ-12dB = 0.783 respectively. Once again, according to Monserud and Leemans (1992), only fair
agreement for κ-14dB is found, while κ-13dB is associated with good agreement and κ-12dB with
very good agreement. This has important implications on the static thresholding approach used
in the batch processing of the ASAR WS data, and should lead to the question whether an
55
Feasibility studies Chapter 3. Mapping inundation
Figure 3.13: Bimodal backscatter distribution of ASAR WS scene from 13 August 2008.
Table 3.10: Confusion matrix comparing ASAR WS threshold classification (t = -12 dB) with
TerraSAR-X threshold classification as reference image.
Reference
Classified Water No water Total User Overall
Water 30 1 31 0.97
No water 12 157 169 0.93
Total 42 158 200
Producer 0.71 0.99 0.94
56
Chapter 3. Mapping inundation Discussion
adaptive thresholding approach might be more beneficial.
3.5 Discussion
Choosing the ideal classification technique is an inherent problem in the field of remote sens-
ing. In effect, this is a choice in model issue - one needs to decide on a classification model that is
to be applied to the imagery. However, the ideal classification model is in fact an idealised one,
as each classification technique has its advantages and disadvantages. Moreover, no single op-
timum classification model exists, it is the quantification and error analysis that should justify
the choice in model. Arriving at the appropriate decision entails an analysis and discussion of
several feasibility studies.
This chapter considers the commonly used density-slicing classification method. A method
needs to be established by which threshold values are set in order to create intervals corre-
sponding to a certain class (Hamandawana et al. 2006). Active-microwave data may be classi-
fied into the two distinct classes of land and water by setting a threshold value that lies between
the two modes of the bi-modal backscatter distribution. However, choosing a single, most ap-
propriate threshold value in a range of many possibilities is subjective. Often the trough value,
i.e. the local minimum, is used, however, a more sophisticated approach is to fit a statistical
model to the data in order to arrive at an optimum threshold value. This method is well suited
to timeseries classification, especially when dealing with an automated classification proce-
dure.
Considering the bimodal nature of the radar backscatter distribution, a double-Gaussian fit
was chosen, and threshold values at µdominant − 3σdominant are considered to result in two sta-
tistically discrete classes. It should be noted, however, that several other statistical models can
be applied to determine the threshold value. Furthermore, the threshold value of tC-band = −14
dB was derived for ASAR WS using several normalised images and the established automated
classification procedure uses this value as a constant. The normalised backscatter distributions
are however image dependent and therefore this static approach may result in classification er-
rors. This issue was investigated in the sensitivity study in section 3.4.1, wherein an optimum
threshold was investigated. The choice of threshold is at the lower end of the optimum thresh-
old range of the random lake sample. However, the error investigation showed that given the
great range of optimum thresholds and the lower signal to noise ratios at higher thresholds,
the average change in mean lake diameter was sufficiently low to use a threshold value of -14
dB. Indeed, given the ASAR WS product’s spatial resolution of 150 m and an average change in
mean lake diameter of less than 22 m, the applicability of the thresholding approach is clearly
justifiable (Trofaier et al. 2012). This is further confirmed in terms of the minimum map unit
(MMU). The MMU gives the size of the smallest feature that can be trusted to be mapped
most reliably. The MMU should therefore be greater than the pixel size [m2], and usually is
within the range of 4 - 16 pixels (Kelly et al. 2011). Choosing an MMU of 4 pixels for the ASAR
WS product gives 4 × 75m × 75m = 22500 m2. In terms of the average change in lake area,
57
Discussion Chapter 3. Mapping inundation
Figure 3.14: From left to right: TerraSAR-X classification (t = -16.76 dB), overlain by ASAR WS
classification (t = -14 dB) (top panel); overlain by ASAR WS classification (t = -13 dB) and (t =
-12 dB) (bottom panel). 58
Chapter 3. Mapping inundation Summary and conclusions
< dA >= 194.5 m2, this change is well below the MMU, more than two orders of magnitude.
Even if one redefined the MMU to be 1 pixel, < dA > would still be much smaller than the
MMU of 5625 m2. Furthermore, the realisation that the optimum threshold is not only image
but also lake dependent provides scope for further debate, not only on how to set the optimum
threshold value, but also whether indeed density slicing delivers the best classification results.
Map accuracy assessments of different classification techniques may then be used to defend
the chosen classification method. Therefore, two map accuracy assessments are presented in
this chapter. In section 3.4.2 it was established that an unsupervised ISODATA classification
does in fact result in a higher map accuracy (91%) compared to the -14 dB threshold classifi-
cation (89%). In addition, in section 3.4.3 the discrepancies in overall map accuracy, derived
using high resolution X-band reference data, arising from a set of threshold values applied to
the C-band ASAR WS data, further illustrates the problem of image dependent threshold val-
ues. This latter study explicitly shows the extent of threshold related classification errors and
in the first instance would advocate an adaptive threshold classification method. Nonetheless,
the difference of 4% between the -14 dB threshold and the determined optimum threshold of
-12 dB in the overall map accuracy of this data fusion II study, the marginal difference in overall
map accuracy of 2% derived for the data fusion I study, the greater computational efficiency
of a simple threshold classification as opposed to the more complex ISODATA clustering tech-
nique and the fact the automated procedure derives temporal composites, do provide enough
justification to use a straightforward, static threshold value density slicing approach. The last
of these reasons is elaborated on in the following chapter and therefore the reader is asked to
take this statement as a given and keep it in mind for later discussion.
Weather affected data are not considered by the automated classification procedure and
therefore result in the reduction of the available, usable data. However, more difficult to rem-
edy are the classification errors arising from vegetation bias. Both these issues need to be con-
sidered for change detection analysis, as is presented in the following chapter.
3.6 Summary and conclusions
The focus of this chapter was on monitoring Arctic surface waters with active microwave satel-
lite data. As such, the chapter sets out to first describe the density-slicing method by which
water bodies are classified, discussing the need to investigate an optimum threshold value as
well as map accuracies, by the traditional means of error matrix analysis using ancillary EO
data (panchromatic ALOS PRISM and X-band TerraSAR-X data) as reference.
• The optimum threshold study showed that a threshold change of 1 dB is associated with
a change in mean lake diameter of approximately 22 m. It was also found that the opti-
mum threshold varies with, and hence is dependent on, both lake and image. Therefore,
the choice of threshold is not a straightforward matter. While for the sample set of im-
ages the mean optimum threshold value was found to be −12.34± 1.22 dB, the choice of
59
Summary and conclusions Chapter 3. Mapping inundation
optimum threshold for the static, temporal composite, classification procedure remains
with tASAR= −14 dB, which is at the lower end of the aforementioned optimum thresh-
old value range, and therefore a more cautious choice, avoiding water body mergers at
the upper end of this range.
• A map accuracy comparison test of an unsupervised ISODATA classification with the
−14 dB density-slicing method of the ASAR WS data to ancillary ALOS PRISM imagery
showed that there was generally good agreement for both classification techniques. Al-
beit the density-slicing classification resulted in a slighly worse overall map accuracy of
89% compared to the ISODATA’s overall map accuracy of 91%. The Kappa coefficients
showed good agreement, 0.60 and 0.63 for density-slicing and ISODATA classifications,
respectively. On the whole, the threshold classification method proved to produce com-
parable results to an unsupervised ISODATA classification, which in addition benefited
from being ’trained’ with a PRISM derived land cover mask, improving the unsupervised
classification through focused clustering.
• A map accuracy test of a threshold derived ASAR WS classification to a threshold derived
TerraSAR-X classification (used as reference data) further demonstrated a fair robust-
ness of the -14 dB density-slicing technique for the ASAR WS data (overall map accuracy
86%). However it only demonstrated fair agreement according to the Kappa coefficient
(κ-14dB = 0.44). Further examination of a set of threshold values, showed that indeed
tASAR= −12 dB was the optimum threshold for this particular image with an overall map
accuracy of 94% and a Kappa coefficient of 0.64. However, having noted the optimum
threshold value and the map accuracy of the resulting classification, it is important to
stress that although far from ideal, the -14 dB density-slicing technique does produce a
sufficiently accurate result, especially given the automated batch classification approach
that is used to create a temporal composite, binary water body product (as described in
the following chapter).
60
Chapter 4
Monitoring seasonally inundated areas
This chapter focuses on the spatial analysis of water bodies and their seasonal changes on
the local scale, as measured by the low-spatial/high-temporal resolution ASAR WS sensor.
As such, the chapter sets out to explain the development of an inundation change detection
method by first elaborating on the automated temporal composite water body classification
method. The temporal composites are the main dataset, forming the basis of this chapter.
Changes in water body extent are first analysed in terms of a continuous water body raster
class (i.e. pixel count). This study constitutes the basis of my first first-author peer reviewed
paper, published in the proceedings of the Tenth International Conference on Permafrost (see
Trofaier et al. (2012)). As such, segments of this paper have been incorporated in this chap-
ter. Where this has occurred, these segments are identified by the aforementioned reference.
Subsequent to the change detection analysis, lake abundance studies are undertaken by trans-
forming the raster datasets into vector files containing discrete polygons for analysis of each
water body individually. The results of these studies were first presented at the fourth ESA
DUE Permafrost workshop at ESRIN in Frascati in February 2014.
4.1 Case 1: Local Scale
This chapter considers the local scale again, the most important study area being ROI 1, for
which change detection analysis has been undertaken. In section 4.3.1 the smaller study area
ROI Terra is used to enable a comparison study between the TerraSAR-X and the ASAR WS
sensor. The boundaries of both study areas were defined in chapter 3, section 3.3 (see figure
3.7). As mentioned above in section 3.3 of the previous chapter, the Russian research camp
Vaskiny Dachi is located at ~70° 17’ N, 68° 54’ E. Active-layer monitoring is ongoing at the
Vaskiny Dachi site but there are no permanent structures (except a small shack) and there is no
meteorological station - the closest meteorological station is Marre-Sale, ~100 km to the south-
west of this area on the coast, which unfortunately means there are no adequate meteorological
data available for local scale analysis.
The entire Yamal peninsula lies within the continuous permafrost zone. Due to the high
61
Spatial analysis of changes in water body extent Chapter 4. Monitoring seasonal inundation
ground ice content found on the Yamal peninsula, this area is inclined to permafrost degrada-
tion in the form of thermokarst erosion. This is a climate-induced process but also arises as a
result of major anthropogenic activities in this region (Kumpula et al. 2010). Despite this area’s
high susceptibility to thermokarst processes - therefore being considered unstable terrain and
unsuitable for large-scale infrastructure (Nelson et al. 2001) - it is the location of exactly such a
development. The Yamal peninsula is a hydrocarbon rich environment, containing the largest
known gas reserves in Russia (Walker et al. 2011). The Bovanenkovo gas field is situated a few
kilometres northwest of Vaskiny Dachi, within the borders of both ROI 1 and ROI Terra. Analy-
ses of aerial photographs in the surroundings of the gas field have shown small scale long-term
expansion as well as decline of thaw lakes (Sannikov 2012). Furthermore, changes in land use
and land cover have been established over the entire peninsula (Kumpula et al. 2012). The fact
that the local scale analysis is centred around an area of known anthropogenic influence needs
to be considered in a discussion of spatial changes in water body extent.
4.2 Spatial analysis of changes in water body extent
Having in section 3.4 discussed the feasibility of classifying active microwave data, in particu-
lar Envisat ASAR WS data, using the density-slicing technique, the method of using temporal
composite water body classifications of several radar images is presented here. Thereafter fol-
lows a brief discussion on the quality control of these temporal composites which are then used
for change detection techniques. The method of using temporal composites allows easy han-
dling of a large and frequent data timeseries, enabling a straightforward analysis of seasonal
changes in water body extent.
4.2.1 Temporal composite water body classification
Temporal composite water body classifications are derived from the ASAR WS radar data, ap-
plying a threshold of -14 dB to each image. This batch processing method is one of the reasons
for choosing a static threshold despite the fact it has been shown above in section 3.4.3 that the
-14 dB threshold is not suitable for every ASAR WS radar image. However, computational effi-
ciency need also be considered. In practice, running the temporal composite classification algo-
rithm with a static threshold value is already greatly time consuming. If the code were to be re-
written to include an adaptive thresholding method that explores the backscatter distribution
of each image individually this would most certainly cumulate in computational inefficiency.
In effect, a clear decision on the ‘cost-benefit’ of an adaptive threshold method was taken. There-
fore, carefully considering the results from section 3.4.1, where it was found that the optimum
threshold is not only image but also lake specific and a threshold change of 1 dB was shown to
be associated with a change in mean lake diameter of approximately 21.84 m (this value being
well below the MMU of 4 pixels, which for the ASAR WS sensor is 4× 75m× 75m = 22500m2),
it was decided to continue working with a static threshold.
62
Chapter 4. Monitoring seasonal inundation Spatial analysis of changes in water body extent
Table 4.1: Classifications and respective time periods for each year.
Time period for each year
Classification Start date End date
A 01 July 14 July
B 14 July 31 July
C 01 August 14 August
D 14 August 31 August
Advanced data handling is required for change detection applications, which ideally rely
on a high number of sample data. In order to adequately deal with the abundant radar data,
temporal composites over given time periods were created. The algorithm classifies each radar
image in this time period according to the threshold conditions as a binary raster and then
sums each of these raster classifications to create a single temporal composite classification of
water body extent. This method also allows the unambiguous determination of the number
of measurements (i.e. images) used to represent each pixel of the temporal composite, as the
pixel values of the water bodies are the sum of each of the binary threshold classifications. In
effect, only two classes are present, water and no water, but the water class consists of ‘sub-
classes’ which are the number of images used to classify these water bodies. The temporal
composite procedure is called, first defining the spatial coordinates of the region and the time
period as arguments which are then passed to the procedure. The summer months July and
August for the years 2007, 2008, 2009 were analysed. The temporal composite shows each
water body at its greatest extent within the given time period. The time period interval was
chosen to be 14 days, i.e. water at its greatest extent within 14 days is given in a single temporal
composite classification with water pixel values lying in the range from 1 to 14. Hence, there are
4 temporal composite classifications for July and August of each year. To allow computational
efficiency, the procedure produces 0.1°×0.1° tiles (pixel size 1/1500°×1/1500°). These are then
mosaicked separately by running a Java program packaged in an executable JAR file. The final
mosaics are the temporal composites, which are then used for change detection analysis.
4.2.2 Temporal composite quality control
As explored in section 3.4 of the last chapter, error analysis is an important part of producing
EO products. Accuracy and quality control of classified data need to be undertaken and pre-
sented to arrive at an exhaustive study. However, although this statement contains a problem
setting that seems simple enough to solve, it is in fact a more complicated matter when consid-
ering multi-temporal classifications. Liu and Zhou (2004) state the need to combine the map
accuracy of each individual classification used in a temporal composite in order to produce a
63
Spatial analysis of changes in water body extent Chapter 4. Monitoring seasonal inundation
measure of uncertainty for the temporal composite classification. Often statistical methods are
applied, or an error propagation model is developed (Burnicki et al. 2007). The simplest way to
achieve a temporal composite map accuracy assessment is to undertake an error matrix analy-
sis for each image that contributes to the temporal composite and to then average the overall
map accuracies of each classification (Liu and Zhou 2004). This, however, works on the basis
that each image is classified separately, requiring manual intervention in both the classifica-
tion procedure, as well as the post-classification error analysis. Another method is to take the
product of the individual map accuracies used to produce the temporal composite classifica-
tion (Burnicki et al. 2007). This again works on the assumption that the temporal composite
is comprised of several individually classified images. The present study, however, has imple-
mented an automated procedure that classifies a data timeseries according to its pixel values.
At no stage are individual images manually inspected and classified. The output is therefore
an automated multi-temporal classification. A new means of error propagation needs to be
applied to evaluate the quality of the end product. One of the forms of quality control is the
assessment of the number of images that were used to classify each pixel in the temporal com-
posite. However, this is not a measure of accuracy. Currently, there is no established method
for evaluating map accuracy of multi-temporal classifications. Therefore, an attempt at creating
such a method is put forward here. This theoretical error propagation model is based purely
on statistical data reduction, applicable to the present automated classification method and is
still open for discussion.
In sections 3.4.2 and 3.4.3, map accuracy studies data taken on two different days were
undertaken. Using the -14 dB threshold value, the overall map accuracies were 89% and 87%,
respectively. Thinking in terms of a two-time comparison method, and therefore averaging
these two map accuracies, would give an overall map accuracy of 88%. This, of course, is
a slightly flawed analysis, as the study areas of the two studies, ROI ALOS and ROI Terra,
do not overlap. However, both study areas are part of the larger ROI 1 and as this area is
fairly homogeneous with respect to water body abundance, and given the fact that both map
accuracies are similar, it is assumed that this two-time overall accuracy is viable for establishing
an error propagation model for multi-temporal -14 dB threshold classifications.
An additional issue for the error propagation model of the temporal composites created
from the ASAR WS data, is that the classification condition is weighted to improve water body
extent. The sensitivity study in section 3.4.1 and also the data fusion II study in section 3.4.3,
both showed that t = −14 dB is a cautious threshold value, and indeed in section 3.4.2 it was
found that no water surface was misclassified as land surface according to the reference data,
i.e. there is no user’s accuracy for the water class. Similarly, this occured in section 3.4.3 for the
same threshold value. This does indeed show a significant degree of reservation in the choice
of threshold value. However, as the temporal composite is the sum of all classifications over a
number of observations (Nobs) where the pixel values fall within the threshold conditions, the
probability that the temporal classification accurately represents water over the time period of
observation should increase, as, inferred from the aforementioned user’s accuracy, it is highly
64
Chapter 4. Monitoring seasonal inundation Spatial analysis of changes in water body extent
unlikely that any pixels below t = −14 dB are part of the land class. Assuming therefore
a logarithmic probability distribution, where the probability that water has been accurately
classified at t = −14 dB increases with Nobs until it plateaus out (when no further observation
will change the spatial distribution of water, although increased Nobs will still contribute to the
quality control - the higher the number of observations, the higher the probability of accurate
classification). Therefore, the temporal composite classifications have a higher probability of
appropriately classifying water (at its greatest extent), and should therefore have a higher map
accuracy to just a single ASAR WS image classified by the -14 dB threshold method.
It is, furthermore, important to realise that the temporal composites illustrate water at its
greatest extent within the chosen time period. Therefore, this theoretical error propagation
model only deals with the error for water at its greatest extent being accurately classified. It
does not, however, consider any errors that may have occured if water bodies exhibit loss
within the chosen time period (in this case 14 days). For this reason, the frequency of the
temporal composite classifications needs to be appropriately set, preferably as short as possible,
to capture the true onset of change without compromising the benefits of a multi-temporal
classification. Within the context of this research’s change detection analysis, this means that
any changes that are identified are changes from one temporal composite to another and do not
provide dates on which these changes occur but a time period in which these changes occur.
The concept of water body change detection is elaborated in the following section.
4.2.3 Water body dynamics
In order to determine the seasonal dynamics of surface water extent over the two summer
months July and August in 2007, 2008 and 2009, changes in the temporal composites are calcu-
lated. First, these temporal composites were reclassified, setting each water pixel value equal
to one. Then, changes in surface water extent were calculated by taking the difference between
each classification and its temporally preceding classification, thereby arriving at three raster
difference maps for changes in surface water extent that have occurred over July and August
(difference maps I=A-B, II=B-C and III=C-D). This approach is similar to Plug et al. (2008) (al-
beit differing as their study deals with annual changes in lake extent) who overlaid classified
images in a Geographic Information System (GIS), in order to analyse lake areas prevalent in
each image. This enables a straightforward delineation of lake shrinkage (classified water only
present in older images) and lake growth (classified water only present in more recent images).
Other EO studies of decadal permafrost thaw lake change use shoreline detection algorithms
to evaluate rates of changes between vector files, e.g. Jones et al. (2011). The approach taken
in the present study was to keep calculations simple by relying on change detection analysis
between raster classifications. Vectorisation is undertaken at a later stage. First, it was decided
to investigate net loss and gain in pixel count for classifications I, II and III of each year.
Pixel count of water body area is highly dependent on the number of measurements avail-
able from Envisat ASAR for each classification. Fortunately, the pixel count for each classifi-
65
Spatial analysis of changes in water body extent Chapter 4. Monitoring seasonal inundation
Figure 4.1: Pixel counts for each classification period in each year as a function of Envisat ASAR
WS measurements.
cation as a function of ASAR measurements is sufficiently high for ROI 1, as demonstrated in
figure 4.1. However, this may not always be the case, as is also discussed below in section 4.2.4.
It is also noted here that less usable data are available for ROI 1 in 2009, especially in the first
time period A. Any related issues are also discussed below in section 4.2.4.
Having, nonetheless, established that the Vaskiny Dachi region ROI 1 does have ample data
available over the summer months July and August in the years 2007 and 2008, the changes in
pixel count are determined from classifications I, II and III. These changes in pixel count were
also calculated for 2009 taking heed not to misinterpret the change detection results following
the limitations of the 2009 time period A dataset. The cumulative changes in water pixel count
for ROI 1, for both loss and gain, are given in figure 4.2, where loss and gain in pixel count
are represented by negative and positive values, respectively. The greatest change in loss in
both regions occurs in classification I. In 2007, a 25% and in both 2008 and 2009, a 28% change
in pixel count due to loss in ROI 1 were determined. In each case, loss in water pixel count
decreases steadily thereafter. In addition to loss, gain in water pixel count was also identified.
This gain is primarily around the margins of the water bodies and could be attributed to mixed
pixels. Furthermore, wind conditions seem to play a crucial role, in particular concerning large
lakes. At high wind speed an increase in the surface roughness of these lakes will result in a
loss of separability from their surroundings. Therefore, these data will not be classified as water
bodies. The classifications, and hence the results for the change analysis, will be affected by this
problem. Consequently, gain in lake extent is also associated with an increase in data (i.e. better
66
Chapter 4. Monitoring seasonal inundation Spatial analysis of changes in water body extent
Figure 4.2: Cumulative change in pixel count due to water pixel loss and gain for ROI 1, as
presented in Trofaier et al. (2012).
wind conditions), resulting in what shall be referred to as false positives. Indeed, as pointed
out above, it can be seen from figure 4.1(a) that there are a lot less ASAR WS measurements for
time period A in 2009. Hence a greater gain in pixel count in classification I (18%) , relative to
the previous years 2007 (7%) and 2008 (8%), is found, as depicted in figure 4.3(b). The affects of
wind and an increase in data per classification time period are discussed in more detail in the
next section below.
4.2.4 Wind biases
Having established that wind may bias an analysis of change in temporal water classifica-
tion composites, the degree of resultant misinterpretation is investigated. This wind bias was
first noted by Trofaier et al. (2012) for a study area on the Yamal peninsula (~ 67° 56’ - 68°
35’ N, 68° 44’ - 69° 32’ E; referred to as ROI 2). It is not so much a misclassification, rather a
‘no data/misinterpretation’ problem related to the classification method. When the bimodal
backscatter distribution is lost, the SAR image will not be classifiable under the threshold
conditions. Hence, no classification data will exist for images taken on windy days. It was
found that this classification problem is noticeable for ROI 2, over July and August in 2008 and
2009. Hence, ROI 2 data were not used for change detection analysis. Figure 4.4 highlights
the issue of data availability in terms of classifiable imagery for 2008. Each of these classifica-
tions is a temporal composite of maximum surface water extent over 14-days. When there are
fewer ASAR WS measurements for a preceding classification than there are for the subsequent
classification, due to ameliorated wind conditions, then a false gain in water body extent is
67
Spatial analysis of changes in water body extent Chapter 4. Monitoring seasonal inundation
Figure 4.3: Maps of pixel loss (drainage) and pixel gain (expansion) for classifications I,II and
III of region ROI 1 in 2007, 2008 and 2009.
68
Chapter 4. Monitoring seasonal inundation Spatial analysis of changes in water body extent
Figure 4.4: Number of ASAR WS measurements and their associated pixel count for ROI 2 in
2008, as presented in Trofaier et al. (2012).
registered; this complicates the analysis of the dynamics of surface water extent. Changes in
surface roughness are particularly noticeable on larger lakes. This is even noticeable for ROI 1
in 2009, where the lower right corner of the map of lake expansion in figure 4.3(b) seemingly
shows the expansion of a large lake when in fact this gain in pixel count is a result of improved
weather conditions and greater data availability in the second classification period B. There-
fore, restating the underlying premise, as it is absolutely crucial to this research, the number
of ASAR WS measurements that contribute to each classification pixel need to be examined
for any lake change detection analysis to avoid a wind biased interpretation of the data. In
addition, access to local meteorological data would be beneficial for assessing the quality of the
classification. Unfortunately, local meteorological data are not available for the Vaskiny Dachi
research camp, however, relating these data to the number of measurements used to produce
each temporal composite would clearly establish the reasons for classification bias, which are
currently interpreted to be due to wind. Of course, the meteorological data would not improve
the classification procedure itself but they would provide additional support to quality control.
69
Water body statistics Chapter 4. Monitoring seasonal inundation
4.3 Limnicity, lake number and lake size distribution analyses
An investigation of lake abundance as well as lake size distribution was undertaken for the
region covered by the TerraSAR-X scene (ROI Terra). The region was chosen, so that a compar-
ison study between the high and medium spatial resolution radar data could be conceived.
The concept of lake abundance and its importance with respect to aquatic processes (in par-
ticular intense carbon processing) was introduced in chapter 2, and studies of limnicity (the
ratio of water surface to land surface in a given study area) have long been used as input for
a myriad of scientific assessments from different disciplines (e.g. hydrology, biogeochemistry,
ecology) (Lehner and Döll 2004). Considering carbon exchange with the atmosphere, the sig-
nificance of accounting for limnicities is typified by the need to upscale and model carbon
emissions from freshwater ecosystems. In the context of permafrost landscapes of the North-
ern Hemisphere, across which abundant surface waters are strewn, the immediate link of per-
mafrost thaw lakes, as well as Arctic wetlands, to methane emissions is not to be neglected
(Downing 2009). Crude calculations on total carbon exchange may be determined from these
limnicity studies.
Limnicities are straightforward calculations, especially to binary classifications of a remotely
sensed dataset, as the limnicity of a classification is the ratio of the two classes. However, lake
number and size distributions need also be considered, as many processes such as carbon ef-
flux, are indeed directly proportional to lake size (Telmer and Costa 2007). Small lakes therefore
play a significant role in hydrological and biogeochemical cycles, and the percentage of total
water body area comprised by lakes that are smaller in size than 0.1 km2 is large (Grosse et al.
2008). It is for this reason a lake abundance study should also incorporate a size-frequency anal-
ysis. Karlsson et al. (2014) have highlighted the magnitude and severity of this phenomenon,
especially in Arctic permafrost zones, where the surface hydrological system is dynamic and
fragmentation of large lakes into smaller lakes is common. In the context of seasonal varia-
tions in surface water extent, this point is all the more prevalent as drainage and runoff of
spring floods will lead to fragmentation of these floods in the image timeseries. Therefore, the
concept of water body abundance analysis (in terms of limnicity analysis) is further narrowed
down to the practice of size-frequency distribution analysis.
This section firstly deduces the limnicities for ROI Terra, comparing both TerraSAR-X and
ASAR WS data, followed by a size-frequency comparison analysis. The results of these studies
reflect the differences that arise due to varying spatial resolutions. Thirdly, the temporal com-
posites of the ASAR WS data are used to examine the seasonal changes in the size-frequency
distributions and the fluctuations that arise inter-annually over the three years 2007, 2008 and
2009.
70
Chapter 4. Monitoring seasonal inundation Water body statistics
Table 4.2: Main statistical parameters of ASAR WS and TerraSAR-X water body data from 13
August 2008.
Parameter ASAR WS TerraSAR-X
Limnicity [%] 11.95 20.03
Total area of surface water [km2] 161.15 251.88
Maximum water body size [km2]† 6.63 8.31
Mean water body size [km2]† 0.99 0.29
Number of water bodies†‡ 1471 7384
† After removal of truncated water surfaces and river branches.
‡ Equivalent to number of polygons.
4.3.1 TerraSAR-X and ASAR WS: Limnicities
As explained above, limnicity calculations are straightforward to determine for land cover clas-
sifications, especially those, as in the present study, that consist only of water and non-water
classes. The limnicities of both ASAR WS and TerraSAR-X data were determined using the
classifications of two images that were presented in section 3.4.3. The choice of these data was
for the same reason given in section 3.4.3, namely the fact the images’ are simultaneous (13
August 2008). The limnicities were calculated from each raster classification and are given in
table 4.2. The limnicity of this study area gives insight into the extent of surface hydrology,
however, a comparison of limnicities for both high and medium spatial resolution radar data,
in fact, also shows the effectiveness of using medium spatial resolution data, such as ASAR WS
data. The calculations of limnicity for both sensors determine that the same region (ROI Terra)
is 1.6 times more abundant in water bodies for the TerraSAR-X data than it is for the ASAR WS
data. Given that a similar approach (density-slicing technique with optimised threshold values
for each dataset) was carried out for each classification, greater limnicity values determined by
the TerraSAR-X data are indicative of the higher spatial resolution of the TerraSAR-X sensor
rather than a discrepancy that could have arisen if different classification techniques had been
undertaken.
4.3.2 TerraSAR-X and ASAR WS: Size-frequency comparison analysis
Another common method for capturing lake abundance is to analyse size-frequency distribu-
tions (Likens 2009). Scale is an important component of this analysis, since small water bodies
are more frequent than large ones. Wetzel (1990) undertook a global analysis of freshwater
bodies, finding a relationship between the number of lakes and their areal, as well as volumet-
ric extent. Their data indicated that the total surface area of small, shallow freshwater bodies
comprised a greater overall areal extent than that of the large lakes. Meybeck (1995) structure
71
Water body statistics Chapter 4. Monitoring seasonal inundation
their data according to the number of water bodies in a given size interval per unit area, for
which they find an empirical equation. This equation follows a probability density function
(pdf) in the form of a Pareto distribution given by:
pdf(a) =
{
ckca−(c+1) for a ≥ k
0 for a < k
(4.1)
where the size of the lakes are given by a, and k and c are parameters related to both location
and shape, respectively. In fact, the exponent, γ = (c+1), is often referred to as the Pareto shape
parameter (McDonald et al. 2012). The Pareto distribution is a power law that after a certain
cut-off point (represented by the parameter k) drops to zero. Several authors (e.g. Lehner
and Döll (2004) and Downing et al. (2006)) use the Pareto distribution to model lake sizes and
abundances in order to extrapolate a respective estimate of size and frequency of small water
bodies.
The usefulness of the Pareto distribution to lake size-frequency analysis is a contentious
matter (Seekell and Pace 2011). The issue remains with its application for extracting these
small water bodies whose biogeochemical processes are yet poorly quantified on the landscape
scale (Downing 2010). The questions are how to establish the value at which the power-law
cut-off takes place, how to further extrapolate smaller lakes that are not represented by the
power-law, and whether indeed the the Pareto distribution is a good model for extrapolation
at all? In addition to the uncertainty in extrapolation, a further problem is encountered when
the input values of the Pareto model are predominantly large and linear regression analysis,
on a base 10 logarithmic scale, is carried out. This will result in unrepresentative Pareto shape
values despite high correlation coefficients (Gan et al. 2006, Perline 2005). Although the upper
end of the lake distribution function does fit the power-law well, the lower end is expected to
roll off (due to the Pareto cut-off) rather than continue following this power-law. However, the
slope of the regression line (representing the Pareto shape parameter) will typically be steeper
than expected if only larger lakes are used for the least-squares fit. This matter will be revisited
in the discussion section of this chapter.
Since the aforementioned studies deal with regional and global lake distributions, it was
decided to investigate such a lake size-frequency analysis on the local scale. Furthermore, the
advantage of using both a high spatial resolution satellite sensor (TerraSAR-X) as well as a sen-
sor of coarse spatial resolution (ASAR WS), provides an opportunity to investigate scale issues
and the associated differences in applying a Pareto model. Here follows a comparative study
of water body size-frequency distributions by exploring the Pareto approach for both ASAR
WS and TerraSAR-X data over the ROI Terra study area (1509.4 km2) on the Yamal peninsula.
4.3.2.1 Method
As previously mentioned, the TerraSAR-X data were classified, according to the filtering method
put forward in section 3.4.3, in which it was also shown that the optimum threshold for the
ASAR WS data from 13 August 2008 is t = −12 dB. Therefore, the −12 dB threshold clas-
72
Chapter 4. Monitoring seasonal inundation Water body statistics
sification was used for this size distribution in order to maximise the results of this study, by
applying the a priori knowledge of an optimum threshold. The raster water body classifications
were then vectorised. The number and size parameters are given in table 4.2. Only untruncated
lakes were included in this analysis - a total surface water area of 4.89 km2 and 7.54 km2 were
removed for ASAR WS and TerraSAR-X, respectively. Furthermore, rivers were also manually
removed from the ASAR WS (4.93 km2) and TerraSAR-X (3.16 km2) data. A bar chart of the
total lake surface area for each lake size interval is given in figure 4.6. Furthermore, linear re-
gression analysis was carried out on the two datasets, arriving at values for slope and intercept
of the regression line; these are the Pareto parameters.
4.3.2.2 Results
Using an exponent, γ, close to unity would indicate that each size interval should comprise
roughly the same total surface lake area (Downing et al. 2006). However, figure 4.6 shows that
the total lake surface area for each size interval for the current study does differ per interval (for
both TerraSAR-X and ASAR WS datasets). Indeed, in both cases, the total lake area is directly
proportional to the lake size interval, i.e. the greater the lake size interval, the greater the total
lake area. Therefore, the greatest contribution to the high limnicities (from each sensor) found
in section 4.3.1 comes from large lakes. Nonetheless, total lake area for each lake size interval
is of the same order of magnitude up to O
(
103 m2
)
.
Linear regression analysis of the abundance-size distributions on a base 10 logarithmic scale
resulted in a high correlation coefficient for the ASAR WS data (R2 = 0.89, p-value = 0.038) but
a low correlation coefficient for the TerraSAR-X data (R2 = 0.26, p-value = 0.138), the regression
lines following the function y = 4.44 − 0.47x and y = 3.31 − 0.18x, respectively (where y and
x are the base 10 logarithms of both lake frequency and lake size, respectively). Therefore, the
Pareto shape parameter, γ, is 0.47 and 0.18 for ASAR WS and TerraSAR-X, respectively.
With respect to a comparison in lake statistics between the two sensors, a total number of
209 water bodies were found to be ≥ 0.1 km2 for the ASAR WS dataset. In addition, ~47% of
the number of these water bodies ≥ 0.1 km2 were found for the TerraSAR-X data (N = 307).
However, the largest difference between these two number distributions arises for water bodies
O
(
103 m2
)
, where TerraSAR-X identifies 1628 water bodies, while ASAR WS can only distin-
guish 773 water bodies from the backscatter of the surroundings. Therefore, a difference of 855
water bodies between the two sensors is determined for water bodies O
(
103 m2
)
. For com-
parison purposes, considering only lakes > 103 m2, the number of lakes O
(
103 m2
)
amounts
to 52.5% and 56.04% for ASAR WS and TerraSAR-X, respectively. Whereas the difference in
water bodies of the same order of magnitude (N diff = 855) represents 58.12% of the total water
bodies identified by ASAR WS, and only 30.48% of the total water bodies (> 103 m2) identified
by TerraSAR-X. However, the percentage difference with respect to the total number of water
bodies identified by each sensor respectively, for water bodies O
(
103 m2
)
, is only about 3.54%.
The total water body surface area per lake size interval is given in the bar chart in figure 4.6(a).
The difference between both ASAR WS and TerraSAR-X total lake surface area is depicted by
73
Water body statistics Chapter 4. Monitoring seasonal inundation
Figure 4.5: A base 10 log-log plot of the lake frequency (dL, the number of lakes in a given size
interval) as a function of lake area for TerraSAR-X (blue) and ASAR WS (red). Linear regression
lines are shown for TerraSAR-X and ASAR WS also in blue (left panel) and red (right panel),
respectively.
the dashed line below in figure 4.6(b). This difference in total water body surface area between
the two sensors is 90.73 km2. Table 4.2 shows the lake statistics for both sensors. There is a
large difference in the maximum water body size between the two sensors (2.32 km2). This is a
curious result and will be discussed further below.
74
Chapter 4. Monitoring seasonal inundation Water body statistics
(a)
(b)
(c)
Figure 4.6: The total lake area within each base 10 log width lake size interval for TerraSAR-X
(blue) and ASAR WS (red) is shown in panel (a). The difference in total water body extent
per lake size interval between these two sensors is depicted by the black line in panel (b). The
number of water bodies in each lake size interval is given in panel (c), where the difference in
the total number of water bodies, Ndiff, is 1433.
75
Water body statistics Chapter 4. Monitoring seasonal inundation
4.3.3 ASAR WS: Early, mid and late season size-frequency distribution
4.3.3.1 Method
A similar approach to the previous section regarding the analysis of size-frequency distribu-
tions for the ASAR WS datasets in 2007, 2008 and 2009 was undertaken although multiple
intra-annual size-frequency distributions are presented here. The analysis was done for each
ASAR WS classification time period (early (A), mid (B and C) and late (D) season) for each
year over the ROI 1 study area. Contrary to before, truncated water bodies were not elimi-
nated from the present study. This allows a fully comparable analysis of each year without
the need to relate the large vector datasets to one another; a step that would have to be done
in order to ensure the exact same water body polygons are eliminated from each dataset. The
least-squares lines were calculated for the frequency data, which were then plotted on a base
10 log-log scale. NANs, resulting from the logarithm of bin counts smaller than zero, were
omitted from the model calculation.
4.3.3.2 Results
The statistics for each dataset are given in table 4.3. From these numbers, it can be seen that the
total water surface extent in 2007 and 2008 are similar in the early and mid season periods A
and B. The lower value in 2009 is attributed to the aforementioned lack in data for this period.
Interestingly, there is a small increase in water surface extent in 2007 in the classification period
C, whereas a decrease is found in 2008. Inspection of the number of water bodies equally iden-
tifies a substantial increase over this time period in 2007. This relative gain compared to 2008
was also identified in section 4.2 and is illustrated in figure 4.2. This gain may be attributed to
the higher quality of the 2008 classification with respect to 2007, associated with the number of
scenes available for the mid season time period C (see figure 4.1(c)) resulting in the aforemen-
tioned false positives. Furthermore, the number of water bodies is greater in period D than A
in 2007, despite there being a clear decline of 90 km2 in total water surface extent (14.3%). The
mean lake size values are 1.81×10−2 km2 and 1.41×10−2 km2 for A and D respectively. On av-
erage, the lakes in the late season period D are 4×103 m2 smaller than those in the early season.
Therefore, these additional water bodies are associated with small polygons that are due both to
data availability as well as possible discrepancies in the temporal composites, arising from the
density-slicing classification technique. In 2008, a steady decline in total surface water extent is
found over the early and mid season periods. A declining pattern is also found for the number
of water bodies over the classification periods A, B and C. However, an increase in the number
of water body polygons (and an associated increase in total surface water extent) is registered
in the late season period in 2008, as was the case in 2007. As mentioned previously, the 2009
data behaved differently to the two precursor years. This behaviour is also illustrated in the
statistics of the number and sizes of water bodies. Interestingly, the number of water bodies in
the early season period 2009 is far higher than those in 2007 and 2008. In spite of the abundance
of water bodies, the total surface water extent in period A in 2009 is 14.3% and 19.05% lower
than in 2007 and 2008, respectively. The histograms of water body extent for early, mid and
76
Chapter 4. Monitoring seasonal inundation Water body statistics
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
Figure 4.7: Histograms of water body extent for the early (A), mid (B and C) and late (D) season
periods in 2007 (a, b, c, d), 2008 (e, f, g, h) and 2009 (i, j, k, l).
late season in 2007, 2008 and 2009 are given in figure 4.7. The frequencies are given on a base
10 logarithmic scale in order to facilitate the interpretation of the size distributions with their
power-law characteristics. In each period, the highest contribution comes from water bodies
that are O
(
103 m2
)
(~99%). Therefore, throughout the active period, ROI 1 is dominated by
tundra ponds. Comparing 2007 to 2008, very similar size-frequency distributions are found,
deviating only in the spring flood period A, where 0.06% more water bodies are found to be
1.6−1.8×107 km2 large in 2007. 2009 gives a slightly different picture for period A, where only
0.02% are found in the above mentioned size category. The differences in the size-frequency
distributions of year 2009 and its two precursors are to be expected due to the lower number
of usable ASAR WS measurements used in the temporal composite classifications in 2009, as
mentioned in section 4.2.3. High R2 values are found for each period in all three years, ranging
from 0.91 to 0.97. Therefore, no significant deviation from linearity is established.
77
Water body statistics Chapter 4. Monitoring seasonal inundation
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j) (k) (l)
Figure 4.8: Size distributions of water body extent for the early (A), mid (B and C) and late (D)
season periods in 2007 (a, b, c, d), 2008 (e, f, g, h) and 2009 (i, j, k, l).
78
Chapter 4. Monitoring seasonal inundation Discussion
Table 4.3: Main statistical parameters of ASAR WS and TerraSAR-X water body data from 13
August 2008.
Parameter Period 2007 2008 2009
Total area of surface water [km2]
A 629.5 656.7 531.6
B 514.8 523.3 470.8
C 521.3 455.3 556.9
D 539.5 500.8 397.1
Number of water bodies‡
A 3482 3800 5403
B 3424 3316 3106
C 4018 2988 3977
D 3837 3284 3405
‡ Equivalent to number of polygons.
4.4 Discussion
The extent and dynamics of the hydrologically active Arctic spring/summer season have until
now remained unstudied on the landscape-scale. The fundamental issue is the need for con-
tinuous, comparable, frequent data. Here, the inherent trade-off between spatial and temporal
resolution is addressed by employing medium/low spatial resolution active microwave data
to frequently monitor the spring dynamics of open surface waters and quantify their extent.
It was shown in chapter 3 that the thresholding approach may be used as a time efficient
means to produce a fairly accurate classification. The classification method is taken further by
creating an automated procedure. Furthermore, rather than comparing changes in surface wa-
ter extent on a near-daily basis (subject to usable data availability), temporal composite water
body classifications are produced using the developed automated thresholding algorithm be-
fore change detection analysis is undertaken. This approach has twofold virtue, allowing (i) a
computationally efficient processing chain over large-scale regions of Siberia, while at the same
time, as pointed out in section 4.2.2, enabling (ii) the reduction of overall classification error.
4.4.1 Change detection analysis
The ASAR WS change detection analysis of lake surface extent on the Yamal peninsula shows
that there are two regions in which a sudden decrease in the number of water pixels in the first
few weeks of July is visible (Trofaier et al. 2012). These two regions are labeled ROI 1 and ROI 2
(see figure 3.7 in previous chapter). Raster analysis of water pixel count shows that the greatest
change in loss in both regions occurs between the first two classifications, i.e. for change classi-
fication I. Loss in water pixel count decreases steadily in classification II and III. This holds for
79
Discussion Chapter 4. Monitoring seasonal inundation
ROI 1. Similarly, this behaviour is also found for ROI 2 in 2007. However, unfortunately the
number of measurements available that meet the threshold conditions was greatly limited in
2008 and 2009 for ROI 2 (Trofaier et al. 2012). The limited number of measurements in ROI 2 ex-
emplifies the issues that wind may cause for this type of study. A change detection analysis of
raster data relies on the completeness of the dataset. The raster classifications must accurately
represent the actual state of the phenomenon, in this case water body extent, whose changes
are to be examined. If the preceding classification is misrepresented due to for example a lack
of data (as is the case here), then change detection analysis results in false positives. A phan-
tom gain in surface water extent is identified as weather conditions ameliorate, enabling more
usable data to be incorporated in the temporal composite classifications. The results show that
the regular monitoring of these water bodies with the aid of ASAR WS is restricted by the
impact of local wind conditions. The high variability in wind speed across this ‘local-scale’
tundra landscape produces skewed classification results, which hinder a comprehensive anal-
ysis based solely on SAR data. Bartsch et al. (2012) first discussed the effects of wind on this
classification technique and the classification problem is distinct for ROI 2 in both 2008 and
2009, when the number of ASAR WS measurements is greatly limited (Trofaier et al. 2012).
These restrictions have implications for continuous monitoring of lake dynamics. Advanced
knowledge of weather conditions from meteorological records provides complementary infor-
mation and allows a pre-selection of wind-free datasets. However, such records are not always
available. Further, this knowledge will have no impact on the number of ASAR WS measure-
ments. The ASAR WS sensor is a highly capable device for frequent monitoring, but severe
wind conditions common to high-latitude regions impinge on its temporal resolution (Trofaier
et al. 2012).
Furthermore, these wind biases are manifested in a higher fragmentation of the classifica-
tion, an issue that becomes significant in the lake abundance studies, as fragmentation results in
a higher number of polygons when converted to vector format. Indeed, this point is illustrated
in section 4.3.3, where the lake statistics show that, despite having similar total water surface
areas (approx. 606±50 km2), the number of water bodies identified in ROI 1 in the early spring
period 2009 is much higher than those found in the same season in 2007 and 2008 (1.6 and
1.4 times higher, respectively). This fragmentation does not pose a problem per se if analysis
proceeds with caution. The number of ASAR WS measurements contributing to each pixel in
the temporal composite classifications (in a 14-day classification pixel values range between 0
and 14) are to be used as a type of quality control before converting each of these classifications
into binary rasters comprised of two classes, water and land. Once again, the variability of
wind speeds across the study area will determine whether the imaged data are wholly usable
(no wind), usable with caution (localised wind) or unusable (regionally prevailing wind) for
classification.
The presence of vegetation is not to be neglected. Arctic spring is not only a hydrologically
active season, but it is of course also growing season. It is therefore anticipated that vegetation
and its phenology are also observed by change detection analysis. The radar data do not permit
80
Chapter 4. Monitoring seasonal inundation Discussion
Figure 4.9: Vegetated lake basin with arctophila emerging from the lake surface. Photograph
taken on 03 July 2012 by AM Trofaier.
the delineation of a distinct vegetation class. Vegetation falls within the dominant mode of
the backscatter distribution and the density-slicing classification method will therefore only
discern water bodies from their surroundings. However, as the growing season gets underway,
the change detection method should observe an increase in landmass and therefore a decrease
in open surface water extent. That is not to say that there is a decrease in inundation, but
only in open surface waters. Arctophila grasses, emerging from lake surfaces and growing
on the shorelines of these lakes, are prevalent on the Yamal peninsula. This is depicted in
figures 4.9 and 4.10. Indeed, the left panel in figure 4.3 identifies landmass gain around most
shorelines. It is, however, not possible to identify to what extent these are true representations
of vegetation phenology or result from incomplete or erroneous raster classifications without
appropriate field validation data. The fact that a similar shoreline pattern is also recognised in
the detected gain in water body extent (figure 4.3, right panel) would favour the latter. Further
evidence, in the form of ground based observations (including repeat photography of a chosen
lake basin), taken continuously over the growing season, would be necessary for appropriate
field validation of these classifications. As yet, this has not been possible, but will be considered
for future work.
Nonetheless, with respect to the change detection analysis, it is hypothesised that the eval-
uated changes in ROI 1 are related to a distinct seasonal pattern in lake drainage (Trofaier et al.
2012). Moreover, it is put forward that these may in turn result from river-lake interactions.
However, due to spatial resolution constraints of the sensor, these interactions cannot be inves-
tigated by the current method. Further research into this phenomenon is necessary and will
be discussed in the succeeding chapter 5, where stream channel networks are modelled and
analysed in the context of these potential seasonal patterns of spring floods.
81
Discussion Chapter 4. Monitoring seasonal inundation
Figure 4.10: Arctophila emerging from the surfaces of two lakes. Photographs taken on 01 July
2012 by AM Trofaier.
4.4.2 Lake size-abundance distributions
It is imperative that an analysis of lake change includes a discussion on size and number dis-
tributions. Knowledge of the abundance of freshwater ecosystems is vital to the analysis of
the carbon cycle since inland water bodies play an important role as conduits of terrestrial car-
bon (Seekell and Pace 2011). Landscape-scale limnological studies, therefore, rely on metrics
that enable upscaling of biogeochemical processes. The potential in using EO products for this
matter is undeniable. Previous studies have investigated size-frequency distributions on the
global scale (Meybeck 1995, Lehner and Döll 2004, Downing et al. 2006), on the regional scale
(Seekell and Pace 2011, Seekell et al. 2013, McDonald et al. 2012), and in the Arctic (Grosse et al.
2008, Muster et al. 2013, Karlsson et al. 2014). These studies all share a common attempt to
extrapolate abundances of lakes, unrepresented in the data. It is generally thought that size-
frequency distributions follow a power-law in the form of a Pareto distribution. Indeed, from
the literature it would seem attempts to fit lake area frequencies to this Pareto function have
become common practice. However, as it was noted by Seekell and Pace (2011), extrapolation
resulting from a Pareto fit leads to overestimation of the abundance of small lakes (< 1 km2), as
the Pareto distribution misrepresents these data at its lower tail. This result is seconded by Mc-
Donald et al. (2012), who find that the lake size-frequency distribution is badly represented by
such a Pareto distribution. Indeed, the linear regression analysis of the ASAR WS data clearly
shows this bias towards overestimation when compared to the TerraSAR-X dataset for which
water bodies as small as O
(
100 m2
)
are classifiable.
In general, very small lakes are underrepresented in digitised (and printed) maps, as well as
lake databases such as the Global Lakes and Wetlands Database (GLWD) compiled by Lehner
and Döll (2004). However, these data have primarily been used for establishing these frequency-
size relationships in the past. Therefore, an element of uncertainty is introduced with respect
to lakes that have a surface extent < 0.1 km2 (Downing et al. 2006). The ASAR WS data were
82
Chapter 4. Monitoring seasonal inundation Discussion
able to classify water bodies as small as O
(
103 m2
)
. However, these classified water bodies are
uncertain as their values are one order of magnitude smaller than the spatial resolution of the
ASAR WS sensor. Given that the Pareto distribution (equation 4.1) is a power-law, it will be lin-
ear on a base 10 fully logarithmic scale and therefore simple regression models can be applied.
Applying linear regression analysis to retrieve the Pareto shape parameter, γ, represented by
the negative slope of the regression line resulted in γ = 0.47 and γ = 0.18 for ASAR WS and
TerraSAR-X, respectively. From figure 4.5 it can be seen that the TerraSAR-X data clearly show
a discrepancy from a power-law, rolling off at lake sizes of approximately 100 -1000 m2. This
is what would be expected from a Pareto distribution: A sudden decline in the abundance of
small lakes. It is, however, not clear what the value of the Pareto cut-off parameter, k, should
be. The lake size-frequency function may in fact possibly be associated with a combination of
multiple Pareto distributions.
Karlsson et al. (2014) find that for a sub-basin (1.2 × 103 km2), located in the Nadym and
Pur catchments of West Siberia, linear regression resulted in γ = 0.94 for lakes that are >
0.01 km2. McDonald et al. (2012) discuss how the fitted Pareto shape parameters vary with
location and ecoregion. However, given that the sub-basin analysed by Karlsson et al. (2014)
is of the same order of magnitude in size as the study area Terra ROI, and also lies within the
region of the West Siberian Lowland, a closer relation between the Pareto shape parameters
would have been anticipated. This may of course also be due to the fact they only consider
lakes > 0.01 km2. When excluding the lakes of O
(
103 m2
)
, linear regression analysis for the
ASAR WS data resulted in a γ value of 0.61. The emphasis again is on overestimating the
amount of small water bodies, if only larger lakes are considered for extrapolation using the
Pareto model. In addition to this issue, the sub-basin from the Karlsson et al. (2014) study
lies below the Yamal peninsula in the discontinuous permafrost zone. The impact different
permafrost conditions have on lake distribution may be relevant for such an analysis and it is
encouraged that such studies should be undertaken in the future to appropriately discuss the
contributing factors permafrost has on lake size and abundance.
In the continuous permafrost zone, in the Vaskiny Dachi area, linear regression analysis of
both ASAR WS and TerraSAR-X datasets confirmed the aforementioned outcome that a Pareto
distribution overestimates water body abundance for small lakes when the input data only
contain larger lake sizes (such as the ASAR WS data). Muster et al. (2013) discuss exactly this
matter, pointing out that extrapolation of small water bodies according to a Pareto fit would by
far overestimate small water bodies O
(
100 m2
)
despite high R2 values. And Seekell and Pace
(2011) emphasise that the lake size-frequency distribution does not follow a Pareto distribution
as the lower end of the pdf badly represents the true nature of lake sizes and abundance. In a
subsequent study, Seekell et al. (2013) discuss the fractal nature of lakes to explain the conflict-
ing results that are found when analysing the size-frequency distributions of water bodies at
different spatial scales. These observations conform with the results from the ASAR WS and
TerraSAR-X data. As the spatial resolution, at which water bodies are represented, improves,
the fractal concept related to the water body shorelines results in the aforementioned short-
83
Summary and conclusions Chapter 4. Monitoring seasonal inundation
comings of the Pareto fit, where the lower tail of the distribution is badly represented (Seekell
and Pace 2011). However, the TerraSAR-X data do show evidence of a scale of Pareto cut-off
values, where the lake size-frequency distributions starts to roll-off. As both TerraSAR-X and
ASAR WS data follow a similar trend at the upper end of the size-frequency distribution, it
may be assumed that the roll-off itself is not an artefact of TerraSAR-X’s finer resolution but is
indeed representative of the point at which lake abundance starts to drop. It is important to
note, however, that there may be a range of cut-off values (rather than one definitive threshold
value) associated with a combination of a number of Pareto distributions.
In the light of these results, the repeated emphasis that a Pareto distribution may not be the
ideal model for extrapolation (Seekell and Pace 2011, Muster et al. 2013) and the need to address
lake abundance studies from an alternative point of view (Seekell and Pace 2011), no further
extrapolation attempts are presented here. There is therefore scope for exploration of a more
appropriate model - which may indeed be a combination of a range of Pareto distributions.
In terms of the changes in water body size-abundance distribution over the course of the
spring/summer months, the distinct seasonal pattern discussed above in section 4.4.1, is visible
in the histograms in figure 4.7. The linear regression analysis of these data, illustrated by the
regression lines in figure 4.8, demonstrates the retreat of spring runoff, which is mirrored by
the values of the correlation coefficients. Apart from larger scale (O
(
107 m2
)
) spring floods no
other ’real’ changes in water surface extent take place. In 2009, deviations in frequencies to the
two preceding years arise due to fragmentation effects.
4.5 Summary and conclusions
The studies presented in this chapter are centered on what has been defined as the local scale in
chapter 3 (section 3.3). Th local scale study area is approximately 8 000 km2 large and is referred
to as Vaskiny Dachi, named after a Russian research camp that is situated within this region,
immediately South of the Bovanenkovo gas field. Having discussed the feasibility of using ac-
tive microwave data for mapping water bodies in the previous chapter, this chapter explores
the method of creating temporal composite water body classifications, which may then be used
for change detection analysis. Issues, such as weather affected imagery (i.e. wind biases) are
highlighted. Finally, the maps of water body extent are used to calculate limnicities and de-
termine lake size-frequency relationships. The main results of each of the studies presented in
this chapter are itemised below.
• Temporal composites depicting water body extent at its maximum (according to a -14
dB threshold value) within a given time period of 14 days were created. This resulted
in 12 temporal composite classifications for July and August 2007, 2008 and 2009. This
automated classification procedure is the main reason a static thresholding technique is
applied. The algorithm classifies each pixel that has a backscatter value below -14 dB in
each image as water. An adaptive classification method has so far not been investigated
84
Chapter 4. Monitoring seasonal inundation Summary and conclusions
as the procedure would be computationally inefficient given the high quantity of ASAR
WS data that are available. Nonetheless it is noted here that there is scope for adapting
this method, first running it on a small sample dataset and then comparing the results to
the static procedure. Unfortunately, this reaches beyond the goals of this PhD but will be
addressed in future work.
• Change detection analysis in the form of difference maps was undertaken. Three dif-
ference maps per year were created. The gain and loss in pixel count of each of these
difference maps were analysed. The greatest change in pixel count loss happened for
each year over the first two time periods A and B. Difference map I showed a 25% loss
in water body extent for 2007 and for both 2008 and 2009 a 28% loss. In 2007, 2008 and
2009 a 7%, 8% and 18% gain in water body extent was found for this classification I,
respectively. In 2009, therefore, an extra 10% gain in water body count was registered.
Closer inspection showed that this was related to the number of scenes that were used to
classify A and B. More scenes were available for B in 2009 resulting in a ’phantom’ gain
in water body extent. Therefore, it is found to be crucial to have high data availability,
without which the change detection method would be mis-representative. This may be
an obvious statement, but it is certainly not a trivial one. Although ASAR WS provides
daily satellite imagery, many of these images are weather affected. Wind biases the clas-
sification method, reducing the amount of usable data. Therefore, it is imperative that
the number of images used to create the original water body classifications (which are
outputted separately by the classification algorithm) are also explored.
• Limnicities derived from both TerraSAR-X and an ASAR WS water body classifications
over the same areal extent (ROI Terra) were determined to be 20% and 12%, respectively.
The 8% difference between these two is linked to TerraSAR-X’s higher spatial resolution,
being able to map water bodies with sizes below O
(
100 m2
)
. The numbers show that
TerraSAR-X is able to detect 1.6 times more water surface area than ASAR WS over ROI
Terra.
• An analysis of total surface water extent per size interval shows that the greater the lake
size interval the greater the surface water extent. This is indicative that despite the abun-
dance of small lakes, the large lakes make the greatest contribution to the limnicity of this
area. This holds true for both sensors. The difference in the total number of lakes iden-
tified by both sensors is 1433, the greatest discrepancy occurring at ASAR WS’s smallest
lake size O
(
103 m2
)
. The difference in the number of lakes in this bin is 855, which corre-
sponds to 58.12% of the total number of water bodies identified by the ASAR WS sensor.
• A size-frequency analysis for both datasets shows these functions follow a power-law
up to a cut-off value. Linear regression analysis of ASAR WS data overestimates water
bodies that are < 103 m2. This is especially visible if only large lakes (> 104 m2) are
incorporated in such an analysis leading to high R2 values. These correlation coefficients
85
Summary and conclusions Chapter 4. Monitoring seasonal inundation
badly represent the true nature of the lake-frequency distribution. The presented analysis,
comparing TerraSAR-X and ASAR WS data, further underlines the scale issues involved
in lake-abundance model parameter estimation.
• The TerraSAR-X data clearly display a roll-off from a power-law distribution, which
would be indicative of a Pareto distribution. However, there may not just solely be one
Pareto cut-off value that dictates where the abundance of small lakes starts to decline, but
it is possible that the lake size-frequency distribution follows a function that is made up
of a combination of a range of Pareto distributions. This model has the potential for suf-
ficient extrapolation of small lakes, which are important for quantifying intense carbon
processing. However, further research on this topic is needed before the model may be
successfully applied.
• Early, mid and late (active) season lake size-frequency statistics were analysed for the
ASAR WS data over the study area ROI 1 in 2007, 2008 and 2009. The high discrepancy
between the number of water bodies found in the early season of 2009 relative to the
same season in 2007 and 2008 typifies the fragmentation effect that localised wind con-
ditions have, especially over larger lakes. Spatial flood patterns associated with great
surface extent and low frequency are clearly established from the abundance-size distri-
butions. Investigating the abundance-size distributions on a fully base 10 logarithmic
scale showed these follow linearity with high correlation values (R2 ranges from 0.91 to
0.97).
86
Chapter 5
Modelling stream channel networks in
aid of a spring flood assessment
Having developed a method to identify changes in surface inundation over the spring and
summer weeks of 2007, 2008 and 2009, it is proposed to investigate whether these inundated
areas arise through hillslope-stream interactions and hence, whether these floods are due to
relief. This is a study that tries to tie an analysis of seasonal water body dynamics to possible
hillslope-stream channel connections. Without access to hydrometric data, the present study
aims to give an indirect assessment of the factors that may be contributing to these seasonal
floods, hypothesising that hillslope-runoff plays an important role in the accumulation of melt-
waters.
Therefore, this chapter considers the use of a GIS based surface water flow model that is
used to arrive at a stream channel network. The model is embedded in the ArcGISTM software
and uses a digital elevation model (DEM) as input, relying on the influence of gravity (water
follows the steepest slope) to derive flow accumulation maps. These are then used to derive the
stream channel network. Network density is explored in the context of spring floods as derived
from the difference maps presented in chapter 4. Furthermore, the topographic wetness index
(TWI) is calculated from the DEM derivatives of flow accumulation and slope gradient. The
stream channel network is then explored with respect to the TWI. The final study presented
in this chapter is an empirical investigation of the location of these seasonally inundated areas
using a panchromatic reference image from 1993 to identify old thermokarst depressions, so-
called alasses, in which spring runoff tends to accumulate. The normalised backscatter values
of these alasses are then explored in more detail with an emphasis on any temporal changes
that may indicate a certain trend in wetting/drying of the ground.
However, in order to quantify the relationship between hillslope-runoff and seasonally in-
undated areas, the quality of the data used to undertake any hydrological modelling need to
be assessed. Therefore, three DEM validation studies are presented here prior to the section
on hydrological modelling. These studies are presented to determine whether the quality and
vertical resolution of the chosen DEM are good enough to continue with further topographic
87
Case 1: Local scale, cont. Chapter 5. Modelling stream channels
analysis. The validation studies make use of a Soviet map, as well as a second DEM. These
datasets are used as reference data. This study is based on Local Scale analysis, focussing on
the Vaskiny Dachi area.
The chapter itself may roughly be divided into three parts. The objectives of each of these
parts are i) to verify the suitability of the chosen DEM for calculating further topographic
derivatives, ii) to derive flow accumulation maps and stream channel networks and quanti-
tatively investigate whether any statistical relationship to the seasonally inundated areas exist
and iii) to explore soil moisture conditions after flood retreat, inferred from the SAR backscatter
values, in order to interpret surface water dynamics over the summer season. The core of this
chapter discusses the effects of relief on seasonal inundation through terrain-based hydrolog-
ical modelling. The aim of this chapter is to underpin observations from the field on topog-
raphy by exploring Soviet maps, undertaking GIS hydrological modelling and exploring SAR
backscatter values in alasses and their relationship to soil moisture over the spring/summer
season.
5.1 Case 1: Local scale
The Yamal Peninsula is a flat, hummocky terrain; a gently rolling tableland that is divided by
a number of alluvial-lacustrine-marine plains (Walker et al. 2011). Terraces strewn with lakes
at different elevations make this permafrost area a unique but complex hydrological system.
The lower terraces are thought to be alluvial, originating from the Mordy-Yakha (Ìîðäûßõa)
and Se-Yakha (C¼ßõa) rivers. The IIIrd terrace is of alluvial marine/lacustrine origin and is at
an elevation of up to 26 m a.s.l. The IVth and Vth terraces are of coastal-marine origin and lie
at about 40-45 m and 58 m a.s.l., respectively. Vaskiny Dachi lies in the watershed of both the
Se-Yakha and Mordy-Yakha rivers (Leibman et al. 2012).
Although aware of the rolling terrain of the Yamal peninsula, the ability to visualise this
landscape was greatly improved by visiting the field in July 2012. Hilltops and cliffs, a land-
scape dissected by river valleys and their tributaries, strewn with lakes on both uplands and
in basins (sometimes single but often coalesced). So far this PhD research has, however, only
considered the dynamics of surface waters that are made visible through change detection
techniques. It became clear that the two-dimensionality of a remotely sensed image will not
account for topographic controls that may play a considerable role in runoff, spring floods and
drainage mechanisms. An analysis of seasonal surface water extent, therefore, needs also to
consider topography. This issue is addressed in this chapter by modelling stream channel net-
works on the basis of a DEM. Hence, the present study was undertaken in order to investigate
the effects of relief on the seasonal dynamics of surface water in the Vaskiny Dachi area of the
Yamal pensinsula. The exact area of investigation was chosen to be ROI Terra, as defined in
chapter 3 (section 3.3) .
88
Chapter 5. Modelling stream channels Data
5.2 Data
5.2.1 Digital Elevation Models - types and features
Topography is fundamental to the geosciences, and most particular to geomorphological stud-
ies. However, topography also plays a role in the movement and settlement of peoples in past,
present and future. Analysis of topography is done across disciplines, from archaeology to
those in urban management and planning.
In this digital age, the need to examine topography has lead to the development of elevation
models in digital formats. Modern day geographic information science relies on the input
of these digital elevation models (DEMs), which are also essential data used for geometric
correction and orthorectification of remotely sensed imagery, as discussed above in section
3.1.2. The pursuit to develop DEMs started more than 30 years ago (Hirano et al. 2003). Spatial
resolutions and vertical accuracies have been improving ever since, although the quest is far
from achieved. Global DEMs have been available since the early 1990s at spatial resolutions
of approximately 1km, occasionally finer (Rees 2012). These DEMs are either compiled from
topographic maps (e.g. GTOPO30) or derived from satellite sensors (e.g. SRTM and ASTER).
Less common, due to the high costs (financial, logistical and time intensive) that are involved
in such research, are ground (e.g. dGPS) or airborne (e.g. LIDAR sensors) surveys with higher
spatial and vertical resolutions. Nonetheless, an increasing number of researchers have found
the time and money to produce these ’home-made’ DEMs which in an ideal world would be
made available to the entire scientific community. However, neither of these two methods are
available for the present study. Here follows, therefore, a short account of the global DEMs that
are available and have been used for this research.
5.2.2 ASTER GDEM2 and ancillary materials
This study takes advantage of the Advanced Spaceborne Thermal Emission and Reflection Ra-
diometer (ASTER) Global Digital Elevation Model Version 2 (GDEM2), which are available and
downloadable free of charge in seamless 1°× 1° tiles via the United States Geological Survey’s
Earth Explorer (http://earthexplorer.usgs.gov/). ASTER GDEM2 was released in October 2011
and retains its predecessor’s nominal resolution of 30 m. GDEM2 has an absolute vertical ac-
curacy of 17 m at a 95% confidence level (ASTER GDEM Validation Team 2011). An additional
260 000 scenes and improved water body masking have strengthened GDEM2 with respect to
version 1 globally, however, the ASTER GDEM Validation Team also noted that at high lati-
tudes there is decreased stereo coverage compared to GDEM1. In addition, the GDEM2 water
mask is derived from the Shuttle Radar Topography Mission (SRTM) which does not cover
latitudes above 60° N (ASTER GDEM Validation Team 2011).
The poor stereo coverage combined with the issue of the SRTM water mask footprint needed
to be addressed first, as it was established that the inland water bodies in the Vaskiny Dachi
area were associated with artefacts in the form of spike-anomalies. These spike-anomalies are
89
Data Chapter 5. Modelling stream channels
Figure 5.1: Example of a spike anomaly. Surface plot showing unrealistically high elevation
values (left panel) associated with low stereo coverage (right panel).
artefacts of erroneously increased elevation compared to their low surroundings. They result
in spike-like features scattered throughout the landscape and are found to be associated with
areas of low stereo coverage (see figure 5.2.2). Quality assessment is done by inspection of
ASTER stacking numbers (Rees 2012). For the Vaskiny Dachi area stacking numbers lie be-
tween 1 and 12. Masking pixels with stacking numbers ≤ 4 eliminated spike-anomalies over
water bodies.
In addition to the ASTER GDEM2, this study also makes use of a Soviet topographic map
(publication date 1987, scale 1:200 000) of this area. This map was acquired from the Cambridge
University Library, scanned and georeferenced in ArcMap, transforming from the original da-
tum of Pulkovo 1942 to the more widely used WGS84. Furthermore, a second DEM is used
to validate the GDEM2. This DEM was compiled under the ESA Data User Element (DUE)
Permafrost project from digitised Russian Topographic Maps at a 1:200,000 scale. Santoro et al.
(2011) have made this dataset available on Pangaea, an open access database hosted by the Al-
fred Wegner Institute for Polar and Marine Research (AWI). The dataset covers latitudes >55°
N and and is provided in equiangular (WGS84 Lat/Lon) projection. Similarly to the GDEM2, it
is provided in 1°×1° tiles. The pixel posting is 3 arcsec (i.e. 0.0008333333°), which is equivalent
to approximately 90 m at the equator. And lastly, a panchromatic SPOT image from July 1993
is used in section 5.5 to visualise and empirically analyse flood locations within alas-systems.
90
Chapter 5. Modelling stream channels ASTER GDEM2 validation study
5.3 ASTER GDEM2 validation study
As mentioned in the introduction to this chapter, the focus of the chapter is on using the ASTER
GDEM2 data to undertake GIS-based hydrological modelling. As a first step, however, it was
decided to validate the GDEM2 for the Vaskiny Dachi area since a valid DEM is essential to
hydrological modelling and the large range of stacking numbers may have resulted in vertical
inaccuracies. The validation process was done by generating random points (N = 100) in
ArcMap and comparing the GDEM2 elevation heights to the fully co-registered Soviet map.
Furthermore, a DEM for the North of West Siberia has been made available under the DUE
Permafrost project, and was also used in chapter 6 for characterising the terrain’s relief.
5.3.1 GDEM2 and map spot-heights distribution test
5.3.1.1 Method
First, the ASTER GDEM2 data were re-projected to UTM zone 42N, and resampled to a 18.15
m × 18.15 m grid. Relating the elevation distribution of the GDEM2 to a scanned paper map,
however, was not a mean feat. Full digitisation of contours is in general costly, therefore, ran-
domly generated spot-heights were extracted for both datasets. This of course had to be done
manually for the Soviet map, introducing an element of human error, each point being allo-
cated the value of the nearest contour.
Probability density functions for both sample data (ASTER GDEM2 and Soviet map ele-
vation values) were produced in order to determine any similarities in the distribution of the
elevation data. To quantify the nature of these elevation distributions, a statistical test known
as the Kolmogorov-Smirnov test (K-S test) was done. The K-S test’s null hypothesis is that both
datasets follow the same distribution. The K-S test is a non-parametric test that compares the
empirical cumulative distribution function (ECDF) of the two sample datasets, in this case the
ASTER GDEM2 spot-heights and the map spot-heights.
5.3.1.2 Results
The probability density functions of the two datasets are given in figure 5.2. These functions
already indicate the non-normal behaviour, however, to investigate whether the sample data
are taken from populations with identical distributions a statistical test may be applied. The
ECDFs’ of both datasets are indistinguishable if the K-S statistic (D) is equal to zero. If D
is equal to one, then the ECDFs are completely distinguishable. The plot of the two ECDFs,
including this distance D derived from the two-sided K-S statistic, are given in figure 5.3. D
is computed to be 0.279 with a p-value of 0.00246. D is relatively small, however, given the
extremely small p-value the null hypothesis has to be rejected. The two datasets appear to
be statistically distinct. This outcome is unexpected, as it would be hoped that reference and
validation datasets follow the same distribution, however, it is possible that the discreteness of
the spot-height data, due to the relatively small sample size, biases this result. It was therefore
decided to proceed with a different validation approach.
91
ASTER GDEM2 validation study Chapter 5. Modelling stream channels
(a)
(b)
Figure 5.2: The probability density functions for both (a) ASTER GDEM2 and (b) the Soviet
map spot-height data sets.
92
Chapter 5. Modelling stream channels ASTER GDEM2 validation study
Figure 5.3: The empirical cumulative distribution functions of both ASTER GDEM2 and the
map spot-height data sets.
5.3.2 GDEM2 and map spot-heights direct comparison
Given the aforementioned discreteness of the validation data, a different approach is explored.
A direct comparison of the two spot-height datasets in the form of the root mean square error
(RMSE) in the z-direction was undertaken. The RMSE does not distinguish between random
and systematic errors but indicates the overall general error (Nikolakopoulos et al. 2006). Nei-
ther can it provide any information on any spatial relationship of the error, although it has
previously been found that errors in a DEM are indeed spatially autocorrelated and are cou-
pled to the properties of the terrain and its complexity (Wise 2011, Gao 1997). Nonetheless,
the RMSE is an established method in DEM error analysis that only requires the knowledge of
sample spot-heights. As these spot-heights have been sampled for both the Soviet map and the
GDEM2 datasets, an RMSE analsys is presented here.
5.3.2.1 Method
Since the RMSE provides some information on the degree of correlation between two datasets,
it was calculated for the sample elevation values of the GDEM2, hGDEM2(x), and the map spot-
heights, hmap(x). In order to appropriately account for any biases, it was further decided to
investigate the extreme difference values. These are the values that differed the most from one
another. It was decided to do linear regression for the GDEM2 and the map spot-height data
- as the elevation values should fit a 1:1 line if the data were perfectly matched. As some of
the ASTER data had flag values set to -9999, the data were subseted to only include elevations
greater than 0 m a.s.l, avoiding NANs and minus infinity values when calculating the RMSE.
Extreme difference values were categorised applying the conditional statement that any ASTER
GDEM2 values that were greater or less than the map spot-heights ±6σ, were to be classed as
extreme. Out of the original 100 sample points, 86 could be unambiguously used for the linear
93
ASTER GDEM2 validation study Chapter 5. Modelling stream channels
regression analysis and 33 were classed as extreme. Given this rather small sample dataset, a
t-test was carried out (the null-hypothesis being that the sample means are equal) to further
explore the statistical significance of the correlation in these two variables. Finally, the mean
error in the sample points was calculated. The error in the sample data points, e(x), is the
difference in the elevation between the datasets and is therefore given by
e(x) = hGDEM2(x)− hmap(x) (5.1)
The GDEM2 elevations were then analysed as a function of this sample error, including a linear
regression analysis. A one-tailed t-test was undertaken to examine whether the slope of the
regression fit was significantly different to zero.
5.3.2.2 Results
The straightforward RMSE calculation yielded a value of 12.17 m. It has been claimed that the
RMSE in the z-direction should be around 7 m for high accuracy, and that an RMSE greater
than 15 m is insufficiently accurate (Huggett 2002). According to this statement the GDEM2 is
sufficiently accurate for the Vaskiny Dachi area - albeit almost borderline. Nonetheless, given
the localised application of the GDEM2 with its nominal 30 m resolution, an RMSE of ~12 m
is an acceptable result. Linear regression resulted in a very low R2 value of 0.15, with p-value
< 0.0001 (i.e. the odds for this happening by chance are 1 in 104). The plot of these data is
given in figure 5.5(a). The red error bar indicates the confidence level that is derived from the
standard deviation (σ) of the fit. The extreme difference values are shown in green in figure
5.5(b), the 1:1 line is depicted in black and the linear best fit is illustrated by the dashed blue
line. Intriguingly, linear regression has resulted in an unexpectedly low correlation between
the two sample datasets. Therefore, a t-test was carried out. Setting the degrees of freedom
to 163 and only accepting a significant result if p≥ 0.05, resulted in a critical t of 1.974. The
t-test of the GDEM2 and map spot-height data determined a t of 1.116, which is smaller than
the critical t. Hence the null-hypothesis that the sample means are equal cannot be rejected. A
boxplot of these data is given in figure 5.4, clearly indicating the outliers. Figure 5.6 shows a
scatter plot of this error versus the GDEM2 elevation.
The mean error, < |e(x)| >, in these two datasets is found to be 16 m, however, a bias
towards a greater error may be found due to the anomalous sample point at e(x) = 58 m, as
can be distinctly seen in figure 5.6 and indeed the anomaly is also identifiable by the outliers in
figure 5.4. In fact, linear regression analysis of the GDEM2 sample data versus the sample error
gave an R2 value of 0.1521. The slope being 0.326. The one-tailed t-test resulted in a p-value
of 0.0001, which confirms that the slope of the regression line is significantly different from
zero. This further illustrates the bias that the anomalous sample point introduces into this error
analysis. Removal of the anomaly results in a mean error < |e(x)| >= 8.06 m and an RMSE of
10.505 m (N = 85). This procedure, therefore, minimises the RMSE to a more acceptable value,
well within the limits suggested by Huggett (2002), indicative of fair accuracy of the ASTER
GDEM2.
94
Chapter 5. Modelling stream channels ASTER GDEM2 validation study
Figure 5.4: Boxplot of the ASTER GDEM2 and Soviet map elevation sample data. The means
are not significantly different, although on average the GDEM2 sample data are slightly higher
than the map sample data. The outlier affecting the bias is also clearly visible as the individual
point.
5.3.3 GDEM2 and DUE Permafrost DEM comparison
5.3.3.1 Method
Following on from the spot-height data validation technique, a DEM cross-comparison was
undertaken. The ASTER GDEM2 was compared to a DEM compiled under the ESA DUE Per-
mafrost project (Santoro et al. 2011), henceforth referred to as the D-DEM. The D-DEM data
were reprojected into the same projection as the ASTER GDEM2 data (UTM zone 42N, WGS84)
and the GDEM2 dataset was resampled to the D-DEM’s pixel size of 31.18 m× 92.97 m using a
nearest neighbour resampling algorithm. Both datasets were resized to an entirely overlapping
area, excluding any regions containing NoData values. The RMSE for these data were then cal-
culated, at first without considering the effects of the water body mask that was applied to the
GDEM2 to eliminate the spike anomalies, as described in section 5.2.2, and later accounting
for this issue. This was done to explore to what extent these spike anomalies impact on the
RMSE. Furthermore, as a means to visualise elevation differences, two transects were chosen
for which elevation profiles were created.
5.3.3.2 Results
Excluding the water body mask, the RMSE was found to be 9.75 m (N = 648900). However,
taking the water body mask into consideration, the RMSE is reduced to 9.64 m (N = 506877).
The DEMs of the study area are shown in figure 5.7 for b) D-DEM, c) GDEM2 and d) the
GDEM2 - D-DEM difference. The sample statistics are given in table 5.1. The mean elevation
of the study area is approximately 16 m for both datasets. The standard deviation is greater for
the D-DEM, but not unsurprisingly so given its lower spatial resolution. These sample statistics
95
ASTER GDEM2 validation study Chapter 5. Modelling stream channels
(a)
(b)
Figure 5.5: ASTER GDEM2 points as a function of map spot-heights. The red error bar indicates
the confidence level derived from the standard deviation (σ) of the linear regression fit. The
green symbols are those values classified as extreme difference values, deviating from the linear
fit by ±6σ. The black line is the 1:1 correspondence line.
96
Chapter 5. Modelling stream channels ASTER GDEM2 validation study
Figure 5.6: The difference in GDEM2 and the map spot-heights (i.e. the sample error, e(x)) ver-
sus the elevation taken from GDEM2. Linear regression fit given by dashed blue line. Low R2
value was inspected by carrying out a one-tailed t-test showing that the slope of the regression
line is significantly different from zero.
Table 5.1: DEM sample statistics
Group N Mean [m] Standard deviation [m] Variance [m2]
ASTER GDEM2 506877 15.84 9.73 94.54
DUE Permafrost DEM 506877 16.09 12.79 163.75
GDEM2 - D-DEM difference 506877 -0.25 9.63 92.83
already give a good indication of the comparability of the datasets, which is further affirmed
by the RMSE.
Elevation profiles are given in figure 5.8, where the GDEM2 is given in grey and the D-
DEM in black. The discrepancies between the two are clearly visible. In addition, the GDEM2
exhibits saw-tooth behaviour which is associated with the water body mask - dropping eleva-
tions to zero. For illustration purposes, the water body mask is not considered for the D-DEM,
where a more natural, smooth elevation profile is found.
5.3.4 Discussion: GDEM2 validation
The three validation studies undertaken here would indicate sufficient correlation between the
ASTER GDEM2 and the reference datasets. The K-S test showed rather poor correlation be-
tween the GDEM2 and the Soviet map spot-heights. This is not an encouraging result but
clearly sample size makes a difference. The design of this study was biased towards an unfa-
vorable outcome due to the small sample set and therefore the discreteness of the sample data.
97
ASTER GDEM2 validation study Chapter 5. Modelling stream channels
Figure 5.7: a) D-DEM cover of the Yamal peninsula, the extent of the study area is illustrated by
the green frame, and b) D-DEM, c) GDEM2 and d) difference DEM of the study area (WGS84
Lat/Lon).
98
Chapter 5. Modelling stream channels ASTER GDEM2 validation study
(a)
(b)
(c)
Figure 5.8: (a) GDEM2 and D-DEM (WGS84 UTM zone 42N) illustrating the outline of elevation
profiles in green (A - A’) and blue (B - B’) and (b) the elevation profile A - A’ and (c) the elevation
profile B - B’ for both GDEM2 and D-DEM in grey and black, respectively. Colour stretch
according to previous figure 5.7.
99
GIS-based surface flow modelling Chapter 5. Modelling stream channels
It has been recorded that the K-S test does not perform well for discrete data (Greenwell and
Finch 2004). No doubt a more positive outcome could be achieved if the sample size was in-
creased two-fold or more. However, this is a manually extremely laborious method so another
approach was taken. The RMSE of the GDEM2 and the Soviet map spot-heights provided a
more sensible result, and although the set of sample spot-heights is still rather small, the third
study using the D-DEM with pixel counts O
(
105
)
as reference data resulted in only a slight
improvement in RMSE by 2.64 m, compared to that of the spot-height data. Both results are
within the range suggested by Huggett (2002) for fair DEM accuracy. With this in mind, it
seemed the GDEM2 data would be a good enough data source to use for the following study
on modelling stream channel networks.
As an aside, it is important to note the fact that the GDEM2 does not originally contain a
water mask needs to be addressed first before analysis can take place. This of course, in the
context of the current research, is easily rectified given that mapping water body extent and
producing adequate products that may then be used as water body masks, is intrinsic to this
PhD research. Considering the stacking numbers of the GDEM2, it was found that the spike
anomalies of the DEM were not only related to the water bodies, but also had low stacking
numbers. Hence, pixels that were associated with stacking numbers n ≤ 4 were excluded
from the analysis. Overall, once the spike anomalies have been removed, the ASTER GDEM2
produces a sufficiently accurate DEM that may be used for further topographic analysis.
5.4 GIS-based surface flow modelling
Once the accuracy of the DEM has been deemed sufficient for further implementation, a GIS-
based watershed analysis may be undertaken. The ArcMap Hydrology module tools were
used for this analysis. These tools allow DEM pit removal to create flow direction - also known
as local drainage direction (LDD) - maps which in turn are used to calculate watersheds and
delineate catchments. Watersheds are bound by topography and their outlines may be traced
along ridges between hilltops that act as vertices. The catchment (or drainage basin) is the
area enclosed by the watershed vertices and the drainage outlet. Watershed and catchment
delineation is a common and straightforward procedure that, therefore, solely depends on the
fact that gravity ensures water will flow downhill, following the steepest route. Hence, these
watershed algorithms rely only on topography and therefore need DEMs of high accuracy
as input parameters. In addition, it needs to be ensured that all pits are removed. Pits are
DEM sink-cells; in a raster they are represented by pixels that indicate a depression, being
surrounded by higher elevation and hence cells of higher pixel values. These pits can cause
problems for hydrological modelling as they trap flowlines and allow water to accumulate.
Obviously, some of these pits may be natural; they may be the location of lakes. However,
most pits are gridding and resampling artefacts which may be related to interpolation errors
(Huggett 2002). It is therefore customary to run pit-removal algorithms first.
100
Chapter 5. Modelling stream channels GIS-based surface flow modelling
5.4.1 Flow accumulation and channel network analysis
5.4.1.1 Background and Method
The DEM is then used to produce an LDD map showing the flow of water across and the
connectivity between each pixel cell (Conolly and Lake 2006). LDD maps are therefore flow
direction matrices which are the basis for calculating further derivatives. In this case, the LDD
matrix was first used to establish the catchment boundaries for the Vaskiny Dachi area. The
software examines the LDD matrix and establishes drainage basins for connected matrix cells
by automatically selecting pour points (the watershed outlet points) at the edge of the grid and
then identifying the cells that contribute to each of these grid drains. This is a fairly crude
method that may result in a number of smaller catchments towards the grid edges since pour
points are not selected manually. Nonetheless, it provides ample information for the purpose
at hand. The focus is on the two central and adjacent catchment areas that are associated with
the Mordy-Yakha river in the South and the Se-Yakha river in the North.
Seasonal inundation and changes in lake surface extent in the Mordy-Yakha and Se-Yakha
catchments are analysed for each catchment respectively. While the Envisat ASAR WS data
coverage are almost entirely confined by the derived catchment boundaries (all but the very
tip of the Se-Yakha catchment (~7.8 km2) are not covered by the ASAR WS lake classification),
~95 km2 and ~301.3 km2 of the Mordy-Yakha and Se-Yakha catchments are not covered by
the acquired TerraSAR-X scene. However, the changes in water body extent are derived from
Envisat ASAR WS data and therefore an analysis of seasonal inundation for both catchment
areas in their near entirety may be undertaken.
Moreover, morphometric properties of the Mordy- and Se-Yakha drainage basins are calcu-
lated in order to further investigate landscape components. The landscape properties can be
divided into three basic levels (Huggett 2002):
one dimensional - Linear: Stream channel networks
two dimensional - Areal: Catchment characteristics
three dimensional - Relief
This three-tier system analysis enables the establishment of physiographic regions related to
catchment runoff. The stream channel network is a planar dendritic-system, characterised by
its interconnectivity; stream channels flow into larger streams, intersecting at junctions that are
defined by their surroundings.
This channel network is, then, a function of topology, i.e. landform structures; these possess
geometrical properties, such as the length of each channel network link, the surrounding area
and relief, as well as the orientation of the network. The network is defined by each of its
branches, known as segment links. These links can be hierarchically organised, a common
ordering-system being the Strahler system, which orders upstream network termination links
101
GIS-based surface flow modelling Chapter 5. Modelling stream channels
as a first-order segment. Second-order segments are created by the interconnection of two first-
order segments, third-order segments arise at the junction of two second-order segments and so
on (Huggett 2002). Before the stream channel network can be modelled, the flow accumulation
matrix must be calculated. The flow accumulation matrix is a LDD derivative. The cell values
of the flow accumulation matrix are the total number of upstream cells that contribute to flow
into each downslope cell. Hence, high accumulation values are regions of concentrated flow
from which stream channels may be identified. A simple script (in this case coded in Python
syntax) can transpose the flow accumulation matrix into a stream channel map with cells that
only contain watercourses; this is a straightforward Boolean operation for which the pseudo-
code is:
Streams = if ( flow accumulation value > or = threshold value for number of up-
stream cells then 1 else 0)
The crucial input parameter is the number of upstream cells that defines each watercourse.
This parameter is a function of precipitation, land cover and geology. Nonetheless GIS derived
stream channel maps are good models, often comparable to, and less work intensive than, wa-
tercourse digitisation from paper maps (Conolly and Lake 2006). The threshold value enables
catchment partitioning; if the number of upstream cells that contribute to each watercourse
is high, then the stream network will be less dense, and there will be fewer subcatchments
(Huggett 2002). This was kept in mind when deciding on the threshold value. Given both
rather coarse spatial resolutions of the GDEM2 and the ASAR WS data, it seemed appropriate
to use a relatively high threshold value. The present study uses the conditional statement that
a contribution of > 500 upstream cells is needed to form a stream channel. This is equivalent to
a minimum areal runoff contribution of approximately 0.16 km2.
Line density values of stream channels, in a 10 km neighbourhood-radius of all stream
channel grid cells, were calculated. Line density units are therefore meters of stream channel
length per square meters of neighbourhood area. Further, mean stream channel densities were
calculated within each surface water change polygon; the underlying thought being to investi-
gate whether areas of seasonal inundation are related to a high density of stream channels, and
therefore, whether seasonal floods are associated with hillslope runoff rather than lake basin
overflow.
The stream channel network was separated into stream links that were classified according
to the Strahler method. One of the most important morphometric properties related to the
one-dimensional landscape level is the drainage density, Dd, of a basin with area A. This was
defined by Horton (1932, 1945) to be:
Dd =
n∑
i=1
Li
A
(5.2)
where Li is the ith stream segment link and i = n is the highest order of stream segment links.
102
Chapter 5. Modelling stream channels GIS-based surface flow modelling
The drainage density is, therefore, the sum of the lengths of each channel segment within the
drainage basin divided by the area of this basin (Melosh 2011). The inverse of this property, is
known as the length of overland flow; it is approximately the mean distance water must flow
before it encounters and joins a defined channel (Horton 1932).
lo ∼ 1
2Dd
(5.3)
This may therefore be considered as the mean hillslope-to-channel length. Drainage density is
thought to be closely linked to climatically driven erosional processes (Tucker et al. 2001) as it
is immediately associated with the infiltration capacity of the ground. As such, drainage den-
sity is an important geomorphic parameter that provides key information on surface flow and
channel concentration, and its associated implications on sediment and mass transport (Bishop
et al. 2012). The current study will investigate channel networks derived from a flow accumu-
lation map, focussing the results on the statistics of the hillslope-to-channel length parameter
calculations.
5.4.1.2 Results
The borders of the Mordy-Yakha and Se-Yakha drainage basins were determined from the LDD
map (figure 5.9). According to these calculations the two basins cover an area of 506.2 km2 and
1087.9 km2, respectively. Using the ASAR WS and corresponding lake drainage data from
section 4.2, the loss in lake surface extent was analysed for each drainage basin. It was decided
only to include loss in extent ≥ 0.1 km2. This corresponds to ~18 pixels, which is set as the
minimum threshold for surface water change detection.
At the beginning of July 2007, limnicity (the areal ratio between surface water and land) for
the Mordy-Yakha basin was ~21.6% and for the Se-Yakha basin was ~19.3%. Limnicities are
based on the ASAR WS water body classification (cell size 0.000667° × 0.000667°; converting
the WGS84 geographic projection to UTM Zone 42 (same datum) and resampling the data to
arrive at an x-pixel size of 24.96 m and a y-pixel size of 74.38 m).
Applying the Strahler system to the derived stream channel network resulted in seven
stream orders. A topographic map of the two basins overlain by the stream channel network
is shown in figure 5.10. This has resulted in a dendric drainage density that may define areas
of high seasonal inundation. The drainage density for both Se-Yakha and Mordy-Yakha basins
individually were calculated according to equation 5.2. The former was found to be 1.986 km−1
and the latter 1.901 km−1. These low numbers are explained by the gently sloping terrain of
this area. From equation 5.3 the mean overland flow distance, lo, in this area is found to be
approximately 0.25 km and 0.26 km for the Se-Yakha and Mordy-Yakha basins, respectively.
Irrespective of stream segment order, this would mean approximately 4 stream branches are
found per 1 km radius. A histogram showing the stream frequencies for each basin is given in
figure 5.11.
103
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Figure 5.9: TerraSAR-X water body classification with outlines of the Mordy-Yakha and Se-
Yakha catchments (left panel) and legend (top right panel). Yamalo-Nenets Autonomous Dis-
trict borders according to Stolbovoi and McCallum (2002) with Vaskiny Dachi area extent indi-
cator (bottom right panel).
104
Chapter 5. Modelling stream channels GIS-based surface flow modelling
Figure 5.10: Derived stream channel network, with associated Strahler segment link order. The
ASTER GDEM2 is used as the background layer. Anomalously high pixel values were not
removed; it is known from the literature that the maximum elevation above sea level in this
area is just under 60 m a.s.l. (Leibman et al. 2012). 0.07% of the pixels were higher than 60 m
a.s.l., resulting in the depicted colour stretch.
105
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Figure 5.11: The frequency of stream segments for each drainage basin
The drainage density for each stream order for both the Se-Yakha and Mordy-Yakha basins
is given in tables 5.2 and 5.3, respectively. The higher the stream order, the lower the drainage
density. The method for computing the changes in water surface extent is given in section 4.2
and the statistics of these changes were discussed in the previous chapter in section 4.3.3. The
current study investigates period A-B changes for both of the derived catchment areas. In 2007,
these changes in surface water extent amounted to 17.7 km2 and 19.74 km2 within the Mordy-
Yakha and Se-Yakha catchments, corresponding to a change in limnicity of ~3.5% and ~1.8%,
respectively.
These are fairly large changes which may almost certainly be linked to snow meltwaters
and a thawing landscape at the start of the warm season. The interesting point to consider is
whether these spring floods are related to hillslope runoff or overbank flooding of lakes and
basin overflow. For this reason the stream channel network was investigated. Moreover, the
density of stream channels within each seasonally inundated area in the Mordy-Yakha and
Se-Yakha catchments was analysed.
As discussed above, line density values are calculated based on a 10 km neighbourhood
radius. The mean stream channel density is first investigated as a function of flood size. This
analysis showed a weak, non-linear correlation for both the Mordy-Yakha and the Se-Yakha
basins, with R2 values of 0.0167 and 0.0068 respectively. Although this already provides some
insight, further assessments are needed before the null hypothesis (in this case, that seasonal
inundation is related to hillslope runoff) may be rejected.
106
Chapter 5. Modelling stream channels GIS-based surface flow modelling
5.4.2 Topographic Wetness Index analysis
5.4.2.1 Background and Method
Having used the DEM to model surface flow across the tundra landscape, another raster ma-
nipulation may be applied to explore cells of high water accumulation by taking local slope into
consideration. The topographic wetness index (TWI) is often used to spatially model the satu-
ration of soil moisture based on terrain. Modelled wetness distribution analyses have shown
that the distribution of wetness is associated with downslope topographic attributes (Rodhe
and Seibert 1999, Crave and Gascuel-Odoux 1997). The TWI is therefore a static terrain-based
wetness index, intrinsically dependant on topography, used to explore the links between topo-
graphic properties and soil moisture (Temimi et al. 2010). As in the study outlined in section
5.4.1, the present analysis is founded on the basis that surface flow, in this case specifically
due to the saturation of soil and/or an excess water input greater than the rate of infiltration,
is also controlled by gravity (Stichbury et al. 2011) and therefore may be explored with the
aid of DEM manipulation. Furthermore, in order to analyse spring floods appropriately, the
spatial and temporal variability in soil moisture plays an important role. Saturated soils lead
to seasonally inundated areas. EO measurements of soil moisture rely on passive and active
microwave sensors with typically large footprints (De Jeu et al. 2009, Loew et al. 2013). Most
commonly scatterometer measurements are used in soil moisture studies (Bartalis et al. 2007,
Hahn et al. 2012, Wagner et al. 2013). SAR measurements of soil moisture are as yet in their ex-
perimental stage but have been used to improve spatial resolution (Loew et al. 2006, Pathe et al.
2009). Simple terrain-based wetness indecies, derived from static DEMs, may be of significance,
complementing soil moisture analysis. Correlation methods for linking the EO soil moisture
analysis to a terrain-based wetness index have been suggested by Temimi et al. (2010). Soil
moisture analysis will be briefly touched upon by examining the ASAR WS backscatter values
in section 5.5, but it will not be investigated in great depth, as currently there are other groups
working on exactly this topic (see PAGE21 21 Project at http://page21.org/), using their exper-
tise to produce soil moisture products from scatterometer data that will become freely available
online in the future. Therefore a more complete analysis of soil moisture in the Se-Yakha and
Mordy-Yakha basins is to be kept in mind for future work once access to these data has been
provided.
The present study considers the TWI in reference to the stream channel network derived in
section 5.4.1, as well as the spring flood events that were determined through change detection
of the ASAR WS data in section 4.2.
In order to calculate the TWI, it is necessary to first determine the slope gradient from the
DEM. This can be easily done using the formula
tan(θ) =
√(
∂z
∂x
)2
+
(
∂z
∂y
)2
(5.4)
where θ = θ(i, j) is the local slope gradient of each raster cell.
107
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Commercial software programmes often will have an inbuilt tool box to simplify the proce-
dure, and indeed the present study makes use of the slope tool in ArcMap. Further steps were
scripted in IDL.
Once the slope gradient has been calculated, the TWI may be derived from both this local
slope gradient as well as the flow accumulation area.
TWI = ln
(
a
tan(θ)
)
(5.5)
where a is the local catchment area (i.e. upstream contributing area) per cell size. Cell size
for GDEM2, and therefore all its derivatives, remains 18.14785 m × 18.14785 m = 329.345 m2.
Therefore flow accumulation, which as established in section 5.4.1 assigns an integer value to
each grid cell equal to the number of cells that contribute flow into it, needs to be scaled by the
cell size itself to arrive at the upland contributing area [m2].
The use of topography for exploring soil moisture has become an established discipline
within GIScience. Many studies on TWI as well as modifications to said TWI may be found
in the literature. As an example, (Sørensen and Seibert 2007) explore the depreciating effects
on TWI from DEMs at decreasing spatial resolutions, while Temimi et al. (2010) include the
leaf area index as a parameter of the TWI in their study, in order to account for the difference
in soil moisture between vegetated and non-vegetated areas. It is important to note, however,
that no prior knowledge on vegetation type and its attributes or even hydraulic conductivity of
the active layer are considered in the present study. Terrain-based studies often only consider
local slope, as it is thought to be representative of the downslope hydraulic gradient itself,
and therefore it is commonly assumed that groundwater tables follow topography (Grabs et al.
2009, Rodhe and Seibert 1999). This model may of course not be entirely representative of
permafrost environments, where ground water movements are almost entirely constrained to
the active layer as well as flow through taliks (van Everdingen 1990). A more complex model
would need to take this into consideration, using knowledge of ground temperatures and soil
type (Frampton et al. 2011), as well as possibly upscaling active layer depth and hydraulic
head measurements to investigate sub-surface water flow (Cooper et al. 2011). Furthermore,
the distribution of snow before the onset of thaw needs to be considered, as this not only has
an insulating effect on permafrost, and therefore ground temperatures (Gisnås et al. 2014), but
also snow water equivalent measurements (Langer et al. 2013) which could give an indication
of the amount of water available for spring flood events. Snow is not considered in this chapter,
but will be elaborated on in more detail in chapter 6. Woo et al. (2008) emphasis the need
to integrate a coupled analysis of ground water, soil moisture and surface water flow into
models of permafrost hydrology. Nonetheless, the development/integration of such a model is
beyond the scope of this study (and indeed this PhD research) as this chapter focuses on surface
water modelling from DEM inputs. More elaborate modelling approaches are not evaluated
here - although there is scope for future work on this topic. The current study is a terrain-
based investigation, using the traditional static approach for catchment analysis of wetness, by
108
Chapter 5. Modelling stream channels GIS-based surface flow modelling
Table 5.2: Se-Yakha stream channel statistics
Order N < Length >
∑
Length < Channel slope > Drainage density < TWI >
[m] [m] [°] [km/km2]
1 5825 0.1867 1087.47 5.608 9.9×10-4 0.3818
2 2453 0.2207 541.49 5.577 4.9×10-4 0.3799
3 1174 0.2149 252.37 5.728 2.3×10-4 0.3821
4 724 0.1748 126.59 5.548 1.2×10-4 0.3773
5 305 0.2197 66.99 4.993 6.2×10-6 0.3819
6 181 0.2422 43.83 4.708 4.02×10-6 0.3723
7 130 0.3199 41.59 1.615 3.8×10-6 0.3775
Table 5.3: Mordy-Yakha stream channel statistics
Order N < Length >
∑
Length < Channel slope > Drainage density < TWI >
[m] [m] [°] [km/km2]
1 2126 0.22 470.59 5.239 9.3×10-4 0.3803
2 1056 0.2214 233.82 5.094 4.6×10-4 0.381
3 447 0.3011 134.62 4.746 2.7×10-4 0.3807
4 274 0.2625 71.94 4.263 1.4×10-4 0.3797
5 102 0.3235 32.99 4.119 6.4×10-5 0.3796
6 80 0.2292 18.34 3.271 3.6×10-5 0.3857
exploring the standard TWI derived from both local slope and upslope contributing area of the
Mordy-Yakha and Se-Yakha catchments.
5.4.2.2 Results
As explained above, the slope of the DEM was calculated. A map of the slope gradient is given
in figure 5.12. As would be expected, the slope gradient associated with each stream channel
decreases with increasing Strahler order. Drainage density tends to decrease with increasing
slope gradient. An anomaly is found in the Se-Yakha basin, where stream channels of the 3rd
order have the highest mean channel slope with 5.728°. The mean channel slopes for each order
for both the Se-Yakha and Mordy-Yakha basins are given in tables 5.2 and 5.3, respectively.
Furthermore, the TWI for the Se-Yakha and Mordy-Yakha basins was calculated and nor-
malised, so that the index would range from zero to one. NaNs and infinity values, that arise
from equation 5.5 related to both the fraction as well as the natural logarithm, were excluded
from further analysis. The map of the TWI of the Se-Yakha and Mordy-Yakha basins is given
in figure 5.13. The histograms of the normalised TWI are shown in figure 5.14. The mean
109
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Figure 5.12: Slope gradient of the ASTER GDEM2.
110
Chapter 5. Modelling stream channels GIS-based surface flow modelling
Table 5.4: TWI statistics at the location of the flooded areas identified from the 2007 difference
map I.
Basin TWI statistics at floods
min max mean
Se-Yakha 0.058 0.996 0.382
Mordy-Yakha 0.082 0.974 0.389
TWI values for the Se-Yakha and Mordy-Yakha basins were found to be 0.38059 and 0.38472,
respectively. Over both catchments the mean TWI is 0.382. A K-S test confirms the inference
that the TWI values are sampled from the same population (D = 0.0239,p < 22 × 10−16). The
empirical cumulative distribution functions are given in figure 5.15. Both compare well and no
biases are found, illustrating a certain consistancy in landscape with respect to the TWI for both
catchment areas. Scatter plots of the mean slope gradient versus the mean TWI for each stream
channel are given in figure 5.16. No linear relationship can be determined for these variables
(R2 = 0.07, p-value = 0.46 and R2 = 0.29, p-value = 0.156 for the Se-Yakha and Mordy-Yakha
basins, respectively). The scatter plots also show the drainage density of each channel order
as a function of slope. It is found that drainage density of each stream channel order increases
with increasing slope gradient.
Furthermore, the statistical properties of the TWI over the areas that have been identified
by the change detection analysis in section 4.2.3 of chapter 4 to be flooded in early July each
year in 2007, 2008 and 2009 (although the last year is biased by wind as discussed in the afore-
mentioned chapter). As a representation of flood extent, only the 2007 difference map I was
used to evaluate the TWI over these flooded areas. The statistics are given in table 5.4. The
spatial variation in TWI is characterised by high values were these floods occur, although the
mean TWI values over the flooded areas are still of the magnitude of the mean TWI values of
the entire catchment areas.
The results show that, although the Yamal peninsula is a wetland environment, the vast
hummocky terrain leads to runoff into the valleys where TWI values are high. In fact, the TWI
clearly maps out inundated areas where the spike anomalies associated with water bodies had
been removed. These anomalies were set to zero, which means that elevation differences were
not badly affected by the removal procedure and indeed low slope gradients are found around
these water bodies. Therefore, it is to be expected that these areas are classified as wet, and it
is to be expected that water runoff finds its sink in these lakes. Nonetheless, according to the
frequency distributions of the TWI, both the Se-Yakha and Mordy-Yakha catchment areas may
overall be classified as moderately wet, rather than high. This indeed is representative of the
rolling terrain and can be confirmed from field experience.
111
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Figure 5.13: Map of TWI for the Se-Yakha and Mordy-Yakha basins (top panel), with spring
floods depicted for 2007, 2008 and 2009 from left to right (bottom panel).
112
Chapter 5. Modelling stream channels GIS-based surface flow modelling
(a)
(b)
(c)
Figure 5.14: Histograms of normalised TWI for the Se-Yakha basin (a), the Mordy-Yakha basin
(b) and both catchment areas combined (c).
113
GIS-based surface flow modelling Chapter 5. Modelling stream channels
Figure 5.15: The empirical distributions functions of the TWI for the Se-Yakha (blue) and
Mordy-Yakha (green) basins.
Figure 5.16: Mean TWI as a function of slope gradient (a), as well as drainage density for
each stream channel order also as a function of slope gradient (b) for the Se-Yakha (black) and
Mordy-Yakha (green) basins. Spline fit for drainage density vs slope.
114
Chapter 5. Modelling stream channels GIS-based surface flow modelling
5.4.3 Discussion: Surface water modelling
The Vaskiny Dachi area is an active area in which seasonal water body change is occurring.
It was shown in chapter 4 (section 4.2) that a clear spring flood pattern can be determined
(over the years 2007 - 2009) using the Envisat ASAR WS sensor. The spatial resolution of the
instrument is, however, such that no distinct interactions between the complex hydrological
system can be deduced. It was therefore decided to try to model the stream channel network to
analyse whether any interactions between the floods and the drainage system could be inferred.
Statistical analysis of the modelled terrain-based surface water flow showed that the Vaskiny
Dachi area is less prone to wetting than would be expected. The hypothesis was that the de-
tected spring floods were related to hillslope runoff, promoted by the limited infiltration capac-
ity of the continuous permafrost environment (Granger et al. 1984). This would be in line with
Romanovskiy (1961), who explain how runoff of spring meltwaters occurs through thermoero-
sion and thermokarst depressions. However, the locations of the observed ASAR WS spring
flood events only corresponded weakly to the regions were terrain-based GIS modelling pre-
dicted stream channels. Indeed, only a fair density of stream channels as well as moderately
high TWI values were associated with seasonal inundation. Neither of the aforementioned cal-
culations showed particularly high values over the flooded areas, but rather correspond to the
mean range of density and TWI values that were found over the entire Se-Yakha and Mordy-
Yakha catchments. If the null-hypothesis, that spring floods in the Vaskiny Dachi area are due
to the terrain-derived stream channel network and hence due to relief, were correct then high
stream density and TWI values should have been found over seasonally inundated areas. This,
however, was not the case.
Moreover, an analysis of line density as a function of flood size did not prove fruitful either.
Indeed only weak correlation for the line density values and flood sizes were found. Further-
more, hillslope-to-channel length is such that approximately 4 stream channels are found in
a 1 km radius. Drainage densities per Strahler order as a function of slope follow the typical
behaviour, increasing drainage density with increasing slope. However the total drainage den-
sity of the two catchment areas is less than 2 km−1 per catchment. This is rather low and a
reflection of the rolling landscape. The null-hypothesis is rejected.
It is believed that seasonally inundated areas are not solely related to hillslope runoff but
indeed a different mechanism must also account for the spring flood events, given that only
insignificant ties between the stream channel network and the observed ASAR WS flood events
could be established. Therefore, runoff must follow a number of different pathways, which of
course may include the routing through (but is not solely restricted to) the stream channel
network. Apart from surface flow originating from meltwaters in rills that adjoin to streams
(Howard 1971), subsurface flow within the thawed active layer need also be considered.
According to a study by Wang et al. (2009), active layer thawing processes were found
to have a considerable impact on the spring runoff regime. Higher topsoil temperatures in
shallow active layers were found to increase surface runoff, but this relationship could not be
115
GIS-based surface flow modelling Chapter 5. Modelling stream channels
identified for active layer depths greater than 60 cm. This would imply that increases in ac-
tive layer depth allow water infiltration and subsurface flow rather than surface runoff. Active
layer depth measurements are ongoing at the Vaskiny Dachi research camp, with 4 established
grids. In the year 2007 the average active layer depths at these sites ranged from 72 to 112 cm
(Leibman et al. 2012). These values would therefore favour infiltration and subsurface flow.
Furthermore it was noted during the field excursion in 2012 that although the landscape gave
the appearance of being relatively dry in patches (as inferred from the vegetation cover), the
topsoil itself was found to be moist when TDR measurements were taken. These measure-
ments, however, were only taken as a preliminary exploration - food for thought - before an
experiment could be set up. It is hoped that in future a study on the degree of wetness as seen
from both a VIR and a microwave sensor will be undertaken.
Carey and Woo (2001) discuss how modifications in the runoff pathways mirror an in-
creased coupling between the hydrological system over the course of the summer season, as
the active layer starts to thaw. In particular, diffuse subsurface flow within organic rich ac-
tive layers is a common phenomenon. They find that hillslopes comprised of an organic layer
overlying mineral soils display a two-layer flow resulting from alterations in the hydraulic con-
ductivity at the soil interface. Runoff occurrs through the porous organic layer. It is well known
that snowmelt infiltration has a warming effect on the active layer as water flow results in ad-
vection (Hinkel and Outcalt 1994). This in turn facilitates lateral water movements through the
active layer according to the hydraulic gradient reflected by the terrain’s slope (Scherler et al.
2010). The present study’s predicted channel network cannot account for any subsurface flow
due to the nature of the terrain-based model, however, the fact that spring flood events have
clearly been measured by the ASAR WS sensor implies the potential of subsurface flow activ-
ity. Leibman and Streletskaya (1997) and Streletskiy et al. (2003) discuss active layer runoff as
a mechanism for washing out saline deposits and ion migration in the Vaskiny Dachi area. It
therefore seems justified to argue that a combination of both surface and subsurface routing of
snow meltwaters are plausible within the Vaskiny Dachi area. However, without the appropri-
ate field observations and measurements, unfortunately no further discussion on lateral flow
through the active layer can be made.
With respect to overland water flow, it should be noted that high surface runoff is directly
linked to the erosional potential of hillslopes (Cherkauer et al. 2013). Indeed, the Vaskiny
Dachi area is well known to be a landslide-prone terrain (Leibman and Streletskaya 1997), with
active-layer detachment slides being a common phenomenon (see figure 5.17). Leibman and
Kizyakov (2007) discuss how this cryogenic landslide activity stands in direct relation to water
and sediment runoff, the development and evolution of stream channels that in turn dissect
the landscape, creating valleys. The importance of these landslides on the stream network and
hillslope runoff are not to be neglected, however their implications have not been investigated
here as such a study would require extensive field research. The possibility for collaboration
with Russian colleagues on such a project will be considered for future work.
116
Chapter 5. Modelling stream channels Empirical lake basin analysis
Figure 5.17: Active layer detachment slide (left panel) and landslide body (right panel) in the
Vaskiny Dachi study area. Willow shrubs are good indicators of landslide activity, as tall thick-
ets of these willows are found on old landslide bottoms due to greater nutrient availability
(Walker et al. 2011). Approximate landslide body and willow thicket outlined with dashed and
solid black lines, respectively. Photographs taken by A.M. Trofaier on 4 July 2012.
5.5 Empirical lake basin analysis
Having investigated terrain-based stream channel networks by modelling these in a GIS en-
vironment, it was decided to explore the flood locations and their relationship to lakes with
the aid of a panchromatic image taken by the SPOT satellite. An empirical approach was
taken, where the optical image was overlain by the ASAR WS derived floods. These were then
manually inspected in order to establish the locations of the active floodplains within drained
lake basins and possible interactions with the stream channels. Furthermore, the ASAR WS
backscatter values of the seasonally inundated areas as well as a bufferzone around these areas
was inspected as a means to evaluate the soil moisture content of said areas. It was mentioned
in section 5.4.2 that SAR soil moisture retrieval is yet in its experimental stage (Bartsch et al.
2007c). This is of course still the case and more research on this subject needs to be under-
taken, however, the present study does not aim to arrive at any absolute soil moisture values,
rather it is hoped that the ASAR WS normalised backscatter values of the selected regions will
give some relative indication of soil moisture, as radar data may be used to investigate ground
conditions by analysing trends in and temporal changes of the backscatter values.
5.5.1 Background and Method
Drained lake basins are a common phenomenon in Arctic permafrost landscapes. In Siberia,
they are often referred to as alasses, and on the Yamal peninsula they are known as khasyreis
(Leibman et al. 2012). These khasyreis resulted from thermokarst development during the
early Holocene. Morgenstern et al. (2013) discuss the possible alas development, thought to
117
Empirical lake basin analysis Chapter 5. Modelling stream channels
have originated at around 13 ka BP, by reconstructing permafrost degradation and dynamics
that could have led to the alas development in their study area in Eastern Siberia.
Alasses act as floodplains. They are depressions towards which overland runoff may drain,
and lakes may overbank during spring thaw activities. Hayashi et al. (2003) consider the runoff
of water into depressions. They note that lake basin depressions in permafrost environments
have a large water storage capacity which controls the input of meltwaters to the stream chan-
nel system during the active season. Therefore, khasyreis are surface water stores of snow
meltwaters. It is hence proposed to take a different approach and to investigate the location of
the ASAR WS floods using optical imagery and to explore the nature of flooded depressions,
as accumulation of meltwaters in these depressions is often also linked to overbank flooding of
lakes.
In addition to spring meltwater runoff, Kirpotin et al. (2008) present their findings on lake
drainage from small lakes on higher levels surrounding large lakes on bottom levels in alasses.
This topic is also touched upon by Morgenstern et al. (2013), and it is noted here that field
observations clearly indicate lakes lying on different levels and terraces on the Yamal peninsula,
again highlighting the scope for developing and applying a hydrological model in this region.
The present study, however, is an empirical investigation of seasonal inundation and the
associated alas areas that were flooded during snow melt. Using the vectorised difference map
I from 2007, floods in these alasses were inspected with reference to a panchromatic image from
1993 taken by the SPOT satellite. To avoid detected changes resulting from classification mis-
matches due to vegetation or wind biases (as discussed in section 4.2.3) and in order to focus
on the most significant flood events, only water bodies in difference map I greater than 0.1 km2
were considered.
Furthermore, it has been shown by Ulaby and Batlivala (1976), and in a follow-up experi-
ment by Ulaby et al. (1978), that a near linear relationship between the radar backscatter coef-
ficient and surface layer soil moisture exists. This claim, though, is greatly biased by surface
roughness and vegetation. Therefore a straightforward interpretation of the radar signal in
terms of soil moisture is not possible. However, keeping the linear relationship in mind the
simplistic approach of radar backscatter is proportional to soil moisture can be used for inves-
tigating soil moisture conditions. The backscatter values of the drained lake basins as well as a
surrounding bufferzone, representing the lake banks, were inspected.
5.5.2 Results
In ROI 1 a total of 109 polygons associated with spring floods were found to be greater than
0.1 km2 in the 2007 difference map I. Of these 109 water bodies, only 91 were located within
the SPOT scene and could hence be used for further empirical investigation. The greatest flood
(ID 181) was found to be 12.84 km2 large. It is illustrated in figure 5.18. Of particular interest
is a water body (ID 892) that had shrunk to a tenth of its 1993 size, when its surface area was
approximately 0.65 km2 compared to the August 2007 size of 0.06 km2. The basins in which
118
Chapter 5. Modelling stream channels Empirical lake basin analysis
these floods occurred were manually inspected. From the SPOT imagery it was quite clear that
all floods were located in alas depressions. However, only four of these basins could clearly be
classified as coalesced alas-systems.
The water body ID 145 is notable, as it is one of the largest lakes in this region (its surface
extent in July 2007 being ∼ 17 km2 but by the end of August 2007 it has completely drained).
Similar conditions apply for the water body ID 535; its surface area being about 4 km2. This
area is approximately the same as is visible in the SPOT image of July 1993. However, in
August 2007 the complete water body had drained. Furthermore, it cannot be ruled out that
this particular water body also drained over the active season in 1993. Nor can it, however, be
affirmed. It is nonetheless a good example for the need of attentiveness in decadal time slice
comparison studies. This will be discussed in more detail in the next chapter. Unfortunately,
the water body ID 145 lies slightly below the SPOT image scene and therefore no comparison
to the SPOT image can be made. The inundated region with ID 675 has been classified as a
water body, however, according to the SPOT imagery no water body is to be found at those
coordinates but an alas, exactly representing the shape and size of the water body ID 675 was
visible in 1993.
The mean backscatter value of the lakes in ROI 1, as well as the mean backscatter value of
a surrounding bufferzone, was analysed over the given time period. Investigating changes in
the backscatter of the surrounding areas allows for an insight into soil moisture content and
possible vegetation changes. The backscatter of three prominent water bodies is given in figure
5.19 and their location is indicated in figure 5.18. The backscatter values of each of these water
bodies follows a similar trend. Once the water has drained the backscatter values range from
−7 to −10 dB. The bufferzones have backscatter values within the same range as the drained
basins. In all three cases a dip in the backscatter in the second week of August can be seen.
Bufferzones ID 535 and ID 675 display lower backscatter values at the beginning of July 2007.
This could be interpreted to have resulted from a wet-snow/slush at the beginning of the thaw
season. According to the Marre-Sale weather station the start of thaw occurred on 15 June
in 2007 and Vaskiny Dachi is 100 km further North and inland. The thaw date will hence be
slightly retarded and the snowpack may still be present at the end of June. Park et al. (2010)
investigate the freeze/thaw transition in the Yakutia region in Eastern Siberia using ASAR
operating in Global Mode (GM). They find that the GM thaw date falls within ±1 week of the
beginning of snowmelt. In the following chapter, the effects of the snowpack and the timing
of the snow-off date will be explored in more detail, using ancillary products derived from
scatterometer data.
5.5.3 Discussion
In cold-climate environments interactions between surface and ground water are generally con-
sidered to be negligible due to the large layer of permafrost that isolates any suprapermafrost
water fluxes, separating the suprapermafrost water table from the ground water table (Grosse
119
Empirical lake basin analysis Chapter 5. Modelling stream channels
Figure 5.18: SPOT image of the Vaskiny Dachi area, overlain by 2007 difference map I.
120
Chapter 5. Modelling stream channels Empirical lake basin analysis
(a) Mean backscatter value over time for water body ID 145
(b) Mean backscatter value over time for water body ID 535
(c) Mean backscatter value over time for water body ID 675
Figure 5.19: Changes in mean backscatter values over the spring/summer season 2007 for se-
lected water bodies (left panel) and a surrounding bufferzone (right panel).
121
Empirical lake basin analysis Chapter 5. Modelling stream channels
et al. 2013). Lake recharge, and seasonal inundation are hence related to summer precipita-
tion and snowmelt runoff. The depleted snowpack recharges the water storage of water bodies
before it can significantly interact with the stream channel network and become a dominant
factor in contributing to stream flow. It is generally appreciated that the snow and ice and their
associated meltrate in the spring season constitute a significant factor contributing to stream
flow dynamics in a cold-climate catchment (Cherkauer et al. 2013). Snowmelt is the dominant
hydrological flux that recharges depleted water storages (Grosse et al. 2013). The floods identi-
fied by the ASAR WS sensor are equally likely associated with meltwaters from the snowpack.
Meteorological data from the Marre-Sale weather station show that in 2007 there was relatively
low winter (149.9 mm) and summer precipitation (147.6 mm) compared to the following year,
when in 2008 winter precipitation amounted to 294.3 mm and summer precipitation was as
high as 319.3 mm. These are curious numbers, as there does not seem to be a great difference
between the flooded areas in 2007 and 2008, as was established by the change detection tech-
nique discussed above in section 4.2. It is possible that these precipitation values reflect the
inadequacy of using meteorological data from a weather station that is situated about 100 km
to the southwest of the study area on the coast. Unfortunately, no weather station has been
established closer to the Vaskiny Dachi site, a fact that should be considered in future project
proposals, especially considering there is much ongoing research in this area that would ben-
efit from accurate meteorological data, and particularly so as monitoring changes in recharge
patterns will provide important information on the impacts of permafrost degradation on the
ecosystem (Grosse et al. 2013).
Floods are found in alas-systems, a few of which are coalesced basins. Many of these floods
are located within the Bovanenkovo gas field. The top panel of figure 5.20 shows the SPOT
image, zoomed to the gas field’s major infrastructure. The bottom panel is a high-resolution
false colour Quickbird-2 composite taken from Forbes and Kumpula (2009) which depicts the
gas field’s infrastructure in further detail. The Se-Yakha river flows right through this devel-
opment. The stream channel network overlying the SPOT image is given in the top right panel
of the figure. Clearly the sink of the detected floods is the Se-Yakha river, but artificial changes
to the stream channel network are most probable given the location of the Bovanenkovo com-
plex. Many of the floods occur in the immediate vicinity of roads and buildings with close links
to the river network. Zakharova et al. (2011) mention how human activities are impacting on
the drainage network in the Poluy-Nadym-Pur-Taz area, which lies to the South of the Yamal
peninsula. This illustrates the impacts human activity can have on the hydrological system,
which are certainly not to be ignored. The potential anthropogenic influences, as found in the
Vaskiny Dachi study area and on the Yamal peninsula in general, are discussed in more detail
in the next chapter.
The SAR backscatter values are used as a proxy for soil moisture conditions in bufferzones
around the main water bodies.
Analysis of the backscatter values of floods that appear to drain over the course of the
first two weeks in July according to the change detection analysis, showed a steep increase
122
Chapter 5. Modelling stream channels Empirical lake basin analysis
Figure 5.20: SPOT image zoomed to the Bovanenkovo gas field, overlain by 2007 difference
map I (top panel) and by the modelled stream channel network (top right panel). False colour
Quickbird-2 scene of the Bovanenkovo gas field compound from Forbes and Kumpula (2009)
(bottom panel).
123
Summary and conclusions Chapter 5. Modelling stream channels
in backscatter values over this period. At the end of the first two weeks, mean backscatter
values are about −7 dB. This is due to the change in scattering mechanism from specular re-
flection to diffuse scattering off the moist surface. Over the course of the active season there is a
steadily decreasing trend in backscatter value. The mean backscatter values of the bufferzones
surrounding the flooded areas are of similar magnitude to those areas that the bufferzones sur-
round once the water body has drained. It is interesting to note that at the beginning of July the
mean backscatter values of the bufferzones are relatively low (approx. −10 dB). This may be
due to the fact that a wet snowpack is still present, as the scattering mechanism off wet snow is
similar to that off open surface waters (Bartsch 2010). It has been modelled by Schneider et al.
(1997) that a change in liquid water content from 0 to 0.5% resulted in a change from −6 dB to
−10 dB for the C-band ERS satellite. These values would indeed indicate a wet snowpack on
the banks of the floods. After the snowpack is depleted, a rise in the mean backscatter values
can be seen. This is due to the wetting of the soil; wet soil having a higher backscatter value.
However, thereafter, the temporal trend in the mean backscatter value again decreases steadily.
The decrease in backscatter value is most likely interpreted as a drying of the soil due to evap-
oration over the course of the summer. This is what would be expected according to the claim
by Ulaby and Batlivala (1976) that the backscatter value is proportional to soil moisture.
5.6 Summary and conclusions
This chapter has elaborated on static terrain-based methods to explore whether the seasonal
changes identified by the Envisat ASAR WS sensor are related to hillslope runoff or rather are
a product of spring meltwaters accumulating in lake basins due to overbank flooding. The
study made use of a digital elevation model, the ASTER GDEM2, to derive stream channel
networks. As a first step, the appropriateness of the GDEM2 for this task was evaluated by
undertaking a validation analysis. This was then followed up with the derivation of the stream
channel network of the Se-Yakha and Mordy-Yakha basins and a statistical analysis thereof.
The main results are given below.
• The validation study of the ASTER GDEM2 compared with the reference spot-height data
resulted in an RMSE of 12.17 m (N = 86). The poor linear regression fit (R2 = 0.15, p <
0.0001) lead to the decision to investigate extreme difference values. A threshold of ±6σ
is used to classify these extreme difference values. 38.5% of the data points were classified
as extreme. A t-test on the linear regression fit of the GDEM2 versus e(x) showed that
the slope of this fit was significantly different to zero. An anomalous point was found to
have an unusually high error associated with it. Removal of this error anomaly resulted
in an improved RMSE of 10.5 m (N = 85).
• The validation study of the ASTER GDEM2 compared with the reference D-DEM resulted
in an RMSE of 9.64 m (N = 506877). This is better than the RMSE of the spot-height study,
as is to be expected given the larger sample. However, an improvement of only 0.87 m
124
Chapter 5. Modelling stream channels Summary and conclusions
is established. The standard deviation of the error of this study is similar to the standard
deviation of the GDEM2, 9.63 m and 9.73 m, respectively. Both validation studies show
that the GDEM2 is a fairly good DEM, as long as water body masking is undertaken
beforehand.
• A watershed analysis delineated two major catchments, referred to as the Se-Yakha and
Mordy-Yakha basins. A flow accumulation map and a stream channel map derived from
the former were explored over the two catchment areas. The stream channel network was
classified according to the Strahler system, arriving at 7 stream orders. Drainage density
of the Se-Yakha and Mordy-Yakha basins were 1.986 km−1 and 1.901 km−1, respectively.
The mean overland flow distance is found to be 0.25 km and 0.26 km, respectively. This
can be interpreted as roughly 4 tributaries within a 1 km neighbourhood radius. The
fairly low drainage densities are a result of the gently rolling terrain.
• The hypothesis that the seasonal changes identified by ASAR WS are associated with
hillslope runoff was investigated by determining the stream channel densities for each
flood location. These line densities were calculated within a 10 km radius and were then
investigated as a function of flood size. This study showed a weak correlation (R2 =
0.0167 and R2 = 0.0068) for the Se-Yakha and Mordy-Yakha basins, respectively.
• In the light of the above result, the normalised TWI was calculated for both catchment
areas. The mean TWI at the location of each channel is of the same order of magnitude
as the mean normalised TWI of the entire study area (TWI = 0.38). Again, no clear corre-
lation between hillslope runoff, as established through the TWI, and seasonal inundation
was found.
• The floods were manually inspected with reference to a SPOT image from 1993. Floods
are found to occur in khasyreis some of which are coalesced alas-systems. Most intrigu-
ingly, many floods could be identified in the near vicinity of anthropogenic activities
within the Bovanenkovo gas field complex.
• The temporal changes in the normalised backscatter values of major floods displayed a
clear signature of drainage associated with a sudden increase in backscatter value within
the first two weeks of July. This increase was then followed by a steadily decreasing trend
in backscatter value over the course of the subsequent weeks. The bufferzones around
these floods exhibited the same trend. At the beginning of July, however, they exhibited
lower backscatter values which are most probably interpretable as a wet snowpack and
its ongoing ablation.
125

Chapter 6
Development and application of
Surface Water Dynamics Algorithm
The previous chapters have thus far focused on the Vaskiny Dachi area on the Yamal peninsula.
This chapter, however, is a regional scale study, considering water body change in the Yamalo-
Nenets Autonomous District (or Okrug in Russian, hence the acronym YNAO). With respect to
water body monitoring a new approach was taken, using a land surface water product, that is
provided under the ESA STSE ALANIS methane project, as input data, in order to develop an
automated change detection procedure which may be run over large regions. The algorithm
was published and discussed in the paper Trofaier et al. (2013a), this chapter being an extended
version of that paper.
In recent years there has been an increased interest in studies of the Yamalo-Nenets Au-
tonomous District, their prime concern being the ongoing land-use and land-cover changes
through industrial development. The impacts human activity have in the Vaskiny Dachi area
due to the study area’s proximity to the Bovanenkovo gas field were already touched upon in
the previous chapter. Recent studies have focused on quantifying these impacts. Disturbances
in vegetation have been investigated through standard NDVI techniques (Walker et al. 2009b)
and Kumpula et al. (2010, 2011, 2012) have discussed the impacts of hydrocarbon exploration
in this area as measured by various sensors (ASTER VNIR, Quickbird-2, SPOT, Landsat TM
and ETM+), each of different spatial resolution. They also take an interdisciplinary approach,
using anthropological techniques to analyse these impacts on the indigenous Nenets people.
Further, Bartsch et al. (2010) used scatterometer data to analyse the effects of snow conditions
on reindeer husbandry. Tundra lake studies have been presented by Kravtsova and Tarasenk
(2011) and, on Yamal, Sannikov (2012).
The north of West Siberia is a region of rapid development. It is, however, a region suscep-
tible to environmental change on two accounts. On the one hand, gas exploration is leaving its
mark on the environment and the stability of the frozen ground. On the other hand, a number
of modelling studies have predicted a warming climate which will have a large impact on this
permafrost region (Anisimov and Reneva 2006). Lakes in permafrost regions are used as indi-
127
Case 2: Regional scale Chapter 6. Development of SWDA
cators of a changing climate (Yoshikawa and Hinzman 2003, Smith et al. 2005, Plug et al. 2008),
however, so far short term changes and other direct impacts have been overlooked and remain
unstudied.
The objective of this study is to provide a starting point for investigating seasonal dynamics
of open water bodies in the north of West Siberia, the Yamalo-Nenets Autonomous District. An
algorithm that detects surface water extent and its retreat over time has been developed and is
presented here. It is decided to refer to the process of spring inundation drainage and runoff as
dynamic behaviour, considering the one-directional changes in water extent as dynamics. It is
anticipated that this method may be applied to future studies over consecutive years in order to
appropriately investigate seasonal dynamics. Water body classifications are derived from syn-
thetic aperture radar (SAR), an active microwave sensor. The focus lies on quantifying seasonal
inundation due to spring floods from snow-melt waters and river discharge. High temporal
resolution data are necessary for this task. The trade-off between spatial and temporal resolu-
tion is addressed by using the European Space Agency’s Envisat Advanced Synthetic Aperture
Radar (ASAR). This instrument operates at medium-low spatial resolution in ScanSAR mode
(with a product resolution of 150 m in wide swath (WS) mode) with high temporal frequencies.
Previous studies have investigated the feasibility of using ASAR data for this type of study
and have shown that its coarse spatial resolution is able to capture dynamics in tundra regions
(Bartsch et al. 2012, Trofaier et al. 2012). These data have been used to create a wetlands product
in northern Eurasia within the framework of the ESA STSE ALANIS methane project (Bartsch
et al. (2012), Reschke et al. (2012); in the following referred to as ALANIS dataset).
The present study evaluates the applicability of these ALANIS methane local wetland datasets
by developing a post-processing algorithm that identifies and quantifies spring flood extent of
water bodies and discusses their nature in the context of their geographical setting.
6.1 Case 2: Regional scale - The Yamalo-Nenets Autonomous District
The Yamalo-Nenets Autonomous District (YNAO) is a federal subject of the Russian Federa-
tion. Its borders enclose the West Siberian Lowland, which lies between the Ural mountains
and the Middle-Siberian Tableland. Both the Ob and the Yenisey rivers flow north through this
territory, along its western and eastern margins respectively, into the Arctic Ocean. Continu-
ous permafrost (in the north above the Arctic circle) high ground ice content and the low relief
of this area encourage poor drainage and flooding which in turn promotes peatland develop-
ment. Quantification of the parameters that characterise this low relief is done by exploring
the standard deviation of the DEM (Santoro and Strozzi 2012) for each study area. This DEM is
the D-DEM used in chapter 5 (section 5.3.3). The standard deviation of the D-DEM for the Taz,
Yamal and Gydan study areas are 9.6 m, 19.7 m and 18.4 m respectively, while the respective
inter-quartile ranges of the D-DEM are 12 m, 30 m and 31 m.
The West Siberian Lowland is an Arctic wetland environment. Indeed, it is thought to be the
world’s largest Arctic wetland (Kremenetski et al. 2003). The West Siberian Lowland embodies
128
Chapter 6. Development of SWDA Case 2: Regional scale
Figure 6.1: Permafrost extent in Eurasia. Distribution of continuous and discontinuous per-
mafrost according to Brown et al. (1997). The black frame indicates the extent of figure 6.2. The
borders of Russia’s administrative regions are outlined in grey (Stolbovoi and McCallum 2002);
the Yamalo-Nenets Autonomous District is highlighted by a line-fill. The magenta frames rep-
resent the ALANIS methane local wetlands product coverage across Eurasia with their associ-
ated identification numbers.
Figure 6.2: Map of the Yamalo-Nenets Autonomous District of the Russian Federation (NW
Siberia) showing the location of the three study areas: Subregion 1 (Taz), subregion 11a (Yamal),
subregion 11b (Gydan). Permafrost extent according to Brown et al. (1997). The borders of
Russia’s administrative regions are outlined in grey (Stolbovoi and McCallum 2002).
129
Case 2: Regional scale Chapter 6. Development of SWDA
roughly 16% of the territory of the Russian Federation and over 50% of this lowland encloses
a wetland ecosystem underlain by permafrost (Solomeshch 2005). The extent of permafrost in
the district is depicted in figure 6.2. The developed algorithm was run for data covering the
entire YNAO. The resultant classification exceeded 0.8 gigabyte. A data volume issue arose
when attempting to vectorise this raster classification, as this would have resulted in over 2
million polygons. It was therefore decided to manually inspect the classification and analyse
subregions of the YNAO. The discussed subregions correspond to the areas with the highest
magnitude of dynamics. They are labeled by their respective ALANIS dataset region identifi-
cation number - when multiple subregions exist within one region then the identification will
be of an alphanumeric type (figure 6.1).
6.1.1 Subregion 1 - Taz river basin near Mangazeya (Ìàíãàçeß)
The Taz river basin has a surface extent of about 150 000 km2 (Gordeev 2006). The river flows
north into the Taz river estuary (Tazovskaya Guba) approximately 250 km long running into
the Ob estuary (Obskaya Guba). The study area lies in the Taz river floodplain south of the
estuary, a few minutes in latitude north of the Arctic circle. The historic trade city of Mangazeya
is located to the east of its borders. The size of this study area is 2345.5 km2 which is an order
of magnitude smaller than the other two study areas.
The deposits on the Taz peninsula are of marine-alluvial origin. The upper layers are allu-
vial sands and peat layers (Mahaney et al. 1995, Astakhov and Nazarov 2010) This area lies in
the discontinuous permafrost zone.
6.1.2 Subregion 11a - Yamal peninsula
The study area on the Yamal peninsula is the largest of the three chosen subregions of the
YNAO. It covers an area of approximately 60929 km2, which is roughly 36% of the 700 km long
and 240 km wide peninsula. This area is subject to ongoing anthropogenic stressors due to
past and current gas exploration. The development of the gas fields, the largest of which is the
Bovanenkovo gas field in the northwest of the study area, are leading to land use and land cover
changes on the entire peninsula (Kumpula et al. 2012). Analyses of aerial photographs in the
surroundings of the gas field have shown small scale long term expansion and decrease of thaw
lakes (Sannikov 2012). Previous investigations of ASAR WS have demonstrated seasonally
reoccurring variations (Bartsch et al. 2012, Trofaier et al. 2012). Indeed, it was shown in chapter
4 (section 4.2.3 and further noted in chapter 5 (section 5.5) that extensive open surface water,
located at approximately 70°N 68°E, drains each year after the first two weeks of July in 2007,
2008 and 2009.
The surficial deposits are slope aeolian, alluvial, lacustrine and marine sandy to clayey
deposits washed out within the active layer and of high salinity within the permafrost. Some of
these deposits are also saline in the active layer (Walker et al. 2009b) resulting from active slope
processes removing active-layer deposits and exposing permafrost (Leibman and Streletskaya
130
Chapter 6. Development of SWDA Data
1997).
The Circumpolar Active Layer Monitoring (CALM) site at research station Vaskiny Dachi
(70° 17’N 68° 54’E) is situated in the proximity of the Bovanenkovo gas field, within the wa-
tershed of the Se-Yakha (C¼ßõa) and Mordy-Yakha (Ìîðäûßõa) rivers. This site is on the IIIrd
alluvial-marine terrace of the Yamal peninsula which is at a maximum height of 36 m asl. Ac-
tive layer depth and temperature monitoring is ongoing and records date back to 1993. In 2007,
the maximum active layer depth at this site (i.e. summer thaw depth) was 135 cm (Leibman
et al. 2012).
6.1.3 Subregion 11b - Gydan peninsula
The Gydan peninsula is, in common with the Yamal peninsula, a hydrocarbon rich environ-
ment. The study area is approximately 11274 km2 large, encompassing the untapped Salmanovskoye
gas field. In the immediate south of this region lies the CALM site Parisento (70° 07’N 75° 35’E).
Here, we find sands and loamy sands. In some places these are enriched by peat (Pavlov 1998).
This site is less than 30 km south of the study area, lying on the same lacustrine-alluvial terraces
and within the floodplain of the river Yuribei (Þðèáåé) - note, same name but different river
to that on the Yamal peninsula over which the Yuribei Bridge was built by the gas industry,
renowned for being the Arctic’s longest bridge to date.
Active Layer depth measurement via CALM are scarce, and last measurements are from
1995 when maximum thaw depth was 185 cm.
6.2 Data
6.2.1 ALANIS Methane local wetlands product
In addition to the original ASAR WS data, this study takes advantage of the local wetlands
product of the ESA STSE ALANIS (Atmosphere LANd Interaction Study) Methane project
(http://www.alanis-methane.info). This wetlands product is a suite of data on water body
extent. The classification of water body extent is derived from the aforementioned ASAR WS
data. It is a ten-day composite over the time period April to September 2007 provided in the
Universal Polar Stereographic North (WGS84) coordinate system. As discussed in chapter 3
(section 3.2.1), specular reflection of the incident radar beam results in a clear separation be-
tween open water body and surroundings. This property is used to produce the ALANIS
10-day composite. The same density-slicing technique that was used to create the fortnightly
water body classifications in chapter 4 (section 4.2) is used to arrive at a 10-day composite clas-
sification of water body extent. The same threshold value of -14 dB that was used to derive the
fortnightly temporal composites is used for the ALANIS methane product, as discussed earlier
this threshold provides good results for separating water bodies from surroundings according
to the bimodal distribution of backscatter values.
131
Data Chapter 6. Development of SWDA
Additional metafiles associated with each composite are also provided under the ALANIS
dataset. Information regarding the number of acquisitions, the number of days that have
passed since the last update as well as the seasonal statistics (Reschke et al. 2012) may be used
as a quality indicator. The former data on sensor acquisition numbers play an important role in
the control of the quality of water body classification (Trofaier et al. 2012). The ALANIS dataset
covers extensive areas of Northern Eurasia which are illustrated in figure 6.1.
The ALANIS dataset covers the entire Yamalo-Nenets Autonomous District. As mentioned
in chapter 3 (section 3.1), the source of the ALANIS dataset, Envisat ASAR, is a C-band (centre
wavelength λ ∼ 5.6 cm) instrument. The ScanSAR resolution in ASAR’s Wide Swath (WS)
mode is 120 m for all sub-swaths (Closa et al. 2003). An undersampled pixel spacing of 75 m is
used.
ASAR WS data are acquired on request; this limits the archived material. However, more
than 10 acquisitions per month are available over the study area in 2007 (Bartsch et al. 2012).
Data were taken in single polarisation HH (horizontally emitted and horizontally received).
The local incidence angle of the backscatter coefficient (σ0) has been normalised to a reference
angle of 30° (Sabel et al. 2011). The normalisation technique, which was noted in chapter 3 (sec-
tion 3.1.2) has been chosen to reduce the angular dependency of the radar backscattering coef-
ficient on the local incidence angle. This step allows a comparable analysis of measurements
taken at different incidence angles and has been successfully applied to SAR studies such as
Bartsch et al. (2007a, 2008) and Reschke et al. (2012), who have investigated wetland dynamics
and extensively cross-validated their datasets, including the ALANIS methane regional wet-
land product. Sabel et al. (2012) discuss the pre-processing in more detail. The digital elevation
model (DEM) used for geometric correction of the SAR data is taken from digitised Russian
Topographic Maps at a 1:200 000 scale (Santoro and Strozzi 2012).
Only datasets from the snow- and freeze-free period have been selected. Snowmelt usually
finishes in mid-June over the Yamal peninsula (Bartsch 2010).
6.2.2 QuikSCAT snowmelt product
The Ku-band scatterometer Seawinds on the QuikSCAT satellite is an active microwave sen-
sor (centre wavelength λ ∼ 2.1 cm). Its data have successfully been applied to snow studies
(Bartsch et al. 2007b, Tedesco and Miller 2007, Wang et al. 2008); the dielectric properties of
water making microwave sensors suited for monitoring the hydro and cryosphere. Snow dis-
plays interesting backscatter characteristics. Its backscatter values are associated with three
different mechanisms: (i) reflection at the surface of the snow, (ii) volume scattering within the
snowpack and (iii) reflection at the transition boundary from snowpack to soil or ice (Scherer
et al. 2005). Reflection off the soil-snow interface will result in further volume scattering. The
backscatter properties of snow are hence dominated by both surface and volume scattering
(Ulaby et al. 1981). Backscatter values are low before snow arrival but increase with snow ac-
cumulation. Dry snow has a low dielectric constant ranging between 1.2 and 2. This of course
132
Chapter 6. Development of SWDA Data
is dependent on snow density. For the aforementioned values snow density is taken as 0.1 to 0.5
g cm-3 (Scherer et al. 2005). Rising air temperatures lead to the formation of liquid water in the
snow pack. Therefore, a sudden decrease in backscatter is measured at the onset of snowmelt
(Bartsch 2010). In poorly drained areas ponded melt waters are associated with low backscatter
due to a change from volume to surface scattering. However, in general the snowoff date may
be related to an increase in backscatter as the permittivity and reflectivity of thawing ground
increases as well (Kidd et al. 2004). These dielectric properties of snow phenology are used for
generating snow-melt products from active microwave data. The start, duration and end of
snow-melt may be established. The QuikSCAT snow-melt product used in the current study is
based on the method developed by Bartsch et al. (2007b). This method calculates diurnal differ-
ences in backscatter that arise due to freeze-thaw effects associated with a lossy, wet snowpack
in the evenings and a dry, refrozen snowpack in the mornings. These diurnal variations of
backscatter may account for a difference of up to 6 dB (Kidd et al. 2004). Diurnal difference
indicators are established, the period of the most prominent indicator relating to the onset and
duration of snow melt. This method allows the determination of a snow-off date (Bartsch et al.
2007b). The value of the QuikSCAT snowmelt product used for this study gives us the last day
of snow cover, the last snow-on date.
The left panel of figure 6.3 shows a boxplot of the QuikSCAT data for each study area. The
median QuikSCAT product value over each region is given by the thick black line. The mean
value was calculated separately. For subregion 1 the median and mean date of the last snow-on
day are 152 and 153 respectively. Both median and mean are the same for subregions 11a and
11b, 165 and 166 respectively. According to these auxiliary data the snow-off date in subregion
1 happened after 3 June 2007. This is the most southerly of the three study areas. For the other
two study areas snow cover does not disappear for almost another two weeks, after 15 and 16
June.
This may also be seen in a MODIS image from 29 June 2007 (see right panel of figure 6.3),
which clearly shows lake ice covers on the Yamal and Gydan peninsulas. As mentioned before,
the threshold classification method of the radar data cannot unambiguously be applied to ice
covered lakes and hence, despite the QuikSCAT data indicating the last snow-on day mid-June.
The MODIS image affirms the choice of investigation time periods for subregions 11a and 11b
at a later date since lake ice covers are still present at the end of June.
133
Data Chapter 6. Development of SWDA
Figure 6.3: Boxplot of QuikSCAT data for each study area. The median is indicated by the
thick black line (left panel). MODIS image (courtesy of NASA Earth Observatory) of the Yamal
peninsula on 29 June 2007. Frozen lakes on both the Yamal and Gydan peninsulas are clearly
visible (right panel).
134
Chapter 6. Development of SWDA Development of the Surface Water Dynamics Algorithm
Figure 6.4: Lake in a partially drained basin on the Yamal peninsula (left panel) and lake with
arctophila emerging from the water surface and growing at its margins (right panel). Photos
taken by A.M. Trofaier on 1 and 3 July 2012, respectively.
6.3 Development of the Surface Water Dynamics Algorithm
The problems for mapping inundation with C-band SAR are related to the effects of waves on
the water surface and emerging vegetation, as discussed in Bartsch et al. (2012, 2009). Emerging
vegetation is common for Arctic and particularly sub-Arctic water bodies. The aquatic and
waterborne vegetation in wetlands and those at the margins of lakes will affect the classification
method by establishing a point of difference from a smooth water surface. Figure 6.4 shows a
photograph of an Arctic lake on the Yamal peninsula with reeds, such as arctophila grasses,
growing in the water and emerging from the lake surface. It is important that change analysis
also considers the effects of waterborne and marginal vegetation as the classification method
can only be used for identifying the drainage and runoff dynamics of open surface waters.
The ALANIS dataset is a thematic map consisting of five classes. Mapping of water bodies
is done using the procedure first put forward by Bartsch et al. (2007a) and which has subse-
quently been used for this type of investigation (Bartsch et al. 2012, Trofaier et al. 2012). An
algorithm that uses the ALANIS dataset for identifying surface water dynamics was estab-
lished. Henceforth it shall be referred to as the Surface Water Dynamics Algorithm (SWDA).
As a first step the SWDA transforms the product into a binary classification; water bodies are
set to the class value 1, the surroundings to 0. This enables a straightforward calculation for
determining any changes in surface water extent. In a further step, the SWDA sums each bi-
nary classification for each study area over a given time period (see equation 6.1), resulting in
a single classification of changes in surface water extent Xij .
Xij =
n∑
k=1
(xij)k (6.1)
135
Results Chapter 6. Development of SWDA
where (xij)k is the kth classification for the region within the given time period, i and j are
respective raster rows and columns and n is six for a two-month period - there are three 10-day
ALANIS Methane water body classifications per month. The current study’s time period is
chosen as July-August 2007. Previous research has shown that water bodies are still subject to
a late-spring ice cover at the end of June (Trofaier et al. 2012). It was also chosen to limit the
analysis time period to two months in order to facilitate computational efficiency.
The SWDA results in a classification of changes in surface water extent: High class val-
ues represent permanent inundation, low class values identify change. Therefore we find that
changes in inundation occur within the range 0 < Xij < n. This states that grid cells equal
to n are determined to be permanent inundation and those equal to zero are the surroundings
with no surface inundation. However, the exact value for seasonal inundation needs to be
determined by inspection of both the SWDA classification and the radar data, since applying
the SWDA as a stand-alone, fully automated procedure, would lead to misclassification due
to wind biases. Trofaier et al. (2012) have discussed this issue in more detail and note that
availability of ample data is essential for this method. This was covered in chapter 4, where
14-day temporal composites were used to examine changes in water body extent, noting that
the maximum number of images used to create a composite is one per day. This, of course, is an
idealised scenario which is not always achievable due to strong wind conditions. On average,
Trofaier et al. (2012) find that 10 image scenes for every 14 days are available. Therefore, it is
found that over a classification time period of a number of t days, the number of scenes needed
to create a temporal composite is about N > 23 t. To exclude the possibility of misclassification
of changes due to weather-affected imagery and lack of data, the number of usable images
available for the given region was inspected. The number of radar scenes used for classifying
each pixel has been determined from metafiles of the ALANIS dataset; the post-processed wa-
ter body product is analysed in combination with the radar images and auxiliary data on the
number of radar images that were available for each pixel in the given time period.
The time period of investigation was cautiously chosen to cover the mid Arctic summer pe-
riod, in order to avoid misclassification due to lake ice cover. A QuikSCAT snow-melt product
(Bartsch et al. 2007b, Bartsch 2010) was used to support this choice.
6.4 Results
The YNAO of the Russian Federation was investigated for changes in surface inundation over
the summer period July-August 2007. Surface hydrological activity, in the form of spring runoff
and drainage after flooding, is quantified by applying the SWDA.
An accuracy test of lake change detection was done for the Yamal pensinsula (subregion
11a) by choosing 45 random lake change polygons and inspecting a timeseries of five original
Envisat ASAR WS images. 36 of these polygons corresponded to lake change; the remainder
did not. The user’s accuracy of the SWDA polygons against a manual interpretation of lake
change, done by inspection of a radar timeseries, is determined to be 86%. In certain cases wind
136
Chapter 6. Development of SWDA Results
Table 6.1: Main parameters of study areas and their lakes in 2007
Parameter 1 11 a 11 b
Area [km2] 2345.5 60929.4 11273.7
Water body extent † [km2] 268.6 7550.6 972.2
Change in water body extent‡ [km2] 42 302 61
Limnicity † [%] 11.45 12.39 8.62
Limnicity change [%] 1.79 0.49 0.54
†at the beginning of July
‡above 0.2 km2 during July-August
action played a role. The error of commission of this simple accuracy test is 27%. Unfortunately,
no ground validation data are available for the present study. Previous studies on monitoring
open water bodies have assessed the accuracy of the water body classification method. Bartsch
et al. (2008) found the Kappa coefficient to be 0.82 with an overall classification accuracy of
95%. An optimum threshold sensitivity analysis, examining the mean water body perimeter
and area as a function of backscatter threshold, was undertaken in chapter 3 (section 3.4.1) and
presented in Trofaier et al. (2012). Reschke et al. (2012) assessed Envisat ASAR WS’s capabilities
for mapping inundation through validation with multiple datasets.
It is found that many changes registered by the SWDA occur around the margins of lakes.
These are not attributed to seasonal changes in hydrology but are associated with the higher
backscatter response caused by vegetation that emerges over the growing season along the
shorelines and in the shallows of the lakes. An extremely high percentage of these changes are
smaller than 0.2 km2 (see figure 6.5). These changes are mostly related to the lake shorelines.
Given that the spatial resolution of the sensor is only a magnitude smaller than 0.2 km2, with
the MMU of 4 pixels, as established in chapter 3 (section 3.5), being 0.0225 km2, this study
focuses on the statistical outliers. Therefore, a change of a minimum of 36 pixels (pixel size
74.5 m × 74.5 m), corresponding to a ninefold MMU, is deemed acceptable for the successful
detection of seasonal inundation. This relates to all changes in water extent that are greater
than 0.2 km2. Any changes below this value are neglected from further analysis. The detected
changes in water extent are given in table 6.1. The resultant SWDA changes for subregion 1 are
shown in figure 6.6b. This study area is 2345.5 km2 large and was covered by 268.6 km2 of water
at the beginning of July 2007. We can also see from figure 6.6b that the main changes occur
around the grid locations B2, C2-3 and D4. The calculated water body extent further shows
that subregion 1 exhibits the highest change in limnicity (1.79%) over the two summer months
July-August, despite being the smallest of the three study areas. This is not surprising, as most
of the study area lies within the Taz river floodplain. Lake change and seasonal inundation are
137
Discussion Chapter 6. Development of SWDA
Figure 6.5: Histogram of changes in water area in subregion 1. Frequency given on a base 10
logarithmic scale. A strip chart of each change is given just above the x-axis.
primarily related to spring floods from the Taz river itself.
6.5 Discussion
6.5.1 SWDA quality control
Data availability is crucial for accurate analysis. Envisat ASAR only operated in wide swath
mode on request. Therefore, archived ASAR WS material is limited. Fortunately, data avail-
ability is proportionately high for the area of interest (Bartsch et al. 2012). The number of
archived scenes that were used to create each 10-day composite is investigated. The informa-
tion for each pixel of each classification is stored in so-called num-files. These fully geocoded
files are provided under the ALANIS methane project and may be used to statistically evaluate
the accuracy of water body classification. Especially important are the minimum number of
scenes used for each classification, as these may give an indication of whether enough data
were available to produce the water body classification and whether this may indeed have led
to misinterpretation.
In subregion 11a, no data were available for 0.45% of pixels in the first 10-day time period
(1 - 10 June). This represents a total of 57050 pixels of which only 31115 pixels were classed
as water bodies. The SWDA did not highlight any changes in this area. Inspection of the
respective 10-day composite revealed that these pixels were on the northeast coast of the Yamal
peninsula, the major part being in the sea. A total lake area of 9.2 km2 was not represented in
this first 10-day time period. Excluding these pixels, a minimum of 4 images were used for
this classification. Furthermore, in the last time period (21 August - 31 August) only 0.23% of
the pixels were classified using one ASAR WS scene. The mean number of images used for
the last time period are 6, the maximum being 10. Figure 6.7 shows a graph of the minimum,
maximum and mean number of ASAR WS acquisitions used for each 10-day composite in each
138
Chapter 6. Development of SWDA Discussion
Figure 6.6: (a) Comparison of classifications of water body extent for subregion 1 for June and
July 2007. (b) SWDA classification for July-August 2007. (c) Radar timeseries of disappearing
open water body June-July 2007.
study area.
Fewer acquisitions are available for subregion 1. On average 6 images were used to classify
water bodies over 2 months in this region; the minimum number of scenes used for SWDA
classes 1,3 and 4 being 2. This should really be the lower limit for accurately classifying water
bodies. Minimum, maximum and mean number of images used to classify water bodies in
each study area for SWDA class 1 are given in table 6.2.
In general, the number of ASAR WS scenes available for water body classification in the
Yamalo-Nenets Autonomous District is sufficiently high for mapping inundation and may
therefore be used for analysis of seasonal water body dynamics.
There are six ALANIS methane water body classifications in 2 months. Hence, for two
Table 6.2: Number of Envisat ASAR WS acquisitions used to classify water bodies over the first
time period 1 - 10 July 2007 (SWDA class 1).
Subregion
Number of acquisitions
min max mean
1 2 3 2.8
11a 4 11 7.2
11b 5 9 7.3
139
Discussion Chapter 6. Development of SWDA
Figure 6.7: Minimum (top), maximum (middle) and mean (bottom) number of Envisat ASAR
WS acquisitions used to classify water bodies for each 10-day composite over the summer pe-
riod July - August for each study area.
140
Chapter 6. Development of SWDA Discussion
Figure 6.8: Surface hydrology classification of study area 11a. Seasonal changes in water ex-
tent for July-August 2007 are shown in red. The Obskaya-Bovanenkovo Railway was partially
traced from GoogleEarth (gaps were left where the spatial resolution did not allow the identi-
fication of the railway tracks). In the north, the southern part of the Sannikov (2012) region is
outlined.
months there are six SWDA classes associated with inundation. However, only the lower
SWDA classes are associated with surface hydrological changes. The threshold between per-
manent and seasonal inundation needs to be determined manually by inspection. Seasonal
inundation is determined to be open surface water that is classified in up to two 10-day clas-
sifications inclusively within 2 months. Anything above these classes is thought to represent
permanent inundation. This assumption is verified by inspection of a timeseries of the original
radar images and holds true for each region.
6.5.2 Assessment of floods detected by SWDA
The Envisat ASAR WS data and associated ALANIS methane local wetlands product clearly
identify inundation changes within the YNAO in the summer of 2007. The three study ar-
eas vary greatly in size, however in spite of this, limnicity does not (see table 6.1). The two
northerly study areas lie within the continuous permafrost zone; similar sedimentary deposits
(marine-alluvial sands and clay, as well as peat) are found in these ground ice rich regions of
the West Siberian North. According to Brown et al. (1997), subregion 1 lies in the discontinuous
permafrost zone at the borders to the continuous zone, its sedimentary deposits are, however,
not dissimilar to the former two subregions.
141
Discussion Chapter 6. Development of SWDA
The changes in surface inundation in subregion 11b mostly occur along small branches of
the Yuribei river and are also to be associated with its floodplain and the respective river spring
floods. One particular lake, located a few kilometers northwest of the river, shows a change
of approximately 13.4 km2 in surface water extent. Closer inspection of optical data clearly
shows the lake’s location within a flood basin through which a stream flows that connects to
the Yuribei river.
The situation in subregion 11a is of a different nature. Myriad meandering rivers and
streams can indeed be found throughout the entire Yamal peninsula. However, Vaskiny Dachi,
in the northwest of the study area, also lies just a few kilometers south of the Bovanenkovo gas
field. Closer inspection of the seasonal changes in inundation seem to show a trend towards
areas where anthropogenic land-use change is underway. Kumpula et al. (2012) have discussed
changes in land-cover caused by the gas industry. Although, these impacts are detectable with
medium resolution satellite imagery, finer resolution data from the commercial Quickbird-2
and GeoEye satellites provide the best evidence of anthropogenic change. Their study showed
the expansion of a visibly affected area over nearly 30 years of gas field development, and
discussed the negative affects on the freshwater system and drainage networks. Moreover, Za-
kharova et al. (2011) have discussed how human activities have been identified as resulting
in changes to the primary hydrological system in the Poluy-Nadym-Pur-Taz (PNPT) region,
south of Yamal, to the west of subregion 1. Thermoerosion is the main process by which sur-
face drains are created, being related to a disturbance in ground covers, exposing easily washed
out sandy deposits. The melting of ground ice on flat surfaces leads to ground subsidence and
thermokarst depressions (French 2007). Processes of thermoerosion and thermokarst are pro-
moted by human activity and infrastructure; from dirt tracks of all-terrain vehicles via gas
pipelines to railway tracks and container settlements, they all contribute to land-cover changes
which may result in the development of new primary drainage networks (Zakharova et al.
2011). The present study has identified summer changes in inundation. On Yamal, these occur
most dominantly in the west of the peninsula, following the route of the Obskaya-Bovanenkovo
railway (see figure 6.8). Further analysis is beyond the scope of the present study but it demon-
strates the potential of the SWDA to studies of inter-annual changes in open water bodies.
6.5.3 The time-slice conundrum
Previous studies have discussed decadal lake change in the region of the Yamalo-Nenets Au-
tonomous District (Smith et al. 2005, Karlsson et al. 2012). In connection with these studies it is
important to point out that seasonal lake change also needs to be considered. In order to inves-
tigate lake change using high-medium resolution satellite imagery, the date for scene compari-
son needs to be around the same time. If the temporal frames are not similar, the changes that
are identified may be due to seasonal dynamics. Karlsson et al. (2012) compare thermokarst
lake change in the years 1973, 1987-88 and 2007-09 in the Nadym and Pur catchments. Many
images in 1973 were taken in June, whereas the images of later years were primarily taken in
142
Chapter 6. Development of SWDA Summary and conclusions
July-September. They supplement their analysis with hydrological data and compare their re-
sults to a study by Smith et al. (2005). They find that their results differ from the latter study.
One of the reasons for these findings, which they put forward in their discussion, is the selec-
tion of time slices. This underlines the point that inter-annual comparisons of remotely sensed
imagery may need to pay greater heed to temporal frame selections. High temporal resolution
imagery may be a useful tool for such an investigation. Figure 6.6c shows a radar timeseries
of an open water body that disappears over a couple of days in the Taz region in June 2007.
Incidentally, subregion 1 overlaps with an area where a disappearing lake was found in the
Smith et al. (2005) study (see figure 6.9), which was a survey of lake change, creating a lake
inventory using Landsat-1 MSS data (spatial resolution 80 m) as well as multispectral data
from the RESURS-1 satellite (spatial resolution 150 m). They state in their supplementary data
that they classify lakes larger than 40 ha to be stable since, to their knowledge and findings,
these did not exhibit any seasonal variations. Floodplain areas were excluded from this anal-
ysis to avoid seasonal affects. The SWDA has detected that most spring floods happen in the
floodplain areas, as is to be expected, which would be in agreement with the premises put for-
ward by Smith et al. (2005). However, after manual inspection of the original data, we also find
that certain open water bodies, which do not lie within floodplains, also disappear within the
2 monthly period. The lake shown in figure 6.6c is one of those water bodies lying outside a
floodplain. Incidentally it is located within the same area that Smith et al. (2005) state to have
found a disappearing lake. This lake is about 1.8 km2 (180 ha) in extent. On Yamal and Gy-
dan we also find water bodies that are larger than the 40 ha criterion and disappear over the
two summer months. Future analysis of consecutive years is needed to investigate whether
these are indeed of an ephemeral nature. However, given the results of this study, it does seem
categorisation of water bodies into stable/unstable landscape features according to size may
not be that straightforward. It also underpins the need for investigating seasonal dynamics
using high temporal resolution data. It is proposed that future work should apply the SWDA
to analyse possible seasonality of these lakes by investigating additional years.
6.6 Summary and conclusions
This study tests the application of a change detection algorithm for analysing spring surface
water dynamics in the Yamalo-Nenets Autonomous District in the north of West Siberia. En-
visat ASAR WS active microwave data and water body classifications provided under the ESA
STSE ALANIS methane project were used for this analysis. These high-temporal resolution
radar data and classifications are ideally suited to illustrating seasonal open water body dy-
namics. The results indicate changes due to melt waters from a thawing landscape and spring
floods from river systems. The limnicities of the study areas are of similar magnitude (low-
est for Gydan), however, the highest change in limnicity is recorded for the Taz study area.
The seasonal inundation in the Taz area is almost certainly associated with the Taz river flood-
plain and increased river discharge after snow-melt. Similar characteristics, albeit to a lesser
143
Summary and conclusions Chapter 6. Development of SWDA
Figure 6.9: Map of the YNAO (black frame) overlain by the region investigated by Smith et al.
(2005). Red squares indicate a disappearing lake, one of which is within subregion 1 in the
discontinuous zone. According to the figure presented in Smith et al. (2005) the location of
this disappearing lake seems to coincide with the seasonal lake shown in figure 6.6c. Magenta
frames are the ALANIS methane subregions. Permafrost extent uses the same colour scheme
as in figure 6.1.
144
Chapter 6. Development of SWDA Summary and conclusions
extent, are found on Gydan where the Yuribei river plays an important role as a source of
spring floods. Whereas these summer changes in water body extent appear to be natural phe-
nomena, major anthropogenic stressors may contribute to these changes on Yamal. Changes
to the drainage network through human-induced thermokarst processes have been discussed
by e.g. Zakharova et al. (2011). Further examination of these open water body dynamics on an
inter-annual basis is needed, providing more comprehensive results on spring inundation and
whether these flooded areas are of seasonal nature. The information will be of value to per-
mafrost researchers and modellers in the light of global warming trends and Arctic wetlands’
increased carbon contribution to the atmosphere.
The main outcomes therefore are as follows:
• The application of the SWDA may be used to explore the extent and the retreat of floods,
with attention to the number of ASAR WS images that were used to create the water body
classifications. This change detection technique is one-directional and can only account
for flood retreat, which is the main concern of this type of seasonal water body study.
Only changes greater than 0.2 km2 are thought to be associated with ’true’ water body
retreat. The high percentage of pixels of water body change below 0.2 km2 (98.77%) hap-
pens along lake shorelines and is therefore considered to be due to vegetation growth at
the margins over the active season, that will impact on the classification procedure. These
water body pixels are therefore excluded from the analysis.
• The QuikSCAT snowmelt product is used to decide on the period of investigation to
avoid snow and ice cover on land and water, respectively. The product gives the last
day of snow cover as established from the most prominent QuikSCAT diurnal difference
indicators. Snow-off dates are found to be 3, 15 and 16 June 2007 for subregions 1, 11a
and 11b, respectively. This underlines the choice of the study time period to begin in July
2007.
• Limnicity changes associated with flood retreat are greatest in subregion 1 (1.79%), the
smallest of the study regions. Subregion 11a and 11b exhibit changes in limnicity of 0.49%
and 0.54%. The greatest change in limnicity in subregion 1 is identified to occur due to
floodplain dynamics.
• On the Yamal peninsula (subregion 11a) the impacts of human activities may not be un-
derestimated. Spring floods are found to happen in areas under anthropogenic influence,
particularly in the Vaskiny Dachi area near the Bovanenkovo gas field but also along
the Obskaya-Bovanenkovo Railway. Previous studies have mentioned how alterations
to the drainage system have arisen due to human activities. These modifications may
be resulting in greater flood risks, and therefore a concentration of spring floods in the
Bovanenkovo region.
• Time-slice choice is an important issue in remote sensing of Arctic water body change.
145
Summary and conclusions Chapter 6. Development of SWDA
The present study has shown that large water bodies changes may happen within a mat-
ter of weeks. To avoid misinterpretation of decade-scale comparison studies, floodplains
should be excluded from such a study. However, it is noted here, that this may not be
enough. In addition, meteorological and hydrological discharge data should be evalu-
ated to examine the possible magnitude of spring flood events. Furthermore, frequent
radar data can identify areas of high flood risk. The combination of these data would
allow a holistic approach to seasonal water body change analysis, arriving at a better
understanding of Arctic water body change in general.
146
Chapter 7
Conclusions
The aim of this PhD research was to develop an operational EO monitoring procedure using
frequent active microwave data to map spring/summer changes in water body extent in a sen-
sitive permafrost environment in the North of West Siberia. The motivation for this research
lay primarily in the idea of exploring a new way of frequently mapping spring surface wa-
ter dynamics. The current lack of spatial data on seasonal inundation and its retreat over
the spring/summer months needed to be addressed. Using such information to complement
longterm lake change studies in the Arctic permafrost regions would greatly benefit such an
analysis, as it is necessary to establish the point in time at which the landscape may be seen
as stable; when alas systems are no longer filled with spring meltwaters and the water body
signatures measured at the satellite sensor are representative of true lake extent, rather than
including flood waters. In the past meteorological and hydrological data were used to comple-
ment such an analysis (e.g. Karlsson et al. (2012)), however, given that the data for geospatial
mapping of the transient nature of spring floods are archived and freely available for scien-
tific research, it seems curious that such an attempt has not been undertaken in the past. It is
believed that the reasons for this deficiency lie with the attitude that coarse-scale spatial reso-
lution data, such as ASAR WS data, are not representative of the tundra landscape - failing to
account for the abundant thaw ponds that lie below the sensor’s spatial resolution. However, it
is stressed here that process-dependent choice in resolution (both spatial and temporal) is more
important than insisting on acquiring the highest possible spatial resolution data available. In
order to examine the extent of spring floods, high temporal resolution data are required - these
come at the cost of spatial resolution. This PhD research highlights the applicability of using
medium/coarse spatial resolution data for monitoring spring flood extent. Incorporating these
periodic spatial data on spring flood extent and retreat in an analysis of longterm thaw lake
change in the Arctic would provide the best possible results for accurately accounting thaw
lake dynamics.
147
Research synthesis and achieved objectives Chapter 7. Conclusions
7.1 Research synthesis and achieved objectives
Here follows a summary of the key results and the achieved aims (bold bulletpoints) of this
PhD research.
i) To establish an optimum C-band SAR mapping procedure of Arctic open surface water
bodies.
In order to take full advantage of an EO water body monitoring procedure, an optimum clas-
sification method needed to be established. Here, optimum refers to the least complex and
hence most efficient classification procedure that enables exemplary results for further change
detection analysis. Therefore, an intermediate goal was to establish uncertainties in mapping
water body extent using a simple threshold classification method. The results of a threshold
sensitivity test and two map accuracy studies were discussed in chapter 3. It was found that
the normalised bimodal backscatter distribution is not only image dependent but the optimum
threshold value varied from lake to lake within each image. Therefore, each image in its en-
tirety, as well as each lake in each image, has a respective optimum threshold value which may
differ from the suggested - 14 dB. This complicated the matter and a decision on threshold
choice would maintain a certain degree of subjectivity. To support the choice, statistical map
accuracy tests were done. Both a sensitivity and two map accuracy tests were therefore pre-
sented in chapter 3. The sensitivity test showed that a change in the threshold of 1 dB resulted
in a change in mean lake diameter of 22 m. This is almost 7 times lower than ASAR WS’s
spatial resolution of 150 m. Furthermore, map accuracy tests using two types of reference data
(panchromatic ALOS PRISM and radar TerraSAR-X data) observed good agreement (overall
map accuracies were 89% and 86%) for the -14 dB ASAR WS classification. After careful con-
sideration (exploring the results of both the threshold sensitivity analysis and the two map
accuracy studies), the choice in C-band threshold value (-14 dB) was overall deemed sufficient
to proceed with developing the monitoring procedure.
ii) To automate the classification procedure, establishing temporal composite water body
classifications.
The aforementioned and tested water body classification procedure was done on an image by
image basis. However, in order to analyse over 60 images in a two-monthly period (due to the
sensors capability of imaging at different incidence angles, the true re-visit period of the satel-
lite is complemented by these additional images which are then normalised to be comparable),
an automated classification procedure needed to be developed. Moreover, to establish the sea-
sonal dynamics in water body extent, an efficient method for handling the data needed to be
conceived before change detection analysis could be carried out. It was deemed appropriate
to create temporal composites, showing water body extent at its greatest magnitude within a
specified time period. This time period was chosen to be 14 days. 4 classifications of water
body extent were produced automatically for each year (2007, 2008 and 2009).
148
Chapter 7. Conclusions Research synthesis and achieved objectives
iii) To investigate change detection analyses that may be used for frequent mapping of
seasonal inundation.
These temporal composites were then used as input for change detection analysis (see chapter
4). In chapter 6 a second approach for change detection is presented using a water body prod-
uct that shows water body extent at its greatest magnitude within 10 days. The two change
detection techniques are listed below.
- Change detection was undertaken manually through difference mapping in a GIS
(chapter 4). Three difference maps were created. Flood retreat was quantified for
each of the difference periods in each of the three years (2007, 2008 and 2009) in
the form of water pixel loss. The greatest loss in water body extent occurred in the
difference map I of each year. In 2007 and 2008 this loss was 25% of the total water
body extent and in 2009 the loss was 28% although the increase is thought to be
related to wind-biases arising in the classification procedure.
- Change detection was undertaken semi-automatically by developing an algorithm
that would account for flood water retreat over a defined period of time (chapter
6). The surface water dynamics algorithm (SWDA) uses temporal composite data
(in this study in the form of a water body product provided under the ALANIS
methane project). The SWDA turns these into binary raster classifications and then
sums each classification to arrive at one-directional water body change classifica-
tions. This procedure is semi-automated as metafile information on number of
measurements used per classification still need to be inspected manually. How-
ever, the SWDA enables a fast and efficient analysis of flood retreat over large ar-
eas. High limnicities and limnicity changes, representative of the extent of runoff
dynamics, were calculated. The greatest change in limnicity, found for a study area
in the Taz river floodplain, amounted to 1.79%, as opposed to the other study areas’
limnicity changes of 0.49% (Yamal) and 0.54% (Gydan).
iv) To analyse the distribution of water bodies, as well as the distribution of spring floods.
Size-frequency distribution were explored. At first, a comparison study between TerraSAR-X
and ASAR WS data was undertaken to evaluate the amount of water body signal lost due to
ASAR WS’s limiting spatial resolution. TerraSAR-X was found to be 1.6 times more efficient
in imaging water body extent, being able to map water bodies with sizes below O
(
100 m2
)
,
while ASAR WS was only able to map water bodies as small as O
(
103 m2
)
. In addition to this
3 magnitude difference, it needs to be noted that these water bodies are also below the chosen
MMU of 22500 m2. Downing et al. (2006) state, that in general, lakes below O
(
105 m2
)
are
underrepresented in maps and water body databases such as the GWLD compiled by Lehner
and Döll (2004). Hence, given the chosen MMU for the ASAR WS water body mapping, the
149
Research synthesis and achieved objectives Chapter 7. Conclusions
ASAR WS data actually provide a sufficient representation of water bodies; one magnitude
better at representing water bodies than the official GWLD.
In terms of the seasonal size-frequency statistics of water bodies, high R2 values in each
period for each year (2007, 2008 and 2009) indicate a close linear relationship. The lowest R2
values for each year are found in period A (0.93, 0.91 and 0.95, respectively), which is to be
expected given spring flood extent is at its highest during this period; their abundance being
low but with large areal coverage.
The reason for examining these size-frequency distributions is to evaluate whether these
may be used to extrapolate lake sizes-frequencies smaller than the measured distribution by
applying a model (given the linearity in the base 10 logarithm of abundance and the associated
high R2, the standard Pareto fit first suggested by Wetzel (1990) and Meybeck (1995), and later
applied by Muster et al. (2013) and Karlsson et al. (2014) in an Arctic tundra context, was cho-
sen). Such an extrapolation would be a great service to modelling carbon fluxes from inland
freshwater bodies. It was, found that the standard Pareto fit is not ideal if only larger lakes
are used for extrapolation. In this case, the power-law distribution would be biased towards
overestimating small lake abundances. In addition, it is necessary to establish the Pareto cut-off
point at which the power-law becomes discontinuous and (according to Pareto) drops to zero.
In reality, the lake size-frequency distribution will not be zero after this cut-off point but should
roll-off. This is indeed found for the TerraSAR-X data and may be evidence of a cut-off scale
of the Pareto distribution, where the lake size-frequency function is a combination of a range
of Pareto distributions. Significantly, this studies stresses the demand for further research on
water body extrapolation techniques so that carbon upscaling can be done at higher confidence
levels.
v) To establish links between relief and the seasonally inundated areas measured by the
ASAR WS sensor and identified by the change detection methods.
Terrain effects on surface water flow and accumulation were explored through temporally-
static hydrological modelling. In order to ensure the ASTER GDEM2, used as input, was suf-
ficiently accurate and so as not to bias the modelled stream channel network, DEM validation
studies were undertaken. A topographic map as well as another DEM (D-DEM) were used as
reference data. The RMSE of the DEM comparison study was found to be 9.64 m, which is
within the range of fair agreement suggested by Huggett (2002).
A flow accumulation map for the GDEM2 input and hence a stream channel network was
established to investigate whether the flood events detected by the change detection analysis
were related to hillslope runoff. No significant correlation between the two could, however, be
established. the mean overland flow distance was found to be 0.25 and 0.26 km for the Se-Yakha
and Mordy-Yakha basins, respectively. This may be interpreted as approximately 4 tributaries
per 1 km radius, associated with the gently rolling landscape on the Yamal peninsula.
150
Chapter 7. Conclusions Future work
vi) To explore the spatial flood patterns in the context of ongoing land-use and land-cover
change on the Yamal peninsula, in the North of West Siberia.
Manual inspection, using optical reference data (panchromatic SPOT image from July 1993),
showed that floods accumulate in alas systems in proximity of areas of high human activity.
This was also established by the SWDA which was run over a much larger area of the Yamal
peninsula. Here, floods seemed to occur along the Obskaya-Bovanenkovo railway line. The
links between these floods and infrastructure can only be inferred at this stage, however, previ-
ous studies by Zakharova et al. (2009, 2011) have highlighted the impact human activities have
on changes to the drainage system. Adequate field work is required to confirm the role of this
anthropogenic component.
vii) To stress the significance of rapid hydrological phenology in the Arctic on EO studies.
Lastly, the chief objective of this PhD, synthesising the above mentioned objectives (i - iii),
was to highlight the importance of rapid hydrological phenology for remote sensing of permafrost
lake studies. The term was coined during this PhD to stress the importance of incorporating
and the need to assess spring flood dynamics in inter-annual analyses of lake change studies.
This issue was addressed by presenting a frequent monitoring technique that uses SAR data
to identify water body extent and its spring dynamics (i, ii). Flood retreat and runoff may be
observed by applying simple change detection techniques to the water body product (iii). This
was at first done manually, creating difference maps that identified water pixel loss and gain.
Subsequently, a semi-automated method was proposed. The SWDA was developed and its
applicability to monitoring spring flood retreat assessed.
This PhD research hopes to illustrate the significance of spatially visualising rapid hydro-
logical phenology through dynamic water body mapping techniques. Given the easy access to
frequent, continuous and comparable SAR data provided in the ESA’s archives it would seem
remiss to not make use of these data, including these data on rapid hydrological phenology in
a longterm study of water body change. In particular it should be noted that the upcoming
Sentinel-1 mission will ensure continuity, providing a longer timeseries and therefore facilitat-
ing a more detailed inter-annual analysis of permafrost thaw lake dynamics.
7.2 Future work
7.2.1 Development of a dynamic thresholding algorithm
It was described in chapter 3 of this PhD dissertation that the current method by which SAR
data are classified operates on a static approach. An optimum threshold was chosen and is
applied to each normalised SAR image. This methodology enabled a straightforward, compu-
tationally efficient classification algorithm to be developed. It was further discussed in chapter
4 (section 4.2.2) how the temporal composite method was beneficial to improving the quality
151
Future work Chapter 7. Conclusions
of the classifications. However, having noted this, there is scope to develop a dynamic clas-
sification algorithm that evaluates the normalised backscatter distributions of each image of
a specified area in a given time period and determines the individual optimum threshold for
each of these SAR images. In relation to such an algorithm development, the decision on how to
statistically choose an optimum threshold may need revision. It may be of value to apply more
complex statistical models, such as a Neyman-Pearson criterion, which is a likelihood ratio
method for binary hypothesis testing (Jarabo-Amores et al. 2013). It remains to be investigated
whether such a complex statistical model would be superior to a more simplistic approach,
applying either a trough value or the 3σ rule to a double-Gaussian backscatter distribution.
7.2.2 Case 3: Pan-Siberian scale - Arctic Russia
The ALANIS methane local wetlands product coverage across Eurasia will be investigated.
The SWDA has in fact already been applied to the entire dataset. However, an analysis of the
results goes slightly beyond the scope of the PhD research, as the focus of this research was on
the Yamalo-Nenets Autonomous District, with an emphasis on mapping seasonal inundation
on the Yamal peninsula. It is hoped that the pan-Siberian SWDA study will be appropriately
investigated in due course, including not only data from 2007 but also data from consecutive
years to account for seasonality. The aim is for this study to be published in a high level journal.
7.2.3 Field work - Implementation of a seasonal inundation monitoring site
Despite all the efforts of this PhD research, there are no field data on spring flood extent and
retreat on the Yamal peninsula, which would establish a clearer picture of the seasonal dynam-
ics of surface hydrology in this area. Given ongoing research in this area, and the possibility
of the Vaskiny Dachi site being proposed as a so-called ’cold spot’ - an area for which the ESA
would prioritise acquisitions of satellite data (this was discussed at the last DUE Permafrost
User Workshop at ESRIN in Frascati in February 2014), it would be crucial to implement cer-
tain monitoring facilities (if the appropriate funds can be raised, and permissions granted):
- A remote meteorological station at the Vaskiny Dachi researh camp. Currently, the closest
weather station is Marre Sale, situated 100 km to the southwest of the Vaskiny
Dachi area on the coast of the Yamal peninsula. Key to a local analysis of seasonal
water bodies in this permafrost environment would be local data on precipitation
and air temperatures.
- Repeat photography. To monitor meltwater accumulation and flood retreat on the
ground, repeat photography would be an ideal tool. By setting up a camera system
pointing at calibrated reference poles, positioned in an alas-system known to act
as a flood basin (as determined from this research), the hydrological activity over
the entire spring period may be documented. The amount of snow accumulation
before the onset of thaw may be inferred from these calibrated poles in combination
152
Chapter 7. Conclusions Significance of this research
with the local meteorological data. The monitoring photos would further illustrate
the accumulation of these melt waters in the alas-system and visually record their
retreat over the summer.
- Invesigate changes to drainage system. To monitor potential changes human activities
have on the drainage system more field data are needed. It is yet unclear what the
best field data for such an investigation would be. This is a more complex issue
which will need careful consideration, and indeed would best be tackled through
collaboration with colleagues more experienced in setting up field experiments and
with local knowledge of the Yamal peninsula.
7.3 Significance of this research
This research presents a new approach for investigating Arctic lake dynamics. Understanding
these dynamics by determining and quantifying the processes that drive them is a key concept
needed to infer potential impacts on environmental change, as well as their significance to local
communities. Infrastructure may come under threat of periodic flooding as potential changes
to the drainage system arise due to disturbances in the ground thermal regime in these per-
mafrost environments. Furthermore, Arp and Jones (2009) have raised the need for analysing
lake abundance and dynamics in order to investigate their importance on local freshwater re-
sources and determine which ecosystem services they provide and hazards they pose to local
communities and economies. As such, this PhD research may only be a piece in a larger puzzle,
but due to the innovative approach of frequent monitoring, it is at the forefront of permafrost
lake research.
Furthermore, this research offers the potential for other studies on Arctic surface hydrol-
ogy, beyond permafrost research itself. For example, Johansson and Brown (2012) assess the
application of SAR data to map supra-glacial lakes on the Greenland ice sheet. Applying the
water body mapping procedure developed during this PhD to monitoring supra-glacial lakes
may be useful to modellers (e.g. Banwell et al. (2012) and Arnold et al. (2013)) interested in the
seasonal melt waters, accumulating in depressions on the ice sheet. Drainage events, and hence
the evolution of these melt water storages over the melt season, could potentially be monitored
and compared to the locations of the modelled lakes, by applying a similar technique presented
during this PhD dissertation.
In general, this PhD research may be applied to monitoring seasonal water body change.
Moreover, in the vast and remote Arctic environment, this kind of EO study may be used to
monitor water body seasonality on the landscape scale. Due to the role of Arctic freshwater
bodies as conduits of terrestrial carbon, this research topic reaches beyond the permafrost com-
munity, with significance to carbon and climate modellers - collaborative work that is hoped to
be undertaken in the future. This research may, therefore, be of interest to various stakeholders
(scientific, governmental and commercial).
153

Epilogue
A matter of scale
Two rather philosophical issues that this PhD research has encountered over the last few years
are related i) to the cartographic dilemma how best to represent change in a static map and ii)
to the definition of the optimum scale at which to represent both space and time. The former
was addressed by creating change detection maps, each class representing a change in flood
extent. This has become the common procedure for change detection analysis (e.g. Congalton
and Green (2009) and van Oort (2007)). The latter is less straightforward, but requires knowl-
edge of the role spatial and temporal variability play in the process that is to be investigated.
Knowledge thereof is essential for mastering the contentious matter of scaling (Sposito 2008).
Despite change detection procedure’s established place within spatial data analyses, prob-
lems of error propagation remain, and no standardisation in the methods used to carry out
change detection accuracy assessments exists. Error in the original classification may be wrongly
associated with change (Khorram et al. 1999) and therefore any changes detected over time may
not necessarily be due to real changes in the field (Foody 2002). This was an issue ever present
during this PhD research, the added matter of starting off with a temporal composite classifi-
cations rather than classifications derived from individual images, complicated the procedure
further. The problem of change detection and temporal composites was discussed in chapter 4
(sections 4.2.1 and 4.2.2) and will hence not be pondered further here.
Far more intriguing is the matter of appropriate choice in scale (both temporally and spa-
tially), as scale will remain precisely that: a choice. Any kind of map is just a representation of
the real phenomenon depicted in the map. Moreover, the map is in fact the map maker’s choice
in representation (Black 1998). The key questions that need to be asked are: What is it that needs
to be represented? How can this representation be achieved? Which landscape features and
properties need to be included to provide a truthful representation? Furthermore, careful con-
siderations concerning process-dependencies need to be made. Processes and events happen at
a certain frequency over a defined area. In order to adequately map these processes, temporal
and spatial boundaries need to be selected according to the events they represent and the type
of research that is undertaken (Inkpen and Wilson 2013).
In the permafrost community, it has become an established procedure to use high spatial
resolution satellite and airborne imagery that enable geomorphological features to be imaged
155
in detail (e.g. Grosse et al. (2005), Grosse and Jones (2011) and Rudy et al. (2013)). The reason
for this choice being the importance of capturing spatial extent over the rather long geomor-
phological timescales that carve these into the landscape. However, hydrological processes
happen on shorter timescales. In the Arctic tundra biome, snow accumulation will take place
throughout the winter season, but snow melt happens over a couple of weeks.
Capturing the hydrological extremes associated with melt waters requires instruments ca-
pable of measuring this change periodically. Hydrometric data may be used for such studies,
however, the point source nature of these measurements does not provide enough information
about its spatial variability. Arp et al. (2011) emphasise the difficulties involved in interpreting
these hydrometric mearurements over time due to rapidly changing water balances. Further-
more, temporal averages give no indication of the true timescale and respective magnitude of
spring flood events. To quote a recent paper by Zakharova et al. (2011) on the role of snowmelt
in spring floods and surface water flow:
“Spring flood can start and end at different dates and its duration may vary significantly
from one year to another, and thus some of the interannual variability of snowmelt is hidden
when monthly or annual averages are used.“
This PhD research has introduced a new approach to monitoring spatial patterns of sur-
face hydrology. The importance of including such a study into an analysis of thaw lake change
should be evident. Thaw lake change, being a geomorpholocial process, occurs over the timescale
of decades. However, the means for monitoring this change (with high-spatial resolution satel-
lite imagery) is compromised by the fact that these data are just a snapshot in time, giving no
indication on rapid hydrological phenology. Additional data are required to make such an anal-
ysis of thaw lake change convincing. What better way than to include data illustrating the
frequency and magnitude of rapid hydrological phenology spatially?
Bibliography
Aber, J. S., Pavri, F., and Aber, S. W. (2012). Wetland environments: A global perspective. Wiley-
Blackwell, Chichester.
Abramoff, M. D., Magalhaes, P. J., and Ram, S. J. (2004). Image Processing with ImageJ. Biopho-
tonics International, 11(7):36 – 42.
Alessa, L., Kliskey, A., Busey, R., Hinzman, L., and White, D. (2008). Freshwater vulnerabil-
ities and resilience on the Seward Peninsula: Integrating multiple dimensions of landscape
change. Global Environmental Change, 18:256 – 270.
Anderson, D. (2004). Glacial and Periglacial Environments. Hodder Education, London.
Anisimov, O. and Reneva, S. (2006). Permafrost and Changing Climate: The Russian Perspec-
tive. AMBIO, 35(4):169 – 175.
Anisimov, O. A., Shiklomanov, N. I., and Nelson, F. E. (2002). Variability of seasonal thaw depth
in permafrost regions: a stochastic modeling approach. Ecological Modelling, 153:217 – 227.
Arnold, N. S., Banwell, A. F., and Willis, I. C. (2013). High-resolution modelling of the seasonal
evolution of surface water storage on the Greenland Ice Sheet. The Cryosphere Discussions,
7(6):6143 – 6170.
Arp, C. and Jones, B. (2009). Geography of Alaska Lake Districts: Identification, Description
and Analysis of Lake-Rich Regions of a Diverse and Dynamic State. US Geological Survey
Scientific Investigations Report 2008-5215, USGS National Center: Reston, VA, USA, 2009.
Arp, C. D., Jones, B. M., Lu, Z., and Whitman, M. S. (2012). Shifting balance of thermokarst lake
ice regimes across the Arctic Coastal Plain of northern Alaska. Geophysical Research Letters,
39(16).
Arp, C. D., Jones, B. M., Urban, F. E., and Grosse, G. (2011). Hydrogeomorphic processes of
thermokarst lakes with grounded-ice and floating-ice regimes on the Arctic coastal plain,
Alaska. Hydrological Processes, 25:2422 – 2438. 1099-1085.
Astakhov, V. and Nazarov, D. (2010). Correlation of Upper Pleistocene sediments in northern
West Siberia. Quaternary Science Reviews, 29:3615 – 3629.
157
BIBLIOGRAPHY
ASTER GDEM Validation Team (2011). ASTER Global Digital Elevation Model Version
2 - Summary of Validation Results. http://www.jspacesystems.or.jp/ersdac/GDEM/
ver2Validation/Summary_GDEM2_validation_report_final.pdf.
Banwell, A. F., Arnold, N. S., Willis, I. C., Tedesco, M., and Ahlstrøm, A. P. (2012). Modeling
supraglacial water routing and lake filling on the Greenland Ice Sheet. Journal of Geophysical
Research: Earth Surface, 117(F4):F04012.
Bartalis, Z., Wagner, W., Naeimi, V., Hasenauer, S., Scipal, K., Bonekamp, H., Figa, J., and An-
derson, C. (2007). Initial soil moisture retrieval from the METOP-A Advanced Scatterometer.
Geophysical Research Letters, 34(L20401).
Bartsch, A. (2010). Ten Years of SeaWinds on QuikSCAT for Snow Applications. Remote Sensing,
2:1142 – 1156.
Bartsch, A., Kidd, R. A., Pathe, C., Scipal, K., and Wagner, W. (2007a). Satellite radar imagery
for monitoring inland wetlands in boreal and subarctic environments. Aquatic Conservation:
Marine and Freshwater Ecosystems, 17:305 – 317.
Bartsch, A., Kidd, R. A., Wagner, W., and Bartalis, Z. (2007b). Temporal and spatial variability
of the beginning and end of daily spring freeze/thaw cycles derived from scatterometer data.
Remote Sensing of Environment, 106:360 – 374.
Bartsch, A., Kumpula, T., Forbes, B. C., and Stammler, F. (2010). Detection of snow surface
thawing and refreezing in the Eurasian Arctic with QuikSCAT: implications for reindeer
herding. Ecological Applications, 20:2346 – 2358.
Bartsch, A., Pathe, C., Scipal, K., and Wagner, W. (2008). Detection of permanent open water
surfaces in central Siberia with ENVISAT ASAR wide swath data with special emphasis on
the estimation of methane fluxes from tundra wetlands. Hydrology Research, 39(2):89 – 100.
Bartsch, A., Pathe, C., and Wagner, W. (2007c). Relative soil moisture from C- and L-band
SAR time series. In Proc. of the First Joint PI symposium of ALOS Data Nodes for ALOS Science
Program, Kyoto.
Bartsch, A., Trofaier, A. M., Hayman, G., Sabel, D., Schlaffer, S., Clark, D., and Blyth, E. (2012).
Detection of wetland dynamics with ENVISAT ASAR in support of methane modelling at
high latitudes. Biogeosciences, 9:703 – 714.
Bartsch, A., Wagner, W., Scipal, K., Pathe, C., Sabel, D., and Wolski, P. (2009). Global mon-
itoring of wetlands - the value of ENVISAT ASAR Global mode. Journal of Environmental
Management, 90:2226 – 2233.
Bhatt, U., Walker, D., Raynolds, M., and Comiso, J. (2007). The Relationship Between Sea
Ice Variability and Arctic Tundra on the Pan-Arctic, Regional and Site Scales. In AGU Fall
Meeting Abstracts, volume 1, page 0612.
158
BIBLIOGRAPHY
Bishop, M. P., James, L. A., Jr., J. F. S., and Walsh, S. J. (2012). Geospatial technologies and
digital geomorphological mapping: Concepts, issues and research. Geomorphology, 137(1):5 –
26. Geospatial Technologies and Geomorphological Mapping Proceedings of the 41st Annual
Binghamton Geomorphology Symposium.
Black, J. (1998). Maps and Politics. Picturing history. University of Chicago Press.
Blikra, L. H. and Christiansen, H. H. (2014). A field-based model of permafrost-controlled
rockslide deformation in northern norway. Geomorphology, 208(0):34 – 49.
Bliss, L., Heal, O., Moore, J., and Programme, I. B. (1981). Tundra Ecosystems: A Comparative
Analysis. International Biological Programme. Cambridge University Press.
Brown, J., Ferrians, J., Heginbottom, J., and Melnikov, E. (1997). Circum-Arctic map of per-
mafrost and ground-ice conditions. Washington, DC: U.S. Geological Survey in Cooperation
with the Circum-Pacific Council for Energy and Mineral Resources. Circum-Pacific Map Se-
ries CP-45, scale 1:10,000,000.
Brown, L. and Young, K. L. (2006). Assessment of Three Mapping Techniques to Delineate
Lakes and Ponds in a Canadian High Arctic Wetland Complex. Arctic, 59(3):283 – 293.
Bunn, A. G., Goetz, S. J., Kimball, J. S., and Zhang, K. (2007). Northern high-latitude ecosystems
respond to climate change. Eos, Transactions American Geophysical Union, 88(34):333 – 335.
Burger, W. and Burge, M. J. (2008). Digital Image Processing. Springer, New York.
Burn, C. (2002). Tundra lakes and permafrost, Richards Island, western Arctic coast, Canada.
Canadian Journal of Earth Sciences, 39:1281 – 1298.
Burn, C. (2005). Lake-bottom Thermal Regimes, Western Arctic Coast, Canada. Permafrost and
Periglacial Processes, 16:355 – 367.
Burn, C. R. and Smith, M. W. (1990). Development of thermokarst lakes during the Holocene
at sites near Mayo, Yukon territory. Permafrost and Periglacial Processes, 1(2):161 – 175.
Burnicki, A. C., Brown, D. G., and Goovaerts, P. (2007). Simulating error propagation in land-
cover change analysis: The implications of temporal dependence. Computers, Environment
and Urban Systems, 31(3):282 – 302.
Callaghan, T., B., O., L., Chapin III, F. S., Chernov, Y., Christensen, T. R., Hutley, B., Ims, R.,
Johansson, M., Riedlinger, D. J., Jonasson, S., Matveyeva, N., Oechel, W., Panikov, N., and
Shaver, G. (2005). Arctic Climate Impact Assessment. chapter 7, Arctic Tundra and Polar
Desert Ecosystems. Cambridge University Press, Cambridge.
Campbell, J. B. (1996). Intoduction to Remote Sensing. Taylor & Francis, London, second edition.
159
BIBLIOGRAPHY
Carey, S. K. and Woo, M.-K. (2001). Slope runoff processes and flow generation in a subarctic,
subalpine catchment. Journal of Hydrology, 253(1–4):110 – 129.
Cherkauer, K., Bowling, L., and Naz, B. (2013). 8.13 watershed hydrology in periglacial envi-
ronments. In Shroder, J. F., editor, Treatise on Geomorphology, pages 151 – 172. Academic Press,
San Diego.
Christiansen, H. H., Etzelmüller, B., Isaksen, K., Juliussen, H., Farbrot, H., Humlum, O., Jo-
hansson, M., Ingeman-Nielsen, T., Kristensen, L., Hjort, J., Holmlund, P., Sannel, A. B. K.,
Sigsgaard, C., Åkerman, H. J., Foged, N., Blikra, L. H., Pernosky, M. A., and Ødegård, R. S.
(2010). The thermal state of permafrost in the Nordic area during the international polar year
2007–2009. Permafrost and Periglacial Processes, 21(2):156 – 181.
Closa, J., Rosich, B., and Monti-Guarnieri, A. (2003). The asar wide swath mode products. In
Geoscience and Remote Sensing Symposium, 2003. IGARSS ’03. Proceedings. 2003 IEEE Interna-
tional, volume 2, pages 1118–1120 vol.2.
Cohen, J. (1960). A coefficient of agreement for nominal scales. Educational and Psychological
Measurement, 20(1):37 – 46.
Cole, J. J., Prairie, Y. T., Caraco, N. F., McDowell, W. H., Tranvik, L. J., Striegl, R. G., Duarte,
C. M., Kortelainen, P., Downing, J. A., Middelburg, J. J., et al. (2007). Plumbing the global car-
bon cycle: integrating inland waters into the terrestrial carbon budget. Ecosystems, 10(1):172
– 185.
Comber, A., Fisher, P., Brunsdon, C., and Khmag, A. (2012). Spatial analysis of remote sensing
image classification accuracy. Remote Sensing of Environment, 127:237 – 246.
Congalton, R. and Green, K. (2009). Assessing the Accuracy of Remotely Sensed Data: Principles
and Practices. Mapping Science Series. CRC Press/Taylor & Francis.
Conolly, J. and Lake, M. (2006). Geographical Information Systems In Archaeology. Cambridge
Manuals in Archaeology. Cambridge University Press.
Cooper, R., Hodgkins, R., Wadham, J., and Tranter, M. (2011). The hydrology of the proglacial
zone of a high-arctic glacier(finsterwalderbreen, svalbard): Sub-surface water fluxes and
complete water budget. Journal of Hydrology, 406(1–2):88 – 96.
Crave, A. and Gascuel-Odoux, C. (1997). The influence of topography on time and space dis-
tribution of soil surface water content. Hydrological Processes, 11(2):203 – 210.
Crossman, J., Bradley, C., David, J., and Milner, A. (2012). Use of remote sensing to identify
areas of groundwater upwelling on active glacial floodplains: Their frequency, extent and
significance on a landscape scale. Remote Sensing of Environment, 123(0):116 – 126.
160
BIBLIOGRAPHY
De Jeu, R., Wagner, W., Holmes, T., Dolman, H., Van De Giesen, N., and Friesen, J. (2009).
Global Soil Moisture Patterns Observed by Space Borne Microwave Radiometers and Scat-
terometers. Surveys in Geophysics, 29:399 – 420.
Dean, W. E. and Gorham, E. (1998). Magnitude and significance of carbon burial in lakes,
reservoirs, and peatlands. Geology, 26(6):535 – 538.
Desnos, Y.-L., Buck, C., Guijarro, J., Suchail, J.-L., Torres, R., and Attema, E. (2000). ASAR -
ENVISAT’s Advanced Synthetic Aperture Radar. ESA bulletin, 102:91–100.
Downing, J. A. (2009). Plenary lecture global limnology: Up-scaling aquatic services and pro-
cesses to planet Earth. Verh. Internat. Verein. Limnol., 30:1149 – 1166.
Downing, J. A. (2010). Emerging global role of small lakes and ponds: little things mean a lot.
Limnetica, 1(29):9 – 24.
Downing, J. A., Prairie, Y. T., Cole, J. J., Duarte, C. M., Tranvik, L. J., Striegl, R. G., McDow-
ell, W. H., Kortelainen, P., Caraco, N. F., Melack, J. M., and Middelburg, J. J. (2006). The
global abundance and size distribution of lakes, ponds, and impoundments. Limnology and
Oceanography, 51:2388 – 2397.
Duguay, C. R., Flato, G. M., Jeffries, M. O., Ménard, P., Morris, K., and Roue, W. R. (2003).
Ice-cover variability on shallow lakes at high latitudes: model simulations and observations.
Hydrological Processes, 17:3465 – 3483.
Duguay, C. R., Zhang, T., Leverington, D. W., and Romanovsky, V. E. (2005). Satellite Remote
Sensing of Permafrost and Seasonally Frozen Ground. In Remote Sensing in Northern Hydrol-
ogy: Measuring Environmental Change, Geophysical Monograph Series 163, pages 91 – 118.
American Geophysical Union.
Eckerstorfer, M. and Christiansen, H. H. (2012). Meteorology, Topography and Snowpack Con-
ditions causing Two Extreme Mid-Winter Slush and Wet Slab Avalanche Periods in High
Arctic Maritime Svalbard. Permafrost and Periglacial Processes, 23(1):15 – 25.
Edwards, M., Walter, K., Grosse, G., Plug, L., Slater, L., Valdes, P., et al. (2009). Arctic
thermokarst lakes and the carbon cycle. Science Highlights PAGES News: Change at the Poles,
17(1):16 – 18.
Engram, M., Anthony, K. W., Meyer, F. J., and Grosse, G. (2013). Characterization of L-band
synthetic aperture radar (SAR) backscatter from floating and grounded thermokarst lake ice
in Arctic Alaska. The Cryosphere, 7(6):1741–1752.
Entekhabi, D., Njoku, E., O’Neill, P., Kellogg, K., Crow, W., Edelstein, W., Entin, J., Goodman,
S., Jackson, T., Johnson, J., Kimball, J., Piepmeier, J., Koster, R., Martin, N., McDonald, K.,
Moghaddam, M., Moran, S., Reichle, R., Shi, J.-C., Spencer, M., Thurman, S., Tsang, L., and
161
BIBLIOGRAPHY
Van Zyl, J. (2010). The Soil Moisture Active Passive (SMAP) Mission. Proceedings of the IEEE,
98(5):704 – 716.
ESA (2007). ASAR product handbook. https://earth.esa.int/pub/ESA_DOC/ENVISAT/
ASAR/asar.ProductHandbook.2_2.pdf.
Evans, T. L., Costa, M., Telmer, K., and Silva, T. S. F. (2010). Using ALOS/PALSAR and
RADARSAT-2 to Map Land Cover and Seasonal Inundation in the Brazilian Pantanal. Se-
lected Topics in Applied Earth Observations and Remote Sensing, IEEE Journal of, 3(4):560 – 575.
Everett, K. R., Vassiljevskaya, V., Brown, J., and D., W. B. (1981). Tundra and analogous soils.
In Bliss, L. C., Heal, O. W., and Moore, J. J., editors, Tundra ecosystems: a comparative analysis.
Cambridge University Press, Cambridge.
Ferrians, O. J., Kachadoorian, R., and Green, G. W. (1969). Permafrost and related engineering
problems in Alaska. United States Geological Survey, professional paper, 678:37pp.
Flessa, H., Rodionov, A., Guggenberger, G., Fuchs, H., Magdon, P., Shibistova, O.,
Zrazhevskaya, G., Mikheyeva, N., Kasanska, O., and Blodau, C. (2008). Landscape controls
of CH4 fluxes in a catchment of the forest tundra ecotone in northern Siberia. Global Change
Biology, 14:2040 – 2056.
Foody, G. (2002). Status of land cover classification accuracy assessment. Remote Sensing of
Environment, 80(1):185 – 201.
Forbes, B. C. and Kumpula, T. (2009). The Ecological Role and Geography of Reindeer (Rangifer
tarandus) in Northern Eurasia. Geography Compass, 3(4):1356 – 1380.
Forouzan, A. and Araabi, B. (2003). Iterative median filtering for restoration of images with
impulsive noise. In Electronics, Circuits and Systems, 2003. ICECS 2003. Proceedings of the 2003
10th IEEE International Conference on, volume 1, pages 232–235 Vol.1.
Forster, P., Ramaswamy, V., Artaxo, P., Berntsen, T., Betts, R., Fahey, D., Haywood, J., Lean,
J., Lowe, D., Myhre, G., Nganga, J., Prinn, R., Raga, G., Schulz, M., and Van Dorland, R.
(2007). Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the
Fourth Assessment Report of the IPCC, chapter 2: Changes in Atmospheric Constitutes and in
Radiative Forcing. Cambridge University Press, Cambridge, UK and New York, USA.
Frampton, A., Painter, S., Lyon, S. W., and Destouni, G. (2011). Non-isothermal, three-phase
simulations of near-surface flows in a model permafrost system under seasonal variability
and climate change. Journal of Hydrology, 403(3):352 – 359.
French, H. M. (2007). The Periglacial Environment. John Wiley & Sons, Chichester, 3rd edition.
162
BIBLIOGRAPHY
Frey, K. E. and Smith, L. C. (2007). How well do we know northern land cover? Comparison
of four global vegetation and wetland products with a new ground-truth database for West
Siberia. Global Biogeochemical Cycles, 21(1):GB1016.
Frohn, R. C., Hinkel, K. M., and Eisner, W. R. (2005). Satellite remote sensing classification
of thaw lakes and drained thaw lake basins on the north slope of alaska. Remote Sensing of
Environment, 97(1):116 – 126.
Gala, T. and Melesse, A. (2012). Monitoring prairie wet area with an integrated LANDSAT
etm+, radarsat-1 SAR and ancillary data from LIDAR. CATENA, 95(0):12 – 23.
Gan, L., Li, D., and Song, S. (2006). Is the Zipf law spurious in explaining city-size distributions?
Economics Letters, 92(2):256 – 262.
Gao, J. (1997). Resolution and accuracy of terrain representation by grid DEMs at a micro-scale.
International Journal of Geographical Information Science, 11(2):199 – 212.
Gisnås, K., Westermann, S., Schuler, T. V., Litherland, T., Isaksen, K., Boike, J., and Etzelmüller,
B. (2014). A statistical approach to represent small-scale variability of permafrost tempera-
tures due to snow cover. The Cryosphere Discussions, 8(1):509 – 536.
Gordeev, V. (2006). Fluvial sediment flux to the Arctic Ocean. Geomorphology, 80:94 – 104.
Goswami, S., Gamon, J. A., and Tweedie, C. E. (2011). Surface hydrology of an arctic ecosystem:
Multiscale analysis of a flooding and draining experiment using spectral reflectance. Journal
of Geophysical Research, 116:G00I07.
Grabs, T., Seibert, J., Bishop, K., and Laudon, H. (2009). Modeling spatial patterns of saturated
areas: A comparison of the topographic wetness index and a dynamic distributed model.
Journal of Hydrology, 373(1):15 – 23.
Granger, R. J., Gray, D. M., and Dyck, G. E. (1984). Snowmelt infiltration to frozen prairie soils.
Canadian Journal of Earth Sciences, 21(6):669 – 677.
Greenwell, R. N. and Finch, S. J. (2004). Randomized rejection procedure for the two-sample
kolmogorov–smirnov statistic. Computational Statistics & Data Analysis, 46(2):257 – 267.
Grosse, G., Jones, B., and Arp, C. (2013). 8.21 thermokarst lakes, drainage, and drained basins.
In Shroder, J. F., editor, Treatise on Geomorphology, pages 325 – 353. Academic Press, San Diego.
Grosse, G. and Jones, B. M. (2011). Spatial distribution of pingos in northern Asia. The
Cryosphere, 5(1):13 – 33. TC.
Grosse, G., Romanovsky, V., Jorgenson, T., Anthony, K. W., Brown, J., and Overduin, P. P. (2011).
Vulnerability and Feedbacks of Permafrost to Climate Change. Eos, Transactions American
Geophysical Union, 92(9):73 – 74.
163
BIBLIOGRAPHY
Grosse, G., Romanovsky, V., Walter, K., Morgenstern, A., Lantuit, H., and Zimov, S. (2008).
Distribution of Thermokarst Lakes and Ponds at Three Yedoma Sites in Siberia. In Ninth
International Conference on Permafrost, pages 551–556, Fairbanks. Institute of Northern Engi-
neering UAF.
Grosse, G., Schirrmeister, L., Kunitsky, V., and Hubberten, H.-W. (2005). The use of CORONA
images in remote sensing of periglacial geomorphology: an illustration from the NE Siberian
coast. Permafrost & Periglacial Processes, 16(2):163 – 172.
Grosse, G., Schirrmeister, L., and Malthus, T. (2006). Application of Landsat-7 satellite data and
a DEM for the quantification of thermokarst-affected terrain types in theperiglacial Lena-
Anabar coastal lowland. Polar Research, 25(1):51 – 67.
Guneriussen, T. (1997). Backscattering properties of a wet snow cover derived from DEM cor-
rected ERS-1 SAR data. International Journal of Remote Sensing, 18(2):375–392.
Hahn, S., Melzer, T., and Wagner, W. (2012). Error assessment of the initial near real-time
METOP ASCAT Surface Soil Moisture Product. IEEE Transactions on Geoscience and Remote
Sensing, 50(7):2556 – 2565.
Hamandawana, H., Eckardt, F., and Ringrose, S. (2006). The use of step-wise density slicing in
classifying high-resolution panchromatic photographs. International Journal of Remote Sensing,
27(21):4923 – 4942.
Han, H. and Lee, H. (2013). Radar Backscattering of Lake Ice During Freezing and Thawing
Stages Estimated by Ground-Based Scatterometer Experiment and Inversion From Genetic
Algorithm. Geoscience and Remote Sensing, IEEE Transactions on, 51(5):3089–3096.
Hayashi, M., van der Kamp, G., and Schmidt, R. (2003). Focused infiltration of snowmelt water
in partially frozen soil under small depressions. Journal of Hydrology, 270(3):214 – 229.
Hendriks, M. (2010). Introduction to physical hydrology. Oxford University Press.
Hinkel, K., Frohn, R., Nelson, F., Eisner, W., and Beck, R. (2005). Morphometric and Spatial
Analysis of Thaw Lakes and Drained Thaw Lake Basins in the Western Arctic Coastal Plain,
Alaska. Permafrost and Periglacial Processes, 16:327 – 341.
Hinkel, K. M. and Outcalt, S. I. (1994). Identification of heat-transfer processes during soil
cooling, freezing, and thaw in central Alaska. Permafrost and Periglacial Processes, 5(4):217 –
235.
Hirano, A., Welch, R., and Lang, H. (2003). Mapping from ASTER stereo image data: DEM val-
idation and accuracy assessment. ISPRS Journal of Photogrammetry & Remote Sensing, 57:356
– 370.
164
BIBLIOGRAPHY
Horritt, M. S., Mason, D. C., Davenport, I. J., and Bates, P. D. (2003). Waterline mapping in
flooded vegetation from airborne SAR imagery. Remote Sensing of Environment, 85:271 – 281.
Horton, R. E. (1932). Drainage basin characteristics. Transactions American Geophysical Union,
14:350–361.
Horton, R. E. (1945). Erosional development of streams and their drainage basins; hydrophys-
ical approach to quantitative morphology. Geological Society of America Bulletin, 56:275 – 370.
Howard, A. D. (1971). Optimal Angles of Stream Junction: Geometric, Stability to Capture, and
Minimum Power Criteria. Water Resources Research, 7(4):863 – 873.
Huemmrich, K. F., Gamon, J. A., Tweedie, C. E., Oberbauer, S. F., Kinoshita, G., Houston, S.,
Kuchy, A., Hollister, R. D., Kwon, H., Mano, M., Harazono, Y., Webber, P. J., and Oechel,
W. C. (2010). Remote sensing of tundra gross ecosystem productivity and light use efficiency
under varying temperature and moisture conditions. Remote Sensing of Environment, 114:481
– 489.
Huggett, R. J. (2002). Topography and the Environment. Prentice Hall, Harlow.
Igarashi, T. (2001). ALOS Mission Requirement and Sensor Specifications. Adv. Space Res.,
28(1):127 – 131.
Iijima, Y., Federov, A., Park, H., Suzuku, K., Yabuki, H., Maximov, T., and Ohata, T. (2010).
Abrupt Increases in Soil Temperatures following Increased Precipitation in a Permafrost Re-
gion, Central Lena River Basin, Russia. Permafrost and Periglacial Processes, 21:30 – 41.
Inkpen, R. and Wilson, G. (2013). Science, Philosophy and Physical Geography. Taylor & Francis.
Intergovernmental Panel on Climate Change (2013). Climate change 2013: The physical
science basis. Cambridge University Press, Cambridge.
Jarabo-Amores, M.-P., De la Mata-Moya, D., Gil-Pita, R., and Rosa-Zurera, M. (2013). Radar
detection with the Neyman - Pearson criterion using supervised-learning-machines trained
with the cross-entropy error. EURASIP Journal on Advances in Signal Processing, 2013(1):1 –
10.
JAXA (2011). Advanced land observing satellite “daichi” (alos). http://www.jaxa.jp/projects/
sat/alos/index_e.html.
Jeffries, M., Morris, K., and Liston, G. (1996). A Method to determine lake depth and water
availability on the North Slope of Alaska with Spaceborne Imaging Radar and Numerical Ice
Growth Modelling. Arctic, 49(4):367 – 374.
Jeffries, M. O., Morris, K., and Kozlenko, N. (2005). Ice Characteristics and Processes, and Re-
mote Sensing of Frozen Rivers and Lakes. In Remote Sensing in Northern Hydrology: Measuring
Environmental Change, pages 63 – 90. American Geophysical Union.
165
BIBLIOGRAPHY
Jia, G. J., Epstein, H. E., and Walker, D. A. (2003). Greening of arctic Alaska, 1981–2001. Geo-
physical Research Letters, 30(20):2067.
Johansson, A. M. and Brown, I. A. (2012). Observations of supra-glacial lakes in west Greenland
using winter wide swath Synthetic Aperture Radar. Remote Sensing Letters, 3(6):531 – 539.
Jones, B., Grosse, G., Arp, C., Jones, M., Walter Anthony, K., and Romanovsky, V. (2011).
Modern thermokarst lake dynamics in the continuous permafrost zone, northern Seward
Peninsula, Alaska. Journal of Geophysical Research, 116:G00M03.
Jones, H. G. and Vaughan, R. A. (2010). Remote sensing of vegetation. Oxford University Press,
Oxford.
Jorgenson, M. T., Harden, J., Kanevskiy, M., O’Donnell, J., Wickland, K., Ewing, S., Manies, K.,
Zhuang, Q., Shur, Y., Striegl, R., et al. (2013). Reorganization of vegetation, hydrology and
soil carbon after permafrost degradation across heterogeneous boreal landscapes. Environ-
mental Research Letters, 8(3):035017.
Kääb, A. (2008). Remote sensing of permafrost-related problems and hazards. Permafrost and
Periglacial Processes, 19(2):107 – 136.
Karlsson, J. M., Lyon, S. W., and Destouni, G. (2012). Thermokarst lake, hydrological flow and
water balance indicators of permafrost change in Western Siberia. Journal of Hydrology, pages
459 – 466.
Karlsson, J. M., Lyon, S. W., and Destouni, G. (2014). Temporal Behavior of Lake Size-
Distribution in a Thawing Permafrost Landscape in Northwestern Siberia. Remote Sensing,
6(1):621 – 636.
Kelly, M., Tuxen, K. A., and Stralberg, D. (2011). Mapping changes to vegetation pattern in a
restoring wetland: Finding pattern metrics that are consistent across spatial scale and time.
Ecological Indicators, 11(2):263 – 273.
Khitun, O., Rebristaya, O., Watson, A. E., and Alessa, L. (2001). Anthropogenic impacts on
habitat structure and species richness in the West Siberian Arctic.
Khorram, S., for Photogrammetry, A. S., and Sensing, R. (1999). Accuracy assessment of remote
sensing-derived change detection. ASPRS monograph series. American Society for Photogram-
metry and Remote Sensing.
Kidd, R. A., Bartsch, A., and Wagner, W. (2004). Development and validation of a diurnal
difference indicator for freeze-thaw monitoring in the SIBERIA II project. In Proceedings of the
2004 ENVISAT & ERS Symposium, Salzburg, Austria.
166
BIBLIOGRAPHY
Kimball, J. S., McDonald, K. C., Frolking, S., and Running, S. W. (2004). Radar remote sensing
of spring thaw transition across a boreal landscape. Remote Sensing of Environment, 89:163 –
175.
Kirpotin, S., Polishchuk, Y., Zakharova, E., Shirokova, L., Pokrovsky, O., Kolmakova, M., and
Dupre, B. (2008). One of the possible mechanisms of thermokarst lakes drainage in West-
Siberian North. International Journal of Environmental Studies, 65(5):631 – 635.
Kozlenko, N. and Jeffries, M. (2000). Bathymetric Mapping o Shallow Water in Thaw Lakes on
the North Slope of Alaska with Spaceborne Imaging Radar. Arctic, 53(3):306 – 316.
Kravtsova, V. I. and Tarasenk, T. V. (2011). The Dynamics of thermokarst lakes under Climate
Change since 1950, Central Yakutia. Êðèîñôåðà Çåìëè, XV:31 – 42. In Russian.
Kremenetski, K., Velichko, A., Borisova, O., MacDonald, G., Smith, L., Frey, K., and Orlova, L.
(2003). Peatlands of the Western Siberian lowlands: current knowledge on zonation, carbon
content and Late Quaternary history. Quaternary Science Reviews, 22:703 – 723.
Kuittinen, R. (1988). Determination of areal snow water equivalent using satellite images and gamma
ray spectrometry. Acta polytechnica Scandinavica. Civil engineering and building construc-
tion series. Finnish Acad. of Technology.
Kumpula, T., B.C., F., Stammler, F., and Meschtyb, N. (2012). Dynamics of a Coupled System:
Multi-Resolution Remote Sensing in Assessing Social-Ecological Responses during 25 Years
of Gas Field Development in Arctic Russia. Remote Sensing, 4:875 – 881.
Kumpula, T., C., F. B., and Stammler, F. (2010). Remote Sensing and Local Knowledge of Hy-
drocarbon Exploitation: The Case of Bovanenkovo, Yamal Peninsula, West Siberia, Russia.
Arctic, 2:165 – 178.
Kumpula, T., Pajunen, A., Kaarlejärvi, E., Forbes, B. C., and Stammler, F. (2011). Land use and
land cover change in Arctic Russia: Ecological and social implications of industrial develop-
ment. Global Environmental Change, 21:550 – 562.
Labrecque, S., Lacelle, D., Duguay, C., Lauriol, B., and Hawkings, J. (2009). Contemporary
(1951–2001) Evolution of Lakes in the Old Crow Basin, Northern Yukon, Canada: Remote
Sensing, Numerical Modeling, and Stable Isotope Analysis. ARCTIC, 62(2).
Landgrebe, D. A. and Malaret, E. (1986). Noise in Remote-Sensing Systems: The Effect on
Classification Error. IEEE Transaction on Geoscience and Remote Sensing, GE-24:294 – 300.
Langer, M., Westermann, S., Heikenfeld, M., Dorn, W., and Boike, J. (2013). Satellite-based
modeling of permafrost temperatures in a tundra lowland landscape. Remote Sensing of En-
vironment, 135(0):12 – 24.
167
BIBLIOGRAPHY
Lee, J. S. (1981). Speckle analysis and Smoothing of Synthetic Aperture Radar Images. Computer
Graphics and Image Processing, 17:24 – 32.
Lehner, B. and Döll, P. (2004). Development and validation of a global database of lakes, reser-
voirs and wetlands. Journal of Hydrology, 296:1 – 22.
Leibman, M., Gubarkov, A., and Khomutov, A. (2012). Research station Vaskiny Dachi. Ex-
cursion guidebook. Tenth International Conference on Permafrost, Salekhard, Russia, June
25-29, 2012. Tyumen: Pechatnik, 50pp.
Leibman, M. and Streletskaya, I. (1997). Land-slide induced changes in the chemical composi-
tion of active layer soils and surface-water run-off, Yamal Peninsula, Russia. In Proceedings
of the International Symposium on Physics, Chemistry and Ecology of Seasonally Frozen Soils, Fair-
banks, Alaska, pages 10 – 12.
Leibman, M. O. and Kizyakov, A. I. (2007). Cryogenic Landslides of the Yamal and Yugorsky Penin-
sulas. Earth Cryosphere Institute SB RAS, Moscow.
Lewellen, W. S. and Yoh, S. (1993). Binormal Model of Ensemble Partial Cloudiness. Journal of
the Atmospheric Sciences, 50(9):1228 – 1237.
Likens, G. E. (2009). Encyclopedia of inland waters. Academic Press, Amsterdam ; London.
Liu, H. and Zhou, Q. (2004). Accuracy analysis of remote sensing change detection by rule-
based rationality evaluation with post-classification comparison. International Journal of Re-
mote Sensing, 25(5):1037–1050.
Loew, A., Ludwig, R., and Mauser, W. (2006). Derivation of Surface Soil Moisture From EN-
VISAT ASAR Wide Swath and Image Mode Data in Agricultural Areas. IEEE Transactions on
Geoscience and Remote Sensing, 44(4):889 – 899.
Loew, A. and Mauser, W. (2007). Generation of geometrically and radiometrically terrain
corected SAR image products. Remote Sensing of Environment, 106:337 – 349.
Loew, A., Stacke, T., Dorigo, W., De Jeu, R., and Hagemann, S. (2013). Potential and limitations
of multidecadal satellite soil moisture observations for selected climate model evaluation
studies. Hydrology and Earth System Sciences, 17(9):3523 – 3542.
Lupascu, M., Welker, J., Seibt, U., Maseyk, K., Xu, X., and Czimczik, C. (2013). High Arctic
wetting reduces permafrost carbon feedbacks to climate warming. Nature Climate Change,
pages 51 – 55.
Lyons, W. B. and Finlay, J. C. (2008). Biogeochemical processes in high-latitude lakes and rivers.
In Vincent, W. F. and Laybourn-Parry, J., editors, Polar Lakes and Rivers: Limnology of Arctic
and Antarctic Aquatic Ecosystems, pages 137 – 156. Oxford University Press, New York.
168
BIBLIOGRAPHY
MacDougall, A. H., Avis, C. A., and Weaver, A. J. (2012). Significant contribution to climate
warming from the permafrost carbon feedback. Nature Geoscience, 5(10):719 – 721.
Macias-Fauria, M., Forbes, B. C., Zetterberg, P., and Kumpula, T. (2012). Eurasian Arctic green-
ing reveals teleconnections and the potential for structurally novel ecosystems. Nature Cli-
mate Change, 2(8):613 – 618.
Mackay, J. R. (1988). Catastrophic lake drainage, Tuktoyaktuk Peninsula area, District of
Mackenzie. In Current Research, Part D, Geol. Surv. Canada, Paper 88-1D, pages 83 – 90.
Mahaney, W., Michel, F., Solomatin, V., and Hütt, G. (1995). Late Quaternary stratigraphy and
soils of Gydan, Yamal and Taz Peninsulas, northwestern Siberia. Palaeogeography, Palaeocli-
matology, Palaeoecology, 113:249 – 266.
Matgen, P., Hostache, R., Schumann, G., Pfister, L., Hoffmann, L., and Savenije, H. H. G. (2011).
Towards an automated SAR-based flood monitoring system: Lessons learned from two case
studies. Physics and Chemistry of the Earth, 36:241 – 252.
Mather, P. and Koch, M. (2011). Computer Processing of Remotely-Sensed Images: An Introduction.
John Wiley & Sons.
McCabe, M., Wood, E., Wójcik, R., Pan, M., Sheffield, J., Gao, H., and Su, H. (2008). Hydrolog-
ical consistency using multi-sensor remote sensing data for water and energy cycle studies.
Remote Sensing of Environment, 112(2):430 – 444. Soil Moisture Experiments 2004 (SMEX04)
Special Issue.
McDonald, C. P., Rover, J. A., Stets, E. G., and Striegl, R. G. (2012). The regional abundance and
size distribution of lakes and reservoirs in the United States and implications for estimates
of global lake extent. Limnology and oceanography, 57(2):597 – 606.
McGuire, A. D., Anderson, L. G., Christensen, T. R., Dallimore, S., Guo, L., Hayes, D. J.,
Heimann, M., Lorenson, T. D., Macdonald, R. W., and Roulet, N. (2009). Sensitivity of the
carbon cycle in the Arctic to climate change. Ecological Monographs, 79(4):523 – 555.
McKnight, D. M., Gooseff, M. N., Vincent, W. F., and Peterson, B. J. (2008). High-latitude rivers
and streams. In Vincent, W. F. and Laybourn-Parry, J., editors, Polar Lakes and Rivers: Lim-
nology of Arctic and Antarctic Aquatic Ecosystems, pages 65–81. Oxford University Press, New
York.
Melosh, H. (2011). Planetary Surface Processes. Cambridge Planetary Science. Cambridge Uni-
versity Press.
Meybeck, M. (1995). Global distribution of lakes. In Lerman, A., Imboden, D., and Gat, J.,
editors, Physics and Chemistry of Lakes, pages 1–35. Springer Berlin Heidelberg.
169
BIBLIOGRAPHY
Monserud, R. A. and Leemans, R. (1992). Comparing global vegetaion maps with the Kappa
statistic. Ecological Modelling, 62:275 – 293.
Morgenstern, A., Ulrich, M., Günther, F., Roessler, S., Fedorova, I., Rudaya, N., Wetterich, S.,
Boike, J., and Schirrmeister, L. (2013). Evolution of thermokarst in east siberian ice-rich per-
mafrost: A case study. Geomorphology, 201(0):363 – 379.
Murton, J. B. (2009). Global Warming and Thermokarst. In Margesin, R., editor, Permafrost Soils
(Soil Biology), volume 16, chapter 13, pages 185 – 203. Springer-Verlag, Berlin Heidelberg.
Muster, S., Heim, B., Abnizova, A., and Boike, J. (2013). Water Body Distributions Across Scales:
A Remote Sensing Based Comparison of Three Arctic Tundra Wetlands. Remote Sensing,
5(4):1498 – 1523.
Muster, S., Langer, M., Heim, B., Westermann, S., and Boike, J. (2012). Subpixel heterogeneity
of ice-wedge polygonal tundra: a multi-scale analysis of land cover and evapotranspiration
in the Lena River Delta, Siberia. Tellus B, 64.
Myers-Smith, I. H., McGuire, A. D., Harden, J. W., and Chapin, F. S. (2007). Influence of dis-
turbance on carbon exchange in a permafrost collapse and adjacent burned forest. Journal of
Geophysical Research: Biogeosciences, 112(G4):G04017.
Nakao, M., Hayakawa, N., Goodrich, L., and Programme, I. H. (1998). Snow and ice science in
hydrology: the IHP training course on snow hydrology. Institute for Hydrospheric-Atmospheric
Sciences, Nagoya University.
Nelson, F. E., Anisimov, O. A., and Shiklomanov, N. I. (2001). Subsidence risk from thawing
permafrost. Nature, 410:889 – 890.
Nghiem, S. V. and Tsai, W.-Y. (2001). Global snow cover monitoring with spaceborne Ku-band
scatterometer. Geoscience and Remote Sensing, IEEE Transactions on, 39(10):2118 – 2134.
Nikolakopoulos, K. G., Kamaratakis, E. K., and Chrysoulakis, N. (2006). SRTM vs ASTER
elevation products. Comparison for two regions in Crete, Greece. International Journal of
Remote Sensing, 27(21):4819–4838.
Oliver, C. and Quegan, S. (1998). Understanding synthetic aperture radar images. Atech House,
inc., Norwood, MA.
Park, S.-E., Bartsch, A., Sabel, D., and Wagner, W. (2010). Monitoring of thawing process using
Envisat ASAR global mode data. In Geoscience and Remote Sensing Symposium (IGARSS), 2010
IEEE International, pages 2031 – 2034.
Pathe, C., Wagner, W., Sabel, D., Doubkova, M., and Basara, J. (2009). Using ENVISAT ASAR
Global Mode Data for Surface Soil Moisture Retrieval Over Oklahoma, USA. Geoscience and
Remote Sensing, IEEE Transactions on, 47(2):468 – 480.
170
BIBLIOGRAPHY
Pavlov, A. (1998). Active layer monitoring in northern West Siberia. Proceedings of the Seventh
International Conference on Permafrost, Yellowknife, 23 - 27 June, pages 1046–1068.
Perline, R. (2005). Strong, weak and false inverse power laws. Statistical Science, pages 68 – 88.
Pienitz, R., Doran, P. T., and Lamoureux, S. F. (2008). Origin and geomorphology of lakes
in polar regions. In Vincent, W. F. and Laybourn-Parry, J., editors, Polar Lakes and Rivers:
Limnology of Arctic and Antarctic Aquatic Ecosystems, pages 25 – 41. Oxford University Press,
New York.
Plug, L. J., Walls, C., and Scott, B. M. (2008). Tundra lake changes from 1978 to 2001 on the
Tktoyaktuk Peninsula, western Canadian Arctic. Geophysical Research Letters, 35:L035028.
Pope, A., Rees, W. G., Fox, A. J., and Fleming, A. (2014). Open Access Data in Polar and
Cryospheric Remote Sensing. Remote Sensing. Submitted.
Prigent, C. (2001). Remote sensing of global wetland dynamics with multiple satellite data sets.
Geophysical Research Letters, 28:4631 – 4634.
Prigent, C., Papa, F., Aires, F., Rossow, W. B., and Matthews, E. (2007). Global inundation
dynamics inferred from multiple satellite observations, 1993 - 2000. Journal of Geophysical
Research, 112:D12107.
Quinton, W., Hayashi, M., and Chasmer, L. (2009). Peatland Hydrology of Discontinuous Per-
mafrost in the Northwest Territories: Overview and Synthesis. Canadian Water Resources
Journal / Revue canadienne des ressources hydriques, 34(4):311 – 328.
Raynolds, M. K., Comiso, J. C., Walker, D. A., and Verbyla, D. (2008). Relationship between
satellite-derived land surface temperatures, arctic vegetation types, and {NDVI}. Remote
Sensing of Environment, 112(4):1884 – 1894. Remote Sensing Data Assimilation Special Issue.
Rees, W. (2001). Physical principles of remote sensing. Cambridge University Press, Cambridge,
second edition.
Rees, W. (2006). Remote Sensing of Snow and Ice. Taylor & Francis, Boca Raton.
Rees, W. (2013). Physical principles of remote sensing. Cambridge University Press, Cambridge,
third edition.
Rees, W. G., W., M., and Vitebsky, P. (2003). Mapping land cover change in a reindeer herding
area of the Russian Arctic using Landsat TM and ETM+ imagery and indigenous knowledge.
Remote Sensing of Environment, 85:441 – 452.
Rees, W. G. (2012). Assessment of ASTER global digital elevation model data for Arctic
research. Polar Record, 48:31 – 39.
171
BIBLIOGRAPHY
Reschke, J., A., B., Schlaffer, S., and Schepaschenko, D. (2012). Capability of C-Band SAR for
Operational Wetland Monitoring at High Latitudes. Remote Sensing, 4:2923 – 2943.
Richards, J. (1993). Remote sensing digital image analysis. Springer - Verlag, Berlin Heidelberg
New York, second edition.
Richards, J. and Jia, X. (2006). Remote sensing digital image analysis. Springer - Verlag, Berlin
Heidelberg, fourth edition.
Riseborough, D. (1990). Soil latent heat as a filter of the climate signal in permafrost. In Pro-
ceedings of the Fifth Canadian Permafrost Conference, Collect. Nordicana, volume 54, pages 199 –
205.
Rivkin, F. (1995). Talik zones in permafrost areas as modern natural models for global warming.
In Ecosystems research report 10: Global change and Arctic terrestrial ecosystems, pages 267 – 272.
European Commission, Brussels - Luxembourg.
Rodhe, A. and Seibert, J. (1999). Wetland occurrence in relation to topography: a test of to-
pographic indices as moisture indicators. Agricultural and Forest Meteorology, 98–99(0):325 –
340.
Romanovskiy, N. N. (1961). Erosion-thermokarst depressions in north of maritime lowlands
of Yakutiya and Novosibirsk islands. In Merzlotnye Issledovaniya, pages 124 – 144. Moscow
University Press, Moscow (In Russian).
Romanovsky, V., Drozdov, D. S., Oberman, N. G., Malkova, G. V., Kholodov, A., Marchenko,
S. S., Moskalenko, N. G., Sergeev, D., Ukraintseva, N. G., Abramov, A. A., Gilichinsky, D. A.,
and Vasiliev, A. A. (2010a). Thermal state of permafrost in Russia. Permafrost and Periglacial
Processes, 21:136 – 155.
Romanovsky, V. E., Smith, S. L., and Christiansen, H. H. (2010b). Permafrost thermal state in
the polar Northern Hemisphere during the international polar year 2007–2009: a synthesis.
Permafrost and Periglacial Processes, 21(2):106 – 116.
Rudy, A. C., Lamoureux, S. F., Treitz, P., and Collingwood, A. (2013). Identifying permafrost
slope disturbance using multi-temporal optical satellite images and change detection tech-
niques. Cold Regions Science and Technology, 88(0):37 – 49.
Sabel, D. (2007). ENVISAT ASAR Processing at IPF. Internal Document.
Sabel, D., Bartalis, Z., Wagner, W., Doubkova, M., and Klein, J.-P. (2012). Development of
a Global Backscatter Model in support to the Sentinel-1 mission design. Remote Sensing of
Environment, 120:102–112.
172
BIBLIOGRAPHY
Sabel, D., Wagner, W., Doubkova, M., Bartalis, Z., Pathe, C., Klein, J.-P., and Hornacek, M.
(2011). Sigma Nought Statistics over Land Activity - Final Report. Institute of Photogram-
metry and Remote Sensing, Vienna University of Technology, Austria. (ESA contract No.
22122/08/NL/JA (CCN-02)).
Salminen, M., Pulliainen, J., Metsämäki, S., Kontu, A., and Suokanerva, H. (2009). The be-
haviour of snow and snow-free surface reflectance in boreal forests: Implications to the per-
formance of snow covered area monitoring. Remote Sensing of Environment, 113(5):907 – 918.
Sannikov, G. S. (2012). Cartometric researches of thermokarst lakes within the Bovanenkovo
gas field territory, Yamal Peninsula. Êðèîñôåðà Çåìëè, XVI:30 – 37. In Russian.
Santoro, M., Beer, C., Cartus, O., Schmullius, C., Shvidenko, A., McCallum, I., Wegmüller, U.,
and Wiesmann, A. (2011). Rerieval of growing stock volume in boreal forest using hyper-
temporal series of Envisat ASAR ScanSAR backscatter measurements. Remote Sensing of En-
vironment, 115:490 – 507.
Santoro, M. and Strozzi, T. (2012). Circumpolar digital elevation models > 55° N with links to
geotiff images. doi:10.1594/PANGAEA.779748.
Scherer, D., Hall, D. K., Hochschild, V., König, M., Winther, J.-G., Duguay, C. R., Pivot, F.,
Mätzler, C., Rau, F., Seidel, K., et al. (2005). Remote sensing of snow cover. In Remote Sensing
in Northern Hydrology: Measuring Environmental Change, pages 7 – 38. American Geophysical
Union.
Scherler, M., Hauck, C., Hoelzle, M., Stähli, M., and Völksch, I. (2010). Meltwater infiltration
into the frozen active layer at an alpine permafrost site. Permafrost and Periglacial Processes,
21(4):325 – 334.
Schmugge, T. J., Kustas, W. P., Ritchie, J. C., Jackson, T. J., and Rango, A. (2002). Remote sensing
in hydrology. Advances in Water Resources, 25(8–12):1367 – 1385.
Schneider, C., Wunderle, S., and Friedrich, M. (1997). Snow cover investigations by means of
ground truth, modelling and ERS-SAR imagery.
Schneider, J., Grosse, G., and Wagner, D. (2009). Land cover classification of tundra environ-
ments in the Arctic Lena Delta based on Landsat 7 ETM+ data and its application for upscal-
ing of methane emissions. Remote Sensing of Environment, 113:380 – 391.
Schneider, M., Lehner, M., Müller, R., and Reinartz, P. (2008). Stereo evaluation of
ALOS/PRISM data on ESA-AO test sites - first DLR Results. In Proc. ALOS PI Symp.
Schneider von Deimling, T., Meinshausen, M., Levermann, A., Huber, V., Frieler, K., Lawrence,
D., and Brovkin, V. (2012). Estimating the near-surface permafrost-carbon feedback on global
warming. Biogeosciences, 9(2):649 – 665.
173
BIBLIOGRAPHY
Schroeder, R., Rawlins, M. A., McDonald, K. C., Podest, E., Zimmermann, R., and Kueppers, M.
(2010). Satellite microwave remote sensing of North Eurasian inundation dynamics: devel-
opment of coarse-resolution products and comparison with high-resolution synthetic aper-
ture radar data. Environmental Research Letters, 5(1):015003.
Schuur, E. A., Bockheim, J., Canadell, J. G., Euskirchen, E., Field, C. B., Goryachkin, S. V.,
Hagemann, S., Kuhry, P., Lafleur, P. M., Lee, H., et al. (2008). Vulnerability of permafrost
carbon to climate change: Implications for the global carbon cycle. BioScience, 58(8):701 –
714.
Schuur, E. A. G., Vogel, J. G., Crummer, K. G., Lee, H., Sickman, J. O., and Osterkamp, T. E.
(2009). The effect of permafrost thaw on old carbon release and net carbon exchange from
tundra. Nature, 459:556 – 559.
Seekell, D. A. and Pace, M. L. (2011). Does the Pareto distribution adequately describe the
size-distribution of lakes? Limnol. Oceanogr., 56(1):350 – 356.
Seekell, D. A., Pace, M. L., Tranvik, L. J., and Verpoorter, C. (2013). A fractal-based approach to
lake size distributions. Geophysical Research Letters, 40:517 – 521.
Senthilnath, J., Shenoy, H. V., Rajendra, R., Omkar, S. N., Mani, V., and Diwakar, P. G. (2013).
Integration of speckle de-noising and image segmentation using Synthetic Aperture Radar
image for flood extent extraction. Journal of Earth System Sciences, 122(3):559 – 572.
Seppälä, M. (2004). Winds as a Geomorphic Agent in Cold Climates. Cambridge University Press,
Cambridge.
Shi, Z. and Fung, K. B. (1994). A Comparison of Digital Speckle Filters. Proceedings of IGRASS
94, August 8 - 12, 1994, pages 2129 – 2133.
Sivia, D. S. and Skilling, J. (2006). Data Analysis - A Bayesian Tutorial. Oxford University Press,
Oxford.
Smith, L. C., Sheng, Y., and MacDonald, G. M. (2007). A First Pan-Arctic Assessment of the
Influence of Glaciation, Permafrost, Topography and Peatlands on Northern Hemisphere
Lake Distribution. Permafrost and Periglacial Processes, 18:201 – 208.
Smith, L. C., Sheng, Y., MacDonald, G. M., and Hinzmann, L. D. (2005). Disappearing Arctic
Lakes. Science, 308:1429.
Smith, M. W. and Riseborough, D. W. (2002). Climate and the limits of permafrost: a zonal
analysis. Permafrost and Periglacial Processes, 13(1):1 – 15.
Smithson, P., Addison, K., and Atkinson, K. (2008). Fundamentals of the Physical Environment.
Routledge, New York, 4th edition.
174
BIBLIOGRAPHY
Solberg, R. and Andersen, T. (1994). An automatic system for operational snow-cover monitor-
ing in the Norwegian mountain regions. In Geoscience and Remote Sensing Symposium, 1994.
IGARSS’94. Surface and Atmospheric Remote Sensing: Technologies, Data Analysis and Interpreta-
tion., International, volume 4, pages 2084 – 2086. IEEE.
Solomeshch, A. (2005). The World’s Largest Wetlands. chapter 2, The West Siberian Lowlands.
Cambridge University Press, Cambridge.
Sørensen, R. and Seibert, J. (2007). Effects of DEM resolution on the calculation of topographical
indices: TWI and its components. Journal of Hydrology, 347(1):79 – 89.
Sposito, G. (2008). Scale Dependence and Scale Invariance in Hydrology. Cambridge University
Press.
Stephani, E., Fortier, D., Shur, Y., Fortier, R., and Doré, G. (2014). A geosystems approach to
permafrost investigations for engineering applications, an example from a road stabilization
experiment, beaver creek, yukon, canada. Cold Regions Science and Technology, 100(0):20 – 35.
Stichbury, G., Brabyn, L., Green, T., and Cary, C. (2011). Spatial modelling of wetness for the
Antarctic Dry Valleys. Polar Research, 30(0).
Stolbovoi, V. and McCallum, I. (2002). Land Resources of Russia. CD-ROM, Insternational
Institute for Applied Systems Analysis and the Russian Academy of Science, Laxenburg,
Austria.
Stow, D. A., Hope, A., McGuire, D., Verbyla, D., Gamon, J., Huemmrich, F., Houston, S., Racine,
C., Sturm, M., Tape, K., Hinzman, L., Yoshikawa, K., Tweedie, C., Noyle, B., Silapaswan, C.,
Duglas, D., Griffith, B., Jia, G., Epstein, H., Walker, K., Daeschner, S., Petersen, A., Zhou, L.,
and Myneni, R. (2004). Remote sensing of vegetation and land-cover change in Arctic Tundra
Ecosystems. Remote Sensing of Environment, 89:281 – 308.
Streletskiy, D., Streletskaya, I., Rogov, V., and Leybman, M. (2003). Redistribution of ions within
the active layer and upper permafrost, Yamal, Russia. In Proceedings of the Eighth International
Conference on Permafrost, volume 21, page 25.
Ström, L. and Christensen, T. R. (2007). Below ground carbon turnover and greenhouse gas
exchanges in a sub-arctic wetland. Soil Biology & Biochemistry, 39:1689 – 1698.
Tedesco, M. and Miller, J. (2007). Observations and statistical analysis of combined active-
passive microwave space-borne data and snow depth at large spatial scales. Remote Sensing
of Environment, 111:382 – 397.
Telmer, K. H. and Costa, M. P. (2007). SAR-based estimates of the size distribution of lakes in
Brazil and Canada: a tool for investigating carbon in lakes. Aquatic Conservation: Marine and
Freshwater Ecosystems, 17(3):289 – 304.
175
BIBLIOGRAPHY
Temimi, M., Leconte, R., Chaouch, N., Sukumal, P., Khanbilvardi, R., and Brissette, F. (2010). A
combination of remote sensing data and topographic attributes for the spatial and temporal
monitoring of soil wetness. Journal of hydrology, 388(1):28–40.
Töyrä, J., Pietroniro, A., and Martz, L. W. (2001). Multisensor Hydrologic Assessment of a
Freshwater Wetland. Remote Sensing of Environment, 75:162 – 173.
Trofaier, A. M. (2010). Thermokarst thaw lakes in Polar regions - Evaluating the need for pan-
Arctic monitoring of periglacial environments. Unpublished MSc thesis, Vienna University
of Technology.
Trofaier, A. M., Bartsch, A., Rees, W. G., and Leibman, M. O. (2013a). Assessment of spring
floods and surface water extent over the Yamalo-Nenets Autonomous District. Environmental
Research Letters, 8(4):045026.
Trofaier, A. M., Rees, W. G., Bartsch, A., and Leibman, M. O. (2013b). Arctic water body dy-
namics: a pan-Siberian study of seasonal changes in inundation - stage 1: The Yamalo-Nenets
Autonomous District. Proceedings of the ESA Living Planet Sympsium, Edinburgh.
Trofaier, A. M., Rees, W. G., Bartsch, A., Sabel, D., and Schlaffer, S. (2012). Feasibility Study of
Using Active Microwave Data for Examination of Thaw Lake Drainage Patterns over the Ya-
mal Peninsula. Proceedings of the Tenth International Conference on Permafrost, Salekhard, 1:413–
418.
Tucker, G. E., Catani, F., Rinaldo, A., and Bras, R. L. (2001). Statistical analysis of drainage
density from digital terrain data. Geomorphology, 36(3–4):187 – 202.
Tyrtikov, A. (1964). Perennially frozen ground and vegetation. In Principles of Geocryology,
pages 399 – 421.
Ulaby, F. and Batlivala, P. (1976). Optimum Radar Parameters for Mapping Soil Moisture.
Geoscience Electronics, IEEE Transactions on, 14(2):81 – 93.
Ulaby, F., Moore, R., and Fung, A. (1981). Microwave Remote Sensing: From theory to applications.
Remote sensing. Addison-Wesley Publishing Company, Advanced Book Program/World
Science Division.
Ulaby, F. T., Batlivala, P. P., and Dobson, M. C. (1978). Microwave backscatter dependence on
surface roughness, soil moisture, and soil texture: Part I-Bare soil. Geoscience Electronics, IEEE
Transactions on, 16(4):286 – 295.
Ulrich, M., Grosse, G., Chabrillat, S., and Schirrmeister, L. (2009). Spectral characterization of
periglacial surfaces and geomorphological units in the arctic lena delta using field spectrom-
etry and remote sensing. Remote Sensing of Environment, 113(6):1220 – 1235.
United Nations Environment Programme (2007). Global Outlook for Ice and Snow.
176
BIBLIOGRAPHY
van Everdingen, R. (1990). Ground-Water Hydrology. In Prowse, T. and Ommanney, C., edi-
tors, Northern Hydrology: Canadian Perspectives, pages 77 – 102. National Hydrology Research
Institute, Saskatoon.
van Oort, P. (2007). Interpreting the change detection error matrix. Remote Sensing of Environ-
ment, 108(1):1 – 8.
Wagner, W., Blöschl, G., Pampaloni, P., Calvet, J.-C., Bizzarri, B., Wigneron, J.-P., and Kerr, Y.
(2007). Operational readiness of microwave remote sensing of soil moisture for hydrologic
applications. Nordic Hydrology, 38(1):1 – 20.
Wagner, W., Hahn, S., Kidd, R., Melzer, T., Bartalis, Z., Hasenauer, S., Figa-Saldaña, J., de Ros-
nay, P., Jann, A., Schneider, S., Komma, J., Kubu, G., Brugger, K., Aubrecht, C., Züger, J.,
Gangkofner, U., Kienberger, S., Brocca, L., Wang, Y., Blöschl, G., Eitzinger, J., Steinnocher, K.,
Zeil, P., and Rubel, F. (2013). The ascat soil moisture product: A review of its specifications,
validation results, and emerging applications. Meteorologische Zeitschrift, 22(1):5 – 33.
Wagner, W., Pathe, C., Doubkova, M., Sabel, D., Bartsch, A., Hasenauer, S., Blöschl, G., Scipal,
K., Mart´inez-Ferna´ndez, J., and Löw, A. (2008). Temporal Stability of Soil Moisture and Radar
Backscatter Observed by the Advanced Synthetic Aperture Radar (ASAR). Sensors, 8:1174 –
1197.
Wagner, W., Scipal, K., Pathe, C., Gerten, D., Lucht, W., and Rudolf, B. (2003). Evaluation of
the agreement between the first global remotely sensed soil moisture data with model and
precipitation data. Journal of Geophysical Research: Atmospheres, 108(D19):4611.
Walker, D., Forbes, B., Leibman, M., Epstein, H., Bhatt, U., Comiso, J., Drozdov, D., Gubarkov,
A., Jia, G., Kaarlejärvi, E., Kaplan, J., Khomutov, A., Kofinas, G., Kumpula, T., Kuss, P.,
Moskalenko, G., Meschtyb, N., Pajunen, A., Raynolds, M., Romanovsky, V., Stammler, F.,
and Yu, Q. (2011). Eurasian Arctic Land Cover and Land Use in a Changing Climate. chapter
9, Cumulative Effects of Rapid Land-Cover and Land-Use Changes on the Yamal Peninsula,
Russia. Springer, Dordrecht Heidelberg London New York.
Walker, D. A., Bhatt, U. S., Raynolds, M. K., Comiso, J. C., Epstein, H. E., and Jia, G. J. (2009a).
The Arctic: land: vegetation [in “State of the Climate in 2008”]. volume 90, pages 104 – 105.
Walker, D. A., Leibman, M. O., Epstein, H. E., Forbes, B. C., Bhatt, U. S., Raynolds, M. K.,
Comiso, J. C., Gubarkov, A. A., Khomutov, A. V., Jia, G. J., Kaarlejärvi, E., Kaplan, J. O.,
Kumpula, T., Kuss, P., Matyshak, G., Moskalenko, N. G., Orekhov, P., Romanovsky, V. E.,
Ukraientseva, N. G., and Yu, Q. (2009b). Spatial and temporal patterns of greenness on the
Yamal Peninsula, Russia: interactions of ecological and social factors affecting the Arctic
normalized difference vegetation index. Environmental Research Letters, 4(4):1–16.
177
BIBLIOGRAPHY
Walter, K. M., Edwards, M. E., Grosse, G., Zimov, S. A., and Chapin III, F. S. (2007). Thermokarst
Lakes as a Source of Atmospheric CH4 during the Last Deglaciation. Science, 318(5850):633 –
636.
Walter, K. M., Zimov, S. A., Chanton, J. P., Verbyla, D., and Chapin III, F. S. (2006). Methane bub-
bling from Siberian thaw lakes as a positive feedback to climate warming. Nature, 443(7):71
– 75.
Walvoord, M. A., Voss, C. I., and Wellman, T. P. (2012). Influence of permafrost distribution on
groundwater flow in the context of climate-driven permafrost thaw: Example from Yukon
Flats Basin, Alaska, United States. Water Resources Research, 48(7):W07524.
Wang, G., Hu, H., and Li, T. (2009). The influence of freeze - thaw cycles of active soil layer on
surface runoff in a permafrost watershed. Journal of hydrology, 375(3):438 – 449.
Wang, L., Derksen, C., and Brown, R. (2008). Detection of pan-Arctic terrestrial snowmelt from
QuikSCAT, 2000 - 2005. Remote Sensing of Environment, 112:3794 – 3805.
Watts, J., Kimball, J., Jones, L., Schroeder, R., and McDonald, K. (2012). Satellite Microwave
remote sensing of contrasting surface water inundation changes within the Arctic–Boreal
Region. Remote Sensing of Environment, 127:223 – 236.
Weise, O. R. (1983). Das Periglazial. Gebrüder Bontraeger, Berlin - Stuttgart.
West, J. and Plug, L. J. (2008). Time-dependent morphology of thaw lakes and taliks in deep
and shallow ground ice. Journal of Geophysical Research: Earth Surface (2003–2012), 113:F01009.
Wetzel, R. G. (1990). Congress in munich 1989. volume 24, chapter Land-water interfaces:
Metabolic and limnological regulators, pages 6 – 24. Schweizerbart’sche Verlagsbuchhand-
lung, Stuttgart.
Wise, S. (2011). Cross-validation as a means of investigating dem interpolation error. Computers
& Geosciences, 37(8):978 – 991.
Woo, M. (1990). Permafrost Hydrology. In Prowse, T. and Ommanney, C., editors, North-
ern Hydrology: Canadian Perspectives, pages 63 – 76. National Hydrology Research Institute,
Saskatoon.
Woo, M.-k., Arain, M. A., Mollinga, M., and Yi, S. (2004). A two-directional freeze and
thaw algorithm for hydrologic and land surface modelling. Geophysical Research Letters,
31(12):L12501.
Woo, M.-K., Kane, D. L., Carey, S. K., and Yang, D. (2008). Progress in permafrost hydrology in
the new millennium. Permafrost and Periglacial Processes, 19(2):237 – 254.
178
BIBLIOGRAPHY
Woodhouse, I. H. (2006). Introduction to Microwave Remote Sensing. CRC Press, Taylor & Francis
Group, Boca Raton, Florida.
Yoshikawa, K. and Hinzman, L. D. (2003). Shrinking Thermokarst Ponds and Groundwater
Dynamics in Discontinuous Permafrost near Council, Alaska. Permafrost and Periglacial Pro-
cesses, 14:151 – 160.
Zakharova, E. A., Kouraev, A. V., Biancamaria, S., Kolmakova, M. V., Mognard, N., Zemtsov,
V. A., Kirpotin, S. N., and Decharme, B. (2011). Snow Cover and Spring Flood Flow in the
Northern Part of Western Siberia (the Poluy, Nadym, Pur and Taz Rivers). Journal of Hydrom-
eteorology, 12:1498 – 1511.
Zakharova, E. A., Kouraev, A. V., Kolmakova, M. V., Mognard, N. M., Zemtsov, V. A., and
Kirpotin, S. N. (2009). The modern hydrological regime of the northern part of Western
Siberia from in situ and satellite observations). International Journal of Environmental Studies,
66:447 – 463.
Zhou, W. and Huang, S. L. (2004). Modeling impacts of thaw lakes to ground thermal regime
in northern Alaska. J. Cold Reg. Eng., 18(2).
Zimov, S., Vorapaev, Y. V., Semiletov, I. P., Davidov, S. P., Prosianikov, S. F., Chapin III, F. S.,
Chapin, M. C., Trumbore, S., and Tyler, S. (1997). North Siberian Lakes: A Methane Source
Fueled by Pleistocene Carbon. Science, 277:800 – 801.
179