Show simple item record

dc.contributor.authorRogers, Martin
dc.date.accessioned2022-05-30T08:52:25Z
dc.date.available2022-05-30T08:52:25Z
dc.date.submitted2020-11-01
dc.identifier.urihttps://www.repository.cam.ac.uk/handle/1810/337589
dc.description.abstractCoastal communities and land covers are vulnerable receptors of erosion, flooding, or both in combination. The accurate, automated, and wide-scale determination of shoreline position, and its migration at the engineering scale (10-1 – 102 km), is imperative for future coastal risk adaptation and management. The recent increase in the acquisition and availability of Big Datasets, including multispectral remote sensing imagery, is providing new opportunities to monitor engineering scale rates of shoreline change and other constituents of coastal risk, including changes to human coastal population densities. This increase in data availability comes with novel challenges to devise and utilise methods to store, process, analyse and extract information from these Big Datasets. This thesis assesses the suitability of different Big Data approaches, namely Machine Learning (ML) and non-ML based tools, for the automated extraction of the coastal vegetation edge in remote sensing imagery. Compared to the instantaneous waterline, few vegetation edge methods have been developed and analysis of the coastal zone processes that can be detected using the shoreline proxy remain understudied. This thesis initially investigates whether non-ML methods are suitable for the extraction of the coastal vegetation edge from multispectral remote sensing imagery. A novel non-ML tool is introduced and applied, CoasTool, which considers the proximity of the instantaneous water line during vegetation edge extraction. CoasTool performance is compared to the outputs derived from well-established threshold contouring techniques and kernel-based methods as well as one form of ML, Support Vector Machines (SVM). Limitations in the performance of these tools, particularly along shorelines with discontinuous or graded vegetation boundaries, provide justification for the application of a separate form of ML, convolutional neural network (CNN), to this task. A novel CNN, VEdge_Detector, is trained and applied to extract the coastal vegetation edge and its outputs are compared to ground-referenced measurements and manually digitised vertical aerial photographs. VEdge_Detector is applied to a time series of vi images to detect annual to decadal scale shoreline dynamics discernible using the coastal vegetation edge. Shoreline change constitutes one element of coastal risk, and this thesis subsequently investigates the viability of integrating multiple ML-derived datasets, pertaining to different aspects of risk, to calculate relative coastal population exposure to shoreline change. The Guiana coastline, northern South America, is one of the most dynamic stretches of coastline in the world and a region where greater than 90% of its population live below 10 m elevation. The identification of locations where coastal populations are at greatest risk to coastal retreat in this region is thus very important to inform coastal risk management decisions. Accordingly, decadal-scale rates of shoreline change calculated using VEdge_Detector derived shoreline positions are combined with secondary, ML-derived, population datasets (WorldPop). The integration of the two ML-based datasets aids the identification of population exposure hotspot locations and discover, previously unpublished, locations where forced migration due to shoreline change has occurred. In concluding, the relative merits and drawbacks of using ML verses non-ML techniques to detect the coastal vegetation edge are discussed as well as considering the suitability of the coastal vegetation as a proxy of shoreline position. Further discussion is given on the different considerations coastal stakeholders will have when choosing the most suitable tool to use in shoreline detection tasks, including tool performance, speed, transparency, and ease of use. Remaining research gaps and future research requirements are emphasised, including the need for collaboration between different research institutions to suitably train and apply ML tools in the geosciences.
dc.description.sponsorshipUKRI NERC/ESRC Data, Risk and Environmental Analytical Methods (DREAM) Centre for Doctoral Training, Grant/Award Numbers: NE/M009009/1, NE/R011265/1
dc.rightsAll Rights Reserved
dc.rights.urihttps://www.rioxx.net/licenses/all-rights-reserved/
dc.subjectMachine learning
dc.subjectRemote sensing
dc.subjectShoreline change
dc.subjectCoastal vegetation
dc.subjectConvolutional Neural Networks
dc.titleMachine Learning and remote sensing applications to shoreline dynamics
dc.typeThesis
dc.type.qualificationlevelDoctoral
dc.type.qualificationnameDoctor of Philosophy (PhD)
dc.publisher.institutionUniversity of Cambridge
dc.date.updated2022-05-26T08:41:55Z
dc.identifier.doi10.17863/CAM.84999
rioxxterms.licenseref.urihttps://www.rioxx.net/licenses/all-rights-reserved/
rioxxterms.typeThesis
pubs.funder-project-idNatural Environment Research Council (2106394)
cam.supervisorBithell, Mike
cam.supervisorSpencer, Tom
cam.depositDate2022-05-26
pubs.licence-identifierapollo-deposit-licence-2-1
pubs.licence-display-nameApollo Repository Deposit Licence Agreement


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record