Edited by: Joana R. Xavier, University of Porto, Portugal
Reviewed by: Michael Vecchione, National Oceanic and Atmospheric Administration (NOAA), United States; Jackson Chu, University of Victoria, Canada
This article was submitted to Deep-Sea Environments and Ecology, a section of the journal Frontiers in Marine Science
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
The ecosystem dynamics of benthic communities depend on the relative importance of organism reproductive traits, environmental factors, inter-specific interactions, and mortality processes. The fine-scale community ecology of sessile organisms can be investigated using spatial analyses because the position of the specimens on the substrate (their spatial positions) reflects the biological and ecological processes that they were subject to in-life. Consequently, spatial point process analyses (SPPA) and Bayesian network inference (BNI) can be used to reveal key insights into the ecological dynamics of these deep-sea communities. Here we use these analyses to investigate the ecology of deep-sea glass sponge dominated community “The Forest of the Weird” (2,442 m depth, Ridge Seamount, Johnston Atoll, Pacific Ocean). A 3D reconstruction was made of this community using photogrammetry of video stills taken from high-resolution ROV video. The community was dominated by two genera of Hexactinellids: Farreidae
Sponge and coral dominated deep-sea communities are complex habitats, providing habitat, and refuge for other benthos creating biodiversity hotspots in the deep ocean (
Ecological analyses of sponge populations and communities have mostly focused on specific ecological processes, such as the importance of episodic recruitment events (
SPPA capitalizes on the spatial position of organisms on the substrate because there are only four different sets of processes which can influence these positions: (1) interactions with environment, (2) dispersal limitation, (3) interactions such as facilitation and competition within and between taxa populations, and (4) density-dependent mortality processes (
When considering interactions between pairs of taxa, auto-correlation from chains of interactions need to be eliminated to ensure only causal relationships are reported. For example, if taxon A directly interacts with taxon B and B interacts directly with taxon C, we could expect a non-random spatial distribution between A and C. However, this non-random distribution would not be the result of a direct interaction, but would just be a reflection of the two interactions between A and B and B and C. One approach to finding only realized dependencies between taxa is using BI) to reconstruct the ecological network of the community (
As part of the National Oceanic and Atmospheric Administration CAPSTONE field campaign, throughout 2015–2017 a series of expeditions was conducted to collect data from the previously unexplored deep-water habitats within the Pacific Remote Islands Marine National Monument (PRIMNM) region in the Pacific Ocean (
In this study we have focused on the “Forest of the Weird” (FW) community, which was recorded on Dive 11 of Okeanos Explorer expedition EX1706 of the Johnston Atoll Unit in PRIMNM (
Different taxa identified within this study. Red laser dots are 10 cm apart.
Screenshots of the 3D reconstruction.
Summary table.
194 | 0.33 | 1.73 | 102 | 0.38 | 3.48 | 91 | 0.29 | 1.27 | |
68 | 0.11 | 0.61 | 27 | 0.10 | 0.92 | 41 | 0.13 | 0.57 | |
7 | 0.01 | 0.06 | 2 | 0.01 | 0.07 | 5 | 0.02 | 0.07 | |
17 | 0.03 | 0.15 | 5 | 0.02 | 0.17 | 12 | 0.04 | 0.17 | |
Holdfast disc | 44 | 0.07 | 0.39 | 15 | 0.06 | 0.51 | 29 | 0.09 | 0.41 |
Holdfast disc with stalk | 101 | 0.17 | 0.90 | 47 | 0.17 | 1.60 | 54 | 0.17 | 0.76 |
33 | 0.06 | 0.29 | 14 | 0.05 | 0.48 | 19 | 0.06 | 0.27 | |
16 | 0.03 | 0.14 | 8 | 0.03 | 0.27 | 7 | 0.02 | 0.10 | |
Other corals | 18 | 0.03 | 0.16 | 9 | 0.03 | 0.31 | 10 | 0.03 | 0.14 |
Other sponges | 8 | 0.01 | 0.07 | 5 | 0.02 | 0.17 | 3 | 0.01 | 0.04 |
23 | 0.04 | 0.21 | 8 | 0.03 | 0.27 | 15 | 0.05 | 0.21 | |
Small undetermined. | 63 | 0.11 | 0.56 | 30 | 0.11 | 1.02 | 33 | 0.10 | 0.46 |
All |
269 | 0.45 | 2.40 | 131 | 0.48 | 4.47 | 137 | 0.43 | 1.92 |
All |
118 | 0.20 | 1.05 | 52 | 0.19 | 1.77 | 66 | 0.21 | 0.92 |
With the advent of high-resolution ROV cameras and high-powered computers, it is now possible to reconstruct entire benthic communities using photogrammetry to centimeter or millimeter scale (
The video recorded from the ROV was 720p five mega-bit per second resolution (
The start of the FW community was taken to be from the ROV track starting at 14°28′24.156″ N, 170°51′18.231″ W until 14°28′24.198″ N, 170°51′17.553″ W.
The edges of the community were taken to be the edges of the ridge. There were organisms outside this region, but the reconstruction did not have sufficient tie point matches to reconstruct the region accurately.
This segment of video was sampled at a capture rate of 1 frame/second using ffmpeg—an open source video encoder/decoder.
Out of focus and shaky frames were removed, as were frames that were close up of individual specimens and/or had other objects obstructing the view in the frame.
The files were then imported into Agisoft Metashape.
Chunks of 100 frames were auto-aligned to form a sparse point cloud.
These chunks were then aligned into the complete community which was formed from 583 out of a total 600 input frames and 131,322 tie points.
A dense cloud (6 million points) and mesh (4.398 million faces) was created from the combined sparse point cloud.
The frames were matched to the ROV GPS tracks by using the epoch timestamp in the GPS log and matching it (using a “join query” in Access) to the image frames, which had been renamed to match. Distances were checked using photographs where 2 red laser dots (10 cm apart) to correctly scale the reconstruction.
Individual organisms were marked up and identified on the frames within the 3D reconstruction. Note that the resolution of the videos was less than that of the high-resolution photographs taken which hinders resolution for smaller specimens and also meant that very small (∼<2 cm) specimens may not have been marked.
Specimen positions were checked from multiple viewpoints.
Screen shots of the 3D reconstruction show both the strengths and weakness of the reconstruction (
In order to get the 2D map for our analyses, the 3D reconstruction was exported into Geomagic Wrap, and a best-fit plane was fit to the Ridge and Ridge Crest areas, respectively. The 2D maps for each area were created by rotating the specimen positions to the best-fit planes for each section separately, which were then exported then rejoined as a single 2D projection (see SI for data).
Four different types of spatial analyses were performed on the data: (1) Random labeling analyses were used to investigate the mortality dynamics within the community. (2) Univariate SPPA were used to investigate the population ecology of each taxon. (3) BNI was used to identify primary dependencies between taxa. (4) Bivariate SPPA was used to determine the most likely underlying processes to the primary dependencies found using BNI.
The simplest scenario within SPPA analyses is that all the points (here different organisms) are randomly distributed within the study area. This random distribution is known as complete spatial randomness (CSR), and can be modeled as an homogeneous Poisson model (
Density-dependent mortality processes are best investigated using a subset of SPPA called Random Labeling Analyses (RLA) (
To investigate mortality processes, we investigated how a state of an organism (dead or alive in this study) changes within organism locations, using RLAs (
Spatial distributions are commonly described using PCFs which describe how the density of points (i.e., sponge specimens) changes as a function of distance from the average specimen (e.g.,
The Quotient test calculates the bivariate PCF between groups relative to the pattern of both groups taken together (the joined pattern), where PCF12 is the bivariate distribution of group 2 relative to group 1 and PCF 21 is the bivariate distribution of group 1 relative to group 2. For joint patterns, PCF 1,1+2 is the bivariate distribution of group 1 relative to both groups together, and PCF 2,2+1 is the bivariate distribution of group 2 relative to the joint pattern. Thus, the Quotient test is the calculation of the distribution: PCF 1,1+2 - PCF 21/PCF 2,2+1 where PCF 12/PCF 1,1+2 - PCF 21/PCF 2,2+1 > 0 indicates that group 2 is mainly located in areas with high density of the joint pattern, and group 1 is in low density areas (i.e., group 2 has more neighbors than group 1). If this quotient is significantly non-zero, then the process underlying the characters is density-dependent; for example, dead specimens occur more commonly in high-density areas, indicating density-dependent mortality.
We test three null hypotheses of mortality spread using RLA of dead/alive state with the organism positions:
H0Comm: The spatial distribution of dead specimens are randomly distributed within the living community of corals and sponges.
H0Asp: The spatial distribution of dead specimens of
H0Bol
Establishing whether the null hypotheses of the Difference and Quotient Tests should be rejected or not is complicated, because there is a lack of independence of the spatial points (organism positions) and a variety of different point pattern distributions (
The following RLAs were conducted using Programita software across three different areas: All mapped area, the Ridge and the Ridge Crest (
To test H0Comm all living specimens of both coral and sponge taxon were contained in the alive group and, and all brown specimens (including the dead
To test H0Asp the white specimens of
To test H0Bol the holdfast brown discs with long stalks were assumed to be the dead
Each hypothesis was tested by running 999 Monte Carlo simulations for each group in order to generate simulation envelopes around the random PCF value (i.e., PCF 11− PCF 22 = 0).
Initial data exploration and data visualization were performed in R (
The following analyses were performed on each abundant taxon (
To test whether a taxon PCF exhibited CSR, 999 Monte Carlo simulations were run on a homogeneous background and the simulation envelopes chosen to be the 49th highest and lowest values (
If a taxon was not randomly distributed on a homogeneous background, and was aggregated, the random model on a heterogeneous background (HP model) was tested by creating a heterogeneous background from the density map of the taxon under consideration, being defined by a circle of radius R over which the density is averaged throughout the sample area. Density maps were formed using estimators in 0.10 m increments over the range of 0.1 m <
The PCF and L functions (
Best-fit Thomas cluster processes (
If the model did not describe the observed data well, the lines were refitted using just the PCF. If that fit was also poor, then only the L-function was used.
999 simulations of this model were generated to create the simulation envelope. The 49th highest and lowest simulation values were chosen in line with previous work to be the limits of the simulation envelopes (
If there were no excursions outside the simulation envelope and the
The best-fit TC model was simulated on the best fit HP model following points 4–5 to simulate a Thomas cluster model on a heterogeneous background (ITC).
The Bayesian network inference (BNI) algorithm used in this study requires discrete data which ensures data noise is masked and only the relative densities of each taxon are important (
The BNI software used was Banjo v2.0.0, a publicly available Java-based algorithm (
To convert the spatial positions of specimens into discretized data suitable for BNI analyses the following steps were taken (following (
The discretized data for the BNI analyses are given in
Discretized data for BNI.
Zero | 8 | 10 | 24 | 32 | 20 |
Low | 18 | 21 | 13 | 6 | 15 |
High | 15 | 10 | 4 | 3 | 6 |
The direction of the edge between nodes in the network indicates which node (taxon) has a dependency on the other node (taxon); this direction is indicated in the network by an arrowhead. For each edge, the directionality was taken to be the direction which occurred in the majority of bootstrapped networks. Where there was no majority (directional edges have a probability between 0.35 and 0.65) an edge was said to have bi-directionality, or mutual dependency was indicated; these are shown without arrows.
The influence score (IS) can be used to gauge the type and strength of the interaction between two nodes. Positive dependencies have an IS > 0: that is, a high density of taxon 1 corresponds to a high density of taxon 2, and so the dependency arrow would point from taxon 2 to taxon 1. Negative dependencies have an IS < 0: a high density of taxon 1 corresponds to a low density of taxon 2. Where the dependency is non-monotonic the IS = 0 so that nature of the dependency changes with taxon abundance. The mean IS for each edge was calculated for each site.
For every dependency found between taxa by BNI, bivariate SPPA were used to infer the most likely underlying processes (cf.
Bivariate PCFs were calculated from the population density using a grid of 10 cm × 10 cm. To minimize noise, smoothing was applied to the PCF dependent on specimen abundance. The amount of smoothing required depends on the number of points in each taxon’s population, with smaller sample sizes requiring more smoothing to ensure that outliers do not overly change the distribution (
For each taxon pair which displayed aggregation (bivariate PCF > 1), CSR, HP, Linked clusters (LC) and shared parent (SP) models were fitted to the data. For segregated bivariate distributions (bivariate PCF < 1), CSR, HP, hard and soft core models (HC), and hard/soft core on heterogeneous background (HCHP) models were fitted as follows. The magnitude of the PCF reflects the intensity of underlying biotic and abiotic processes: two taxon populations with a PCF = 4, for example, are four times more aggregated than if they exhibited CSR; thus, the relative magnitudes of the PCFs can be used to compare relative strengths of interactions and associations.
If a taxon was aggregated, the random model on a heterogeneous background (HP model) was tested. Creation of the heterogeneous background was as for the univariate distribution, but instead of a single taxon, the density map of the joint distribution of the two taxon was used.
For each taxon pair, two best-fit Linked cluster models (LC) were found. First, Taxon 1 was kept constant and Taxon 2 was modeled as a Thomas Cluster aggregation around Taxon 1 (following the same procedure as fitting univariate TC models). Then Taxon 2 was kept constant and Taxon 1 was modeled as an aggregation. The shared parents models (SP) modeled both taxa distributions as Thomas Cluster models around randomly distributed shared points, with the respective Thomas Cluster models fitted as per the univariate models.
Segregated distributions were modeled on both homogeneous background and on heterogeneous background. Three different models were determined: Segregation around Taxon 1, around Taxon 2 and around both Taxon 1 and Taxon 2. For each of these the radius was increased from 0.10 m to 1.00 m in 0.10 m increments. The severity of the segregation was modeled in two ways: hard and soft-core. For hard-core models (corresponds to
The mapped community consisted of 592 specimens over 100.7 m2 of substrate. The community was dominated (69.7%) by two genera of Hexactinellids: Farreidae
Spatial maps.
Analyses to detect edge effects (
Random labeling analyses.
H0A dv | All | 0.991 | 0.915 | 0.911 | 0.301 | None | 0.892 |
H0A dv | Ridge Crest | 0.751 | 0.811 | 0.346 | 0.870 | None | 0.205 |
H0A dv | Ridge | 0.265 | 0.412 | 0.387 | 0.278 | None | 0.260 |
H0C omm | All | 0.411 | 0.436 | 0.672 | 0.790 | None | 0.363 |
H0C omm | Ridge Crest | 0.833 | 0.617 | 0.885 | 0.208 | None | 0.822 |
H0C omm | Ridge | 0.498 | 0.158 | 0.726 | 0.982 | None | 0.192 |
H0A sp | All | 0.618 | 0.873 | 0.429 | |||
H0A sp | Ridge Crest | 0.862 | 0.997 | 0.112 | |||
H0A sp | Ridge | 0.546 | 0.826 | 0.221 | 0.755 | None | 0.240 |
Within the whole community there was no evidence of significant deviations from zero for any of the RLAs (
The RLAs of the
Summary table of univariate PCF analyses.
A | No | 0.001 | 0.395 | 0.23 | NA | 20 | 18.694 | 0.104 | 37 | NA | NA | ||
RC | No | 0.002 | 0.227 | 0.25 | NA | 10 | 19.048 | 0.042 | 38 | NA | NA | ||
R | 1.2–2.1 m | 0.001 | 0.505 | 0.158 | NA | 10 | 19.048 | 0.042 | 38 | NA | NA | ||
A | No | 0.704 | 0.001 | 0.001 | NA | 50 | 2.262 | 4.798 | 5 | NA | NA | ||
A | No | 0.012 | 0.232 | 0.211 | NA | 20 | 8.737 | 0.135 | 17 | NA | NA | ||
RC | >1 m | 0.004 | 0.003 | 0.245 | NA | 20 | 8.737 | 0.135 | 17 | NA | NA | ||
R | No | 0.27 | 0.498 | 0.249 | NA | 30 | 8.737 | 0.135 | 17 | NA | NA | ||
A | No | 0.001 | 0.253 | 0.352 | NA | 20 | 5.351 | 0.555 | 11 | NA | NA | ||
RC | >3 m | 0.021 | 0.093 | 0.102 | NA | 20 | 7.174 | 0.468 | 14 | NA | NA | ||
R | >1.3 m | 0.007 | 0.201 | 0.154 | NA | 20 | 5.351 | 0.555 | 11 | NA | NA | ||
A | No | 0.032 | 0.772 | 0.43 | NA | 20 | 5.499 | 1.097 | 11 | NA | NA | ||
RC | No | 0.001 | 0.761 | 0.632 | 0.316 | NA | 20 | 5.499 | 1.097 | 11 | NA | NA | |
R | <1.5 m | 0.117 | 0.117 | 0.093 | 0.066 | 20 | 5.499 | 1.097 | 11 | 5 | 0 |
The Bayesian network inference found three significant dependencies between four of the five abundant taxa (
The bivariate distribution for
Summary table of bivariate PCF analyses.
A | 0.013 | 0.2 | 0.026 | 0.409 | NA | NA | 0.105 | 0.164 | 5.217 | 0.2951 | 10 | 50 | NA | NA | NA | NA | NA | |||
R | 0.128 | 0.2 | 0.001 | 0.688 | NA | NA | 0.131 | 0.319 | 5.217 | 0.2951 | 10 | 50 | NA | NA | NA | NA | NA | |||
RC | 0.001 | 0.2 | 0.146 | 0.262 | NA | NA | 0.244 | 0.424 | 5.217 | 0.2951 | 10 | 50 | NA | NA | NA | NA | NA | |||
A | 0.011 | 0.1 | 0.001 | NA | NA | 0.15 | 0.031 | 0.334 | 10.55 | 0.1588 | 21 | 17 | NA | NA | NA | NA | NA | |||
R | 0.117 | 0.1 | 0.001 | NA | NA | 0.005 | 0.009 | 0.206 | 10.55 | 0.1588 | 21 | 17 | NA | NA | NA | NA | NA | |||
RC | 0.001 | 0.1 | 0.001 | NA | NA | 0.221 | 0.003 | 0.459 | 10.55 | 0.1588 | 21 | 17 | NA | NA | NA | NA | NA | |||
A | 0.032 | 0.3 | 0.495 | 0.604 | 0.066 | NA | NA | NA | NA | NA | NA | NA | NF | 0.5 | NF | 1 | 1 | |||
R | 0.146 | 0.3 | 0.377 | 0.426 | 0.017 | NA | NA | NA | NA | NA | NA | NA | NF | 0.5 | NF | 1 | 1 | |||
RC | 0.001 | 0.1 | 0.274 | 0.111 | 0.164 | NA | NA | NA | NA | NA | NA | NA | NF | 0.3 | NF | 1 | 1 |
This study reconstructed the 592 specimens over 100.7 m2 of the benthic community of “The Forest of the Weird” from still frames taken from a single stream of video footage using photogrammetry methods (
This study has demonstrated how SPPA and BNI provide two new types of insights into the ecology of deep-sea benthic systems. First, SPPA can detect many different types of interactions and associations without any
A high proportion (37.5%) of the specimens mapped were dead glass sponge skeletons (
The
Random labeling analyses plots showing the Quotient test.
The Janzen-Cornell hypothesis is an explanation of how high diversity is maintained in tropical forests and coral reefs, whereby species-specific predators or pathogens are attracted to adults and/or high density areas resulting in high mortality surrounding the adults or in the high-density areas (
Mass-mortality events and sponge diseases are on the increase (
The FW community showed a clear increase in density and change of composition on the Ridge (seen in the 3D reconstruction at the edge of
The strength of SPPA lies with describing how the density of different taxa change over the fine scale (here 0–4 m), and then using model fitting to determine the most likely underlying processes. When spatial models such as heterogeneous Poisson models, heterogeneous Thomas cluster models or hard/soft-core models are the best-fit to observed spatial patterns, this suggests that abiotic or habitat heterogeneities are the strongest influence on the studied populations (
Univariate PCF for the abundant taxa.
In contrast
The population ecology of the two sponges
Evidence for competition between marine sponges leading to elimination of a species are relatively rare, and most commonly involve outcompeting for a limited substrate, overflowing or chemical mediation (
Community composition for deep-sea sponge and coral communities is strongly driven by abiotic factors over broad scales (∼1–100 km) (
Bayesian networks of associations between abundant sponge and corals. Dependencies between taxa are indicated by the lines connecting the two taxa, the width of which indicates the occurrence rate in the bootstrap analyses (wider lines indicate higher occurrence). Arrows indicate non-mutual dependence between two taxa; A- > B indicates that B depends on A. Mean interaction strengths of the correlations are indicated; positive interaction strengths indicating aggregation, negative interaction strengths indicating segregation. For more details, see section “Materials and Methods.”
The nature of these changes depends on the different biotic and abiotic intra and inter-specific interactions. For this community, the strength of the inter-specific interactions (as calculated by the PCF value;
Bivariate PCF analyses for the taxa with non-independent spatial distributions. The gray area is the simulation envelope of complete spatial randomness (CSR) for 999 Monte Carlo simulations. The
To our knowledge, this study is the first to analyze the mortality, population and community dynamics of a deep-sea sponge community using SPPA. Our results provide the first insight into the variety of ecological behaviors within this benthic community, and show how these different organisms have developed diverse responses for the biotic and abiotic gradients within their habitat. We have demonstrated how ecological interactions change with substrate topography, while the community composition remains relatively constant. BNI has demonstrated the connectivity of the abundant taxa, identifying
The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below:
EM conceived this study, processed the 3D model to spatial map, analyzed the data, and wrote the manuscript. SH developed the protocol for creating the 3D model from the video data and reconstructed the 3D model. Both authors contributed to the article and approved the submitted version.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We wish to thank Chris Kelley both as the lead scientist for the cruise and for his help with the data.