Project Highlights

Global Seamounts Project

Geomorphic Proxies for Benthic Biodiversity on Seamounts

The GSP plans to estimate biomass and species abundance over wide areas of benthic terrain adjacent to seamounts through the development of geomorphic proxies, as inputs into GSP ecosystem models. Testing and validation of biophysical proxies derived from benthic substrate image visualization, calibrated with lab analysis of samples, require a substantial volume of visual data for analysis. Global Oceans’ recent acquisition of the 6000-meter Ocean Explorer 6000 towed vehicle, and our plan to rebuild the system with high resolution cameras, video, multibeam echosounder (MBES), and precision control will contribute to the collection of visual data at depth as proposed here.


Jesse van der Grient, PhD 

Post-doctoral Researcher, Department of Oceanography, University of Hawaii at Manoa, Honolulu, HI, USA


Orhun Aydin, PhD

Spatial Statistics Research, Esri, Inc., Redlands, CA, USA


Through evolutionary history, organisms have adapted to their surroundings, and these adaptations influence the distribution and abundance of organisms today. Seafloor characteristics, such as sediment composition, geomorphology, substrata type and complexity, and topographic relief, are just a few examples of physical variables which are known to influence the structure of the benthic biological community (Jennings et al. 1996; Levin et al. 2001; Curley et al. 2002; Thrush et al. 2005; Anderson et al. 2009).


The development and use of biophysical proxies to estimate deep-sea benthic biodiversity and structure has made initial progress as a method for investigating these habitats at a larger geographic scale (Anderson, et al, 2011; Roff, et al, 2003). Exploring the utility of this approach for the Global Seamounts Project (GSP) becomes relevant when thinking about how to estimate seamount system-wide benthic biodiversity and biomass for use in populating ecosystem models. A separate GSP Working Group on geomorphic proxies has been formed to assess the value of using this approach for the project; to develop appropriate field methods; and to develop strategies for data analysis, including with the use of statistical machine learning approaches.


Studies that investigate these complex biophysical relationships are often performed at a fine spatial scale to explore relationships between physical drivers and the biological community. Sampling strategies often consist of collecting video and still image data and can also contain sediment samples and biological samples. These data sources can be used to relate and classify specific environmental and biological variables based on common appearance.


However, the vast sources of data also pose a challenge: incorporating direct and indirect measurements of the environment to predict biological indicators. For the Global Seamounts Project, we propose to utilize a statistical machine learning approach to enable classification tools to predict the composition of benthic biological communities (e.g. biodiversity and biomass) based on analysis of wide-area high-resolution photographic transects of selected seamount terrain. Representative benthic core samples will be processed for biological analysis to create a set of training data for supervised analysis of geomorphic proxies. A deep-neural network is proposed to relate biophysical proxies to biological indicators.


This approach is proposed as a method for estimating benthic biodiversity and biomass on surveyed seamounts, correlated with mapped habitats, that will feed into multiple ecosystem models developed for the project.


Deep-neural networks, also known as perceptron, allow incorporating vast data types without any ad-hoc manipulation, thus reducing the subjectivity (see Figure 1). Workflows utilizing deep neural networks also possess the flexibility for computational scale-up, allowing characterizing vast extents of area at a fine scale in a feasible amount of time (Lary et al., 2016). Such methods can be optimized on a cloud topology for mass training and prediction for large areas. In addition, the decoupled training and prediction processes allow cloud-based training to explore relationships and enabling an optimized model using an edge computing device that can be installed on-board the deep-sea survey vehicle for real-time classification of biological indicators.


As biophysical relationships are known from the fine-spatial studies, a general-purpose classifier enables generalizing these relationships to larger areas. Effective characterization is achieved by mapping the ocean floor at an adequate resolution to preserve physical heterogeneity in the model to model realistic biophysical relationships.


Currently, acoustic technologies are improving at an unprecedented rate, and large areas of the ocean floor can now be mapped to a much higher resolution to create habitat maps allowing for the testing of biophysical relationships at a much larger spatial scale. 


Complementary high-resolution video and photographic survey data of selected benthic terrain will be used in this project as the primary basis for correlating environmental variables with estimates of biological composition. State-of-the-art process-based models for validating analytical tools are limited. Thus, we use neural networks to capture patterns between physical and biological indicators. Neural networks allow modeling uncertainty in the biological indicators as well. In this study, both epistemic and measurement-based uncertainty exist.


Secondly, the quality of the data source measuring the physical environment can cause uncertainty, and sources of uncertainty are not independent. For example, visual data resolution can determine whether a physical variable is picked up or not, and whether the variation in this variable is adequate to discriminate between different niches; in other words, whether a classification is possible. A physical variable that appears to work well as a predictor may require further validation to assess the degree of correlation with specific fine-scale heterogeneity, which may not have been picked up in the visual photographic information, to ensure the physical variable is not masking the true relationship between biology and the environment.


Thus, a neural network approach is proposed to rigorously learn relationships, both weak and strong, and account for dependence between predictors. A proposed deep-neural network will be tested against field visualization data and calibration sediment cores to quantify accuracy. In addition, we propose to use a stochastic network topology to account for uncertainty in data. We propose to map the semblance of the neural network to quantify the local importance of every physical predictor. The semblance analysis (i.e. which local variables are most important for representing local biological variables) will guide further data collection and reveal uncertainty in predictions.


The tests and validation of biophysical proxies require substantial data volume, which is provided by data from videos and still images. These can take a long time to analyse, however as classification algorithms are optimized and project datasets are expanded these analytical tools should become more effective as biological predictors. The proposed method is naturally scalable and fault resilient. In addition, the method we propose to use lends itself to offline classification via edge-computing, enabling real-time prediction in situ. Edge-computing technology could be readily integrated on the new Ocean Explorer 6000 towed vehicle (Figure 2) that Global Oceans plans to deploy for the Global Seamounts Project.


Caption for Figure 1: Conceptual application of deep neural networks for geomorphic feature detection and habitat delineation. The illustration on the left is a temperature profile for a location off the coast of California; the middle schematic is a conceptual depiction of a neural network; and the map on the right represents Ecological Marine Units. Direct observations (in this case temperature profile) are used to map biodiversity using a machine learning process. Image: Orhun Aydin, PhD, Esri, Inc.


References


Jennings, S., and N. V. C. Polunin. "Effects of fishing effort and catch rate upon the structure and biomass of Fijian reef fish communities." Journal of Applied Ecology (1996): 400-412.


Levin, Lisa A., et al. "Environmental influences on regional deep-sea species diversity." Annual Review of Ecology and Systematics 32.1 (2001): 51-93.


Curley, B., Kingsford, M. J., and Gillanders, B. M. (2002). Spatial and habitat-related patterns of temperate fish assemblages: implications for the design of Marine Protected Areas. Marine and Freshwater Research 53, 1197–1210.


Thrush, Simon F., et al. "Multi-scale analysis of species–environment relationships." Marine Ecology Progress Series302 (2005): 13-26.


Anderson, T.J., Syms, C., Roberts, D.A., Howard, D., 2009. Multi-scale fish-habitat associations and the use of habitat surrogates to predict the organisation and abundance of deep-water fish assemblages. J. Exp. Mar. Biol. Ecol. 379, 34–42.


Lary, D. J., Alavi, A. H., Gandomi, A. H., & Walker, A. L. (2016). Machine learning in geosciences and remote sensing. Geoscience Frontiers, 7(1), 3-10.


Anderson, Tara J., et al. "Deep-sea bio-physical variables as surrogates for biological assemblages, an example from the Lord Howe Rise." Deep Sea Research Part II: Topical Studies in Oceanography 58.7-8 (2011): 979-991.


Roff, John C., Mark E. Taylor, and Josh Laughren. "Geophysical approaches to the classification, delineation and monitoring of marine habitats and their communities." Aquatic Conservation: Marine and freshwater ecosystems 13.1 (2003): 77-90.