Species Distribution Modelling of Corals and Sponges in the Eastern Arctic for Use in the Identification of Significant Benthic Areas

Published: 21-01-2019| Version 1 | DOI: 10.17632/mcb726kcbx.1
Lindsay Beazley,
Francisco Murillo,
Ellen Kenchington,
Javier Guijarro-Sabaniel,
Camille Lirette,
Tim Siferd,
Margaret Treble,
Emily Baker,
Marieve Bouchard Marmen,
Gabrielle Tompkins-MacDonald


Species distribution modelling using a random forest (RF) machine learning approach was used to predict the probability of occurrence and biomass of sponges, sea pens, and large and small gorgonian corals in the Hudson Strait portion of Fisheries and Oceans, Canada's (DFO) Hudson Bay Complex Biogeographic Zone (sponges only), and in the eastern extent (Davis Strait and Southern Baffin Bay) of the Eastern Arctic Biogeographic Zone. A suite of 54 environmental predictor variables from different data sources were used. Models utilized catch records from the DFO multispecies trawl surveys and DFO/industry northern shrimp surveys collected between 2006 and 2014. For each taxonomic group in each region, both presence-absence random forest models using data collected across gear types (Alfredo, Campelen, and Cosmos trawls), and biomass random forest models using data collected within gear types were run. Most presence-absence models had good predictive capacity with cross-validated Area Under the Receiver Operating Characteristic Curve (AUC) values ranging from 0.643 to 0.894. The lower AUC was produced from the Hudson Strait sponge model, which also had poor sensitivity and specificity relative to the models performed in the Eastern Arctic Biogeographic Zone. The random forest biomass models performed inconsistently within taxa by gear type, with the models for sponges using data from Alfredo and Campelen trawl surveys perfoming best (R2 = 0.327 and 0.480 respectively). Generalized additive models (GAMs) were developed to predict the biomass distribution of each taxonomic group and serve as a comparison to the RF models. Aside from providing continuous prediction maps of significant benthic taxa for these regions, our results will be useful in ecosystem management decision-making processes. In particular, good SDM models could be used to refine the outer boundaries of significant concentrations of these organisms identified by kernel density analyses and identify new suitable habitat not sampled by the trawl surveys in areas of extrapolation.