A Comparison of Spatio-temporal Species Distribution Modelling Frameworks Using Cucumaria frondosa as a Case Study

Abstract

Cucumaria frondosa—the most abundant sea cucumber species in the Maritimes—is the target of a relatively new offshore fishery on the Scotian Shelf. In an effort to ensure the long-term sustainability of the stock, spatial reserves for C. frondosa were set aside for the 2018/2019 fishing season. However, the expected habitat map used to design the reserve boundaries was acknowledged to be somewhat coarse at the time. Furthermore, high-resolution environmental data layers (including several from a digital elevation model) have recently become available for the area. We therefore incorporated these new data layers into three spatio-temporal species distribution models, with the aim of re-examining the distribution of C. frondosa on the Scotian Shelf. Each model accounted for spatio-temporal autocorrelation differently, allowing us to compare their underlying methodologies.

Specifically, we fitted a generalized additive model with spatio-temporal smooths and two generalized linear mixed models with spatio-temporal random effects, one of which implemented the random effects using Gaussian Markov random fields, and the other with nearest-neighbour Gaussian processes. All three were fitted to catch data (recorded 2000-2019) from Fisheries and Oceans Canada’s annual Research Vessel and Snow Crab surveys. To accommodate the large number of zero catch values, each model consisted of two parts; the first part modelled the probability of sea cucumber presence using all tows, whereas the second used the Gaussian distribution to model positive catch for presence tows. Together, the three models suggested that several environmental covariates were possibly predictive of sea cucumber habitat and abundance in the region, including log-transformed depth and bottom stress. Furthermore, they indicated that considerable C. frondosa aggregations existed on Middle Bank, Banquereau Bank, and Sable Bank as of 2019. The models’ predictions were generally very similar, although they provided somewhat different inferences regarding covariate effects; we discuss the implications of this result for the practitioner.