Selecting environmental descriptors is critical for modelling the distribution of Antarctic benthic species

Abstract

Species distribution models (SDMs) are increasingly used in ecological and biogeographic studies by Antarctic biologists, including for conservation and management purposes. During the modelling process, model calibration is a critical step to ensure model reliability and robustness, especially in the case of SDMs, for which the number of selected environmental descriptors and their collinearity is a recurring issue. Boosted regression trees (BRT) was previously considered as one of the best modelling approach to correct for this type of bias. In the present study, we test the performance of BRT in modelling the distribution of Southern Ocean species using different numbers of environmental descriptors, either collinear or not. Models are generated for six sea star species with contrasting ecological niches and wide distribution ranges over the entire Southern Ocean. For the six studied species, overall modelling performance is not affected by the number of environmental descriptors used to generate models, BRT using the most informative descriptors and minimizing model overfitting. However, removing collinear descriptors also helps reduce model overfitting. Our results confirm that BRTs may perform well and are relevant to deal with complex and redundant environmental information for Antarctic biodiversity distribution studies. Selecting a limited number of non-collinear descriptors before modelling may generate simpler models and facilitate their interpretation. The modelled distributions do not differ noticeably between the different species despite contrasting species ecological niches. This unexpected result stresses important limitations in using SDMs for broad scale spatial studies, based on limited, spatially aggregated data, and low-resolution descriptors.

Incorporating mesopelagic fish into the evaluation of conservation areas for marine living resources under climate change scenarios

“…These three SDMs are commonly used to predict species distribution in the SO (Duhamel et al 2014 ; Freer et al 2019 ; Hindell et al 2020 ; Pinkerton et al 2010 ; Ran et al 2022 ; Woods et al 2023 ). No environmental variables were excluded from the models due to collinearity for the following reasons: (1) machine learning (RF, BRT, and Maxent) can effectively deal with collinearity and account for the complex interactions among environmental variables (Charlène et al 2020 ; Ellis et al 2012 ; Hapfelmeier et al 2014 ; Phillips et al 2006 ); and (2) Including more biologically relevant predictors often results in better predictive performance (Duhamel et al 2014 ; Freer et al 2019 ; Hindell et al 2020 ; Pinkerton et al 2010 ; Ran et al 2022 ; Xavier et al 2016 ).…”

“…Although the ecological importance of meso-fish is widely recognized, they remain the least studied ecosystem components (Dowd et al 2022 ; Woods et al 2022 ). Over preceding decades, an increasing number of studies has been conducted to model the spatial distribution of Antarctic species, mainly for higher predators (e.g., seals, seabirds and penguins) (Hindell et al 2020 ), Antarctic krill (Sylvester et al 2021 ), benthic organisms, e.g., sea urchins and sea stars (Charlène et al 2020 ; Fabri-Ruiz et al 2020 ), cephalopods (Xavier et al 2016 ), and copepods (Pinkerton et al 2010 ). However, relatively little research has been done on meso-fish biomass, life history, and special distribution under environmental and climate change scenarios (Dowd et al 2022 ; Duhamel et al 2014 ; Freer et al 2019 ; Kaartvedt et al 2012 ; Loots et al 2007 ; Ran et al 2022 ).…”

Fine‐scale spatial and temporal distribution patterns of large marine predators in a biodiversity hotspot

“…In general, BRTs are reasonably robust to correlation between predictor variables (Charlène et al, 2020; Elith & Leathwick, 2009); though highly correlated variables can complicate the interpretation of the results, and their inclusion may not greatly increase the predictive power (Leathwick et al, 2006). Here, multiple steps were undertaken to produce parsimonious models for each species (i.e.…”

Benthic habitat mapping: A review of three decades of mapping biological patterns on the seafloor

“…Several other forms of geospatial data are measured and implemented as predictor variables for benthic habitat mapping. Spatial measurements such as longitude and latitude coordinates, or distances from geographical features such as coastline, islands, or geological phenomena may serve as surrogates for benthic habitat drivers such as sediment transport, physical or chemical oceanographic parameters, dispersal, or habitat connectivity (McArthur et al, 2010;Giusti et al, 2014;Vassallo et al, 2018;Charlène et al, 2020). These variables also may enable leveraging of spatial autocorrelation of the response variable in order to increase predictive capacity of geospatial models -either by capturing relevant information on unmeasured environmental variables, or by modelling spatial relationships that arise as a function of symbiotic or community processes (Legendre & Fortin, 1989).…”