Broad-scale species distribution models applied to data-poor areas

Abstract

Species distribution models (SDMs) have been increasingly used over the past decades to characterise the spatial distribution and the ecological niche of various taxa. Validating predicted species distribution is important, especially when producing broad-scale models (i.e. at continental or oceanic scale) based on limited and spatially aggregated presence-only records. In the present study, several model calibration methods are compared and guidelines are provided to perform relevant SDMs using a Southern Ocean marine species, the starfish Odontaster validus Koehler, 1906, as a case study. The effect of the spatial aggregation of presence-only records on modelling performance is evaluated and the relevance of a target-background sampling procedure to correct for this effect is assessed. The accuracy of model validation is estimated using k-fold random and spatial cross-validation procedures. Finally, we evaluate the relevance of the Multivariate Environmental Similarity Surface (MESS) index to identify areas in which SDMs accurately interpolate and conversely, areas in which models extrapolate outside the environmental range of occurrence records. Results show that the random cross-validation procedure (i.e. a widely applied method, for which training and test records are randomly selected in space) tends to overestimate model performance when applied to spatially aggregated datasets. Spatial cross-validation procedures can compensate for this over-estimation effect but different spatial cross-validation procedures must be tested for their ability to reduce over-fitting while providing relevant validation scores. Model predictions show that SDM generalisation is limited when working with aggregated datasets at broad spatial scale. The MESS index calculated in our case study show that over half of the predicted area is highly uncertain due to extrapolation. Our work provides methodological guidelines to generate accurate model assessments at broad spatial scale when using limited and aggregated presence-only datasets. We highlight the importance of taking into account the presence of spatial aggregation in species records and using non-random cross-validation procedures. Evaluating the best calibration procedures and correcting for spatial biases should be considered ahead the modelling exercise to improve modelling relevance.

Benthic ecoregionalization based on echinoid fauna of the Southern Ocean supports current proposals of Antarctic Marine Protected Areas under IPCC scenarios of climate change

“…The congruence between present ecoregions and general biogeographic patterns of the SO benthos organisms can be explained first by the contribution of abiotic factors. Typically, depth, seafloor temperature, and sea ice concentration are important contributors to ecoregions delineation; they are also important drivers of the benthos distribution at large spatial scale (David et al, ; Guillaumot et al, ; Gutt, ; Pierrat et al, ). General biogeographic patterns are also the legacy of common evolutionary events triggered by climate history and the paleogeography of the SO (Crame, ; Saucède, Pierrat, Brayard, & David, ; Saucède et al, ).…”

“…Species records were aggregated to a pixel size of 0.1° × 0.1°, a scale determined by the resolution of environmental data available (see below). Duplicates of species occurrence were removed from each pixel as occurrence duplication may bias model outputs (Guillaumot et al, ). Individual ENM were generated for each selected species before using GMM to assemble individual models (Figure S3).…”

“…To limit the effect of this sampling bias on model performance, the selection of pseudo‐absences was weighted based on a Kernel Density Estimation (KDE) map used as a proxy of the sampling effort. The KDE map was computed with all echinoid occurrences of the database using Spatial Analyst in ArcGIS v10.2 (ESRI, ) and following Guillaumot et al (). To limit the effects of the spatial autocorrelation (SAC) that breaks the ‘independent errors’ assumption, a major issue in spatial analysis and ecological modeling (Guillaumot et al, ; Legendre, ), several replicates of pseudo‐absences were generated for the model calibration.…”

“…The KDE map was computed with all echinoid occurrences of the database using Spatial Analyst in ArcGIS v10.2 (ESRI, 2011) and following Guillaumot et al (2019). To limit the effects of the spatial autocorrelation (SAC) that breaks the 'independent errors' assumption, a major issue in spatial analysis and ecological modeling (Guillaumot et al, 2019;Legendre, 1993), several replicates of pseudo-absences were generated for the model calibration. SAC of residuals was then quantified using the Moran I index computed with the ape R package (Paradis et al, 2008).…”

Diversity of Antarctic Echinoids and Ecoregions of the Southern Ocean

“…Our knowledge of echinoid species distribution in the Southern Ocean is still fractional and biased by uneven sampling efforts and the spatio-temporal aggregation of data collection (Gutt et al, 2012;Guillaumot et al, 2018Guillaumot et al, , 2019. Statistical tools and modeling approaches however have been developed to address this issue and determine the genuine factors that determine species distribution (Newbold, 2010;Barbet-Massin et al, 2012;Hijmans, 2012;Tessarolo et al, 2014;Guillaumot et al, 2018Guillaumot et al, , 2019Valavi et al, 2018). Ecological Niche Models (ENM), also known as Species Distribution Models offer a baseline for detecting, monitoring and predicting the impact of climate change on species and biota (Gutt et al, 2015;Kennicutt et al, 2015).…”