Climate envelope modeling for ocelot conservation planning: peering inside the black box

Abstract

Climate envelope models have been used to evaluate the predicted impacts of climate change on species of concern and can be a useful planning tool in determining the long-term suitability of current habitat and potential introduction sites. However, due to model complexity, these models have generally been seen as "black boxes" when it comes to understanding why they make the predictions they do. In this study, we examined current potential ocelot (Leopardus pardalis) habitat using publicly available ocelot records and CHELSA bioclimatic variables combined in an ensemble model approach. We then predicted future potential ocelot habitat under three different emission scenarios using five climate models produced at a 30 arcsec grid cell size. To better understand what was driving model predictions, we used a variety of model interpretability approaches, including variable importance values, Shapley additive explanations, interaction detection, and a local surrogate model. These approaches revealed that ocelot potential habitat was associated with at least 233 mm of precipitation during the warmest quarter (BIO18) and at least 146 mm of precipitation during the wettest month (BIO13). Shapley additive explanations, applied to four protected areas across the species range, were valuable for understanding which climatic predictors drove predicted ocelot habitat at a local scale, with the protected areas with the highest predicted values having consistently more precipitation and higher temperatures. These tools are model agnostic and can be used to assess model predictions for biological validity and to further understand potential underlying relationships. As climate change shifts species range distributions, species distribution models and the tools to interpret them may become increasingly valuable, particularly on the leading edge of range expansion.

Species distribution modeling of the castor bean tick,Ixodes ricinus(Linnaeus, 1758), under current and future climates, with a special focus on Ukraine and Latvia

“…In other words, SHAP values can be integrated into global explanations such as variable importance, but using a completely different from various widely applied so far feature selection methods [64]. Until only now, the use of SHAP for explaining which environmental factors (including climatic) influence the spatial distribution of species occurrences has started to be explored more widely [65][66][67][68][69][70].…”