Evaluating model-agnostic post-hoc methods in explainable artificial intelligence: augmenting species distribution models.

Species distribution models (SDMs) remotely guide conservation programs for endangered species by estimating potential reserve areas based on a set of environmental features. Most SDM research only explains their predictions across the study area (global), effectively disregarding the predictions for specific sites (local) where conservation-related activities are confined. This study aims to address this spatial gap in explainability by applying model-agnostic post-hoc methods in explainable artificial intelligence for SDM at two scopes. These methods explain the importance, effects, and interactions of bioclimatic features on the SDM for Mindoro warty pigs (Sus oliveri), an emblematic yet endangered endemic fauna in Mindoro Island, Philippines. Areas with a high predicted probability of presence coincide with higher elevation, spanning the Mindoro Mountain Range. Global explainability methods-Permutation Feature Importance, Shapley Additive Explanations (SHAP), and Accumulated Local Effect-reveal that annual precipitation mostly accounts for this island-wide trend, with more rain corresponding to higher probabilities. This is also observed using local explainability methods-SHAP, Local Interpretable Model-agnostic Explanations, and Break Down-for the respective predictions on three potential conservation sites. The cumulative effect of bioclimatic features in these ~ 1 km