A deep learning macroecological model to assess the combined effect of habitat loss and climate change on biodiversity

Climate change and habitat loss are major drivers of biodiversity loss, yet their combined effects at local and global scales remain insufficiently understood. To improve our understanding of the response of biodiversity to these intertwined drivers, we introduce a deep learning, spatially explicit macroecological model. Our model builds upon fundamental ecological generalizations, namely species-area relationships (SAR) and environmental niches, to predict local species richness based on habitat type, habitat surface area, and local climate and environmental predictors. Focusing on European plant species, we compile a comprehensive dataset consisting of habitat polygons, along with associated bioclimatic and environmental conditions. This dataset, complemented with opportunistic species occurrence data used to derive local species richness, is employed to train our deep-learning model. The model's accuracy is validated against an independent, plot-based dataset. This validation step not only demonstrates the model's capacity to accurately predict species richness across diverse habitat types and environmental conditions but also facilitates uncertainty estimation. We use explainable AI techniques to dissect the processes guiding the model's predictions, advancing our understanding of the complex relationships between climate, habitat, and biodiversity, and identifying vegetation group-specific drivers of species richness. We further discuss the model's potential for conducting comprehensive biodiversity change attribution studies. Looking ahead, we outline how the approach could be extended to predict functional groups. This is a crucial step towards coupling climate and biodiversity models to capture not only the effects of climate on biodiversity but also the feedbacks provoked by biodiversity on Earth system processes.