A generative deep learning method for global species distribution prediction

Abstract

Anthropogenic pressures on biodiversity necessitate efficient and scalable methods to predict global species distributions. Current species distribution models (SDMs) face limitations with large-scale datasets, complex interspecies interactions, and data quality. Here, we introduce EcoVAE, an autoencoder-based generative model that integrates bioclimatic variables with georeferenced occurrences. The model is trained separately for plants, butterflies, and mammals to predict global distributions at both genus and species levels. EcoVAE achieves high precision and speed, outperforming traditional SDMs in spatial block cross-validation. Through unsupervised learning, it captures underlying distribution patterns and reveals species associations that align with known prey-predator relationships. Additionally, it evaluates global sampling efforts and interpolates distributions in data-limited regions, offering new applications for biodiversity exploration and monitoring.