STEM-LM: Spatio-Temporal Ecological Modeling via Masked Language Model for Multi-Species Distribution

Abstract

Species distribution models (SDMs) are central to biodiversity forecasting and conservation decision-making. As ecological datasets grow in size, dimensionality, and resolution, there is a need for flexible yet scalable SDMs tailored to large-scale species observation data. Inspired by the masked modeling approach in natural language processing and genomics, if we treat each species' presence or absence as a token, and a site's species assemblage together with its spatio-temporal and ecological covariates as analogous to a sentence, we can learn joint co-occurrence structure by reconstructing masked species from their neighboring sites. We propose STEM-LM, a Transformer-based SDM that applies masked language modeling to multi-species distribution modeling with explicit consideration of spatio-temporal structure. By applying dynamic masking rates during training, a single trained model can support both purely spatio-temporal/ecological prediction and prediction conditioned on arbitrary subsets of observed species. On a North American butterfly dataset, a Northeast US demersal fish trawl survey, and a global plant distribution dataset, STEM-LM has the highest discrimination (AUROC) among all compared methods, and its rank-monotonicity (continuous Boyce index) is either the best or close to the best.