Published · Specimen Article
A Language Model for the Birds: Predicting Migration Patterns from Low-Resolution GPS Traces
Martínez, J., Hale, P., & Okafor, T. · Methods in Ecology · 2025
A transformer trained on 400K sparse tracking records predicts migration routes of three songbird species with 87% accuracy, halving the sampling frequency required for field studies.
migrationtransformersecologytrackingsongbirds
§ I. Background
High-resolution GPS tags remain prohibitively expensive for long-term songbird studies. Existing migration models rely on dense samples and break down at the low frequencies most researchers can actually afford in the field.
§ II. Methodology
The authors train a causal transformer on 400K tracking records collected from three species (Catharus ustulatus, Setophaga fusca, Hirundo rustica) and use masked-span loss to force the model to learn spatial autoregression. They benchmark against state-space and LSTM baselines on held-out seasons.