AI and the World's Hard Problems · Life sciences · Nature (computational biology) · 2026-10-20
Bridging the deployment gap in automated moth monitoring through biodiversity foundation models
Species recognition models for automated biodiversity monitoring are typically trained on large citizen-science image repositories, creating a domain gap with field deployment images. The authors built a fine-grained moth classifier from roughly 363,000 citizen-science images covering 2,727 species and optimized the transfer-learning pipeline through biodiversity-specific pretrained representations (BioCLIP2.5), parameter-efficient adaptation, and data augmentation. The pipeline achieved 85.0% Top-1 accuracy on a commercial monitoring system and 89.5% on a low-cost edge device, raising field performance from 38.9% to 69.7% and from 62.3% to 82.3%, respectively.