Bridging the deployment gap in automated moth monitoring through biodiversity foundation models
Year: 2026
Type: article
Abstract: Automated monitoring enables biodiversity surveys across broader spatial and temporal scales, but developing robust species recognition models requires large annotated image datasets that are often unavailable from field monitoring systems. Consequently, models are often trained using large citizen-science image repositories, creating a substantial domain gap between training and deployment images. Here, we developed a fine-grained moth classifier using approximately 363,000 citizen-science images representing 2727 species and evaluated it across two contrasting field monitoring systems encompassing 526 species. To improve deployment transferability, we optimized three components of the transfer-learning pipeline: biodiversity-specific pretrained representations, parameter-efficient model adaptation, and data augmentation. Biodiversity-specific pretrained representations with BioCLIP2.5 substantially improved deployment performance, outperforming conventional convolutional and generic vision-language models. Parameter-efficient adaptation improved deployment transferability, whereas deployment-oriented augmentations consistently enhanced robustness across both monitoring systems. The optimized pipeline achieved 85.0% Top-1 accuracy and 58.1% Macro F1 on the commercial monitoring system, and 89.5% Top-1 accuracy and 68.6% Macro F1 on the low-cost edge device. Optimization increased field performance relative to the source domain from 38.9 to 69.7% on the commercial system and from 62.3 to 82.3% on the low-cost edge device. Our study provides a practical framework for reducing the deployment gap between citizen-science training images and automated field monitoring, facilitating scalable species recognition in automated biodiversity monitoring.
Source: Scientific Reports
Institution: Taiwan Forestry Research Institute
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7222993086
FWCI:
Cites:
Cited by:
Related to: