A deep-learning based approach to detect and classify animals flying near wind turbines using thermal surveillance cameras and open-source software

Rapid growth in wind energy creates a need for wildlife-monitoring tools that can process long-duration video efficiently while preserving ecologically meaningful information on animal activity near operating wind turbines. To address this need, we developed an open-source, proof-of-concept thermal-video workflow combining computer vision and convolutional neural networks (CNNs) to detect and classify moving biological objects near a single wind turbine, and to preserve object-level outputs for downstream analysis. Using held-out imagery from the same wind turbine and camera system used for model development, the final workflow distinguished biological targets from non-biological motion, including moving wind turbine blades and clouds, with 99% accuracy, and classified bats (90%), birds (83%), and insects (69%). This open-source thermal-video workflow allows the recording of object ID, frame number, timestamp, x/y pixel position, predicted class, and class probability for each biological detection, which enables post-processing of apparent biological object flight paths, time spent within the field of view, and relationships with environmental variables. The models presented and described in this work were trained and tested on imagery from a single wind turbine and camera system as a proof of concept and the quantitative performance metrics are specific to this setup, therefore broader deployment will likely benefit from further site-specific retraining or calibration, though the released models and code provide a practical starting point for transfer learning and initial evaluation at new sites. The code and trained models have been released under permissive licenses to support community evaluation, retraining and reuse.