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

The large number of bat fatalities resulting from collisions between flying bats and rotating wind turbine blades is concerning given the vulnerability of certain bat populations and the anticipated growth in wind farm development. Further, given the increasing size of wind turbines, rapidly changing environmental conditions, and uncertainty about factors influencing bat behaviors around wind turbines, it is difficult to predict how bat interactions and fatality risk will change over time. Thermal-imaging surveillance video is a powerful technology for studying bat behavior and could be used to continuously monitoring bat risk and associated factors at wind turbines. However, continuously operating thermal-imaging systems produce more data than is practical for humans to visually review. To realize the potential for real-time thermal-imaging methods to quantifying nocturnal activities of bats at wind turbines, we developed computer vision methods in a deep learning context to automatically detect and classify bats, birds, and insects in thermal-imaging video recorded at wind turbines. Convolutional neural networks (CNNs) derived sufficient saliency of features from the prescreened input data to effectively discriminate flying animals from non-biological objects such as moving clouds and turbine blades with 99% accuracy, as well as classify detected animals with reasonable accuracy of 90% for bats, 83% for birds and 69% for insects. The methods we describe and demonstrate herein do not require specialized computers, can process thermal imagery closer to real time than prior efforts, can be adapted for other imagery types and use cases, can be used to inform turbine management/curtailment decisions and are based entirely on open-source software to encourage and support future tool development by the community. ### Competing Interest Statement The authors have declared no competing interest.