Search for a command to run...
To understand conditions associated with bat fatalities at wind turbines for informing reduction strategies, researchers began using thermal camera video monitoring with a field of view (FOV) focused on the rotor swept area (RSA) to detect fatalities. However, confidently detecting fatalities and correctly classifying when they occur in the RSA is difficult and relatively few have been documented when compared to the hours of recordings produced to date. We conducted the first proof-of-concept study to determine whether a ground-based thermal camera system focused below the RSA, combined with machine learning algorithms, could effectively track and detect bat fatalities at wind turbines. To determine camera success, we equipped two wind turbines in southern Texas each with two cameras and performed standard post-construction monitoring (PCM). We classified a total of 274,051 bat tracks, of which 189 were identified as potential bat fatalities by machine learning algorithms, while the remaining tracks corresponded to bats flying within the camera's FOV. After manual review of 10-minute summary images of all recorded videos, 23 bat tracks identified by algorithms were also manually identified in video as possible fatalities that occurred in a camera's field of view. These 23 tracks also aligned with a searcher discovered bat carcass in the detection space of the corresponding camera with an estimated time of death overlapping with a night of an identified track. Our findings demonstrate that the camera system, paired with our proprietary machine learning algorithm, can process thousands of hours of video to identify individual tracks as possible bat fatalities. We propose that the primary value of this system lies in its ability to automatically process many hours of video data and identify suspected bat fatality tracks which could direct human searchers to specific turbines in an effort to minimize manual searches, as well as its potential application in estimating offshore fatalities in settings where traditional search methods are not feasible.