Bat Fatality Detection System Summary Images

Description

Each image is a ten-minute summary of that contains an identified bat fatality analyzed as part of a research project at a wind farm in southern Texas. The project objective was to conduct a proof-of-concept study to test and validate a thermal camera system capable of accurately measuring the timing of bat fatalities at wind turbines (within seconds) by concentrating the FOV below the RSA. The dates of images are from 15 July 2022 to October 21 2022.

Steps to reproduce

We deployed two Axis Q1952-E thermal cameras with a 35 mm lens on tripods at two wind turbines approximately 75 m from the turbine looking back towards it, one on the northside facing south and one to the southside facing north, such that each covered an approximate 50-meter by 100-meter rectangular area within the search plot around the turbine. We pointed each camera 15 degrees above the horizon. The FOV included part of the turbine tower on the left side and sky as a reference point. We programmed cameras to operate from at least 2 hours before sunset to 2 hours after sunrise of the longest night during the study. We processed all videos using proprietary algorithms in Python programming language to identify bat fatalities by detecting ‘tracks’ of associated detections. We used a background subtractor to automatically detect objects in the video stream. The algorithm detected all ‘new’ objects (e.g., bats, insects, birds) by dynamically modeling the image background, subtracting it from the incoming image, and determining which parts of the image deviated substantially. Those ‘different’ parts of each image are identified as detections. Using pixels from the area of each detection, the algorithm derived information about each object’s pixel location, area, width, length, and other distinguishing features. Next, the algorithm connected detections into tracks, with each track representing a path followed by a single object through time. The algorithm required each connected detection within a track to be both spatially and temporally close and can eliminate ‘noise’ (disconnected detections) in the dataset based on distance. Once tracks were generated, the algorithm quantified associated features, such as velocity and acceleration statistics, as well as the average size of the detections within the track. Because bats are not the only objects that cameras could detect, we classified each track as a bat, insect, or other object using the associated features detailed above. To be able to classify each track, a dataset was built by meticulously going through 10,000s of tracks and manually labeling each one as an object commonly seen in video to create a labeled dataset. This manual training of the algorithm was a part of a Department of Energy Small Business Innovation Research grant (DE‐SC0021867). A model was then built based upon the labeled data to classify the remaining 100,000s of tracks. The model was built using features of detections that can separate each track into its classified category. We further classified each bat track as either activity or fatality based on the detection attributes. Since a bat fatality is a rare event, to get enough training data for this model we created a physics-based model of a falling bat carcass and generated thousands of tracks to simulate a fatality. We then supplemented this synthetic data with tracks we manually idenitified to create the final model.

Categories

Funders

• Renewable Energy Wildlife Research Fund

  Grant ID: B-21

Licence