Improving passive acoustic monitoring applications to the endangered Cook inlet beluga whale

Abstract

A decade after the Cook Inlet beluga was listed as endangered in 2008, its population has shown no signs of recovery. Lack of ecology knowledge limits our understanding of, and ability to manage, potential threats impeding recovery of this declining population. NOAA Fisheries, in partnership with the Alaska Department of Fish and Game, initiated a passive acoustics program in 2017 to monitor beluga seasonal occurrence by deploying a series of acoustic moorings, followed by months’ work of manual validation for the detectors' output. To reduce this labor intensive and time-consuming process, we extracted a series of spectrograms from the sound files containing validated detections, and built 4 deep learning convolutional neural networks (CNN) with fine tuning of parameters. The final model is an ensemble of these individually optimized models, and achieves 96.57% precision and 92.26% recall on testing data. As a comparison, current detectors tend to trigger more false positives, which result in 20%–60% precision when human noise is present and 85%–95% precision in quiet areas. Following the success of the model with its easy generalizations to other acoustic detection problems, our next step will be comparing results from non-validated raw data to adopt this updated analysis process.