ResNet-based bio-acoustics presence detection technology of Hainan gibbon calls

As the most precious primate with only about 30 left, it is highly desirable to monitor the population of Hainan gibbon for the sake of effective protection. The traditional monitoring technique is facing the challenges of statistical hysteresis and a large number of labor costs, which makes it unsustainable. Deep learning based passive acoustic monitoring has shown its great potential in automatic detection of Hainan gibbon calls in Dufourq et al. (2021). Motivated by this progress, a bioacoustics presence detec-tion network (BPDnet) was proposed for Hainan gibbon calls by tailoring acoustic signal pre-processing, feature extraction, label smoothing loss function, as well as data augmentation in the ResNet-based pas-sive acoustic detection framework. In addition, a post-processing procedure was proposed to further improve the passive acoustic presence detection. Numerical results are presented to show that, the pro-posed BPDnet outperforms the baseline scheme of Dufourq et al. (2021) in terms of bio-acoustics pres-ence detection accuracy, callback rate, F1-score by at least 12.9% when using the same 72-h test recording and test method without manual intervention. When manual post-processing is allowed, the proposed BPDnet provides us an almost perfect Hainan gibbon call presence detection scheme. (c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

Monitoring the population of Hainan gibbon is crucial due to its critically endangered status. Traditional monitoring techniques face challenges of statistical hysteresis and high labor costs. A deep learning-based passive acoustic monitoring method has shown potential in automatic detection of Hainan gibbon calls. In this study, a bioacoustics presence detection network (BPDnet) is proposed, which surpasses the baseline method in terms of detection accuracy, callback rate, and F1-score. The proposed BPDnet provides an almost perfect solution for Hainan gibbon call presence detection when manual post-processing is allowed.