Benchmarking Automated Detection and Classification Approaches for Long-Term Acoustic Monitoring of Endangered Species: A Case Study on Gibbons From Cambodia

Abstract

Recent advances in deep learning and transfer learning have revolutionized our ability for the automated detection of acoustic signals from long-term soundscape recordings. Effective automated detection approaches can vastly improve our ability to monitor endangered species, like gibbons. Here, we provide a benchmark for the automated detection of female duet contributions from southern yellow-cheeked crested gibbons (Nomascus gabriellae) recorded in Jahoo, Cambodia. For the benchmarking, we compared the performance of support vector machines (SVMs), a quasi-DenseNet architecture (Koogu), transfer learning with ResNet50 models trained on the “ImageNet” dataset (ResNet), and transfer learning with embeddings from a global birdsong model (BirdNET). Transfer learning models based on BirdNET embeddings had superior performance with a smaller number of training samples, whereas Koogu and ResNet50 models only had acceptable performance with a larger number of training samples (> 200 gibbon samples). We deployed the BirdNET-based model over > 130,000 h of continuous soundscape data, which, after manual review, resulted in > 12,000 verified true positive detections. We found that female gibbon calling events occurred mostly in the early morning hours between 05:00 and 06:00 local time. We had fewer gibbon detections during the monsoon period and found substantial variation in spatial patterns of calling events across months and years that may reflect territorial dynamics of gibbon groups. Reliable automated detection approaches are a critical first step for using passive acoustic monitoring to assess endangered gibbon populations at ecologically relevant temporal and spatial scales.