Deep Audio Detection Networks from Passive Acoustic Monitoring of a Critically Endangered Primate

Biodiversity monitoring is a crucial component of conservation, providing essential information on species occurrence, population dynamics, community composition, and ecosystem structure. Recently, to enhance wildlife protection, passive acoustic monitoring (PAM) technology has been developed based on animal sound characteristics and has become an important tool for wildlife monitoring. The vast amount of data generated by PAM has been transformed by the "big data" revolution, and research methods that integrate artificial intelligence (AI) with efficient and powerful machine learning models are rapidly advancing. The primary challenge is to isolate the vocalization data of target species from this massive dataset and to determine whether the developed methods can be applied to other species. In this study, we focused on the acoustic signals of a critically endangered white-headed langur, a primate species endemic to China. Unlike traditional methods that transform sound frequencies into spectrogram images, our approach uses a Deep Audio Detection Network (DeepADN) that directly converts audio into acoustic features, which are then fed into a convolutional neural network for accurate detection of white-headed langur calls, even in noisy environments. Our method optimized detection performance, achieving a recall rate of 98.22% and reducing manual review workload by 87.07%. Furthermore, DeepADN demonstrated cross-species applicability by successfully detecting François' langur calls, highlighting its potential for broader conservation monitoring efforts.