Abstract:

Aims: Identifying target calls from large volumes of passive acoustic monitoring data commonly requires manual spectrogram review with selective audio playback and is particularly time-consuming in montane forests dominated by wind, stream and cicada noise. Hoolock tianxing has a small and highly fragmented population and urgently requires long-term, non-invasive monitoring methods. This study aimed to develop an automated call-detection framework, evaluate its cross-site generalisation and cross-species transferability, and characterise its diurnal calling rhythm. 

Methods: Two eight-channel PAM systems were deployed in Sudian Township, Yingjiang County, Yunnan. We developed CBSkyNet, which combines PCEN features, a CNN backbone, a convolutional block attention module and a bidirectional long short-term memory network. The dataset was split by spatial source to evaluate cross-site generalisation. 

Results: Across three random-seed runs, PCEN achieved the highest mean recall and F1-score for all four evaluated backbone networks and showed the smallest cross-architecture ranges for both metrics. In the fixed-seed ablation experiment, CBSkyNet used only 6.62 M parameters and achieved 99.30% recall and a 99.60% F1-score on the cross-site test set. At the bout level, all 20 independent calling events in the test set were detected. For approximately 2,008 h of continuous recordings, model inference and manual verification of candidate segments required about 16.1 h in total, making the complete workflow approximately 4-6 times faster than manual spectrogram screening. The model also recovered three low-signal-to-noise calling days missed during initial manual screening. Cross-species transfer showed that CBSkyNet was transferable to a closely related gibbon species, with a clearer benefit from pretraining when target-domain annotations were limited. Calling activity was concentrated in the early morning after sunrise, whereas the between-site difference in first detected calls requires further validation using independent family groups. 

Conclusion: CBSkyNet provides an efficient and reusable framework for screening large-scale recordings of H. tianxing and may facilitate long-term acoustic monitoring of this species and related endangered gibbons. The code and representative data are publicly available at https://github.com/mf0302/CBSkyNet.

Key words: Hoolock tianxing, passive acoustic monitoring, automatic call detection, deep learning, vocalisation rhythm, transfer learning