Mamba-YOLO-SRC: An Automatic Deep Learning Framework for Respiratory Behavior Detection in the Chinese Giant Salamander

The Chinese giant salamander (Andrias davidianus), a species of high ecological and conservation value, shows abnormal respiratory behaviors as early signs of health decline. Accurate assessment of its pulmonary respiration is crucial for improving captive breeding and post-breeding parental care—key strategies for its survival and population recovery. However, its nocturnal and cave-dwelling nature makes traditional observation extremely difficult. Manual monitoring suffers from poor visibility at night, while conventional detection methods often miss subtle respiratory movements, limiting behavioral and health research. To address these challenges, this study presents the first automated method for monitoring respiratory behaviors in this species. We propose Mamba-YOLO-SRC, a novel hybrid detection framework that combines Mamba and YOLO architectures to accurately identify four key behaviors: diving (Dive), head-raising (HeadUP), inhalation (Inhale), and exhalation (Exhale). The proposed model achieves a mean average precision (mAP@0.5) of 0.944, with per-class average precision scores of 0.975 for Dive, 0.925 for HeadUP, 0.948 for Exhale, and 0.928 for Inhale. Mamba-YOLO-SRC provides a feasible and referable technical solution for advancing research on the Chinese giant salamander in both captive and natural settings.