Complementary acoustic-visual monitoring array for Indo-Pacific humpback dolphin habitat use in the Pearl River Estuary.

In-situ visual observation of marine mammal surfacing behavior is critical for understanding behavioral responses to environmental stressors. However, continuous visual monitoring systems remain sparse compared to fixed passive acoustic monitoring (PAM). We deployed a near-real-time passive acoustic-visual array for a full year to monitor the world's largest known population of Indo-Pacific humpback dolphins (Sousa chinensis) in their core estuarine habitat. The system integrated ten paired acoustic buoys with high-resolution cameras mounted on bridge infrastructure, operating in two deep learning-based modes: (1) Acoustic-coupled Visual Detection (AVD), where real-time acoustic cues triggered video recordings processed by YOLOv5l, and (2) Near-real-time Visual Detection (NVD), which continuously analyzed video streams independently of acoustic input. Over a one-year deployment (>78% operational uptime), AVD captured 112 dolphin-containing video clips, whereas NVD recorded 213, which underscores NVD's ability to complement passive acoustic monitoring by maintaining detection continuity during noisy underwater conditions or vocal quiescence. Generalized additive models identified distinct environmental drivers for each modality: acoustic detections were governed primarily by seasonal patterns, whereas visual detections were mainly influenced by wind speed and atmospheric visibility. Both approaches revealed pronounced spatial heterogeneity in dolphin occurrence, with higher detection frequencies near artificial islands, where structural habitat modification and prey aggregation likely enhance localized habitat use. Together, the acoustic and visual systems demonstrated strong methodological complementarity-each capturing events missed by the other-thereby advancing scalable, near-real-time monitoring of endangered cetaceans in urbanized estuaries.