A video-based behavioral classification framework for assessing stress responses in the cultured clam Meretrix taiwanica using a convolutional object detection model

Abstract

Monitoring animal behavior is critical in aquaculture for assessing health status and detecting stress, yet this remains particularly challenging for infaunal clams due to the absence of suitable automated monitoring approaches. This study presents a video-based behavioral classification framework for the cultured clam Meretrix taiwanica, utilizing a YOLOv9 object detection model. Behavioral states were defined based on siphon activity and burrowing status, and classified into four states: siphon (S), hidden (H), exposed with siphon (ES), and exposed (E). Exposure experiments simulating aquaculture stressors, including reduced salinity and elevated temperatures, were conducted to evaluate the effectiveness of the proposed framework. Under low salinity conditions (6 psu), clams showed increased hidden and exposed states, alongside reduced siphon activity, likely reflecting stage-specific behavioral progression and inter-individual variation in coping strategies under sustained osmotic stress. In response to thermal stress, the number of clams with siphon state initially increased under stressful high-temperature conditions (32–36°C), followed by a decline under extreme heat (36–40°C), coinciding with an increase in individuals exhibiting shell exposure and physiological deterioration. Behavioral changes proved more sensitive than wet weight in detecting early stress responses. The detection model achieved high performance, with an overall mean Average Precision (mAP₅₀) of 0.977. In all, this study provides a promising foundation for real-time, non-invasive behavioral monitoring systems in infaunal bivalves, with potential applications in aquaculture management and early warning strategies.