The hard clam Meretrix taiwanica is a vital farming species in Taiwan; the frequent massive mortality damages the clam fishery and causes substantial economic loss. In commercial aquaculture, insights into animal behavior help to assess organisms’ physiological conditions, such as applying behavior monitoring techniques in fish farming to reduce risk and operation costs. However, it isn’t easy to achieve in clam farming due to their infaunal living habits. This study presents a novel video-based clam behavior monitoring system to alert unusual events and reduce economic losses.
Shell exposure and enduring valve closure are the primary reactions of infaunal clams under stress based on our long-term laboratory and clam farm in-situ investigations. Hence, four clam states, Siphon (S), Hidden (H), Exposed (E), and Exposed with a siphon (ES), were formulated as behavior indicators. An automatic clam behavior monitoring system was developed using an underwater camera and deep learning convolutional neural network (CNN) model; the accuracy, precision, recall, and F1-score of clam state classification in laboratory and clam farm scene trained model was 0.924, 0.838, 0.793, 0.804 and 0.930, 0.808, 0.667, 0.704, respectively. The system can individually track each clam, conduct real-time state analysis, and output the results to the website platform. Our results demonstrate the feasibility of utilizing non-invasive visual signals for the behavioral monitoring of infaunal clams.