Fish mortality is a major production-limiting factor in aquaculture. Real or near real-time mortality tracking can provide valuable inputs to farm managers or automated systems in order to initiate procedures to prevent mass mortality events. Additionally, avoiding excessive mortalities blocking outflow from a tank’s bottom drain is critical to maintain normal system operations in recirculating aquaculture systems (RAS). While traditional systems use periodic human operator observation and tracking - often in conjunction with an underwater camera - this study augments this approach with Machine Learning and Internet of Things (IoT) deployed at the Edge Device to provide round-the-clock mortality monitoring and trigger an alarm when mortality thresholds are exceeded. An imaging sensor [Pi camera (model: Pi 4 HQ, Raspberry Pi foundation, Cambridge, UK)] integrated with an edge computing device [model: Pi 4] was customized for underwater application and has been deployed in the tank at 0.6 m above the central drain to collect one image every hour for six weeks. Acquired images will be annotated as live and dead fish in Roboflow (Roboflow, Inc., Des Moines, Iowa, USA) and will be split up into training (70%), validation (20%), and test (10%) dataset to train a custom Yolo V6 mortality model. The accuracy of the model will be evaluated in terms of mean average precision and F1 score. The model predicted daily and cumulative mortality will be compared with the ground truth data, and the reliability of the alarm events will be analyzed and presented.
Keywords: Precision aquaculture; Decision support, Machine learning; RAS; Fish detection