Enhancing aquatic ecosystem monitoring through fish jumping behavior analysis and YOLOV5: Applications in freshwater...

Traditional fish monitoring methods suffer from limited continuity and significant uncertainty in tracking population distribution. This study develops recognition rules using the inherent variability in fish jumping behavior, influenced by habitat differences and physical traits. A comprehensive dataset is constructed on fish jumping behavior (FJB), relevant features are extracted and domain expertise is incorporated into the design process, using the YOLOV5 deep-learning model for target detection. An automatic fish species identification model, Fish-reco, is developed based on YOLOV5, which utilizes jumping behavior data, feature extraction, and recognition rules. Our results demonstrate the capability of the water splash data extraction model to effectively capture video clips depicting fish-water interactions in natural aquatic environments. Notably, both precision and recall exceed 96 % in the validation set. Additionally, a comprehensive feature library is established through feature engineering on 877 dataset samples, which encapsulates the jumping behaviors and resulting ripples of three freshwater fish species, including catfish, bighead carp, and carp, across various jumping stages. Finally, the robust performance of the FJB-based Fish-reco fish species identification model in classifying the above three common freshwater species is demonstrated. The recognition precision of carp, bighead carp, and catfish can reach 0.845, 0.92, and 0.995, respectively, and the mAP@50 can reach 0.918, 0.908, and 0.993. This study intuitively reflects the fish's physiological state and habitat by shifting the observation viewpoint from below to above the water surface, offering valuable insights for fishery resource assessment and water ecology evaluation.