Enhancing eco-sensing in aquatic environments: Fish jumping behavior automatic recognition using YOLOv5

Contemporary research on ichthyological behavior predominantly investigates underwater environments. However, the intricate nature of aquatic ecosystems often hampers subaqueous observations of fish behavior due to interference. Transitioning the observational perspective from subaqueous to supra-aquatic enables a more direct assessment of fish physiology and habitat conditions. In this study, we utilized the YOLOv5 convolutional neural network target detection model to develop a fish jumping behavior (FJB) recognition model. A dataset comprising 877 images of fish jumping, captured via a camera in a reservoir, was assembled for model training and validation. After training and validating the model, its recognition accuracy was further tested in real aquatic environments. The results show that YOLOv5 outperforms YOLOv7, YOLOv8, and YOLOv9 in detecting splashes. Post 50 training epochs, YOLOv5 achieved over 97 % precision and recall in the validation set, with an F1 score exceeding 0.9. Furthermore, an enhanced YOLOv5-SN model was devised by integrating specific rules related to ripple size variation and duration, attributable to fish jumping. This modification significantly mitigates noise interference in the detection process. The model's robustness against weather variations ensures reliable detection of fish jumping behavior under diverse meteorological conditions, including rain, cloudiness, and sunshine. Different meteorological elements exert varying effects on fish jumping behavior. The research results can lay the foundation for intelligent perception in aquatic ecology assessment and aquaculture.