EMGCM: ensemble learning of multiple graph convolutional models for fish skeleton-based swimming behavior recognition

Accurate recognition of fish swimming behavior helps in health assessment and disease prevention. However, behavior similarity, drastic pose variations, and class imbalance pose challenges for efficient fish swimming behavior recognition models. Hence, this paper proposes a two-stage framework for recognizing fish swimming behaviors. In the initial stage, the fish skeleton is defined and a novel Skeleton-based Graph Convolutional Network (SGCN) is proposed to extract spatiotemporal features of fish movement. It reliably extracts the positions and spatial relationships of joints, bones, joint motion, and bone motion during fish swimming and reduces data complexity and noise. In the subsequent stage, aiming to improve the performance of the swimming behavior recognition model further, a Hybrid Ensemble (HE) method is designed. This method integrates multiple models’ strengths, reduces individual models’ prediction bias, and improves prediction robustness. To validate the effectiveness of EMGCM, experimental evaluations on a comprehensive PL-behavior dataset achieved an accuracy of 90.31 %, an F1 score of 81.33 %, and a precision of 87.08 %. These results demonstrate that EMGCM outperforms existing methods and supports swimming behavior monitoring in practical aquaculture applications.