As a crucial means for implementing intelligent fishery supervision， electronic monitoring （EM） relies heavily on image recognition technology. A key challenge lies in deploying high-performance yet lightweight models in edge computing scenarios. This study introduces the Vision Mamba （ViM） model from the field of deep learning. By utilizing the selective state space model （SSM） to construct a bidirectional encoder， the model achieves global modeling of long-range dependencies in images while maintaining linear computational complexity. Based on the Nature Conservancy （TNC） Fisheries Monitoring dataset， a systematic performance comparison was conducted against mainstream models such as ResNet， EfficientNet， and DeiT. The results demonstrate that the ViM model exhibits outstanding performance in both efficiency and accuracy. Among lightweight models， ViM-Tiny achieved a 1.12% increase in accuracy and a 2.19% improvement in F1 score while reducing parameters by 44.28% compared to the ResNet-18 baseline. Among medium-sized models， ViM-Small， with 44.65% fewer parameters than the ResNet-101 baseline， achieved comparable accuracy （0.960 3） and F1 score （0.964 5）. The study indicates that the ViM model maintains strong fishery species classification capability while significantly reducing model complexity， achieving an excellent balance between lightweight design and high accuracy. This research provides a novel technical pathway for constructing efficient and intelligent fishery supervision systems.