Rainbow trout ( Oncorhynchus mykiss ) are highly vulnerable to disease outbreaks and environmental stress during intensive aquaculture, making accurate health-state discrimination essential for timely intervention. However, different causes, such as ammonia nitrogen poisoning, nitrite poisoning, and Aeromonas salmonicida infection, often induce macroscopically similar abnormal behaviors, while their discriminative cues are subtle, local, and temporally dynamic. Existing vision-based methods remain limited in modeling these fine-grained spatio-temporal differences. To address this gap, this study proposes SAM-Net, a hybrid Self-Attention–Mamba network for recognizing similar abnormal behaviors in rainbow trout. By integrating global dependency modeling, efficient long-range temporal representation, and dynamic multi-scale spatio-temporal attention, SAM-Net improves the discrimination of disease-related and environment-induced abnormal states. The proposed framework combines self-attention to capture global spatio-temporal dependencies with an enhanced Mamba block to efficiently model temporal dynamics, thereby improving joint spatio-temporal feature learning. In addition, a Dynamic Spatio-temporal Multi-scale Attention (DSTA) module is introduced to enlarge the receptive field, enhance discriminative feature representation, and suppress interference from water ripples and turbid backgrounds. Experimental results show that the proposed method achieves an accuracy of 97.8% on a self-built dataset, and attains Top-1 accuracies of 83.1%, 72.4%, and 88.7% on Kinetics-400, SSv2, and Diving48, respectively, demonstrating strong generalization capability. These results indicate that the proposed framework provides an effective solution for the cause-oriented discrimination of similar abnormal behaviors in rainbow trout and offers technical support for early non-destructive warning and precise intervention in fish health management.
Meng et al. (Mon,) studied this question.