This paper proposes a metric-based few-shot learning method for fish species identification with limited samples. The method improves upon prototypical networks by introducing an attention module to enhance similarity calculation and classification performance. The model is trained on three public fish datasets: Croatian fish, Fish4Knowledge, and WildFish. Compared to prototypical networks, the proposed method achieves accuracy improvements of 2% to 10%, demonstrating effectiveness in small sample sizes and complex scenarios. The method combines attention feature extraction with nearest-neighbor classification, enabling accurate fish identification. Key contributions include the design of an attention prototype nearest neighbor network, the ability to generalize to new categories, and the application of image recognition technology for efficient fisheries management. The model outperforms existing methods in accuracy and generalization, providing a valuable tool for fisheries resource development and fish biodiversity conservation. The method uses an embedding module with channel and spatial attention mechanisms to improve feature extraction and classification. The classification module incorporates prototypical networks and a nearest-neighbor classifier, achieving high accuracy in fish species identification. The model was tested on three datasets, showing significant improvements in classification performance. The results demonstrate the effectiveness of the proposed method in fish species identification with limited samples.This paper proposes a metric-based few-shot learning method for fish species identification with limited samples. The method improves upon prototypical networks by introducing an attention module to enhance similarity calculation and classification performance. The model is trained on three public fish datasets: Croatian fish, Fish4Knowledge, and WildFish. Compared to prototypical networks, the proposed method achieves accuracy improvements of 2% to 10%, demonstrating effectiveness in small sample sizes and complex scenarios. The method combines attention feature extraction with nearest-neighbor classification, enabling accurate fish identification. Key contributions include the design of an attention prototype nearest neighbor network, the ability to generalize to new categories, and the application of image recognition technology for efficient fisheries management. The model outperforms existing methods in accuracy and generalization, providing a valuable tool for fisheries resource development and fish biodiversity conservation. The method uses an embedding module with channel and spatial attention mechanisms to improve feature extraction and classification. The classification module incorporates prototypical networks and a nearest-neighbor classifier, achieving high accuracy in fish species identification. The model was tested on three datasets, showing significant improvements in classification performance. The results demonstrate the effectiveness of the proposed method in fish species identification with limited samples.