Automatic identification of Kuril Habor seals (Phoca vitulina stejnegeri) in underwater video footage recorded at a salmon set-nets

In Hokkaido's Erimo area, the Ministry of the Environment has championed the installation of rope-grid deterrents in salmon set-nets as a mitigation strategy against seal-induced fishery damage. To evaluate the efficacy of this approach and inform seal population management in the region, continuous underwater video monitoring has been conducted since 2016. This study proposes a lightweight deep learning model for underwater object detection, integrating the YOLOv4-tiny architecture with the Efficient Channel Attention (ECA) mechanism. The model was trained and evaluated using a dataset derived from underwater surveillance videos recorded between 2019 and 2021, focusing on seals, various fish species, and additional marine organisms. To optimize model performance, three distinct attention mechanisms—Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), and ECA—were systematically integrated with the YOLOv4-tiny framework and comparatively assessed using established performance metrics, including Precision (P), Recall (R), and F1-score. The results demonstrate that the ECA mechanism yielded superior performance on the employed dataset. Ultimately, the proposed enhanced model achieved a mean Average Precision (mAP) of 73.23% on the test set, with the Average Precision (AP) for seal detection reaching 97.23%, while maintaining a high inference speed of 236.68 frames per second (FPS). To further validate its practical utility, a 24-minute continuous surveillance video segment was analyzed, comparing the model's automated detection results against manual annotations. The findings revealed a high degree of consistency, indicating that the optimized model holds significant potential to augment or partially automate manual detection tasks in underwater monitoring scenarios.