Hynobius amjiensis automatic recognition based on deep learning

Hynobius amjiensis is a critically endangered species endemic to China, classified as a national first-class protected animal. Understanding the current wild status of Hynobius amjiensis is essential for conservation and regulatory efforts. However, due to the species’ unique ecological habits, direct field tracking is challenging. This paper explores the feasibility of underwater high-definition video surveillance and deep learning to monitor Hynobius amjiensis in the field, and evaluates the impact of data sets on model accuracy. The results indicate that: (1) Variations in time periods, behavioral states, and occlusion levels of target samples significantly affect the model's detection accuracy and recall rate; (2) Incorporating samples of similar species improves the overall performance of the model, enhances species classification precision, reduces the false detection rate, and increases the model's robustness and generalization capability. The optimal model, YOLOv8x-ABCD, demonstrated superior performance across all evaluation metrics and exhibited the highest accuracy in detection results. However, non-living objects in the water, such as aquatic plants, shed egg bands, and branches, were occasionally misidentified as adult Hynobius amjiensis. Experimental results suggest that increasing the number of negative samples effectively reduces such misdetections, lowers the model’s error rate, and enhances detection accuracy and generalization.