With the large-scale application of devices such as infrared cameras and unmanned aerial vehicles in national parks,monitoring data on biodiversity has grown explosively.Traditional manual identification methods are inefficient and difficult to meet the monitoring needs of the big data era,making the promotion of intelligent recognition technology an inevitable choice.To achieve systematic model selection and integration for low-computing-power and multi-task scenarios,this study constructs a modular model training and deployment framework,and systematically compares the performance of MobileNetV3,EfficientNetB0,and EfficientNetB3 on a real-world dataset from national parks.1)MobileNetV3 is most suitable for edge deployment,EfficientNetB0 is appropriate for medium-computing-power scenarios,and EfficientNetB3 is applicable for high-precision offline analysis.2)After deploying MobileNetV3 on infrared camera terminals in the Wolong Area of the Giant Panda National Park and performing optimization and fine-tuning,the validation accuracy increased to 89.2%,and the false positive rate decreased to 3.8%,which verified the feasibility and robustness of lightweight models in field environments.The framework supports incremental learning and multi-source data integration,providing a reusable engineering solution for intelligent monitoring in national parks.