A Lightweight Deep Neural Network for Individual Leopard Identification: Supporting Movement Ecology in the Taihang Mountains of Henan.

Abstract

Individual identification and tracking of leopards (Panthera pardus) are critical for conserving wild populations and understanding their behavioral ecology. Current methods primarily rely on manual coat pattern matching, which is inefficient for large datasets and makes it difficult to identify subadults undergoing growth-related pattern changes. Moreover, individual-based studies on the movement ecology of wild North China leopards (Panthera pardus japonensis) remain scarce. To address these challenges, we propose LeoNet, a lightweight hybrid deep learning model integrating convolutional neural networks and multilayer perceptrons for individual leopard identification. Using 42 868 images collected from 142 captive and wild felids, LeoNet outperformed most of the 11 models in adult identification (best accuracy, 97.42%) and performed robustly in cub identification (95.24%) and the stranger test (96.28%). By applying LeoNet to camera trap data from the Taihang Mountains in Henan, we identified 41 wild individuals, reconstructed their movement trajectories, estimated individual home ranges (1.63-29.09 km²), and analyzed inter-individual social relationships based on behavior over the 2017-2024 monitoring period. Significant sex-based differences in social behavior were detected, with females recorded in a higher proportion of co-appearance events in Jiyuan (p < 0.01) and males in Jiaozuo (p < 0.05). These findings revealed three spatially separated groups, extensive home range overlap, and pronounced habitat fragmentation associated with anthropogenic infrastructure. This study introduces a computationally efficient model and an integrated framework for long-term monitoring and ecological analysis of solitary carnivores, providing individual-level records and ecological insights for evidence-based conservation of the endangered North China leopard.