Low-disturbance UAV-AI monitoring of the endangered Przewalski's gazelle on the Qinghai-Tibetan plateau

The monitoring of endangered species often faces two constraints of behavioral disturbance and limited data availability. This study focuses on Przewalski's gazelle ( Procapra przewalskii ), an endemic endangered ungulate of the Qinghai-Tibetan Plateau. In this article, we propose an integrated, non-invasive monitoring framework that combines vertical take-off and landing (VTOL) unmanned aerial vehicle (UAV) with deep learning techniques to support conservation of this species. Disturbance-response experiments in the study area indicated a clear, graded increase in behavioral reactions as UAV operations became lower and more acoustically prominent, supporting the feasibility of low-disturbance aerial surveys when operated under a conservative and low-impact flight regime. We built a UAV imagery dataset and evaluated multiple object-detection models across varying training-data splits. Our results showed that You Only Look Once version 11 (YOLOv11) provided the best balance between accuracy and efficiency under small-sample conditions (20% training data), achieving a mean Average Precision (mAP) of 0.872 and maintaining strong operational performance, whereas Real-Time Detection Transformer (RT-DETR) attained slightly higher peak accuracy at substantially greater computational cost. UAV-based counts showed strong agreement with ground surveys (a mean relative agreement of 94.52%), underscoring the reliability of UAV monitoring in heterogeneous landscapes. Spatial statistical analysis revealed significant positive spatial autocorrelation (Moran's I = 0.238, p < 0.001), with high-density clusters (>81 individuals/km 2 ) concentrated in the southwestern part of the lakeside meadows of the study area. Fences and transportation infrastructure exacerbated fragmentation and constrained movement corridors. Overall, the UAV–deep learning framework provides actionable evidence for population assessment, habitat management, and conservation planning in alpine ecosystems. • Established a repeatable, non-invasive UAV-AI pipeline for Przewalski's gazelle monitoring, validated against synchronous ground counts. • Characterized disturbance responses across flight profiles and demonstrated robust detection under limited-label training. • Spatial analyses revealed persistent lakeshore clustering and infrastructure-driven fragmentation, motivating connectivity- and permeability-focused management.