Modernizing Alpine Grassland Biodiversity Conservation: Efficient AI-Powered Detection of Large Herbivores in the Qinghai-Tibet Plateau

Monitoring large herbivore diversity in alpine grasslands is a cornerstone for biodiversity conservation and ecosystem management. Traditional remote sensing approaches mainly rely on overhead perspectives, yet they often fall short in resolving species-level details under complex environments. Here, we present a framework characterized by the synchronized dual-network extraction of species occurrences and micro-habitat attributes from a single imagery stream. Using images collected along transects in Sanjiangyuan National Park, we developed an improved YOLOv10 model that incorporates an attention mechanism and a lightweight down-sampling module. Experimental results demonstrate that the improved model reduced parameters by 6.77% and floating-point operations by 6.09%, while increasing mAP@50–95 by 7.02%. On a portable device, it achieved an inference speed of 47 FPS. This ~1,645-fold efficiency gain over manual review (avg. 35 s/image) enables a rapid "Stop-and-Capture" survey mode, significantly reducing field sampling time. Applying the model output to population estimation, we quantified the abundance of four key herbivore species in the Sanjiangyuan National Park area: Tibetan wild ass (1,980), Tibetan gazelle (896), Tibetan antelope (3,321), and Wild yak (435). Furthermore, we utilized the EfficientNetV2_S network to identify habitat types (alpine meadow, alpine steppe, water sources, bare ground) from the imagery. The identified habitat types were subsequently integrated with the animal detection data to perform a habitat quality assessment. This integrated approach is crucial for understanding the influence of landscape heterogeneity on the population distribution and ecological balance of large herbivores. This ground-level AI-based approach provides a practical and cost-effective framework for large-herbivore detection and population monitoring in alpine ecosystems. By integrating artificial intelligence with landscape ecology analysis, our study offers an upgraded supportive tool for biodiversity monitoring and provides fine-scale data to support conservation decisions and ecological research.