YOLO-Pika: A Lightweight Improved Model of YOLOv8n Incorporating Fusion_Block and Multi-scale Fusion FPN and Its Application in the Precise Detection of Plateau Pikas

Yihao Liu¹Zhao Jianyun^1*Changjun Xu²Yuedi Hou³Yuxiang Jiang⁴

• ¹Qinghai University, Xining, China
• ²Qinghai Provincial Key Laboratory of Geospatial Information Technology and Application, Xining, China
• ³Wuhan University, Wuhan, Hubei Province, China
• ⁴Qinghai Institute of Technology, Xx, China

The plateau pika (Ochotona curzoniae), recognized as a keystone species on the Qinghai-Tibet Plateau, has garnered considerable attention in recent years due to its significant ecological impact on grassland ecosystems. Population density, a critical indicator for assessing this impact, is typically estimated based on the number of plateau pika burrows. To enable rapid and cost-effective monitoring of plateau pika population density, this study proposes a lightweight detection model termed YOLO-Pika. Built upon the YOLOv8n architecture, the model introduces two key innovations: (1) The Fusion_Block is integrated into the backbone layer. Leveraging highdimensional mapping and fine-grained gating mechanisms, it enhances feature representation with negligible increase in computational cost. (2) The MS_Fusion_FPN is developed, composed of multiple MSEI modules. This structure enables multi-scale frequency-domain fusion and edge enhancement, thereby improving the model's detection performance. Experimental results demonstrate that on the plateau pika burrow dataset, YOLO-Pika improves mAP50 and mAP50-95 by 3.4 and 5.0 percentage points, respectively. At the same time, it reduces model parameters by 22.7% and FLOPs by 0.01%. Notably, the average precision (AP) for small-, medium-, and largesized targets all show substantial improvements.On the publicly available Brandt's vole hole dataset, YOLO-Pika further achieves a 4.9 percentage point increase in mAP50. It also reduces false detections related to localization errors, redundancy, and background noise by 30-50%, underscoring its strong cross-regional generalization capability. When benchmarked against five state-of-the-art lightweight object detection models, including YOLOv10, YOLO-Pika achieves the highest detection accuracy with the fewest model parameters. By effectively integrating real-time performance, detection precision, and deployment feasibility, YOLO-Pika offers a practical and scalable solution for rodent damage assessment in alpine grasslands and the detection of other rodent burrows.