AUTHOR=Liu Yihao , Zhao Jianyun , Xu Changjun , Hou Yuedi , Jiang Yuxiang 

TITLE=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

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1607492

DOI=10.3389/fpls.2025.1607492

ISSN=1664-462X

ABSTRACT=The plateau pika (Ochotona curzoniae) is a keystone species on the Qinghai–Tibet Plateau, and its population density—typically inferred from burrow counts—requires rapid, low-cost monitoring. We propose YOLO-Pika, a lightweight detector built on YOLOv8n that integrates (1) a Fusion_Block into the backbone, leveraging high-dimensional mapping and fine-grained gating to enhance feature representation with negligible computational overhead, and (2) an MS_Fusion_FPN composed of multiple MSEI modules for multi-scale frequency-domain fusion and edge enhancement. On a plateau pika burrow dataset, YOLO-Pika increases mAP50 by 3.4 points and mAP50–95 by 5.0 points while reducing parameters by 22.7% and FLOPs by 0.01%; AP improves for small, medium, and large targets. On a public Brandt’s vole hole dataset, it achieves a further 4.9-point gain in mAP50 and reduces false detections from localization errors, redundancy, and background noise by 30–50%. Compared with five state-of-the-art lightweight detectors (including YOLOv10), YOLO-Pika attains the highest detection accuracy with the fewest parameters. These results show that YOLO-Pika balances real-time performance, detection precision, and deployment feasibility, offering a practical, scalable solution for rodent burrow detection and alpine grassland damage assessment with strong cross-regional generalization.