You Only Look Once 11 (YOLO11) optimization for lightweight and accurate plant detection in uncrewed aerial vehicle imagery

Detecting plant objects from uncrewed aerial vehicle (UAV) aerial images provides crucial information for applications such as precision agriculture, ecological monitoring, and urban green space management. Optimizing object detection models for this task requires addressing unique challenges, including small object sizes, complex backgrounds, and ensuring lightweight architectures suitable for UAV systems with limited computational resources. YOLO11, the latest official release in the You Only Look Once (YOLO) series, serves as the foundation for our study, where we propose three key modifications to enhance its performance for UAV-based plant detection. First, we add a P2 detection head and remove the P5 detection head to better leverage high-resolution features for small object detection while reducing computational cost. Second, we integrate the convolutional block attention module into the Neck, enhancing multi-scale feature fusion and focusing on critical plant-related features. Third, we replace the original Complete Intersection over Union (CIoU) Loss with Shape-Intersection over Union (Shape-IoU) Loss, improving bounding box regression by incorporating geometric consistency. Additionally, we combine four valuable single-class plant detection datasets into a larger, multi-class dataset, providing a more comprehensive benchmark for UAV-based plant detection. Experimental results on this benchmark indicate a notable reduction in parameter count and computational cost, with accuracy comparable to or marginally higher than state-of-the-art YOLO-based baselines, demonstrating effectiveness in lightweight and resource-constrained scenarios.