YOLO-Citrus: a lightweight and efficient model for citrus leaf disease detection in complex agricultural environments.

Accurate and efficient detection of citrus leaf diseases is crucial for ensuring the quality and yield of global citrus production. However, many existing agricultural disease detection methods face significant challenges, including overlapping leaf occlusion, difficulty in identifying small lesions, and interference from complex backgrounds. These limitations often lead to reduced accuracy and efficiency of object detection. Moreover, current models generally necessitate significant computational resources and possess substantial model sizes, which restrict their practical applicability and operational convenience. To tackle these issues, this study presents a novel model named YOLO-Citrus. It is a lightweight and efficient YOLOv11-based model designed to enhance the precision of detection while simultaneously minimizing computational expenses and the size of the model. This makes it more suitable for practical agricultural applications. The proposed solution incorporates three major innovations: the C3K2-STA module, the ADown module, and the Wise-Inner-MPDIoU loss function. In particular, YOLO-Citrus utilizes Star-Triplet Attention by embedding Triplet Attention into the Star Block to enhance bottleneck performance in C3K2-STA. It also adopts the ADown module as a lightweight and effective downsampling strategy and introduces the Wise-Inner-MPDIoU loss to facilitate optimized bounding box regression and enhanced detection accuracy. These advancements enable high detection accuracy with substantially reduced computational requirements. The experimental results demonstrate that YOLO-Citrus attains 96.6% mAP@0.5, representing an improvement of 1.4 percentage points over the YOLOv11s baseline (95.2%). Furthermore, it reaches 81.6% mAP@0.5:0.95, i.e., an enhancement of 1.3 percentage points compared to the baseline value of 80.3%. The optimized model delivers considerable efficiency gains, with model size reduced by 25.0% from 19.2 MB to 14.4 MB and computational cost decreased by 20.2% from 21.3 to 17.0 GFlops. Comparative analysis has confirmed that YOLO-Citrus performs better than other models in terms of comprehensive detection capability. These performance enhancements validate the model's effectiveness in real-world orchard conditions, offering practical solutions for early disease detection, precision treatment, and yield protection in citrus cultivation.