LMDD-YOLO: A lightweight multi-scale detection model for citrus leaf disease in complex scenarios

A B S T R A C T To address challenges in citrus leaf disease detection, including difficulties in multi-scale feature recognition, large model parameter counts, and insufficient detection speed for real-time applications, this study proposes a lightweight model, LMDD-YOLO, based on YOLOv11n. The model replaces standard convolutions with depthwise convolutions (DWConv), effectively reducing the number of parameters, GFLOPs, and the model size compared to the baseline. However, model compression inevitably causes a loss of feature detail and a decline in accuracy. To restore performance while maintaining lightweight, we incorporate a Dynamic Upsampling Module (DySample) to recover feature resolution and preserve critical disease details information. Second, for citrus disease detection, we propose Lightweight Multi-scale Channel Spatial Attention (LMCSA). This mechanism adaptively and precisely focuses on multi-scale disease regions by capturing and fusing contextual information from different receptive fields in parallel, thereby enhancing feature representation capability. Finally, the Generalized Intersection over Union (GIoU) loss function improves bounding box regression and localization accuracy. Experimental results show that the LMDD-YOLO model achieves 88.2% precision, 94.4% recall, and 89.9% mAP 50 , with only 1.53M parameters, 3.8 GFLOPs, and a model size of 3.22 MB. In a 5-minute continuous test, an average frame rate of 25.6 FPS is achieved on the Raspberry Pi 5B edge device. Compared with the original model, this model resulted in only marginal decreases of 0.9%, 0.2%, and 0.3% in precision, recall, and mAP 50 , respectively, while the number of model parameters, GFLOPs, and model size were significantly reduced by 40.7%, 41.5%, and 38.4%, respectively. These results indicate that the model achieved a good balance between lightweight design and detection performance. This study provides a high-accuracy and cost-efficient lightweight solution for real-time citrus leaf disease detection in complex field environments, offering significant value for the practical deployment of intelligent agricultural disease monitoring systems.