TY  - JOUR
AU  - CHEN, Wanyong
AU  - CHEN, Fengjun
AU  - YU, Xuewen
AU  - ZHANG, Xuankai
AU  - GAO, Chenghao
PY  - 2026
TI  - RCL-YOLOv11n: A lightweight improved YOLOv11n model for detecting rice pests and diseases
JO  - Transactions of the Chinese Society of Agricultural Engineering
SN  - 1002-6819
SP  - 281
EP  - 289
VL  - 42
IS  - 11
AB  - Rice is one of the staple food crops for more than half of the global population in China. Therefore, it is often required to accurately and rapidly detect rice diseases and pests for high yield and national food security. However, multiple challenges remain in real-world field environments: the diversity of disease and pest species with significant scale variations, frequent occlusion among rice plants, and the high visual similarity between lesion/pest features and the background, all of which increase the difficulty of identification. Conventional detection has been limited by the inadequate feature representation and the complex factors, directly leading to the high rates of missed detections and false alarms. Consequently, it is a high demand for detection techniques to meet the requirements of precise prevention and control, particularly for the high rice yield and quality. In this study, an RCL-YOLOv11n framework with an improved YOLOv11n architecture was proposed to detect rice diseases and pests. Six commonly and highly harmful diseases and pests were selected: Rice gall midge, rice water weevil, rice leaf roller, blast, brown spot, and bacterial leaf blight. A series of improvements was utilized for the efficient characteristics of the YOLO series algorithms. The RCL-YOLOv11n model then achieved superior performance. (1) RepConv convolution module was introduced into the backbone network. HGNetv2 architecture was established as a multi-scale fusion backbone network, named RepHGNetv2. The original model was enhanced to capture disease and pest features at different scales. Subsequently, the disease and pest features were more efficiently perceived after optimization of the network structure, effectively reducing the number of parameters and computational complexity of the lightweight model. (2) Context Guided Block_Down module was incorporated with the context guidance mechanism in the neck network. The high accuracy was then achieved to better utilize multi-scale contextual information during feature fusion. Meanwhile, disease and pest features were recognized and distinguished against complex backgrounds, effectively reducing background interference. Furthermore, the LADH detection head was introduced to replace the Detect module of YOLOv11n. Asymmetric multi-level compression was adopted to improve the generalization with the reduced number of parameters for computational efficiency. The detection head effectively reduced the number of parameters and computational complexity. The better performance was obtained in locating and then extracting the multi-scale features of diseases and pests. Experimental results demonstrated that the improved RCL-YOLOv11n model achieved a precision of 93.7%, a recall of 86.4%, an mAP0.5 of 92.8%, and an mAP0.5-0.95 of 71.6%. The number of parameters, computational complexity, and model weight size were 2.2 M, 5.1 G, and 5.0 MB, respectively. The mAP0.5 and mAP0.5-0.95 were improved by 1.8 and 1.4 percentage points, respectively, compared with the original YOLOv11n model, while the number of parameters, computational complexity, and model weight size were reduced by 15.4%, 19.0%, and 9.1%, respectively. Furthermore, the RCL-YOLOv11n-RT model achieved a single-image inference time of only 1.6 ms on the server side with an mAP0.5 of 91.8%, indicating excellent computational efficiency with high accuracy after optimization. The RCL-YOLOv11n model can be expected to effectively balance accuracy and lightweighting. The finding can also provide a valuable reference for real-time detection and lightweight deployment of rice diseases and pests.
UR  - https://doi.org/10.11975/j.issn.1002-6819.202511179
DO  - 10.11975/j.issn.1002-6819.202511179