Research on Lightweight Detection and Identification of Tobacco Diseases Based on RT-YOLOv10 and Drone Remote Sensing Images

Chen Zili; Guo Yan; Wang Mingxin; Lin Wei; Wang Laigang; Yang Xiuzhong; Liu Jianjun; Zheng Hengbiao; Wang Aiguo

doi:10.14083/j.issn.1001-4942.2025.09.016

Abstract：To address issues such as large scale differences of diseased plants,complex backgrounds,and low resolution in tobacco drone remote sensing images,this study proposed a lightweight disease detection algorithm RT-YOLOv10 based on YOLOv10,which was used for high-precision real-time monitoring of tobacco diseases in drone images.First,RFAConv convolution and SimAM parameter-free attention were introduced to optimize the downsampling module,improving the accuracy of feature extraction.DySample upsampling was incorporated to more accurately recover effective information for disease identification,such as color,texture,and edges.Meanwhile,a dual-branch collaborative dense connection neck structure was designed to promote the interaction of multi-level semantic information and enhance the network's ability to learn and express fea-tures of different types of diseases.Second,the improved MS_C2f and MS_CSP modules were used to replace the C2f modules in the backbone and neck networks,respectively,which enhanced the network's capability to capture multi-scale features of diseased plants with different sizes and shapes,as well as the efficiency of fea-ture fusion.Additionally,the attention mechanisms embedded inside these modules focued on key information,highlighted the main targets of diseased plants,and reduced false detections and missed detections caused by interference from complex backgrounds such as weeds and light spots.Finally,the original bounding box loss function was replaced with MPDIoU to further improve the accuracy of disease identification.Experimental re-sults showed that the standard version of RT-YOLOv10,RT-YOLOv10-s,had lower parameter count of 7.4 M and floating-point operations(FLOPs)of 33.3 GFLOPs,achieved 97.2%,93.6%,97.5%,and 83.9%in precision,recall,mean average precision(mAP50),and mAP50-95,respectively,outperforming compara-tive models such as YOLOv10-s,YOLOv9-s,and YOLOv8-s.Although the more lightweight version(RT-YOLOv10-t)showed slight decreases in these metrics,it still reached 95.5%,92.4%,97.2%,and 79.7%in precision,recall,mAP50,and mAP50-95,respectively.Moreover,its parameter count was only 2.3 M,which was 67.1%,86.6%,61.7%,79.3%,76.3%,and 71.6%lower than that of YOLOv5-s,YOLOv6-s,YOLOv7-Tiny,YOLOv8-s,YOLOv9-s,and YOLOv10-s,respectively.In conclusion,RT-YOLOv10 could meet the requirements of practical applications and could be deployed on edge devices with extremely limited computing resources,providing an effective method for large-scale and accurate monitoring of tobacco diseases.