大田环境下的马铃薯早疫病叶片检测和严重程度评估  

Detection and Severity Assessment of Potato Early Blight Ieaves in a Field Setting

作　　者：孟祥国 冯全 MENG Xiang-guo;FENG Quan(School of Mechanical and Electrical Engineering,Gansu Agricultural University,Gansu Lanzhou 730070,China)

机构地区：[1]甘肃农业大学机电工程学院,甘肃兰州730070

出　　处：《林业机械与木工设备》2024年第9期44-53,共10页Forestry Machinery & Woodworking Equipment

基　　金：国家自然基金(32160421);甘肃省教育厅产业支撑项目(2021CYZC-57)。

摘　　要：马铃薯早疫病的病害严重程度监测对于马铃薯的产量影响非常重要,根据早疫病病害严重度调整农药使用,可以有效降低使用成本,避免因农药使用过量对马铃薯造成伤害。而田间病害监测设备需要轻量化模型的部署,为了解决病害评估网络因硬件计算资源有限,在田间设备上难以部署、运行速度慢等问题,实现对田间场景下的马铃薯早疫病病害严重程度的田块级评估,提出了一种针对马铃薯早疫病检测与严重度评估的模型。模型主要包括基于轻量级检测器CS-YOLO的叶片识别、结合DeepSORT跟踪算法的叶片碰线计数、以及基于BP网络对早疫病发病严重度的回归预测。试验结果表明,评估模型在马铃薯叶片(健康与早疫病)检测任务上的参数量相较于YOLOv5s减少了90.03%,准确率达到62.35%;叶片碰线计数的准确度能够达到0.686,能够较好地捕捉早疫病叶片;BP神经网络对早疫病发病严重程度的预测拟合效果较好,决定系数R^(2)达到0.81。为植保机器人在田间场景中的部署提供了一种有效的技术手段,为实现马铃薯早疫病的监测和实时评估提供了可行的解决方案。The monitoring of potato disease degree is very important for the impact of potato yield,and the equipment for field disease monitoring needs to be deployed with lightweight models.In order to address issues such as limited hardware computing resources leading to the hard deployment and slow running speed of disease assessment networks in field devices,this paper proposes a real-time model to assess the severity of potato early blight disease in field scenarios.The model includes the early blight leaf recognition method based on a lightweight detector,and combines with the line counting via DeepSORT tracking algorithm,and a BP neural network for identifying the number of diseased and healthy leaves on the canopy level.Experiment shows that CS-YOLO reduces the parameter count by 90.03%compared to YOLOv5s in the task of early blight detection in potatoes,achieving an accuracy of 62.35%.The accuracy of line counting is found to be 0.686,which means early blight potato leaves could be effectively captured in detection and tracking tasks.The result also shows the BP neural network delivers a fitting result of predicting the disease severity,with a coefficient of determination R^(2)reaching up to 0.81.This method provides an effective detection means for the deployment of agricultural protection robots in field scenarios,offering a viable solution for monitoring and real-time assessment of potato early blight disease.

关 键 词：植物保护 马铃薯早疫病 轻量化 目标跟踪 病害评估 

分 类 号：S435.32[农业科学—农业昆虫与害虫防治]