基于二阶段深度模型的马铃薯晚疫病严重程度分级研究

Severity Evaluation of Potato Late Blight Based on Two-stage Depth Modeling

摘要:

晚疫病是马铃薯生长过程中最具灭绝性危害的病害,精确量化马铃薯晚疫病严重程度是开展晚疫病防控的数据基础。传统的量化方式极度依赖于人工对马铃薯离体叶片面积及病斑面积的统计计算来划分严重等级,但面对海量病害叶片的图像数据时,人工量化的效率和准确度均难以满足实验需求。为解决上述问题,本研究提出一种二阶段深度学习模型YOLO-PIDNet,对马铃薯晚疫病的严重等级进行批量自动分级。模型第1阶段使用改进后的YOLO v8n模型对图像中每枚马铃薯离体叶片进行分割和定位,模型第2阶段使用改进后的PIDNet模型对图像中每枚马铃薯离体叶片上的晚疫病病斑进行精准分割,计算图像中每枚离体叶片上病斑区域像素点数量和该离体叶片区域像素点数量的百分比,依照晚疫病严重程度分级标准自动划分病害严重等级。测试结果显示方法第1阶段模型的mAP50-95、内存占用量、浮点运算量和帧率分别为95.82%、6.8 MB、7.8 × 10?和142.08 f/s,第2阶段模型的像素准确率、Dice系数和平均交并比分别为99.70%、0.843 3和74.55%。上述结果表明,本研究提出的二阶段模型可以实现对马铃薯离体病叶快速、有效地分割与分级,具有轻量化、可迁移性高的特点,可为马铃薯晚疫病精准防治提供技术支撑。

Abstract:

Late blight is the most devastating disease in potato production, and precise quantification of its severity level serves as the data foundation for disease control. Traditional quantification of potato detached leaf area and lesion spot area relies heavily on manual statistics and calculations to determine disease severity levels. However, when dealing with a massive amount of disease leaf image data, manual quantification falls short in both efficiency and accuracy, failing to satisfy experimental requirements. To address these challenges, a two-stage deep learning model named YOLO – PIDNet was proposed for automated batch severity grading of potato late blight. Stage 1 employed an enhanced YOLO v8n model to achieve segmentation and localization of individual detached potato leaves. Stage 2 employed an improved PIDNet architecture to achieve precise segmentation of late blight lesions on each leaf. The model calculated the pixel ratio between lesion areas and total leaf regions, automatically classifying disease severity according to standardized criteria. Experimental results demonstrated that the stage 1 model achieved 95.82% mAP50-95 with 6.8 MB parameter size, 7.8 × 10? FLOPs computational complexity, and 142.08 f/s inference speed. The stage 2 model attained 99.70% pixel accuracy, 0.843 3 Dice coefficient, and 74.55% mean intersection over union. The above experimental results indicate that the two-stage model proposed in this study can achieve rapid and effective segmentation and classification of the detached diseased potato leaves. It is lightweight and highly transferable, and can provide technical support for the precise control of late blight for potatoes.