中国农业科学 ›› 2022, Vol. 55 ›› Issue (8): 1557-1567.doi: 10.3864/j.issn.0578-1752.2022.08.007

• 植物保护 • 上一篇    下一篇

基于图像的水稻纹枯病智能测报方法

韩晓彤1(),杨保军2,李苏炫1,廖福兵1,刘淑华2,唐健2,姚青1   

  1. 1浙江理工大学信息学院,杭州 310018
    2中国水稻研究所水稻生物学国家重点实验室,杭州 311401
  • 收稿日期:2021-11-03 接受日期:2021-12-23 出版日期:2022-04-16 发布日期:2022-05-11
  • 联系方式: 韩晓彤,E-mail: 453116984@qq.com。
  • 基金资助:
    国家重点研发计划(2021YFD1401100);浙江省自然科学基金(LY20C140008);所级统筹基本科研业务费项目(CPSIBRF-CNRRI-202123)

Intelligent Forecasting Method of Rice Sheath Blight Based on Images

HAN XiaoTong1(),YANG BaoJun2,LI SuXuan1,LIAO FuBing1,LIU ShuHua2,TANG Jian2,YAO Qing1   

  1. 1School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018
    2State Key Laboratory of Rice Biology, China National Rice Research Institute, Hangzhou 311401
  • Received:2021-11-03 Accepted:2021-12-23 Published:2022-04-16 Online:2022-05-11

摘要:

【目的】目前水稻纹枯病测报依赖人工调查水稻发病丛数、株数和每株严重度来计算其病情指数,操作专业性强,费时费力且数据难以追溯。本研究提出基于图像的水稻纹枯病病斑检测模型和发生危害分级模型,为水稻纹枯病智能测报提供理论依据。【方法】利用便携式图像采集仪采集田间水稻纹枯病图像,研究不同目标检测模型(Cascade R-CNN和RetinaNet)和特征提取网络(VGG-16和ResNet-101)对水稻纹枯病病斑的检测效果,筛选出具有较好检测效果的模型。针对Cascade R-CNN模型检测纹枯病病斑存在漏检现象,根据纹枯病病斑呈现形状不规则、大小和位置多变的复杂情况,对Cascade R-CNN进行改进,添加OHEM结构均衡难易样本,选择边框回归损失函数,通过精准率、漏检率、平均精度和P-R曲线来评价不同模型的检测效果。在改进的Cascade R-CNN-OHEM-GIOU模型检测结果基础上,分别建立基于病斑面积和病斑数的水稻纹枯病丛发生危害分级模型,通过决定系数(R 2)和Kappa值筛选分级模型。【结果】在相同主干网络条件下,Cascade R-CNN模型较RetinaNet模型对水稻纹枯病具有更好的检测效果,其中Cascade R-CNN-ResNet-101目标检测模型效果最佳,病斑检测准确率为92.4%,平均精度为88.2%,但漏检率为14.9%。改进的Cascade R-CNN-OHEM-GIOU检测模型有效解决了样本不均衡问题,添加边框回归损失函数有效降低了漏检率,较Cascade R-CNN-ResNet-101模型降低8.7%,平均精度提高到92.3%。以人工分级结果作为标准,基于病斑面积的水稻纹枯病发生危害分级模型在0—5级分级准确率分别为96.0%、90.0%、82.0%、76.0%、74.0%和96.0%,平均分级准确率为85.7%,Kappa系数为0.83,基于图像的水稻纹枯病丛发生危害分级与人工分级结果具有较高的一致性。【结论】基于图像的水稻纹枯病智能测报方法可实现病斑自动检测和发生危害自动分级,提高了测报的智能化水平,结果客观且可追溯,也可为其他农作物病害智能测报提供参考。

关键词: 水稻纹枯病, 病斑图像, 智能测报, Cascade R-CNN模型, 危害分级模型

Abstract:

【Objective】At present, the forecast of rice sheath blight relies on the number of diseased clusters, the number of rice plants and the severity of each plant to calculate the disease index based on manual surveys. The method is highly professional, time-consuming and laborious. The data is difficult to trace. The objective of this study is to propose a detection model of rice sheath blight lesions and a damage grading model of rice sheath blight based on images, and to provide a theoretical basis for the intelligent forecasting of rice sheath blight.【Method】Images of rice sheath blight in paddy field were collected by a portable image acquisition instrument. Different detection models (Cascade R-CNN and RetinaNet) and feature extraction networks (VGG-16 and ResNet-101) were developed to test the detection effect of disease lesions. The best model was chosen. However, the Cascade R-CNN model appeared some missing detection of sheath blight lesions. Because the sheath blight lesions are irregular in shape, and variable in size and location, the Cascade R-CNN model was improved through adding OHEM structure to balance the hard and easy samples in the network and choosing the bounding box regression loss function. The precision rate, missing rate, average precision and P-R curve were used to evaluate the detection effects of different models. Based on the detection results of the improved Cascade R-CNN-OHEM-GIOU model, two damage grading models based on the area and number of disease lesions were developed, respectively. The determination coefficient (R2) and Kappa value were used to choose the damage level model of rice sheath blight. 【Result】Under the same backbone network conditions, the Cascade R-CNN model had a better detection effect on rice sheath blight than the RetinaNet model. The Cascade R-CNN-ResNet-101 model had the best detection effect on sheath blight lesions. The precision rate was 92.4%, the average precision was 88.2% and the missing rate was 14.9%. The improved Cascade R-CNN-OHEM- GIOU model effectively solved the problem of sample imbalance, and effectively reduced the missing rate by adding a border regression loss function, which was 8.7% lower than the missing rate of the Cascade R-CNN-ResNet-101 model, and the average precision was increased to 92.3%. With the results of manual disease grading as the standard, the grading model of rice sheath blight at 0 to 5 grades based on the area of diseased lesions had the accurate rates of 96.0%, 90.0%, 82.0%, 76.0%, 74.0% and 96.0%, respectively. The average grading accuracy rate was 85.7%, and the Kappa coefficient was 0.83. The damage grade results of rice sheath blight based on images were consistent with the manual grading results.【Conclusion】The intelligent forecasting method of rice sheath blight based on images can automatically detect the disease lesions and calculate the damage grade. This method increases the intelligence level and the results are objective and traceable. It may also provide an idea for intelligent forecasting of other crop diseases.

Key words: rice sheath blight, disease lesion image, intelligent forecasting, Cascade R-CNN model, damage grading model