Images-based Intelligent Forecasting Methods For Rice Sheath Blight

China is a large rice planting country,and the serious and frequent occurrence of rice sheath blight has caused immeasurable losses to China’s rice yields and economy.Real-time and accurate forecasting of rice sheath blight can effectively control rice sheath blight and reduce pesticide application.Due to the problems of high labor intensity,poor objectivity,low efficiency,non-traceability,and non-real-time nature in manual field surveys,it cannot meet the requirements of agriculturalal development to intelligent and informatization.In order to truly free the forecasters from the survey tasks,it is urgent to establish an intelligent forecasting method for rice sheath blight.This paper studies the image-based intelligent forecasting method for rice sheath blight,proposes an improved Cascade R-CNN automatic detection model for rice sheath blight,and combines the detection model to establish a rice sheath blight disease grading model based on the characteristics of the disease spot area,The main research methods and results include:(1)Research on different detection algorithms of rice sheath blight based on deep learning.Two algorithms of single-stage Retina Net and two-stage Cascade R-CNN are applied to the detection of rice sheath blight.With VGG-16 and Res Net-101 as the backbone network,two models are trained for each of the Cascade R-CNN and Retina Net algorithms,and a total of four models are established.Experiments show that under the same backbone network conditions,Cascade R-CNN has a better detection effect against rice sheath blight.Among them,the Cascade R-CNN-Res Net-101 object detection model has the best effect,with an accuracy rate of92.4% and missed detection rate of 14.9%.Therefore,this paper selects the Cascade R-CNN object detection algorithm as the basic algorithm for detecting rice sheath blight.(2)Research on rice sheath blight detection algorithm based on improved Cascade R-CNN.The rice sheath blight disease spot area occupies a small area in the image,and the location and size are changeable.The missed detection rate of original model is obviously higher.In response to the above problems,this article improves on the basis of Cascade R-CNN,adds an OHEM structure to the network to balance difficult and easy samples,and uses the improved bounding box regression loss function to further improve the accuracy of rice sheath blight detection and reduce the missed detection rate.The missed detection rate of R-CNN-OHEM-GIOU model is6.2%,and AP is 92.3%.The improved Cascade R-CNN model further improves the detection effect of rice sheath blight.(3)Research and realization of rice sheath blight disease grading model.After the improved Cascade R-CNN model is obtained,by studying the correlation between the characteristics of rice sheath blight spots and the classification of the disease,the characteristics of the spots used for the classification model are determined.Select the characteristics of lesion area and number of lesions,fit the disease grade based on the regression analysis,and then calculate the weight coefficients of different lesion characteristics after fitting.By comparing the coefficient of determination,select the characteristic of lesion area as the basis for grading,and establish the rice disease grading model based on the characteristics of disease spot area,the test results of the grading model and the artificial recognition results are compared,the average grading accuracy is 80% and the Kappa coefficient is 0.75,showing high consistency.In this paper,an improved Cascade R-CNN detection model for rice sheath blight is proposed,and a rice sheath blight disease grading model based on the disease spot area is established based on the detection model.The model has a high consistency with the manual diagnosis data.It is helpful for intelligent forecasting of rice sheath blight.