Study On Classification Method Of Rice Leaf Blast Based On UAV Hyperspectral

Rice blast is a disease caused by Botrytis cinerea in the whole growth period of rice.According to the time and location of rice blast,rice blast can be divided into seedling blast,leaf blast and ear blast,among which leaf blast is the most harmful.The existing control measures of rice blast are mainly realized through the application of"early prevention and large dosage"by farmers.This control measure has many disadvantages.The reason is the lack of a rapid,accurate and quantitative method to detect rice blast.Moreover,the research objects of the existing UAV hyperspectral rice blast detection methods are mostly ear blast.The incidence of ear blast is late,and its research significance has some limitations.Based on this,this study aimed at Rice Leaf Blast,took two different varieties of rice as the research object,and based on the data obtained from the experiment of artificially induced rice leaf blast.Firstly,the geographic location information of hyperspectral images was registered to obtain the accurate spectrum of diseased rice.Secondly,a variety of methods are used to preprocess the hyperspectral data of UAV,and the best preprocessing method is selected through classification modeling,so as to reduce the noise generated by the environment and experimental instruments.In order to eliminate the redundant bands of spectral data and improve the accuracy and efficiency of subsequent modeling and classification,three methods are used to reduce the dimension of hyperspectral data.Finally,the hyperspectral rice leaf blast detection model of UAV is constructed by three different classification and modeling methods.In order to explore the feasibility of using UAV hyperspectral detection of rice leaf blast,provide basis for accurate application of drugs,and provide reference for early detection and prediction of rice blast.The specific research contents are as follows:（1）In this study,SG smoothing（SG）,multiple scatter correction（MSC）and first derivative(1^st Der)were used to preprocess the original hyperspectral data,and partial least squares discriminant analysis（PLS-DA）was used to establish the full band detection model of rice leaf blast.It was concluded that the three preprocessing methods improved the accuracy of the model to varying degrees.Among them,1^stDer pretreatment method is the best.For Mongolian rice and Yanfeng 47 rice varieties,the accuracy of the classification model is 66.67%and 53.92%respectively.It shows that in the hyperspectral detection of rice leaf blast by UAV,1^st Der pretreatment method can eliminate the noise caused by instrument and environmental noise to a certain extent and improve the accuracy of the model.（2）In this study,the dimensionality of UAV hyperspectral data is reduced by extracting spectral features,screening spectral feature bands and constructing vegetation index.Through a variety of classification modeling methods,the optimal dimensionality reduction method is determined as competitive adaptive reweighting method（CARS）.Through cars,11characteristic bands of Mongolian rice leaf blast are selected The 14 characteristic bands of Yanfeng 47 leaf blast were 404.1,497.1,503.6,506.8,510.1,513.4,612.7,619.4,629.5,780.9,823.3nm and 506.8,522.8,596.0,616.1,660.0,663.4,683.9,697.3,704.4,749.5,763.4,766.9,819.7 and 833.9nm respectively.（3）In this study,partial least squares discriminant analysis,random forest and support vector machine were used to classify and model the disease grade of different varieties of rice leaf blast.The model classification results show that no matter what dimensionality reduction method is used,the spectral variables after dimensionality reduction are used as the model input,the accuracy of the classification and detection model of rice leaf blast disease of Mongolian rice varieties is better than Yanfeng 47,and the accuracy of the classification and detection model of leaf blast disease constructed by random forest modeling method is better than other modeling methods.Among them,the classification and prediction accuracy of characteristic band input random forest model obtained by competitive adaptive reweighting method for Mongolian rice and Yanfeng 47 varieties are 84.31%and 79.41%respectively,and kappa coefficients are 0.7982 and 0.7426 respectively.