Research On Identification Method Of Verticillium Wilt Of Cotton Based On UAV Images

Cotton is one of the most important cash crops in our country,most provinces and regions are planted.Cotton disease has always been one of the main factors limiting cotton production,and the more serious the disease,the greater the loss of yield.Verticillium wilt is the most serious and widely distributed cotton disease,known as "cancer of cotton".When verticillium wilt appears in cotton,it is easy to reduce or even fail harvest.The traditional cotton disease identification method is mainly through manual sampling investigation,and the disease grade is determined by the proportion of infected leaves and canopy in the cotton field.This traditional method of disease identification method consumes time and energy,and the method has lag.Therefore,it is of great significance to study and propose an automatic identification method for Verticillium wilt of cotton.UAV agricultural remote sensing technology has the characteristics of high spatial resolution,high timeliness and low cost of remote sensing image,which plays an important role in the application of crop disease detection.In this paper,we use the Parrot Bluegrass to collect the RGB and multispectral images of cotton fields.The affine transform band registration method based on ground control points is used to register multispectral images and RGB images.And on that basis,EfficientDet network was used to automatically detect Verticillium wilt leaf in cotton field,and then extract color component and spectral reflection value of RGB image.The classification of Verticillium wilt leaves with different disease severity was realized by machine learning method.This paper mainly completes the following work:(1)The image data of UAV is preprocessed.Image correction mainly includes the color correction and spectral correction.For the problem of color deviation in RGB image,the pixel values in the areas of black,white and gray color correction plate are extracted,and then the pixel values obtained in the standard environment are used for polynomial regression analysis,so as to achieve the color correction of RGB image.At the same time,to solve the problem of spectral reflectance deviation of multispectral image,the pixel values in the area of the spectral correction plate are extracted,and then the input light intensity is calculated by combining the real reflectance information,so as to realize the spectral reflectance correction.(2)Multispectral image registration of unmanned aerial vehicles.Parrot is equipped with an array structure multispectral camera.When the multispectral image is used to analyze the information of different bands in the same cotton field area,it is necessary to register the images of different bands to solve the problem of the corresponding pixel coordinate position deviation in the region.An affine transform band registration method based on ground control points is proposed.This method can effectively reduce the registration error between RGB and multispectral images at different resolutions,so as to achieve the registration target of images with different bands with higher accuracy.(3)Detection of cotton verticillium wilt leaves.The deep learning method was used to detect Verticillium wilt leaves.The Efficientdet and Yolov5 network model were utilized to adjust the configuration of network parameters and test the model performance.The test results showed that the method based on EfficientDet-D2 target detection model has a better performance in Verticillium wilt detection and has a high recognition rate of 91.99 % for Verticillium wilt leaves.(4)Classification algorithm of cotton verticillium wilt leaves.The color features in RGB images and spectral reflection values in multispectral images were extracted as input features of Support Vector Machine(Support Vector Machine,SVM)and Random Forest(Random Forest,RF),The input features were used to construct a leaf disease region identification classifier to identify the disease spots,healthy areas and land areas in the frame diagram of diseased leaves.The disease classification was realized by the area of the disease spot and the ratio of leaf area.The test data were used to test the performance of leaf disease region recognition classifier.The Verticillium wilt disease classification model with color features and spectral features as input was able to carry out high classification recognition on the data set.The classification accuracy based on SVM reached 92.6%,which was able to better identify all levels of Verticillium wilt leaves.The identification accuracy of second class Verticillium wilt was 92.9%.The experimental results show that the classification algorithm based on support vector machine is more suitable for the classification of verticillium wilt leaves,and can be used for the classification of verticillium wilt leaves.