| Stripe rust is one of the important diseases that affect wheat yield and threaten food security.In order to fundamentally solve the harm caused by wheat stripe rust and ensure the high and stable yield of wheat,it is necessary to cultivate new wheat varieties with stripe rust resistance by using crop breeding technology according to the target cultivation environment.In the breeding for wheat stripe rust-resistance,phenotyping is a crucial step in the screening and identification of genotypes.However,the breeding population is often composed of hundreds of genotypes,while the traditional phenotyping method through artificial vision is single and inefficient.In order to solve the above problems,this study takes the breeding population wheat infected with stripe rust(a total of 600 samples and 516 genotypes)as the research object.The canopy image data of the experimental population are collected by UAV,and the vegetation index is extracted after the preprocessing operations of image mosaic,radiation correction,geometric correction,image segmentation and image cropping.Through machine learning,feature selection,time series analysis and other methods,the dynamic changes of wheat infected with stripe rust are accurately quantified to realize the dynamic analysis of stripe rust in wheat population.It provides detailed,rich,accurate and efficient phenotypic data for wheat stripe rust resistance breeding.This study proposed a new method for the field phenotype investigation for wheat stripe rust-resistance breeding,and also provided a reference for the phenotyping of stress resistance breeding of other crops.The specific contents and main conclusions of this study are as follows:(1)The vegetation index collected and extracted by UAV RGB and multispectral imaging technology could realize the dynamic monitoring of wheat stripe rust.The highest cross validation F1-score of the machine learning classification model for the disease stage and disease severity of wheat stripe rust established with the vegetation index as the input features are 0.9612 and 0.7284,indicating that the visible light and multispectral vegetation index collected by UAV could describe the change of the incidence of wheat stripe rust with time;In terms of modeling accuracy and algorithm calculation time,support vector machine(SVM)algorithm is more suitable for the construction of stripe rust disease stage and stripe rust disease severity classification model;Compared with RGB imaging technology,we find that multispectral imaging has greater advantages in realizing the dynamic analysis of the incidence of wheat stripe rust for it has more bands and could carry out radiation correction.Considering the factors such as economy and convenient operation,the RGB imaging technology has certain application value in improving the efficiency of field investigation on the severity level of wheat stripe rust in the later stage of the occurrence of stripe rust in wheat population.(2)Using the responses of CIrededge,CIgreen,GARI,NDVI,NPCI and WI as indicators,the incidence of stripe rust of different wheat varieties(lines)in the experimental population could be dynamically quantified more finely.Feature selection algorithm was used to screen the sensitive characteristics of stripe rust disease stage and disease severity;The SVM classification model of stripe rust disease stage and disease severity based on sensitive features has little difference in the test set,which proved the effectiveness of the screened sensitive features;However,combined with the general law of stripe rust incidence,although some indexes could improve the classification accuracy,they could not give a reasonable explanation from physiological and biochemical changes after wheat infected with stripe rust,so it is not appropriate to use them as the dynamic quantitative indicators of stripe rust incidence in the experimental population and individuals in the population.(3)The resistance grade of wheat stripe rust could be identified,using the time series of vegetation index response collected and extracted by UAV.Based on the two-dimensional images transformed from NPCI GARI,NDVI and WI response time series through Gramian Angle Field,the maximum F1-score of the deep learning classification model in the test set could reach 0.840;Among them,the classification model trained by the images transformed by NPCI and WI response time series is better than the other two;In terms of modeling accuracy,model storage size and the time used to process a single image,the comprehensive effect of Densenet121 obtains the best in the deep learning network selected in this study when the classification feature is the two-dimensional image transformed form the time series of NPCI response,which shows that the feature extraction method of densely connection blocks to realize information reuse proposed by this network is more suitable for the classification of stripe rust resistance grade. |
