Analysis for stripe rust dynamics in wheat population using UAV remote sensing

Breeding for wheat stripe rust-resistance could effectively reduce the harm caused by stripe rust. Accurate and efficient phenotyping is an essential step in crop breeding. Furthermore, UAV multispectral imaging had been used for crops growth monitoring, pathological phenotyping, biochemical index estimation and yield prediction widely for it could overcome the limitation of traditional phenotyping methods which has low efficiency and accuracy. In order to solve the problems of poor means and low efficiency of phenotyping of stripe rust in wheat, a dynamic high-throughput phenotyping framework for field stripe rust in population wheat using the UAV multispectral imaging was proposed. The experiment was carried out in Caoxinzhuang experimental farm, Yangling, Shaanxi Province, China. The breeding wheat population with 516 genotypes was used as experimental materials and planted into 600 plots according to the augmented design. Using UAV low altitude remote sensing and multispectral imaging, canopy multispectral images of experimental population wheat infected with stripe rust naturally were collected. At the same time, the data of disease stage and disease severity of stripe rust were recorded according to the disease time and the traditional manual identification standard of stripe rust. After image mosaic, radiometric correction, geometric correction, segmentation of region of interest and background, image clipping and other operations are carried out on the UAV remote sensing raw data, 22 vegetation indices were calculated to analyze. Using 22 spectral vegetation indices as classification features, we firstly established classification models of disease stage and disease severity in stripe rust, and screen out the vegetation indices sensitive to stripe rust disease stage and disease severity through random forest algorithm. At the same time, the random frog algorithm is also used to filter the original features to reduce the contingency caused by using the random forest algorithm to realize the feature screening merely. Using vegetation index with the selected probability ranking of about the top 1/3 according to the random frog algorithm as the input feature, the Support Vector Machine (SVM) algorithm was used to establish the classification model of disease stage and disease severity level in stripe rust, so as to verify the effectiveness of the features screened by the random frog algorithm. Combining the results of two feature selections, three vegetation indices sensitive to disease stage and disease severity were selected, and the differences of disease dynamics of six reference varieties in the population were analyzed based on the time series of response of these indices. The results showed that 1) for disease stages classification of stripe rust, the F1-score of the random forest classifier based on 22 original features and the SVM classifier based on the features sensitive to disease stage selected by random frog were 0.970 and 0.985 respectively, which proved the sensitive features selected by both random forest and random frog are reasonable and the two machine learning classifiers can be used in the classification of the disease stage of stripe rust in wheat population; 2) for disease severity classification of stripe rust, the F1-score of the random forest classifier based on 22 original features and the SVM classifier based on the sensitive features selected by random frog were 0.741 and 0.780 respectively. Although they are lower than the classifier for the disease stage of stripe rust, they have also achieved relatively satisfactory results, which also proved the features sensitive to disease severity selected by both random forest and random frog are reasonable and the two machine learning classifiers can be used in the classification of the disease severity of stripe rust in wheat population; 3) the response of Difference Vegetation Index RedEdge (DVIRE), Normalized Difference Vegetation Index (NDVI) and Normalized Difference Vegetation Index rededge (NDVIrededge) were more sensitive to both disease stage and disease severity in classification models. Especially, DVIRE was selected as the most sensitive feature to disease stress, and can describe both disease stage and severity of stripe rust in field meanwhile. So, dynamic phenotyping by using multitemporal vegetation indices extracted from UAV multispectral images may accurately and finely quantify stripe rust stress state in wheat population and could be considered as a reliable high-throughput method. 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.