EPIDEMIOLOGY AND ECOLOGY
LIANG Jiahui, JIANG Qian, WANG Hongli, WANG Haiguang
Stripe rust (yellow rust) caused by Puccinia striiformis f. sp. tritici, leaf rust caused by P. triticina, and powdery mildew caused by Blumeria graminis f. sp. tritici are important wheat leaf diseases, which seriously affect the yield and quality of wheat. Timely identification and diagnosis of the diseases are very important for their effective management. Now it is more convenient to acquire plant disease images, but it is difficult to unify image sources in actual disease identification, thus it is very important to explore the effects of images from different sources on image-based plant disease identification. In this study, a total of 1 354 images of wheat stripe rust, wheat leaf rust and wheat powdery mildew (including individual disease images and mixed-infected leaf images) were obtained in the field by using three different image acquisition devices including a smartphone iPhone 5s, a smartphone iPhone 11, and a digital camera Nikon D700. The single lesion images of wheat stripe rust were obtained by using the lasso tool in the software Adobe Photoshop CS6 in combination with the slice tool, and the single lesion images of wheat leaf rust and wheat powdery mildew were obtained by using the slice tool in combination with the K-means clustering algorithm. A total of 9 479 segmented single lesion images were obtained, and then from them, 164 color, shape, and texture features were extracted. By using ReliefF and clustering feature selection methods, a feature combination consisting of 36 features was determined for building disease recognition models. Four training sets, three composed of the images from the three individual sources, respectively, and one composed of the images from the three different sources, were constructed. Based on each training set, disease recognition models were built by using six modeling methods including support vector machine (SVM), K-nearest neighbor (KNN), Gussian naive Bayesian, multilayer perceptron (MLP), label propagation algorithm, and self-training semi-supervised algorithm, with the optimal parameter/parameter combination obtained by using the random search method. Accuracy, precision, recall, and F1 score were used to evaluate the disease recognition performance of the built models. The results showed that, for the SVM, MLP, and self-training semi-supervised models built based on each training set consisting of the individual-source images, high accuracies, precisions, recalls, and F1 scores were achieved on both the training set used for modeling and the testing set with the same image source as the training set used for modeling, however, the corresponding values achieved on the two other testing sets with the different image sources as the training set used for mode-ling reduced, except that the accuracy, precision, recall, and F1 score of the testing set with the same image source as the training set used for modeling and the corresponding values of the testing set consisting of the images acquired by using the smartphone iPhone 11 were not much different for the self-training semi-supervised model built based on the training set consisting of the images acquired by using the smartphone iPhone 5s. Based on the training set consisting of the images from the three different sources, satisfactory recognition performances were achieved by using the built SVM, MLP, and self-training semi-supervised models; acceptable recognition performance was achieved by using the built KNN model; and the built SVM model was optimal, with the accuracies, precisions, recalls and F1 scores not less than 91.31% on all the testing sets. The results indicated that different image sources can affect image-based disease recognition and demonstrated that satisfactory recognition performance on multi-source images can be achieved by using the model built based on the training set consisting of images from different sources. In this study, some basis was provided for further studies on the automatic and intelligent recognition of wheat diseases.