Identification of peach leaf disease infected by Xanthomonas campestris with deep learning

Abstract

This paper utilizes convolutional neural network (CNN) to identify peach leaf disease infected by Xanthomonas campestris. Transfer learning was used to fine-tune AlexNet. Feature visualization from the trained CNN indicate the excellent ability of self-learned features. Three comparative experiments were conducted to compare the performance of CNN with the traditional classification methods including Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Back Propagation (BP) neural network in identifying peach leaves. Confusion matrix of each method displays that CNN can identify the peach leaves affected by Xanthomonas campestris with the accuracy of 100%. ROC (Receiver Operating Characteristic) curves and AUC (Area Under ROC Curve) values, an overall performance measurement, show that CNN achieves higher performance with AUC value of 0.9999. The test of significant experiment shows that CNN is significantly superior to the other three mentioned methods, which the p-values is 0.0343 (vs.SVM), 0.0181 (vs.KNN) and 0.0292 (vs.BP). In a word, CNN is superior to the state-of-the-art in identifying diseased peach leaves.