Investigating the accuracy of classification in unbalanced data in order to diagnose two common potato leaf diseases (early blight and late blight) using image processing and machine learning

Abstract

AbstractThis study aimed to classify potato disease as early blight, late blight, and healthy leaves using color image processing techniques, data imbalance and feature extraction techniques. To this end, two groups of potato disease leaves with similar symptoms and one group of healthy leaves were considered. A new method based on increasing or decreasing image data using Over-Sampling and Under-Sampling techniques was used to balance them. For this purpose, after separating the diseased area from the leaf surface, the features were extracted. In total, 45 color features, 99 texture features were extracted from each of the three color spaces RGB, l*a*b and HSV, and 6 shape features were extracted from the images. Then classification was done by Random Forest. The classification accuracy results for three proposed models including classification with Original, Over-Sampling and Under-Sampling dataset were 87.89, 91.09 and 96.17%, respectively. Also, the most effective features extracted for the diagnosis of potato diseases were selected by the Relief feature selection algorithm. The results show that texture features contribute more to disease classification and data balancing techniques can increase classification accuracy. In addition, the results show that features extracted from different color spaces can improve disease diagnosis performance and help data engineers to increase the amount of features extracted and compare their performance. This study can be used in future research to classify potato diseases and other plant diseases, and its results can help researchers in their efforts.