Misclassification and cluster validation techniques for feature selection of diseased rice plant images

Abstract

Damages of rice crops can be averted by taking corrective measures at an early stage based on the classification of diseases of rice plants. The paper aims at developing an appropriate data mining methodology to extract knowledge about the characteristics of diseases by analyzing the images acquired from the field. Since all features are not equally involved in classifying diseases; selection of optimum features is a challenging task to address the problem. The work is performed in three steps. Firstly, thirty six features of different category are extracted from the diseased plant images using image processing techniques. Then, images with respect to each attribute are clustered by k-means algorithm where k is the number of class labels of the diseases. Cluster validity indices are computed using the proposed and existing techniques. Finally, all the computed indices are fused and based on that score of the attributes is calculated. The attributes having scores greater than average score are considered as optimal feature set. With the reduced feature set, classification accuracy is calculated by applying the proposed method on four hundred fifty infected rice plant images and compared with other classifiers demonstrating effectiveness of the proposed model. © 2012 Springer-Verlag GmbH Berlin Heidelberg.