Automatic vegetable disease identification approach using individual lesion features

Abstract

This paper proposes an automated approach for plant disease identification at leaf surface level using optimized features extracted from pathologically localized disease lesion areas. The automatic identification is achieved using machine learning classifiers built on the features. The disease region is mainly characterized as the combination of two symptom regions, the chlorotic and necrotic lesions that convey distinct characterization patterns. Typical machine learning approaches extract these patterns globally from the disease region as a whole region of interest (ROI), which often results in 10 s to 100 s of redundant features in the final feature vector that degrades the identification performance. Consequently, other methods resolve to use complex optimization algorithms to perform feature selection, which further compounds on the method complexity often without meaningful improvement. This study presents a localized feature extraction method from the individual chlorotic and necrotic lesions minimizing feature redundancy and vector size. Color coherence vector (CCV), a feature that portrays distinct homogeneous patterns relative to the disease progression are extracted from the chlorotic region. On the other hand, local binary pattern (LPB) is extracted from the necrotic region. Concatenating these individual lesion features results in a pathological feature vector for disease identification, thus minimizing feature size and avoiding the potentiality of dealing with descriptors that are superficial. To test the effectiveness of the proposed method, different conventional classifiers are used to test the quality and efficiency of these features in accurately classifying the plant diseases. The efficacy of the proposed approach is confirmed with the achieved high precision and recall rates of over 99% in all cases, and an improved accuracy compared to other reported works. The proposed work in this paper finds application in simplified decision support systems for the automation of plant disease identification and other resource management practices in the field of precision agriculture.