Automated assessment for early and late blight leaf diseases using extended segmentation and optimized features

Abstract

Early and late blight diseases lead to substantial damage to vegetable crop productions and economic losses. As a modern solution, machine learning-based plant disease assessment aims to assess the disease incidence and severity through the disease region of interest (ROI) and its extracted features. In the case of existing conventional classifier methods, extracting the features involves generalized ROI segmentation that loosely follows the disease inference. As a result, accuracy is reduced, and the fuzzy boundary region that carries potential properties for improving feature characterization capability is truncated from the ROI. Besides, most of the existing practices extract only the global features, This leads to redundant and extensive feature vector, which causes increased complexity and underperformance. Furthermore, individual lesion severity is not considered in the assessment. This thesis addresses the issue of the ROI segmentation by using color thresholding based on ratios of leaf green color intensity to incorporate the fuzzy boundary region, denoted as extended ROI (EROI). Secondly, the issue of the feature extraction is addressed by the proposed localized feature extraction method to reduce complexity and improve disease classification performance. Based on the color and texture morphological properties of the individual lesions within the EROI, color coherence vector and local binary patterns features are extracted. As a result, a pathologically optimized feature vector is obtained, which is used to build a support vector machine classifier to classify between the disease types of early blight, late blight, and healthy leaves. lastly, a 2-tier assessment is proposed. The disease type classification is given as the first tier, while the leaf lesion area ratios of the individual lesions are given as severity quantification for the second tier. Overall, the proposed EROI segmentation method reduced under-segmentation by up to 80%. The proposed optimized feature reduced the execution run-time by up to 50% and achieved an average classification performance of up to 99%. Finally, the quantified severity is in close agreement with the ground truth by achieving an average accuracy of 93%.