Application of discrete orthonormal Stockwell transform and local neighborhood patterns for leaf disease classification in banana

Among fruit crops, banana is a significant fruit in agricultural production and trade. However, global banana production is severely affected by diseases, leading to substantial declination in the yield. Hence, timely detection and control of these diseases is necessary. Diagnosis of diseases by manual inspection of initial symptoms is quite challenging for farmers. This paper demonstrates a novel machine learning approach towards classification of three major foliar fungal diseases, Sigatoka, Deightoneilla and Cordana, using discoloration patterns appearing on the leaf blade. The images, after pre-processing and segmentation, are transformed to frequency domain using discrete orthonormal Stockwell transform (DOST). Feature vectors based on local neighborhood patterns such as local binary pattern (LBP), elliptic LBP (ELBP) and its variants, are extracted from DOST transformed images and classified using five prominent image classifiers. A comparative analysis of performance metrics through a ten-fold cross validation procedure is done. Classification accuracy of 95.9% is obtained for ELBP features when classified with artificial neural network classifier (ANN). The technique of integrating DOST with LBP based features has achieved significant accuracy compared to contemporary methods of disease classification in plants.