Enhanced Segmentation and Ensemble Classification for Accurate Plant Disease Detection

Abstract

The majority of the crops are wasted owing to deficiency of transport, plant diseases and lack of storage facilities. Above 15% of crops are worn out in India owing to diseases and therefore it has turned out to be a main concern to be solved. This study introduces an advanced framework for plant disease detection by integrating enhanced image segmentation techniques with robust ensemble classification models. Our methodology begins with the pre-processing of plant leaf images using median filtering and Wiener denoising to reduce noise and enhance image quality. As the next step, the Improved Region Growing Algorithm (IRGA) is deployed for the segmentation of images. Then, features together with ‘Scale Invariant Feature Transform (SIFT), improved Binary Gabor Pattern (IBGP), Haralick features, color features like RGB Color Histogram, disease area and higher order statistical features (Entropy, Skewness, variance and kurtosis)’ are extracted. The improved independent component analysis (IICA) model is then used to choose the best attributes. Lastly, detection takes place using Ensemble classifiers (EC) including Neural Network (NN), modified effective squeeze and excitation block-based deep convolutional neural network (M-ESE-DCNN) and bi-directional gated recurrent unit (BI-GRU). Further, the DCNN weights are optimised via the Colliding Archimedes and Teamwork Algorithm (CA-TWA) model. For the best case with dataset 1, EC + CA-TWA got a high accuracy of 0.94, while EC + BOA, EC + DOX, EC + SSO, EC + TOA and EC + ArOA had lower accuracy. Furthermore, for all schemes, dataset 1 displays superior outputs to dataset 2 and dataset 3. Finally, an evaluation is done to validate this work.