Detection and Classification of Rice Plant Diseases Using Fusion Deep and Texture Features

This thesis demonstrates an intelligent system for the early detection of plant diseases using deep learningbased image processing. The primary objective is to support food security and enhance agricultural productivity. The proposed system utilizes Contrast-Limited Adaptive Histogram Equalization (CLAHE) and a U-shaped convolutional Neural Network (U-Net) for image segmentation, aiming to enhance image quality and accurately localize diseased regions. The feature extraction includes integration of deep features of Convolutional Neural Networks (CNN) and the Residual Network-18 architecture-based model GramNet together with the texture feature extracted from the graylevel co-occurrence matrix. All features are concatenated to create a comprehensive feature representation. A stacking ensemble approach is adopted for classification, utilizing support vector machines, K-nearest neighbors, and extreme gradient boosting as base classifiers. In contrast, the light gradient boosting machine is the final meta-classifier. An experimental evaluation was conducted on a publicly available rice disease dataset containing 11,790 samples, which demonstrates that the model achieves 99.99% accuracy on the training data and 96.90% on the testing data. This work demonstrates the effectiveness of integrating preprocessing, hybrid feature extraction, and ensemble learning in the early detection of plant diseases.