Improved of Rice Disease Image Classification Approach using Hybrid Deep Learning Architectures

Abstract

newlineRice is a staple food for billions, yet its cultivation is threatened by leaf diseases like blast, blight, and brown spot, leading to reduced yields and poor grain quality. Traditional diagnosis methods are manual, error-prone, and hard to scale. This thesis introduces a novel hybrid deep-learning model that combines DenseNet, vision transformers, and lightweight CNNs to enhance feature extraction and classification accuracy. Designed for real-world variability, the system improves reliability and efficiency in disease detection offering a scalable, automated solution that supports timely intervention in precision agriculture. newlineThis thesis introduces a novel hybrid deep-learning system for rice-leaf disease recognition, combining DenseNet s dense feature reuse, vision transformer modules for global context, and lightweight CNNs to refine local details achieving 96.3% accuracy and real-time inference (~25 ms). It begins with a field-scale dataset of over 33,000 rice-leaf images captured, smartphones, and sensors under varied lighting, pre-processed through contrast enhancement, denoising, and augmentation. A learnable weighted stacking ensemble improves class balance and distinguishes visually similar diseases. Training uses PyTorch with adaptive learning rate scheduling and dropout. Compared to DenseNet, ResNet-50, and Inception-V3, the hybrid model shows superior precision, recall, and F1-score. Edge-ready deployment is achieved via pruning and quantization for IoT devices. Extensive robustness analysis confirms resilience to poor lighting, clutter, and occlusion. The work marks a leap in sustainable agriculture by enabling timely, low-resource diagnosis for enhanced crop management and food security. newlineOverall, the thesis marks a significant leap in precision agriculture by showcasing how a hybrid deep learning strategy can overcome limitations of traditional rice disease classification systems. By merging cutting-edge architectural innovations with computational efficiency, the proposed approach deli