Hybrid Image Preprocessing and Deep Learning Pipelines for Early Rice Disease Diagnosis: A Systematic Review and Future Roadmap

Abstract

Rice is a critical staple crop feeding over half of the global population, yet its production faces significant threats from various diseases that can reduce yields by up to 52%. This literature review examines the state-of-the-art in hybrid approaches combining traditional image processing techniques with deep learning methods for early identification of rice plant diseases, covering research published between 2019 and 2024. The review analyzes 30 highly relevant studies that demonstrate the evolution and effectiveness of hybrid methodologies in automated disease detection systems.

The analysis reveals that hybrid approaches consistently outperform standalone methods, with reported accuracies ranging from 94% to 99.99%. These systems typically integrate preprocessing techniques such as Contrast-Limited Adaptive Histogram Equalization (CLAHE), segmentation methods (U-Net, K-means clustering), handcrafted feature extraction (texture features, GLCM, LNEP), and deep learning architectures (CNNs, ResNet, VGG, DenseNet, Transformers). However, while many studies claim to address "early detection," only a limited subset explicitly targets asymptomatic or early-stage disease identification with rigorous validation. The most promising early detection approach utilizes hyperspectral imaging combined with 3D convolutional neural networks, achieving 95.44% accuracy in detecting asymptomatic rice bacterial leaf blight infections. This review identifies key trends, methodological patterns, performance benchmarks, and critical gaps in the field, providing recommendations for future research directions in precision agriculture and disease management.