Improving Generalization and Interpretability in Rice Leaf Disease Detection Using N-Fold Cross-Validated CNN with Grad-CAM Explanation

Abstract

Rice is a critical staple crop that supports food security in the world, but itsproductivity is highly damaged by numerous leaf diseases, including bacterial blight, brown spot and leaf blast. These diseases require early and proper diagnosis in order to reduce crop losses and ensure sustainability of agricultural production. Recent development of deep learning has demonstrated significant potential in automated disease detection. However, most of the current methods use single train-test split, and frequently result in overfitting and limited model generalization. Additionally, deep learning models are not interpretable, which reduces the level of transparency and reliability of automated diagnostic systems.To overcome these limitations, this paper suggests an explainable deep learning framework of rice leaf disease detection with a Convolutional Neural Network (CNN) and N-Fold Cross Validation and Gradient-weighted Class Activation Mapping (Grad-CAM). The framework starts with data augmentation and preprocessing of images in order to diversify the dataset. Subsequently, N-Fold cross-validation methodology is implemented in order to enhance the strength of the model, as well as to ensure proper assessment of the performance by multiple splits of data. The CNN model is able to extract the discriminative features of rice leaf images automatically and categorize them into various types of diseases. Grad-CAM is applied in obtaining visual heatmaps to improve interpretability, and applies attention to those affected by diseases in the model predictions.The results of the experiments demonstrate that the given model improves the quality of classification with interpretableand transparentpredictions. Integration of cross-validation and explainable AI techniques can help improve the accuracy of the deep learning-based approach to detect agricultural diseases and contribute to intelligent crop monitoring systems to manage diseases at the earliest stages.Additionally, Grad-CAMis used to generate visual impact of disease-affected regions in rice leaf images. The outcomes of experimentshow that the proposed framework improves performance of classification and provides transparent model predictions. The system can support intelligent agricultural monitoring systems for the diagnosis of early disease and crop management.