Rice leaf diseases pose a major threat to crop productivity and global food security. Therefore, rapid and accurate disease diagnosis is essential for effective crop management. Despite the advancements made by deep learning algorithms in automated disease detection, the majority of current deep learning-based models are built and evaluated on single-source datasets. Therefore, the generalization capability of these models is still questionable. In order to overcome this problem, this paper presents an interpretable deep learning approach for classifying diseases in rice leaves using ConvNeXt-Tiny and Grad-CAM. The proposed semantic class normalization technique is used to harmonize the heterogeneous data classes to form a seven-class classification system. The model performance was analyzed based on two available public datasets for rice leaf disease and a combination of both. After performing five individual runs, the average classification accuracies were calculated to be 99.13±0.37%, 95.28±0.63%, and 97.27±0.41% for Datasets 1, 2, and the combined dataset, respectively. The confusion matrix analysis indicated minor misclassification errors in the form of false positives for the disease classes which have similarities in appearance. The training and validation curves showed consistent learning with minimum overfitting. Moreover, Grad-CAM analysis confirmed the focus of the model on the disease-specific regions. These results collectively show that the combination of transfer learning, semantic class harmonization, and explainable AI is a stable platform to conduct automated rice disease detection. It is evident from these outcomes that reliable performance can be achieved even under heterogeneous imaging conditions. Consequently, there exists great potential for this approach in agricultural decision-support systems.