Rice leaf diseases pose a major threat to crop productivity, and early visual diagnosis remains challenging due to the subtle and overlapping patterns found across disease types. To address this problem, this study proposes MLA-RiceNet, a multi-level attention convolutional neural network designed to enhance feature discrimination for small scale agricultural image datasets. The model integrates hierarchical feature extraction with channel attention at three different depths, enabling the network to emphasize informative spatial and color-based cues. MLA-RiceNet was trained and evaluated on a dataset of 200 images containing four rice leaf diseases using a five-fold cross validation strategy to ensure robust performance. The proposed method achieved an average accuracy of 99.0 percent and a macro F1 score of 99.0 percent, outperforming two baseline models, ResNet18 and MobileNetV2, which achieved 97.5 percent accuracy. Grad CAM visualizations further show that MLA-RiceNet consistently focuses on meaningful lesion regions, supporting the interpretability of the model. Overall, the findings demonstrate that MLA-RiceNet provides an accurate, efficient, and explainable solution that can support practical rice disease diagnosis in real agricultural environments.