Accurate identification of rice leaf diseases is important for timely disease management, particularly when disease symptoms have similar visual characteristics. This study proposes DSBMA-Net, an attention-based hybrid deep learning model for six-class rice leaf disease classification. The model uses a pretrained DenseNet-169 backbone to extract high-level visual features, followed by global average pooling and a squeeze-and-excitation block for channel-wise feature refinement. The refined representation is processed through two parallel multilayer perceptron branches with ReLU and GELU activations, and their outputs are integrated through attention-based feature fusion before six-class Softmax classification. The experiments used a balanced dataset of 2,928 images, partitioned into 2,100 training, 528 validation, and 300 testing images using stratified random sampling with a fixed seed of 42. Images were resized to 224×224 pixels, and augmentation was applied only to the training set. On the test set, DSBMA-Net achieved 98.86% accuracy, 98.87% precision, 98.86% recall, and 98.86% F1-score. The model obtained a micro-average area under the receiver operating characteristic curve of 0.9997 and an average precision of 0.9987. Inference evaluation yielded a latency of 1.75ms and a throughput of 570.34 frames per second. To support model interpretability, explainable artificial intelligence (XAI) analysis using Grad-CAM was employed to visualize the image regions contributing to the model predictions, with the resulting activation maps primarily highlighting disease-related regions of rice leaves. These results indicate that DSBMA-Net provides accurate six-class rice leaf disease classification with low inference latency and interpretable predictions under the evaluated experimental conditions.
Copyrights © 2026