<![CDATA[Rice leaf diseases can significantly reduce crop productivity and grain quality, making rapid, objective, and accurate identification essential. This study compares the performance of three deep learning architectures, namely ResNet50, ConvNeXtTiny, and Vision Transformer (ViT-B16), for rice leaf disease classification. The dataset consists of 5,932 images representing four disease classes: Bacterial Blight, Blast, Brown Spot, and Tungro. The data were divided into 3,559 training images, 1,186 validation images, and 1,186 testing images using the same random seed. All models were trained using transfer learning with ImageNet pretrained weights, data augmentation applied only to the training set, dropout regularization, and a two-stage training strategy involving feature extraction and fine-tuning. Performance was evaluated using validation loss, accuracy, precision, recall, F1-score, training curves, and confusion matrices. The results show that ResNet50 achieved the best performance, with a validation accuracy of 99.92% and a validation loss of 0.00326, followed by ViT-B16 (99.16%) and ConvNeXtTiny (98.99%). All models classified the Tungro class with high accuracy, while minor misclassifications occurred among the Bacterial Blight, Blast, and Brown Spot classes due to their similar visual characteristics. Overall, ResNet50 proved to be the most effective model for rice leaf disease classification, although validation using real-world field images is still required before practical deployment]]>
Copyrights © 2026
