Publication

Abstract : Accurate classification of potato Leaf diseases are essential for prompt treatment and crop management. However, the presence of imbalanced datasets often biases model predictions toward majority classes. In this work, we present a thorough comparison of transformer-based models-Vision Transformer (ViT) and Swin Transformer-for the task of potato leaf disease classification. This study also evaluates the performance of these models on a real-world, imbalanced agricultural dataset applying the Synthetic Minority Oversampling Method (SMOTE) to balance class distributions. Key evaluation metrics include test accuracy, classification report, confusion matrix, true positives (TP), Cohen's Kappa score, and ROC-AUC. For benchmarking, we also train conventional CNN architectures—VGG16, ResNet50, DenseNet121, InceptionV3, MobileNetV2, and EfficientNetV2B3-under the same conditions. Results demonstrate the superior performance of transformer models, particularly Swin Transformer, which achieved a classification accuracy of 91.4%, while Vision Transformer reached 87.5%.