Enhancing Plant Disease Classification with Transfer Learning and Data Augmentation: A Comprehensive Evaluation of Deep Learning Architectures

Plant diseases are a highly crucial problem faced by the world today, so accurate and efficient early detection of plant diseases is needed. In this study, we evaluate the effectiveness of transfer learning and data augmentation in improving the robustness and generalization of deep learning models for plant disease classification. We conduct experiments using five Convolutional Neural Network (CNN) architectures-AlexNet, MobileNet, ResNet, Xception, and DenseNet-on the PlantVillage dataset. To handle data imbalance, oversampling is performed through augmentation techniques on the training set. The results show that DenseNet outperforms other models, achieving a precision of 0.9976, recall of 0.9976, $\mathbf{F 1}$-score of 0.9974, and an accuracy of 0.9976. The superior performance of DenseNet is attributed to its dense connectivity, which enables efficient feature extraction and propagation. Integrating transfer learning and augmentationbased oversampling significantly improves the model's generalization ability across varying environmental conditions. Statistical significance testing further confirmed that DenseNet's performance advantage is consistent.