End-to-End Transfer Learning for Crop Pest and Disease Classification with Layer-wise Learning-Rate Decay and a One-Cycle Schedule

Early diagnosis of pests and diseases in smart-agriculture environments is essential for improving productivity. In particular, apple mango—a subtropical crop cultivated in Korea—is highly vulnerable to pests and diseases, necessitating reliable automated diagnostic methods. Although transfer learning–based deep learning classifiers have recently attracted significant attention in agricultural image analysis, conventional stepwise approaches suffer from inherent instability during optimization. This study proposes an end-to-end training strategy that combines Layer-wise Learning Rate Decay (LLRD) with a One-Cycle learning-rate schedule. Using this strategy, we effectively adapted pretrained convolutional neural networks (ResNet50, VGG16, and MobileNetV2) for classifying apple-mango pests and diseases. We employed a publicly available AI-Hub dataset comprising 5,400 images labeled as Healthy, Thrips, and Sooty Mold, and applied data augmentation to simulate variations in real agricultural imaging conditions and mitigate overfitting. Experimental results confirm that the proposed method achieves superior performance compared to our prior stepwise transfer-learning baseline across all backbone architectures. In particular, ResNet50 achieved the best performance, with 90.71% accuracy and a Macro-F1 score of 89.21%. These results indicate that the proposed training strategy overcomes the limitations of standard transfer learning and enhances fine-grained discrimination among disease classes. Moreover, its efficiency suggests practical applicability in resource-constrained environments such as drones and smart-farm devices. In the future, we plan to expand the range of crops and classes and explore multimodal fusion to further strengthen field robustness.