Optimized CNN-Based Transfer Learning Approach with ReduceLROnPlateau for Early Potato Leaf Disease Classification

Andi Sunyoto (Universitas Amikom Yogyakarta)
Yoga Pristyanto (Universitas Amikom Yogyakarta)
Anggit Ferdita Nugraha (Universitas Amikom Yogyakarta)

Jurnal RESTI (Rekayasa Sistem dan Teknologi Informasi)
Vol 10 No 5 (2026): October 2026 (In progress)

Potato leaf disease classification is important for supporting early disease detection and improving crop management. This study proposes an optimized CNN-based transfer learning approach for classifying potato leaf images into three classes: Early Blight, Healthy, and Late Blight. Six pre-trained architectures—InceptionV3, Xception, DenseNet121, EfficientNetB0, ResNet50, and MobileNet were evaluated using ImageNet weights. Each model was modified with a custom classification head consisting of Global Average Pooling, a dense layer with 1,024 neurons, dropout, and a Softmax output layer. The models were evaluated using 5-fold stratified cross-validation. A frozen baseline configuration was first established, followed by fine-tuning configurations that varied learning rate, dropout rate, number of unfrozen layers, and the use of ReduceLROnPlateau. Data augmentation was applied only to the training subset in each fold. Performance was assessed using accuracy, AUC, validation loss, confusion matrix, ROC curves, training time, inference time, and GFLOPs. The results showed that optimized fine-tuning improved the average accuracy from 0.9876 to 0.9975 and reduced the average validation loss from 0.0316 to 0.0025. The Xception with the HP3 configuration achieved the best performance, with a mean accuracy of 0.999444 and an AUC of 0.999997. However, EfficientNetB0 achieved a high accuracy of 0.998333 with only 0.80 GFLOPs per image, making it more suitable for efficiency-oriented deployment. These findings demonstrate that fine-tuned transfer learning combined with adaptive learning rate scheduling is effective for potato leaf disease classification. Because the optimized configurations differ from the frozen baseline in the number of unfrozen layers, the dropout rate, the learning rate, and the use of ReduceLROnPlateau simultaneously, the reported gain reflects the combined effect of these components rather than the isolated contribution of ReduceLROnPlateau. The results were obtained on PlantVillage images acquired under controlled conditions and should therefore be regarded as an upper reference rather than an estimate of performance under field condition.