DENSENET 201 TRANSFER LEARNING FOR POTATO LEAF DISEASE CLASSIFICATION AND BENCHMARKING AGAINST ESTABLISHED CNN ARCHITECTURES

Abstract

Background: Potato (Solanum tuberosum) is an important food crop affected by foliar diseases that can reduce yield and quality. Early and accurate identification of early blight and late blight is therefore important for timely disease management in practice. Deep learning provides an automated alternative to visual diagnosis, but model selection and generalization remain important considerations.

Methods: This study developed and evaluated a transfer learning-based framework for potato leaf disease classification using four pretrained convolutional neural networks: DenseNet201, VGG19, ResNet50, and InceptionV3. The models were evaluated for three classes: early blight, late blight, and potato healthy. DenseNet201 was used with its pretrained feature-extraction layers frozen and trained for 50 epochs. Performance was assessed using accuracy, precision, recall, F1-score, Matthew’s correlation coefficient (MCC), ROC-AUC, PR-AUC, and top-2 accuracy, together with confusion-matrix analysis.

Results: Among the evaluated models, frozen DenseNet201 achieved the strongest overall performance, with 99.07% test accuracy, 99.34% macro precision, 97.52% macro recall, and 98.39% macro F1-score. It also achieved 98.35% MCC, 99.98% macro ROC-AUC, 99.83% macro PR-AUC, and 100% top-2 accuracy. Class-wise recall reached 100% for early blight, 99.00% for late blight, and 93.55% for potato healthy leaves. These findings demonstrate the effectiveness of DenseNet201 for automated potato leaf disease classification.