A Comparative Study of a Custom CNN and Pre-Trained Deep Learning Models for Potato Leaf Disease Classification.

Accurate detection of potato leaf diseases is critical for minimizing yield losses and supporting sustainable agriculture. This study evaluates the performance of multiple deep learning architectures for classifying three categories of potato leaves: Early Blight, Late Blight, and Healthy. Six state-of-the-art models, including VGG19, DenseNet121, InceptionV3, MobileNetV3Large, Vision Transformer (ViT), and a Custom CNN trained from scratch were assessed using metrics such as Accuracy, Precision, Recall, and F1-score. A two-stage training strategy was employed, leveraging pre-trained ImageNet weights for transfer learning followed by fine-tuning of higher layers to adapt features to domain- specific patterns. Experimental results demonstrated that VGG19 achieved the highest overall performance, with 97.83% accuracy, 0.9419 precision, 0.9721 recall, and 0.9559 F1-score. DenseNet121 and InceptionV3 also exhibited strong performance, while lightweight and transformer-based models (MobileNetV3Large and ViT) showed moderate results due to the limited dataset size. The Custom CNN achieved high validation accuracy but slightly lower test performance, highlighting the advantage of transfer learning for improved generalization. These findings confirm that deep convolutional architectures combined with transfer learning provide robust and reliable solutions for automated potato leaf disease classification.