Evaluation of Convolutional Neural Network Architectures for Potato Leaf Disease Detection

Abstract

Potato, one of the world's most important staple food crops, is highly susceptible to various foliar diseases that significantly affect its productivity and pose a serious threat to food security, thereby contributing to economic losses and impacting farmers’ income. Therefore, early and accurate detection is essential. Conventional detection methods rely primarily on manual observation, which is time-consuming and requires specialized personnel. In this study, five convolutional neural network (CNN) architectures were evaluated for the automatic classification of potato leaf diseases, including pretrained models (ResNet50, MobileNet, and VGG16) and models trained from scratch (AlexNet and LeNet-5). The dataset was constructed by integrating and selecting images from publicly available Kaggle repositories, resulting in a total of 6,691 images distributed across five classes: early blight, late blight, potato leafroll virus (PLRV), mosaic virus (PVY), and healthy leaves. Multiple experiments were conducted by varying hyperparameters such as batch size, optimizers, and the number of training epochs. The results show that VGG16 achieved the best performance, with an accuracy of 99.87%, outperforming the other architectures. Additionally, a mobile application based on the optimal model was developed for real-time detection. These findings demonstrate the potential of deep learning for intelligent and scalable agricultural diagnostic systems.