DeepCropGuard: A Hybrid DenseNet–Vision Transformer Framework for Potato Disease Detection and Precision Management

Abstract

Potato diseases pose a serious threat to the productivity and quality of potato crops; the most important ones are early blight and late blight, which require fast and accurate automated detection. This study introduces DeepCropGuard, a combined DenseNet–Vision Transformer network to classify potato diseases and support precision management. The framework is a combination of DenseNet based hierarchical feature extraction and Vision Transformer's self-attention to obtain local disease characteristics and spatial relationships. The potato leaf images are pre-processed, enhanced, and classified into Healthy, Early Blight and Late Blight classes. The feature fusion combines the convolutional and the transformer representation, and the explainable artificial intelligence module identifies the disease relevant regions. The results of the illustrative experiment demonstrate that DeepCropGuard’s accuracy, precision, recall, F1-score was 97.56% and its MCC value was 0.963 which is better than DenseNet and Vision Transformer models. Additionally, the framework will include a confidence-based decision support tool for targeted field monitoring. The results highlight the potential of hybrid deep learning to explainable and scalable potato disease management in different agricultural conditions.