Potato viral diseases, especially Potato Virus X (PVX) and Potato Leafroll Virus (PLRV), contribute to an estimated 9–11% decline in global potato yields each year, presenting a major threat to agricultural sustainability. Traditional detection methods, for example enzyme-linked immunosorbent assay (ELISA), which are being currently utilized; however, their reliance on specialized equipment and expertise has limited their accessibility, particularly in resource-constrained settings. This study offers a deep learning-based method for automating the detection and classification of viral potato leaf diseases, applying a modified EfficientNet-B1 architecture. The model is trained on the publicly available Potato Viral Disease Dataset, which contains 1328 images distributed across three categories: healthy, PLRV-infected, and PVX-infected. To address dataset imbalance, several data augmentation techniques are applied, generating 9296 augmented samples. A class-weighted cross-entropy loss function is employed to enhance classification performance. The model accomplishes a state-of-the-art accuracy reaching 100%, outperforming existing convolutional neural network (CNN) architectures. The interpretability of the model is validated through Grad-CAM visualizations, which confirm its ability to focus on disease-relevant leaf regions. These findings demonstrate the prospective of the proposed approach and other similar AI-based approaches in precision agriculture, providing a scalable, cost-effective, and efficient alternative to conventional disease detection methods.