Deep Convolutional Neural Network with Explainable AI for Potato Foliar Disease Recognition

Timely and precise identification of diseases affecting potato foliage plays a vital role in preserving agricultural yields and reducing harvest losses. This research introduces a customized deep learning architecture combined with interpretable AI techniques to sort potato leaf images into three distinct groups: Early Blight, Late Blight, and Healthy specimens. A curated portion of the PlantVillage database containing 2,152 photographs underwent preprocessing and augmentation procedures, resulting in 11,593 samples to improve model generalization capabilities. Our proposed network incorporates six convolutional layers utilizing LeakyReLU activation, batch normalization, max pooling layers, dropout regularization, and global average pooling, optimized through the Adam algorithm. Experimental outcomes demonstrated 99.05% test accuracy, with precision, recall, and F1-scores all surpassing 99%. Grad-CAM visualizations enhanced interpretability by producing heatmaps that identify disease-specific areas. Comparative assessments against prior approaches validate our model's effectiveness and superior results. These findings demonstrate the promise of combining deep learning with explainable artificial intelligence for practical farming applications, delivering trustworthy early disease identification. Upcoming work will investigate real-time diagnostics, compact deployment on edge computing devices, and database enrichment with varied field-acquired photographs.