A VGG-16-Based Deep Learning Approach for Potato Leaf Disease Detection

Potato early blight and late blight were considered in this experiment because both classes are visually close enough to require careful image classification. The work used the potato subset of the PlantVillage image collection. Each leaf image was resized to 224 x 224 pixels and was then given directly to a VGG-16 feature extractor. No separate lesion mask, K-means step, or colour-thresholding stage was used. The final decision layer was a single sigmoid unit for the two-class output. The dataset contained 1000 images of early blight and 1000 images of late blight, divided into training, validation, and testing sets in a 60:20:20 ratio. With 32 epochs and a batch size of 100, the best run produced 99.6% training accuracy and 98.5% test accuracy. The paper also gives precision, recall, F1-score, the confusion matrix, ROC-AUC, and same-split results for ResNet-50, MobileNetV2, and VGG-19.