Detection of Potato Tuber Diseases in Türkiye Using a Deep Learning Methodology

Abstract Visual inspection performed by experts to diagnose plant disease is time-consuming, labor-intensive, and subject to human error. This study presents a deep learning-based approach for detecting common scab in potato tubers ( Solanum tuberosum ), a disease caused by a group of bacteria known as Streptomyces spp., using four different potato varieties: Provento , Othello , Madaline , Bettina. A total of 800 healthy and diseased tubers were collected with detailed records of variety and field of cultivation for each tuber variety. The presence of disease was confirmed through laboratory analysis. RGB and near-infrared (NIR) images were captured in a controlled environment designed to simulate a commercial sorting machine. By applying this approach a cleaned dataset of 1374 images was produced. The data were split into training (70%), validation (15%), and testing (15%) sets, and four datasets (RGB images separated by variety, NIR images separated by variety, RGB images merged across varieties, NIR images merged across varieties) were analyzed. MobileNetV3 achieved the highest test accuracy (90.5%) on the NIR dataset with a loss of 0.415 and training time of 24 min. In comparison, VGG19 reached 95.7% accuracy on RGB images. Analysis of the merged RGB dataset showed a maximum accuracy of 99% for both InceptionV3 and VGG19, while ResNet101 achieved 99.5%. Thus, ResNet101 provides the maximum classification certainty (99.5%) among all tested CNN architectures. Across all datasets tested, MobileNetV3 demonstrated the ideal combination of maximum accuracy (> 99%), optimal low loss (< 0.41), and efficient average training time (24 min). Thus, it is suggested as an optimal candidate for real-time, automated potato disease detection in precision agriculture systems.