Early Detection of Potato Leaf Disease Using Deep Learning

Abstract

Potato cultivation is crucial to global agriculture, but continues to face challenges from various disturbances, including six classes of potato diseases: pests, nematodes, viruses, fungi, Phytophthora, and bacteria. These diseases can spread rapidly across fields, causing potato shortages, resulting in significant losses. Traditional manual monitoring by farmers is also laborious and prone to human error, often resulting in delays or errors in treatment. To address potato leaf diseases, this study develops an advanced automated system for the early detection and localization of these diseases using innovative deep learning techniques. The research begins with image preprocessing, supported by the Contrast Limited Adaptive Histogram Equalization (CLAHE) method. Additionally, the system is trained on diseased potato leaf data using appropriate deep learning models. This training process is expected to enable optimal performance in detecting all classes of potato plant diseases across various environmental conditions and field lighting scenarios, such as the ResNet-18 architecture and the YOLO26 framework, to facilitate real-time object detection and precisely localize areas affected by the 6 class diseases. The results from both models demonstrated excellent results where the ResNet-18 model achieved 96% accuracy, followed by the YOLO26 model enabled highly precise early-stage identification, achieving a Top 1 accuracy of 97.06% and a perfect Top 5 accuracy of 100%. This research provides a powerful and scalable tool for agriculture, empowering farmers to implement highly targeted interventions, reduce pesticide use, and maintain crop quality.