Potato Leaf Disease Detection Using Machine Learning

Abstract

Potato (Solanum tuberosum) plays a crucial role in global food security and is extensively cultivated in Bangladesh, where it serves as both a staple crop and a vital source of income for farmers. Nevertheless, its production is commonly frustrated by foliar diseases like Early Blight and Late Blight that if detected late can result in significant loss of late. The more conventional methods of disease detection are traditionally manual with expert observation and most impractical and inaccessible for farmers in distant or resource-poor regions. This study comes up with a solution to the inadequacy of a rapidly deployable and accurate disease detection system through the introduction of a deep learningbased pipeline based on the YOLO (you only look once) object detection framework. The proposed approach consists of producing a custom-annotated data set of potato leaves and training the three sophisticated YOLO models: YOLOv8m, YOLOv9m, and YOLOv10m. The evaluation was done as regards mean Average Precision (mAP), Inference speed, and computational effectiveness. The highest detection accuracy of YOLOv8m resulted in mAP@50 of 97.5%, mAP@50-95 of 91%, and fast inference time of 11.3 ms. YOLOv9m and YOLOv10m performed rather competitively with mAP@50-95 as 91.4 % and 91.3 %, respectively (YOLOv10m demonstrated superior computational efficiency – 63.4 GFLOPs). The best-performing model was converted to TensorFlow Lite format and integrated into a cross-platform Flutter-based Android application. This mobile app allows users to capture or upload leaf images and receive realtime disease predictions without internet dependency. The system offers an effective, farmer-friendly tool that promotes early disease intervention, reduces unnecessary pesticide usage, and supports sustainable precision agriculture practices in low-resource environments.