Abstract

Plant diseases are threatening global food security and the sustainability of agriculture. The manual inspection methods that are traditionally carried out tend to be inefficient, subjective and can barely be scaled. The proposed project, as it is going to be described in this paper, is a real-time plant disease detection system that is built on the latest object detection architecture, YOLO11 (You Only Look Once). In comparison to the traditional approach, that is applied to learn transfer learnings on the premise of generalized datasets, our model is being trained on a custom-curated dataset of healthy leaves and particular pathological outcomes, these two are Downy Mildew and Leafy-minor infections. It has been demonstrated through experiments that the model mAP 0.50 = 0.585 is an average precision and its overall recall rate of 0.727 and the capacity to minimize the number of undiagnosed diseased cases in the field is high. Despite the fact that the model displays quite promising scores as far as the recognition of the healthy leaves ( Recall: 0.852) and the Downy Mildew ( Recall: 0.872) are concerned, it can be stated based on the review of the confusion matrix that a certain level of inter-class variation is present between the minor stages of the symptoms and the background noise. The findings suggest that a special YOLO11 application may be deployed to provide a scalable and quick system of providing precision agriculture. The following research will focus on creating more precise detection methods based on a class-balancing method and multi-scale feature fusion that will be used to detect lesion in the early stage.