Potato Leaf Disease Detection and Classification With Weighted Ensembling of YOLOv8 Variants

Abstract

The identification and control of potato leaf diseases pose considerable difficulties for worldwide agriculture, affecting both the quality and yield of crops. Addressing this issue, we investigate the efficacy of the lightweight YOLOv8 variants, namely YOLOv8n, YOLOv8s and YOLOv8m, for the automated detection and classification of different potato leaf states. These conditions are categorised into three types: healthy, early blight disease and late blight disease. Our findings show that YOLOv8n achieves a mean average precision (mAP) of 94.2%, YOLOv8s achieves a mAP of 93.4%, and YOLOv8m achieves a mAP of 94%. Building on these results, we propose a novel weighted ensembling technique based on the confidence score (WECS) to combine the predictions of these YOLOv8 variants. The WECS technique efficiently leverages the advantages of each YOLOv8 variant by assigning weights based on the confidence scores of individual model predictions. These weighted forecasts are then combined to produce a final ensemble prediction for each sample. Achieving 99.9% precision and 89.6% recall, the WECS method attains a global mean Average Precision (mAP) of 96.3%, showcasing its robustness in real-world applications.