Stochastic On-the-Fly Augmentation for Mango Leaf Disease Detection Using YOLO and RT-DETR Architectures

Abstract

This study investigated the application of deep learning-based object detection for identifying major foliar diseases in mango leaves. As mango is an economically important crop in Thailand, its productivity is frequently constrained by common leaf diseases that are difficult to diagnose consistently through manual field inspection. A dataset of 750 annotated images covering three prevalent diseases: Anthracnose, Algal leaf spot, and Sooty mold was used to train and evaluate five high-capacity object detection architectures: YOLOv8l, YOLOv11l, YOLOv26l, RT-DETR-L, and RT-DETRv2-L. To mitigate data leakage and storage overhead, training was conducted using an inline, stochastic on-the-fly augmentation strategy (horizontal mirroring and HSV illumination adjustment). Among the five models evaluated under identical experimental conditions, RT-DETR-L achieved the strongest overall performance on the independent test set with a mean Average Precision (mAP@0.5) of 79.8% and an F1-score of 75.9%, followed by RT-DETRv2-L (78.7% mAP@0.5), while YOLOv11l emerged as the strongest convolutional alternative (75.0% mAP@0.5). These findings demonstrate the viability of transformer-based object detectors for mango leaf disease identification while highlighting important trade-offs between architectural design and detection accuracy.