Optimizing watermelon leaf disease detection using Sam-based augmentation with YOLO for practical agricultural solutions

Automated plant disease detection systems are widely adopted in agriculture to increase production effectiveness. In watermelon plantations, where diseases spread rapidly over large areas, these systems reduce the labour and time required for disease inspection. Thus, this research introduces a deep learning-based watermelon disease detection on a field-collected dataset obtained from farms in Thailand. In contrast to lab-collected imagery, field-collected images present different challenges due to uncontrolled environmental conditions. In our field dataset, a clear class imbalance is observed. To mitigate this issue, we propose an efficient augmentation method that integrates the Segment Anything Model (SAM) with Photometric and Geometric (PG) augmentations. The PG process generates augmented images to diversify dataset. To further improve the detection of underrepresented objects, we employ SAM to extract precise object masks from the augmented data using bounding box annotations. A lightweight version of YOLO-based architectures is utilized to assist real-time detection of watermelon diseases. The experimental results show that YOLOv9t achieved the highest detection accuracy of 0.92 mAP50 and 0.71 mAP95. Moreover, we evaluate model performance using an Accuracy per Inference Time (APT) metric to identify efficient models for real-time applications. The results show that YOLOv11 achieved the highest APT score, followed by YOLOv5n, YOLOv9t, and YOLOv8, respectively. • This work introduces a watermelon disease dataset acquired from field in Thailand. • The proposed work mitigates class imbalance present in real-field datasets. • YOLO architectures are employed to enable real-time watermelon disease detection. • The proposed method improves minority class detection in real-field scenarios. • YOLOv11n offers the most effective trade-off between detection accuracy and speed.