Apple scab, caused by Venturia inaequalis, poses a major threat to apple production, particularly in Pennsylvania, where climatic conditions favor rapid disease development. Traditional scab severity quantification methods, including visual estimation and manual segmentation using ImageJ, are time-consuming and limit large-scale assessment. This study introduces ScabDoc, a system for quantitative estimation of apple scab severity providing lesion segmentation and weather information. ScabDoc incorporates two segmentation approaches: Mask R-CNN for automated lesion identification and ScabSAM, a prompt-based model built on the Segment Anything Model (SAM), to improve segmentation flexibility under field conditions. Disease severity, defined as the proportion of leaf area covered by scab lesions, was quantified and compared to ImageJ measurements. ScabDoc achieved a strong correlation with manual ground truth, with an R 2 of 0.7971 at a 60% confidence threshold and 0.1237 at 85%. On external datasets, ScabDoc detected scab presence in PlantVillage images but showed limited lesion segmentation; lesion segmentation was successful in 33% of tested Apple Leaf Disease Dataset images, highlighting persistent domain differences. ScabSAM’s point-prompt capability enabled detection of small lesions while reducing manual effort. ScabDoc was deployed through a Flask-based web platform for apple scab severity assessment and visualization. The platform helps characterize disease dynamics and may support future predictive model development.
Yang et al. (Mon,) studied this question.