Deep learning–based computer vision for rapid and accurate quantitative field assessment of powdery mildew severity: model comparison, validation, and agreement with human raters

Abstract Quantifying foliar disease severity accurately and reproducibly remains a fundamental challenge in plant pathology, as traditional visual assessment is inherently subjective and prone to rater variability. Although computer vision has enabled automated disease detection, validated end-to-end frameworks that integrate reference generation, model training, and field-deployable percent severity estimation are lacking. Here, we developed and validated a computer vision pipeline for quantifying powdery mildew ( Oidium spp.) severity in pea ( Pisum sativum ) using smartphone-acquired images under field conditions, with tomato ( Solanum lycopersicum ) powdery mildew serving as a cross-species validation system. A semi-automated multichannel segmentation approach generated continuous percent severity estimates as reproducible training references. Among evaluated deep learning architectures, a YOLO-based model was selected for its superior balance of predictive accuracy, computational efficiency, and deployment flexibility and was integrated into PowderyComputerVision, a purpose-built web application enabling real-time inference. Under field conditions, the automated system showed substantially higher agreement with reference severity values than trained human raters (concordance correlation coefficient (CCC), 0.952 vs. 0.513). A linear mixed-effects model revealed that human evaluators produced absolute estimation errors 280.96 times greater than the automated system ( p < 0.0001; 95% CI 120.34–655.91). Performance remained robust across variable lighting conditions, imaging angles, and plant structures, supporting applicability beyond controlled settings. The proposed four-stage pipeline comprising image capture, host and disease classification via YOLOv11m-cls, leaf localization via YOLOv11n-seg, and lesion quantification via multichannel segmentation offers a scalable, objective, and reproducible framework for continuous field-based disease severity estimation.