Abstract

Accurate assessment of the presence and severity of plant diseases is essential for effective crop monitoring and management. This study presents a deep learning-based pipeline for quantifying blackleg crown canker disease severity in canola stems by combining image segmentation and severity prediction tasks. Three architectures (ResUNet, UNet and SegFormer) were compared for the first step of stem segmentation to isolate relevant regions. The disease severity scores of four experts, and their aggregated median, were used to train models which were evaluated for label consistency, ambiguity, and model robustness. Among the three segmentation architectures, SegFormer achieved the best performance (mean IoU = 0.939, F1 score = 0.962), outperforming ResUNet and UNet. There was a high correlation in disease severity scores across expert labels, with the median-trained model achieving correlation coefficients of 0.924–0.963 against individual expert assessors on the evaluation dataset. Confusion matrix analysis further demonstrated reliable classification across severity levels. This work highlights the influence of segmentation quality, label aggregation strategies and data imbalances on downstream prediction tasks. This study uses controlled imaging conditions, but the proposed framework provides a strong foundation for future application in field environments. The framework enhances model interpretability by generating severity scores that align closely with expert assessments to support users such as agronomists and plant breeders for better decision-making and help track disease resistance by providing consistent, objective disease measurements over time. Future work will focus on exploring multi-task learning for greater efficiency, alongside validating the approach in field conditions to enable broader adoption.

IPC Classification

Keywords