Phenotypic scoring of Canola Blackleg severity using machine learning image analysis

Abstract Canola blackleg is a fungal disease that causes significant yield loss and plant death of infected canola (Brassica napus L., Brassica rapa L., Brassica juncea L.) fields worldwide. One of the most effective methods for controlling blackleg is through the cultivation of resistant varieties. Consequently, scoring blackleg disease severity of infected plants is a key metric for identifying and selecting resistant varieties. Traditionally, blackleg severity is scored by expert raters who evaluate disease in stem cross sections using established rating scales and reference images; however, human raters are expensive and inconsistent in their scoring. Here, we introduce a machine learning algorithm based on deep learning models that can score blackleg severity from cross-section images of infected plants. We find that expert ratings are largely inconsistent across raters and across years for the same rater, creating substantial noise in susceptibility ratings. Meanwhile, our trained machine learning model performs more consistently than the median rater while maintaining a similar heritability as expert raters for the blackleg susceptibility trait. This model can be used to standardize blackleg susceptibility scoring across locations and years to improve canola breeding outcomes across affected regions. Core Ideas Canola Blackleg is a fungal disease affecting yield of canola, and accurate scoring of Blackleg severity is important for tracking disease and breeding for resistant varieties. The standard practice of utilizing expert raters is expensive, and scores assigned are inconsistent across raters and years. Our deep learning model for assigning blackleg severity scores is more accurate than the median expert rater, opening the door for improved breeding of new resistant varieties. Competing Interest Statement All authors were employed by Corteva Agrisciences at the time of their contributions to this research. No additional contributions were made to the research by any members of the team following the end of their employment with Corteva. As such for authors that have moved to other organizations, only their Corteva affiliation is relevant to the research in this study. Footnotes Abbreviations: CNN (convolutional neural networks); DL (deep learning); EU (experimental unit); G (genotype); GT (ground truth); BB (bounding box); ML (machine learning); BLUP (best linear unbiased prediction)