Abstract
Breeding for quantitative, polygenic resistance is widely considered the most durable, cost-effective, and environmentally safe approach to crop disease control. However, progress in resistance breeding is hindered by the limited capability of current approaches to measure highly quantitative disease phenotypes under field conditions with high precision and sufficient throughput.
Here, we present an imaging protocol and a modular image processing pipeline that enables wheat disease detection and severity estimation directly from very-high-resolution canopy imagery, eliminating the need for physical interaction with the monitored plants as required in previously proposed sensor-based methods capable of symptom-level diagnosis. The pipeline combines deep-learning-based semantic segmentation, keypoint detection, and depth estimation to diagnose and quantify disease symptoms and extract the analyzable reference plant surfaces for severity estimation. By leveraging estimated relative depth and analyzing image texture, well-focused areas with sufficient quality were accurately segmented. Despite the challenging nature of canopy images and frequent symptom ambiguity, symptom detection and segmentation models trained on a new dataset reached a similar performance as already described in more simplified scenarios where detached, flattened leaves were analyzed.
Plot-level severity estimates of Septoria Tritici Blotch, a major wheat disease, obtained using the new method and a precise but more laborious reference method were highly correlated (Pearson R=0.83) across a range of morphologically contrasting cultivars. Validation of the new method on data collected by different operators at different sites demonstrated the robustness of the approach. The ability of the method to process imagery acquired in a contact-free manner can enable deployment on autonomous ground vehicles, paving the way for automated, scalable phenotype acquisition.
Competing Interest Statement
The authors have declared no competing interest.