Poster # 520
Subash Thapa 1*, Hillson Ghimire 2*, Mandeep Singh 1, Dante Koupal 1, Swas Kaushal 1, Jyotirmoy Halder 1, Karanjot Kaur 1, Shaukat Ali 1, Jinfeng Zhang 1, Maitiniyazi Maimaitijiang 2, and Sunish K. Sehgal 1
1. South Dakota State University, Department of Agronomy, Horticulture, and Plant Science, Brookings, SD 57007, USA
2. South Dakota State University, Department of Geography and Geospatial Sciences, Brookings, SD 57007, USA
Corresponding Authors: Maitiniyazi Maimaitijiang, maitiniyazi.maimaitijiang@sdstate.edu; Sunish K. Sehgal, sunish.sehgal@sdstate.edu
*These authors contributed equally to this work and share first authorship.
Thapa, Subash
Reliable phenotyping of Fusarium head blight (FHB) in field nurseries and assessment of Fusarium-damaged kernels (FDK) after harvest are essential for identifying resistant wheat germplasm, yet conventional visual ratings and manual kernel sorting are subjective and labor-intensive. We developed WheatAI, a cloud-based phenotyping platform that processes smartphone-acquired RGB images to generate standardized measurements of FHB severity and FDK percentage. For field FHB assessment, a multistage workflow detects spikes in canopy images, extracts individual spikes, selects suitable frontal views, and uses a YOLOv11-based model with oriented bounding boxes to detect and classify individual spikelets as FHB-symptomatic or healthy. Disease severity is calculated as the percentage of symptomatic spikelets among all detected spikelets on selected spikes. For postharvest grain assessment, a separate YOLOv11-based model detects and classifies individual kernels as Fusarium-damaged or healthy and estimates FDK percentage based on their respective counts. On held-out test images, YOLOv11 achieved strong agreement between predicted and ground-truth FHB severity (R² = 0.962, MAE = 0.037, RMSE = 0.064) and FDK percentage (R² = 0.918, MAE = 0.035, RMSE = 0.053), with MAE and RMSE reported on a normalized 0–1 scale. The platform supports individual-image and batch processing, provides annotated images for visual review, and exports tabular results in CSV or Excel format for downstream analysis and breeding records. Additional modules for spike and spikelet counting, kernel morphometrics, and stomatal phenotyping extend its utility across biological scales. By integrating image acquisition, automated object-level classification, quantitative trait calculation, and data export into a single interface, WheatAI offers an accessible workflow for FHB resistance screening and postharvest grain assessment. Continued evaluation across environments, genotypes, and imaging conditions will be important to establish the robustness and reproducibility of the models for routine use in breeding programs.
© Copyright
2026 by individual authors. All rights reserved. No part of this abstract or paper publication may be reproduced without prior permission from the applicable author(s).