Location: Hard Winter Wheat Genetics Research
Project Number: 3020-21000-012-032-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Jul 1, 2026
End Date: Jun 30, 2028
Objective:
The Cooperator will develop of digital phenotyping technology for wheat stem sawfly resistance screening. This technology will reduce the labor requirements for evaluation of wheat resistance to the sawfly.
Objective 1. Develop tools and protocols to support rapid and precise digital imaging to enable phenotyping WSS resistance traits.
Objective 2. Develop computer vision and artificial intelligence (AI) models, as well as analytical pipelines for automated quantification of various WSS resistance phenotypes.
Approach:
Objective 1. Develop tools and protocols to support rapid and precise digital imaging to enable phenotyping WSS resistance traits.
Task 1-1-. Develop tools and protocols for the rapid and precise collection of kernel images to support kernel size, shape, and thousand kernel weight data measurements in wheat.
Task 1-2. Develop tools for rapid and precise transverse and longitudinal sectioning of wheat stems.
Task 1-3: Develop an imaging platform and protocol for the uniform collection of the stem images from WSS nurseries.
Objective 2. Develop computer vision and AI models, as well as analytical pipelines for automated quantification of various WSS resistance phenotypes.
Task 2-1. Automated kernel shape and weight quantification. Extraction of precise kernel morphological traits will be implemented with a tiered deep learning architecture (Maimaitijiang et al., 2026).
Task 2-2. Automated stem solidness and dimension quantification. Solid-stemmed wheat varieties are characterized by a pith-filled lumen, and stem solidness is primarily assessed using dimensional traits of the stem, lumen, and pith, including diameter and cross-sectional area. To enable rapid, accurate, and automated quantification of these traits, high-resolution smartphone images of transverse sections from 20 stems per genotype will be collected. The analytical pipeline begins with an AI-assisted localization stage, where a specialized object detection model (such as YOLO26 or RT-DETR) is trained to automatically detect each individual stem section. To ensure morphological accuracy, these detections will be refined using OBB to account for the elliptical or irregular orientations of the cut sections. Once localized, a high-fidelity instance segmentation framework leveraging a pretrained SAM-based segmentation model will be deployed to precisely delineate three critical boundaries: the outer stem circumference, the inner lumen (hollow) boundary, and the extent of the pith-filled core. This multi-boundary segmentation enables the automated derivation of precise geometric traits, including stem diameter, total cross-sectional area, lumen dimensions, and pith area ratio.
Task 2-3. Automated quantification of tiller density and spike count. Tiller density and head (spike) count per unit area are critical yield components and high-priority indicators of a genotype’s yield potential. To automate the estimation of tiller density, we will develop a predictive AI framework leveraging advanced Transformer-based architectures (e.g., Vision Transformer (Han et al., 2022) or Swin Transformer (Liu et al, 2021)). These models will be trained on high-resolution smartphone images of wheat canopies captured during the critical tillering to jointing (Feekes 3–6) growth stages. Unlike traditional vegetation indices, these deep learning models will be optimized to extract complex spatial features and canopy textures that correlate with individual tiller emergence, even under varying lighting conditions and overlapping foliage common in early-season field environments.