Pubescence color is routinely recorded along with flower, pod, and hilum color, but visual scoring is labor-intensive. Breeders may rate thousands of rows under poor light or at suboptimal maturity, and pod color can interfere with perception. Previous high-throughput phenotyping (HTP) work using multispectral imagery and a support vector machine (SVM) separated tawny from gray pubescence but struggled with light tawny. The trait is controlled by qualitative genetics, including the T and Td loci, yet field notes remain the practical benchmark. Given these challenges, or because of these issues, in-depth research is needed on automated, transferable classification of soybean pubescence color using aerial imagery and machine learning.
Researchers at the University of Illinois Urbana-Champaign (UIUC) and the Instituto Nacional de Investigación Agropecuaria (INIA) La Estanzuela in Uruguay published (DOI: 10.1016/j.plaphe.2026.100189) the study in Plant Phenomics in 2026 (Volume 8). They used three years of soybean breeding trials, drone-based red, green, and blue (RGB) and multispectral images, and the Random Forest (RF) algorithm to classify gray, light tawny, and tawny pubescence. The study also tested whether models fitted in earlier years could classify rows in an independent 2020 environment.
The experiments were grown near Savoy, Illinois, from 2018 to 2020 and contained F4:5 breeding lines planted in single rows. In 2018, images came from a Phantom 4 Pro with an integrated RGB camera; in 2019 and 2020, an Inspire 2 carried an Altum multispectral sensor, with one 2019 flight using a RedEdge-M camera. The team extracted median pixel values per row and calculated red/blue, blue normalized difference vegetation index (NDVI), green NDVI, and red edge NDVI. Models also included maturity date, breeding block, field annotator, check cultivar, population, and parental pubescence color. Using four flights before or at maturity outperformed a single maturity flight. The best scenario reached 86.55% overall accuracy (Kappa = 0.7976), with sensitivities of 0.893 for gray, 0.788 for light tawny, and 0.915 for tawny pubescence. The red/blue index was the most informative feature; blue NDVI and green NDVI helped separate gray from light tawny. Parental pubescence and germplasm relationships improved accuracy by about 9.3%. In an independent environment, accuracy fell to 65.86% (Kappa = 0.4874), but tawny pubescence remained reliably identified.
The authors said the goal is not to remove breeders from the loop but to make better use of their time. They said the Random Forest (RF) model was strongest when it combined a time series of drone flights with parental pubescence records, rather than relying on one late-season image. They said tawny pubescence was classified most reliably, while gray and light tawny remained the hardest pair, mirroring field conditions. They said this hybrid human-machine workflow could let breeding teams automate clear cases and focus manual notes on ambiguous rows.
Applied at scale, the pipeline could reduce the labor and fatigue associated with scoring thousands of breeding rows. It could flag potential selfing, mislabeling, or purity issues, and help programs decide which populations to advance. Because model transferability declined across years, the authors recommend combining imagery with genotyping for the T and Td loci and pod color, which can interfere with aerial pubescence ratings. New camera systems and additional vegetation indices may further improve gray versus light tawny discrimination. Even imperfect predictions, they argue, can provide actionable information when field notes are limited or postponed to later generations.
###
References
DOI
10.1016/j.plaphe.2026.100189
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100189
Funding information
This work was supported by the Instituto Nacional de Investigación Agropecuaria (INIA, Uruguay), the University of Illinois at Urbana–Champaign Provost Match and Charles Adlai Ewing Fellowships, and the USDA National Institute of Food and Agriculture [Hatch, project 1015014].
About Plant Phenomics
Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.