Leveraging Multi-Omics Data with Machine Learning to Predict Grain Yield in Small vs. Big Plot Wheat Trials

Abstract

Accurate grain yield (GY) prediction is essential in wheat breeding to enhance selection and accelerate breeding cycles. This study explored whether high-throughput phenotyping (HTP) data collected from small plot (SP) trials can effectively predict GY outcomes in later-stage big plot (BP) trials. Genomic (G) data were combined with hyperspectral (H) and multispectral + thermal (M) imaging across the 2022 and 2023 growing seasons at the Plant Science Research and Education Unit, Citra, Florida. A panel of 312 wheat genotypes was analyzed using GBLUP-based models, integrating G + H and G + M data from SP to predict BP yield. SP models demonstrated promising predictive ability, with G + H models achieving moderate within-year (0.43 to 0.51) and across-year (0.43) prediction accuracies, while G + M models reached 0.53 to 0.58 and 0.45, respectively. The Random Forest Regression (RFR) model produced an accuracy of 0.47 when M data from the 2022 SP, combined with G, was used to predict BP yield in 2023. Additionally, the top 25% specificity (coincide index) was evaluated, with models showing up to 47–51% within a year and 43–45% between years overlap in the highest predicted-yielding lines between SP and BP trials, further emphasizing the potential of SP data for early selection. These findings suggest that SP trials can provide meaningful predictions for BP yields, enabling earlier selection and faster breeding cycles.

Plant phenotyping relevance

小区試験の高スループット画像データを用いて大区画の穀粒収量を予測し、複数年・モデル間の予測精度を評価しており、フェノタイピングデータの解析ワークフローが中心的です。

Code and data availability

The paper reports UAV-derived hyperspectral/multispectral/thermal phenotyping and genomic prediction of wheat grain yield. The Data Availability Statement cites two Dryad dataset DOIs (10.5061/dryad.p5hqbzkzr and 10.5061/dryad.wwpzgmsvk), which would be paper-specific public phenotype data assets, but no Dryad URL is a

No evidence-backed public reproduction asset is currently recorded.