Soybean variety development using aerial imagery and machine learning enabled phenomic‐assisted selection

Abstract

Traditional yield trials (YTs) are resource‐intensive and often limit genetic gain. This study assesses uncrewed aerial vehicle (UAV)‐based multispectral sensing for high‐throughput seed yield prediction for phenomic‐assisted selection (PAS) in a soybean [ Glycine max (L.) Merr.] cultivar development program. UAV data were collected at three growth stages across 54,000 soybean breeding plots over 3 years in Iowa using multispectral reflectance. Minimal performance loss was observed across ground sampling distances of 2.89–5.78 cm, suggesting that higher altitude flights are viable. Random forest models were trained and analyzed using raw band and derived multispectral vegetation index features from progeny row (PR) and YT breeding plots separately. Integrating at least two time points, especially later ones, improved predictive power compared to a single time point. Model performance was evaluated on independent field trials to assess PAS. Raw bands and the two vegetation index (VI) feature groups showed similar predictive power with an average R in PRs of 0.61 and YTs of 0.71. Red‐edge and NIR bands, and VIs incorporating them, were the most robust. Models achieved sensitivities of up to 0.57, specificities of 0.93, and accuracies of 0.87 at selection thresholds of 10%–30%. A case study comparing PAS with breeder selections showed average accuracy, sensitivity, specificity, and Spearman correlation values of 0.79, 0.67, 0.85, and 0.71, respectively. These findings show that UAV‐based PAS is a high‐throughput, scalable decision‐support tool that can assist in earlier identification of promising and poor‐performing breeding lines, potentially reducing reliance on end‐of‐season harvest data.