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

Predicting seed yield for phenomic-assisted selection (PAS) in plant breeding programs requires efficient, scalable phenotyping tools that operate across thousands of genotypes and environments. Traditional yield trials are resource-intensive and often limit the pace of genetic gain. This study assesses Uncrewed Aerial Vehicle (UAV)-based multispectral sensing for high-throughput SY prediction in a soybean cultivar development program. UAV data were collected at multiple growth stages across >54,000 soybean breeding plots over three years in Iowa, using multispectral reflectance bands and vegetation indices (VIs) as model inputs. Minimal performance loss was observed across ground sampling distances of 2.89 cm to 5.78 cm, suggesting that higher-altitude flights are viable. Random forest models were trained using raw bands, RGB VIs, and multispectral VIs for progeny rows and yield trials. Integrating at least two time points, especially later ones, significantly improved predictive power compared to a single time point. Model performance was evaluated on independent field trials to assess PAS. Raw bands, multispectral VIs, and RGB VIs showed similar predictive power in progeny rows (R2 = 0.61) and yield trials (R2 = 0.74). At 10–30% selection thresholds, models achieved up to 0.58 sensitivity, 0.93 specificity, and 0.87 accuracy. A case study comparing PAS to breeder selections resulted in mean accuracies, sensitivities, specificities, and Spearman correlations of 0.79, 0.64, 0.85, and 0.62, respectively. These findings validate UAV-based PAS as a high-throughput, robust, and scalable approach for identifying high-yielding soybean lines earlier in the season and suggest the potential to reduce reliance on end-of-season harvest data.