Plant breeding requires phenotyping for multiple traits throughout the growing season in large, multilocation field trials. Unoccupied aerial vehicles (UAVs) equipped with sensors offer potential to assist breeding programs in data collection from these field trials. However, approaches to best use UAV-based imaging to support field trial evaluation remain underexplored.

To initially determine the potential of UAV-based imaging to support in-field phenotyping, a diverse hexaploid wheat (Triticum aestivum L.) nested-association mapping (NAM) population consisting of 1160 recombinant inbred lines was evaluated at three locations in two years. UAV-based multispectral imaging was conducted at 10-15 timepoints throughout the growing season at approximately weekly intervals. Image features, including spectral summary statistics, spectral indices, and texture features, were extracted from plot images at each timepoint.

Models of varying complexity ranging from simple linear regression to gradient boosted decision trees were assessed for prediction accuracy of breeding-relevant traits on test sets. LASSO regression models trained on image feature sets were able to predict days to heading (mean R2 = 0.76), days to maturity (mean R2 = 0.84), plant height (mean R2 = 0.70), and grain yield (mean R2 = 0.64) within testing environments more accurately than all other tested models. Cross-environment predictions were also evaluated, and a combination of image feature and genomic prediction models led to higher prediction accuracies for grain yield than either model alone (mean R2 = 0.39). Image-based prediction models were then applied to durum wheat (Triticum turgidum L. var durum) breeding population field trials evaluated in four environments.

High within-environment LASSO regression prediction accuracies for grain yield (up to R2 = 0.88) indicated the potential for high-throughput phenotyping of complex traits in breeding populations. Genome-wide association mapping of image features from the NAM population provided insight into the heritable information image features were capturing. The texture feature energy detected a marker-trait association near the locus of the wheat height gene Rht-B1 and was associated with the presence of lodging in one environment.

This work is being built on to determine potential applications for high-throughput phenotyping to support the evaluation of breeding populations at earlier stages of the breeding program, and for traits which are difficult to manually measure including response to drought and heat stress.