Phenomic prediction of wheat yield and grain quality using multispectral, hyperspectral, and thermal imagery

Abstract

Grain quality in durum wheat ( Triticum turgidum L. subsp. durum (Desf.) Husn.) remains a major challenge for plant breeders because grain yield and grain protein concentration often show a trade‐off, and both are influenced by grain nitrogen concentration. Unmanned aerial vehicle and airborne‐based imaging platforms offer a high‐throughput approach for estimating wheat yield and grain quality before harvest. This study developed spectral‐derived prediction models using multispectral, hyperspectral, and thermal imagery collected across growth stages from field experiments conducted in northwest Mexico from 2014 to 2017 under two tillage systems, two irrigation levels, and six nitrogen treatments. Five image‐derived predictor sets were evaluated: vegetation indices, canopy temperature, spectral bands, vegetation indices plus canopy temperature, and spectral bands plus canopy temperature. These predictor sets were tested using stepwise regression, partial least squares, and decision tree models. In multispectral data, the spectral‐band model combined with partial least squares achieved R ²  = 0.74–0.77, whereas vegetation‐index‐based models achieved R ²  = 0.84–0.91. Prediction accuracy varied by trait, with R ² up to 0.84 for grain yield, approximately 0.91 for grain nitrogen concentration, and 0.74–0.77 for grain protein concentration. Growth stages GS41 and GS73 were the most informative stages, identifying useful preharvest decision stages for phenomic prediction of wheat yield and grain quality.