Application of multi-layer neural network and hyperspectral reflectance in genome-wide association study for grain yield in bread wheat

Abstract

Grain yield (GY) is a primary trait for phenotype selection in crop breeding. Rapid and cost-effective prediction of GY before harvest from remote sensing platforms can be integrated with practical breeding activities. In this study, a natural population containing 166 wheat cultivars and elite lines was used for time-series prediction of GY using ground-based hyperspectral remote sensing. Canopy hyperspectral data (350–2500 nm) was collected at the flowering, early grain-filling (EGF), mid grain-filling (MGF), and late grain-filling (LGF) stages under four environments. GY was predicted by using full bands reflectance as input of multi-layer neural network. Genome-wide association study (GWAS) was performed using 373,106 markers from 660 K and 90 K single-nucleotide polymorphism (SNP) arrays in 166 wheat genotypes. Prediction accuracy for GY characterized by R 2 values were 0.68, 0.69, 0.76, and 0.65 at flowering, EGF, MGF, and LGF, respectively. Among the 26 loci identified by predicted GY, 13 were located in similar positions to previously reported loci related to yield, and another 13 were potentially new loci. Linear regression ( R 2 ) ranged from 0.87 to 0.94 indicating that distinct cumulative effects of favorable alleles detected by predicted GY were increasing as compared to measured GY. This study highlights the feasibility of combining remote sensing with machine learning for wheat breeding decisions and to understand the underlying genetic basis of crop yield. • The use of hyperspectral reflectance to predict wheat yield is investigated. • Higher accuracy of yield prediction at the mid grain-filling stage. • Predicted yield can detect QTLs that are consistent with the measured yield.