A latent factor approach to hyperspectral time series data for multivariate genomic prediction of grain yield in wheat

TL;DR

This study tackles the challenge of incorporating high-dimensional hyperspectral time-series data into genomic prediction for wheat grain yield. It introduces a latent-factor approach with Procrustes rotation to extract time-consistent, interpretable features from hyperspectral data, which are then used as secondary traits in a multivariate genomic prediction model. Across a CIMMYT wheat trial with three watering treatments, the method yields measurable gains in predictive ability over univariate models and clarifies which timepoints and spectral regions drive predictions. The approach offers a scalable, interpretable framework for leveraging sensor-derived phenotypes in plant breeding, with implications for optimizing phenotyping strategies and multi-environment analyses.

Abstract

High-dimensional time series phenotypic data is becoming increasingly common within plant breeding programmes. However, analysing and integrating such data for genetic analysis and genomic prediction remains difficult. Here we show how factor analysis with Procrustes rotation on the genetic correlation matrix of hyperspectral secondary phenotype data can help in extracting relevant features for within-trial prediction. We use a subset of Centro Internacional de Mejoramiento de Maíz y Trigo (CIMMYT) elite yield wheat trial of 2014-2015, consisting of 1,033 genotypes. These were measured across three irrigation treatments at several timepoints during the season, using manned airplane flights with hyperspectral sensors capturing 62 bands in the spectrum of 385-850 nm. We perform multivariate genomic prediction using latent variables to improve within-trial genomic predictive ability (PA) of wheat grain yield within three distinct watering treatments. By integrating latent variables of the hyperspectral data in a multivariate genomic prediction model, we are able to achieve an absolute gain of .1 to .3 (on the correlation scale) in PA compared to univariate genomic prediction. Furthermore, we show which timepoints within a trial are important and how these relate to plant growth stages. This paper showcases how domain knowledge and data-driven approaches can be combined to increase PA and gain new insights from sensor data of high-throughput phenotyping platforms.

Figures (8)

• Figure 1: A graphical illustration of our dimensionality reduction and multivariate genomic prediction workflow. Starting with a genetic covariance matrix of secondary traits (red), we estimate lower dimensional factor scores and loadings per timepoint. The factor loadings are reordered using Procrustes rotation to trace factor scores across timepoints. Next, we use subset selection to relate factor scores ($\mathbb{\hat{\xi}}$) to the focal trait ($\mathbf{\hat{\mathbf{y}}}_{f}$). The selected factor scores ($\mathbf{\tilde{\Xi}}$) and focal trait ($\mathbf{\hat{\mathbf{y}}}_{f}$) are then used in multivariate genomic prediction for unseen genotypes through a (genomic) relationship matrix ($\mathbf{K}$). Complete explication of the notation and modeling can be found in the sections below.
• Figure 2: Illustration of using smoothing splines per genotype based on factor 1 within the Optimal Flat treatment. In terms of yield, the bottom 10 genotypes (red) diverge from the best 10 genotypes (green) around March. Moreover, the population spline is indicated by the black line at $y = 0$. Hence, one can see the factor space also as a normed space.
• Figure 3: Boxplot (a) and scatter plots (b) of genotypic means (best linear unbiased estimators) of grain yield in tonnes per hectares (t/ha) indicated on the axis of the respective treatment within each scatter plot. Pearson correlation of each pair of treatments displayed as $\rho$.
• Figure 4: Wavelength reflectivity loadings on factor 1 and 2 of hyperspectral reflectivities for each timepoint and treatment before (top panel) and after (bottom panel) Procrustes rotation
• Figure 5: Scatter plots and Pearson's correlation ($\rho$) of genotypic mean of each factor score within each treatment. Factor is abbreviated in axis labels as 'F' for readability