Leveraging unmanned aerial vehicle derived multispectral data for improved genomic prediction in potato (Solanum tuberosum)

Abstract

Multispectral leaf canopy reflectance as measured by unmanned aerial vehicles is the result of genetic and environmental interactions driving plant physiochemical processes. These measures can then be used to construct relationship matrices for modeling genetic main effects. This type of phenotypic prediction is particularly relevant for trials with many entries, such as those used in early generation potato (Solanum tuberosum) breeding. We compared three methods for making predictions in our potato breeding program: first, using multispectral-derived relationship matrices; second, using the traditional approach based on genomic derived relationships; and third, using a combination of both. Multispectral bands were collected at five different time points for two market classes of potato: chipping and fresh market. We modeled genetic main effects for yield and quality traits at each time point and all stages combined. Models with multispectral relationship matrices exhibited better prediction accuracy for yield and roundness than genomic only models and models featuring spectra plus genomic kernels outperformed both single-kernel predictions in terms of accuracy for most traits. Time points were variably informative depending on the trait measured, however, for all traits combining across time points performed as well or better than single time point models. Similarly, using feature selection to limit our models to important variables did not improve prediction accuracy significantly. This work highlights two potential uses for spectral data in genomic prediction: first, as an alternative to genetic data and second, in combination with genetic data to increase precision of selection.

FIGURE 1

Trait best linear unbiased estimates (BLUEs) distribution for (A) chipping clones and (B) fresh market clones.

FIGURE 2

Broad‐sense heritability for both chipping clones and fresh market clones.

FIGURE 3

Comparison of models including genomic (G), genome by environment (GE), and/or multispectral (W) relationship matrices for chip traits across time points. Comparison is based on prediction ability in terms of Pearson's correlation of model‐predicted values to phenotypic observations across both years.

FIGURE 4

Comparison of models including genomic (G), genome by environment (GE), and/or multispectral (W) relationship matrices for traits measured in fresh market clones across time points. Comparison is based on prediction ability in terms of Pearson's correlation of model‐predicted values to phenotypic observations across both years.

FIGURE 5

Comparison of year‐specific prediction ability for models including genome (G), genome by environment (GE), and/or multispectral (W) data combined across time points for chipping clones. Prediction ability is the Pearson's correlation of the model prediction to the observed phenotypic value.

FIGURE 6

The influence of time point of spectral data collection on prediction ability for various traits in (A) chipping potatoes and (B) fresh market clones.

FIGURE 7

Prediction ability for chipping clone traits across time points using subsets of the multispectral variables. SA, simulation annealing; SF, selection by filtering.