Phenomic prediction can be improved by optimization of NIRS preprocessing

Abstract

In recent years, phenomic prediction has emerged as a new method in plant breeding that has been shown to have great potential. However, there are still many open questions regarding its practical application. For example, in the field of spectroscopy, it is standard practice to optimize the preprocessing of spectra, which so far has only been done to a limited extent for phenomic prediction. In this study, we therefore used three different data sets of soybean, triticale and maize to identify the best combinations of Savitzky–Golay filter parameters for preprocessing near‐infrared spectra for phenomic prediction. We tested 677 combinations of polynomial order, derivative and window size and evaluated them with Monte Carlo cross‐validation. Our results showed that the predictive ability can be improved with the right settings. However, there was no global optimum that gave the best results for all data sets. Even for different traits within the same data set, different combinations of parameters were necessary to achieve the highest predictive ability. Nevertheless, we show that some combinations generally result in a very low predictive ability and should not be used for preprocessing. In addition, we used the normalized discounted cumulative gain to assess whether preprocessing affected the ranking of individuals, which revealed no major changes in the top 1%, 10% or 20% of predicted individuals. Taken together, our results show the potential of preprocessing near‐infrared spectroscopy data to improve the phenomic predictive ability, but there appears to be no global optimum of parameter settings across data sets and traits.