Comparing statistical 'phenomic prediction' models for remote-sensing-based phenotyping of maize susceptibility to...

We investigate the potential of phenomic prediction (PP) in remote-sensing-based phenotyping for genetic studies. Rather than relying on a single vegetation index, we utilize all available data collectively to predict the human-assigned visual score (VS). The conceptual motivation is that when a trained model is available, these predictions may provide a more accurate assessment of disease symptoms than the use of a specific vegetation index (VI). To evaluate the PP approach, we employ the predicted VS in a genome-wide association study (GWAS) and consider strength and position of the detected genetic signal. We use two different sets of predictor variables: i) the five basic wavelengths captured by a multispectral and a thermal camera (basic traits model, BT) or ii) all traits (AT), consisting of the five basic wavelengths plus ten vegetation indices. As statistical methods, we compare a) (linear) ordinary least squares regression (OLS), b) (linear) ridge regression (RR), c) (linear) least absolute shrinkage and selection operator (LASSO) d) an artificial neural network (ANN) and e) a gradient boosted regression tree method (GBRT). Our results indicate that the simple linear OLS regression on the five basic wavelengths (BT-OLS) performs on a level comparable to the best individual vegetation index G. The use of all traits in the OLS regression (AT-OLS) leads to overfitting, which was prevented by the regularization in AT-RR and AT-LASSO. The non-linear ANN approach seems to improve the results further, but the differences between the methods were not statistically significant. The strongest improvement for the purification of the genetic signal was observed when genomic estimated breeding values (GEBVs) for the different traits (VS, basic wavelengths, vegetation indices) instead of their adjusted phenotypes were used. Across all approaches, the combination of GEBVs with Ridge Regression or the non-linear ANN provided the best results.