Speaker
Description
Genomic selection has become a powerful tool in sugar beet breeding, but genotyping costs remain a barrier to routine, large-scale application. Phenomic selection, using seed near-infrared spectroscopy (NIRS) and UAV-derived field phenotyping as a cheaper alternative to marker-based prediction, is gaining attention, but rigorously benchmarking such pipelines across many genotypes, environments, preprocessing variants, and prediction models requires large-scale, repeated cross-validation, a computational workload well beyond desktop capacity.
In this study, conducted at the University of Hohenheim in collaboration with Strube D&S GmbH, we benchmarked seed-NIRS-based phenomic prediction for three sugar beet traits (corrected sugar yield, corrected root yield, sugar content) using data from 32 field trial locations across Germany, France, Belgium, and the Netherlands (2025 hybrid breeding program). We compared 190 NIRS profiles (raw + 189 Savitzky-Golay preprocessed profiles) across five prediction models (Elastic Net, LASSO, Partial Least Squares Regression, Ridge Regression BLUP, Random Forest), assessed the contribution of three UAV-derived field-emergence covariates (plant stand count, canopy cover percentage, plant instance area) to phenomic prediction accuracy, and modeled genotype-by-environment interaction under four cross-validation schemes (CV1, CV2, CV00, leave-one-location-out).
All computations ran on bwUniCluster 3.0 at KIT's Scientific Computing Center: more than 370 nodes, including 340+ CPU nodes (Intel Ice Lake, 64 cores/node; AMD EPYC 9454, 96 cores/node) and 28 NVIDIA A100/H100 GPU nodes, on a Lustre parallel file system. Screening 190 profiles × 5 models × 3 traits under 100 repetitions of 5-fold cross-validation produced roughly 1.4 million individual train-test model fits for the core phenomic prediction objective; evaluating the UAV covariates across 3 profiles × 3 models × 3 traits × 4 covariate configurations under 100 repetitions of 10-fold cross-validation added a further approximately 108,000 fits. Genotype-by-environment models were additionally fitted in a Bayesian RKHS framework (BGLR, 12,000 MCMC iterations, 2,000 burn-in per model) across three traits, two NIRS profiles, and four cross-validation schemes. We parallelized this workload as SLURM array jobs, submitting each profile-model-trait combination as an independent task, which made this scale of benchmarking practically feasible within the thesis timeline.
Savitzky-Golay preprocessing consistently improved predictive ability over raw spectra, with narrow window sizes and higher-order derivatives most effective; sugar content remained the hardest trait to predict. Among UAV covariates, plant instance area and canopy cover percentage improved predictive ability for yield traits, while plant stand count generally reduced it. Under the genotype-by-environment model, predictive ability stayed high when genotype or environmental information was available in training but declined substantially when both were absent.
These results show that seed-NIRS-based phenomic selection, supported by UAV-derived covariates, is a promising, cost-effective approach for genotype evaluation in sugar beet breeding. More broadly, they illustrate how bwHPC infrastructure was not merely a convenience but a critical enabler of this thesis: without large-scale parallel computing, systematically benchmarking hundreds of thousands of model fits across preprocessing profiles, models, and cross-validation schemes would not have been feasible.