Feature engineering and parameter tuning: improving phenomic prediction ability in multi-environmental durum wheat...

Area of Science:

• Plant breeding
• Agricultural science
• Spectroscopy

Background:

• Phenomic selection predicts plant performance using spectral data.
• Efficient selection is vital for successful plant breeding programs.
• Near-infrared spectra (NIRS) offer a tool for phenomic selection.

Purpose of the Study:

• To optimize phenomic selection in durum wheat.
• To improve prediction ability using feature engineering and parameter tuning.
• To investigate genotype and environment effects on prediction for agronomic and quality traits.

Main Methods:

• Applied phenomic selection in multi-environmental durum wheat trials.
• Utilized feature engineering and parameter tuning for prediction improvement.
• Employed grid search over Savitzky-Golay filter parameters for preprocessing optimization.

Main Results:

• Preprocessing improved phenomic prediction ability by up to 1500%.
• Optimized preprocessing is dataset, trait, and model-dependent.
• Phenomic prediction showed low-to-moderate abilities, highest when predicting within the same environment.

Conclusions:

• Optimized phenomic selection in durum wheat requires tailored NIRS preprocessing.
• Feature engineering and parameter tuning enhance predictive power.
• NIRS captures genotype and genotype-by-environment interaction variance.