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

Area of Science:

• Genetics
• Plant Science
• Remote Sensing

Background:

• Remote-sensing-based phenotyping is crucial for genetic studies.
• Traditional methods often rely on single vegetation indices, potentially limiting accuracy.
• Phenomic prediction (PP) offers a novel approach by utilizing collective data.

Purpose of the Study:

• To investigate the potential of phenomic prediction for genetic studies using remote-sensing data.
• To compare various statistical methods for predicting visual scores (VS) from multispectral and thermal data.
• To evaluate the impact of PP on the accuracy and signal strength in genome-wide association studies (GWAS).

Main Methods:

• Phenomic prediction (PP) was employed to predict human-assigned visual scores (VS) using remote-sensing data.
• Two sets of predictor variables were used: basic wavelengths (BT) and all traits including vegetation indices (AT).
• Statistical methods included Ordinary Least Squares (OLS), Ridge Regression (RR), LASSO, Artificial Neural Network (ANN), and Gradient Boosted Regression Tree (GBRT).

Main Results:

• BT-OLS performed comparably to the best individual vegetation index.
• Using all traits with OLS (AT-OLS) resulted in overfitting, which was mitigated by regularization (AT-RR, AT-LASSO).
• Non-linear ANN showed potential improvement, though not statistically significant.
• The strongest genetic signal purification was achieved using genomic estimated breeding values (GEBVs) instead of adjusted phenotypes.

Conclusions:

• Phenomic prediction using collective remote-sensing data can enhance genetic studies.
• Regularization techniques are essential to prevent overfitting when using comprehensive trait data.
• Combining GEBVs with Ridge Regression or ANN offers the most promising approach for improving genetic signal detection.