Principal variable selection to explain grain yield variation in winter wheat from features extracted from UAV...

Area of Science:

• Agricultural Science
• Plant Breeding
• Remote Sensing

Background:

• Automated phenotyping accelerates crop breeding but generates vast datasets.
• Reducing the number of secondary traits for phenotyping can significantly decrease data processing efforts.
• Identifying key predictive features is crucial for efficient breeding programs.

Purpose of the Study:

• To identify principal features from UAV imagery and critical growth stages that best predict winter wheat grain yield.
• To reduce phenotyping efforts by selecting a minimal set of highly predictive variables.

Main Methods:

• Collected multispectral and RGB imagery using a UAV system throughout the growing season.
• Extracted 172 variables from vegetation index and plant height maps, including pixel statistics and growth rates.
• Applied LASSO regression and random forest algorithms for variable selection to identify key yield-influencing features.

Main Results:

• Both LASSO and random forest identified plant height variables around grain filling as most important.
• Selected vegetation index variables from early to senescence stages were also crucial.
• Yield prediction using selected variables was comparable or superior to using all 172 variables.
• Prediction accuracy was higher for adapted NE lines (r=0.58-0.81) than other lines (r=0.21-0.59).

Conclusions:

• UAS-based phenotyping enables derivation of detailed features (e.g., within-plot variation) beneficial for breeding.
• Feature selection significantly reduces data complexity while maintaining high grain yield prediction accuracy.
• This approach allows for better allocation of resources in phenotypic data collection and processing for crop improvement.