Leveraging root trait identification for winter wheat yield prediction using machine learning

Area of Science:

• Agricultural Science
• Plant Science
• Soil Science

Background:

• Crop yield prediction is complex due to plant-soil interactions.
• Root traits are crucial for resource acquisition but their depth-specific roles and stability are unclear.
• Quantifying root trait contributions is vital for trait-based yield prediction.

Purpose of the Study:

• To identify key root indicators with stable predictive relevance for winter wheat yield.
• To develop a physiologically interpretable framework for analyzing root trait contributions across soil depths.
• To evaluate the predictive performance of different root trait combinations using machine learning.

Main Methods:

• Analyzed multi-year winter wheat field data using a layer-wise approach (0-60 cm soil profile).
• Employed a trait-prioritization strategy including correlation, stability, sensitivity, and contribution analyses.
• Compared machine learning models under a unified preprocessing and cross-validation workflow.

Main Results:

• Root traits in the 20-30 cm soil layer demonstrated the most stable and informative associations with yield.
• Root dry matter density (RDMD) at 20-30 cm was the most representative single predictive trait.
• Nonlinear ensemble models, particularly AdaBoost, achieved the best predictive performance (R²=0.74), with key predictors concentrated in the 20-30 cm layer.

Conclusions:

• Root traits exhibit functional differentiation across soil layers, impacting winter wheat yield.
• A trait-based, physiologically interpretable framework enhances yield variation analysis.
• The study provides a methodological reference for integrating root traits with environmental and management data for improved crop yield prediction.