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

Introduction: Accurate crop yield prediction remains challenging because yield formation is influenced by complex interactions between plant traits and soil environments across multiple soil layers. Although root traits are widely recognized as important for crop resource acquisition and yield formation, their relative contributions across soil depths and their interannual stability have not been fully quantified. This limits their effective use in trait-based yield prediction. Therefore, identifying key root indicators with stable predictive relevance and developing physiologically interpretable feature frameworks are important for improving the analysis and prediction of crop yield variation. Methods: Based on multi-year field experiment data, winter wheat root traits across the 0-60 cm soil profile were analyzed using a layer-wise approach. To ensure consistency with the scale of grain yield observations, multi-point and multi-layer root measurements were aggregated into plot-year-level feature variables before modeling. A trait-prioritization and feature-organization strategy integrating correlation analysis, interannual stability evaluation, weighting sensitivity analysis, and model-based contribution analysis was used to identify representative root traits and construct a hierarchical root-trait feature framework. These analyses were used to support trait prioritization and feature organization, rather than as prediction-optimized trait preselection prior to cross-validation. Representative machine-learning models from different categories were then compared under a unified preprocessing and year-grouped cross-validation workflow to evaluate the predictive performance of different root-trait combinations. Results and discussion: The results showed clear soil-depth dependency in the relationship between root traits and winter wheat yield. Root traits in the 20-30 cm soil layer showed the most stable and informative associations with yield variation under the present experimental conditions. Root dry matter density in the 20-30 cm layer (RDMD at 20-30 cm) was identified as the most representative single predictive trait, showing relatively strong performance in correlation strength, interannual stability, and model-based contribution. Further analyses indicated that incorporating additional root volume- and surface-area-related traits from the 20-30 cm layer improved predictive performance, while deep root distribution indicators from 30-60 cm provided complementary information. Comparative modeling results showed that nonlinear ensemble models generally performed better than linear models in capturing multi-layer root-trait information, with AdaBoost achieving the best cross-validated performance under the profile distribution-enhanced feature tier, with a mean R