Integrating phenological and morphological traits for yield prediction in lucerne (Medicago sativa) by using machine learning algorithms

Context Accurate prediction of phenological and morphological traits in alfalfa (Medicago sativa L.) is essential to improve yield performance and increase selection efficiency in breeding programs. Aims To identify morpho-phenological traits driving total dry matter yield in local lucerne genotypes and to evaluate the predictive ability of multiple machine-learning (ML) algorithms. Methods Field data were collected over 2 years with three replicates per trait (n = 360 observations). Correlations between yield and traits were assessed using Pearson’s correlation coefficient, r. Four ML algorithms – random forest (RF), elastic net regression (ENet), extreme gradient boosting (XGBoost), and support vector regression (SVR), were trained to predict yield; performance was compared using coefficient of determination (R2) and root mean square error (RMSE). Trait importance was examined across models. Key results Yield was strongly positively associated with winter dormancy, post-harvest regrowth rate and plant resistance (r ≈ 0.67–0.68). RF achieved the best predictive performance (R2 = 0.61; RMSE = 26.49). Five traits (winter dormancy, post-harvest regrowth, resistance, plant form and root-crown bud number) consistently ranked as the most influential predictors. Conclusions RF best captured the partially non-linear, interaction-driven yield structure, and pinpointed a coherent set of morpho-phenological predictors aligned with classical correlation outcomes. Implications Integrating morpho-phenological traits with ML substantially improves yield predictability and provides a practical, data-driven framework to prioritise traits (e.g. winter dormancy, regrowth, resistance, plant form, root-crown buds) for selecting high-yielding, stress-resilient lucerne genotypes.