Hybrid phenological modelling for enhancing plant phenology prediction

Accurate phenological forecasting is vital for ecological monitoring, climate adaptation, and agricultural decision-making. Common modelling approaches, which include Process-Based Models (PBMs) and Machine Learning (ML), offer contrasting strengths: PBMs provide biological interpretability but struggle with ecological complexity and transferability, while ML models are data-driven and adaptable but often lack physiological transparency. This study investigated two hybrid phenology modelling strategies for predicting budburst dates under climate variability. The first strategy applied an ML-derived parameterization of the UNIFORC Process-Based Model, dynamically estimating site- and year-specific parameters. The second combined multiple fixed-parameter UNIFORC variants and machine learning regressors using a linear ensemble metamodel. Both approaches were evaluated over a 15-year period (2001–2015) using a year-forward chaining strategy and windowed climate data. The ML-derived parameterization was found to produce ecologically meaningful parameter dynamics, such as plausible interannual shifts in the thermal forcing threshold (f_crit ranging from ~41 to 48) and forcing onset (t0 ranging from DOY 25 to 29), while maintaining relatively consistent predictive accuracy (mean RMSE = 8.70, MAE = 6.99). In contrast, the ensemble model achieved comparable overall performance (mean RMSE = 8.17, MAE = 6.93), but exhibited a pronounced Regression-to-the-mean effect, evidenced by a lower slope (0.08) and R² (0.02) in observed vs. predicted plots. This study concluded that while both models offer predictive value, the ML-derived parameterization provides superior interpretability and biological transparency, making it more suitable for applications requiring ecological insight and climate adaptability. Keywords: ML-derived Parameterization, Hybrid Modelling, UNIFORC, Year-Forward-Chaining, Parallel Ensemble