Optimizing runoff simulation in three mid-high latitude catchments by integrating terrestrial ecosystem modelling, hybrid machine learning, and causal inference

Study region: Three mid-high latitude catchments include the Krycklan (Boreal Sweden), the Danube (Central Europe) and the Mississippi (USA). Study focus: We develop a hybrid eco-hydrological framework that couples the process-based terrestrial ecosystem model LPJ-GUESS (Lund-Potsdam-Jena General Ecosystem Simulator) with five machine-learning (ML) algorithms through two hybrid model structures. Monthly runoff is simulated for each catchment and benchmarked against both stand-alone LPJ-GUESS and isolated ML models. Model attribution is carried out with the interpretable ML algorithm SHapley Additive exPlanations (SHAP), while causal forests estimate average treatment effects (ATEs) of key drivers on prediction bias, thereby bridging correlation and causation. New hydrological insights for the region: Hybrid models raise Nash–Sutcliffe efficiency by 0.4–1.9 relative to original LPJ-GUESS and reduce peak-timing phase bias, while demanding only modest extra computation. SHAP consistently ranks incoming radiation, temperature and precipitation as the leading factors on runoff, but shows that LPJ-GUESS underweights the radiation effect by 11–12 % in Krycklan and Mississippi catchments. Causal-forest inference analysis confirms a strong radiation-driven bias (ATE: –0.13, –0.27 and –0.59 s.d. for Krycklan, Danube and Mississippi catchment, respectively) even after controlling for other drivers. The results demonstrate missing energy-balance processes as an important source of model error and provide quantitative guidance for future model development in cold-region catchments.