Explaining Great Lakes water level variability through interpretable ensemble machine learning

Understanding the environmental drivers of water-level variability in the Great Lakes is critical for water-resource planning, ecosystem resilience, and binational policy. This study develops an interpretable, multi-model machine learning framework to quantify the immediate and lagged controls of environmental drivers on monthly lake-level fluctuations in Lakes Superior, Michigan, Erie, and Ontario over 1982-2022. Eight tree-based algorithms (Random Forest, Extra Trees, Gradient Boosting Regression Trees (GBRT), Histogram-Based Gradient Boosting (HGBRT), XGBoost, LightGBM, CatBoost, and AdaBoost) were trained using a time-aware cross-validation scheme with lagged predictors up to six months, and their complementary strengths were integrated through a Supervised Committee Machine Learning (SCML) ensemble. Boosting models (XGBoost, LightGBM, HGBRT) consistently outperformed Random Forest and AdaBoost across the upper lakes, while the SCML ensemble delivered the most stable prediction overall, achieving RMSE values as low as 0.118 m and improving test-set performance particularly in challenging periods and high-variance regimes. To unravel the governing processes, SHapley Additive exPlanations (SHAP) were paired with Variogram Analysis of Response Surfaces (VARS), providing a complementary view of direct and lag-dependent driver influence. SHAP revealed that inflow and outflow overwhelmingly dominate lake-level dynamics, with evaporation, runoff, and air temperature acting as secondary but lake-specific modulators. VARS further exposed strong hydrological memory, identifying shifting sensitivities at lags 3-4 and showing that hydrological fluxes (inflow, outflow, runoff) become increasingly influential at longer lags, while atmospheric drivers govern short-term responses. Lake-specific contrasts-such as Ontario's dampened hydrological memory due to outflow regulation and Erie's strong dependence on upstream inflow-highlight how morphometry, climate, and regulation shape predictability. By integrating predictive accuracy with transparent interpretability, this study advances mechanistic understanding of Great Lakes water-level variability and provides a robust diagnostic framework to support adaptive management under climate variability and increasing anthropogenic pressures.