Barzegar et al. (2026) Explaining Great Lakes water level variability through interpretable ensemble machine learning

Identification

• Journal: The Science of The Total Environment
• Year: 2026
• Date: 2026-01-01
• Authors: Rahim Barzegar, Ehsan Raei, Jan Adamowski
• DOI: 10.1016/j.scitotenv.2025.181302

Research Groups

• Groundwater Research Group (GRES), Research Institute on Mines and Environment (RIME), Université du Québec en Abitibi-Témiscamingue (UQAT), Amos, Québec, Canada
• Department of Bioresource Engineering, McGill University, Sainte-Anne-de-Bellevue, Quebec, Canada
• United Nations University Institute for Water, Environment and Health (UNU-INWEH), Richmond Hill, ON, Canada

Short Summary

This study develops an interpretable ensemble machine learning framework to quantify the immediate and lagged controls of environmental drivers on monthly water-level fluctuations in the Great Lakes (Superior, Michigan, Erie, Ontario). It reveals that boosting-based models and an ensemble approach significantly improve predictions, with inflow and outflow being dominant drivers, while temperature, evaporation, and runoff act as secondary, lake-specific modulators.

Objective

• To quantify the immediate and lagged controls of environmental drivers on monthly water-level fluctuations in Lakes Superior, Michigan, Erie, and Ontario using an interpretable multi-model machine learning framework.

Study Configuration

• Spatial Scale: Lakes Superior, Michigan, Erie, and Ontario (Great Lakes region).
• Temporal Scale: Monthly fluctuations over the period 1982–2022, with lagged predictors up to six months.

Methodology and Data

• Models used: Eight tree-based algorithms (Random Forest, Extra Trees, Gradient Boosting Regression Trees (GBRT), Histogram-Based Gradient Boosting (HGBRT), XGBoost, LightGBM, CatBoost, and AdaBoost) integrated into a Supervised Committee Machine Learning (SCML) ensemble. SHapley Additive exPlanations (SHAP) and Variogram Analysis of Response Surfaces (VARS) were used for interpretability.
• Data sources: Environmental drivers used as predictors include inflow, outflow, evaporation, runoff, and air temperature.

Main Results

• Boosting-based machine learning models (XGBoost, LightGBM, HGBRT) significantly improved Great Lakes water-level prediction compared to Random Forest and AdaBoost.
• The Supervised Committee Machine Learning (SCML) ensemble provided the most stable overall predictions, achieving Root Mean Square Error (RMSE) values as low as 0.118 m.
• The SHAP–VARS framework effectively identified dominant drivers and revealed their lagged sensitivities.
• Inflow and outflow are the overwhelming dominant drivers of lake-level dynamics across the Great Lakes.
• Evaporation, runoff, and air temperature act as secondary but lake-specific modulators of water levels.
• Runoff and inflow primarily govern water levels in Lakes Superior and Michigan, while inflow overwhelmingly drives Lake Erie's dynamics.
• Temperature and evaporation exert strong long-lag effects, particularly evident in Lake Ontario.

Contributions

• Development of an interpretable, multi-model machine learning framework for enhanced Great Lakes water-level prediction.
• Quantification of immediate and lagged controls of environmental drivers on lake levels using a novel SHAP–VARS framework.
• Identification of dominant and secondary environmental drivers and their lake-specific lagged sensitivities across the Great Lakes.
• Significant improvement in water-level prediction stability and accuracy through the application of a Supervised Committee Machine Learning (SCML) ensemble.

Funding

• Not mentioned in the provided paper text.

Citation

@article{Barzegar2026Explaining,
  author = {Barzegar, Rahim and Raei, Ehsan and Adamowski, Jan},
  title = {Explaining Great Lakes water level variability through interpretable ensemble machine learning},
  journal = {The Science of The Total Environment},
  year = {2026},
  doi = {10.1016/j.scitotenv.2025.181302},
  url = {https://doi.org/10.1016/j.scitotenv.2025.181302}
}