Comparative evaluation of machine learning models for extreme river water level forecasting in Bangladesh: Implications for flood and drought resilience

Abstract

Reliable forecasting of extreme river water levels is vital for managing flood and drought risks in Bangladesh, a deltaic nation highly vulnerable to climate change. This study compares nine machine learning (ML) models for predicting monthly maximum and minimum water levels at three key stations along the Old Brahmaputra River using a 34-year dataset (1990–2024). Performance was assessed using ten metrics including RMSE, R², and NSE. Random Forest Regression (RFR) consistently outperformed other models, achieving the highest accuracy for both maximum (RMSE: 0.64–0.77 m; R²: 0.87–0.92) and minimum water levels (RMSE: 0.49–0.66 m; R²: 0.82–0.92), while linear models underperformed in capturing nonlinear patterns. A PCA-based framework further validated RFR's robustness, with average normalized composite scores of 1.00 (maximum) and 0.99 (minimum), significantly surpassing Ensemble Regression (0.89/0.84), Support Vector Regression (0.88/0.88), and other models. Spatially, midstream stations showed higher accuracy (R² > 0.90) due to stable hydrodynamics, while downstream performance decreased from tidal effects. Key innovations including autoregressive lag features, sliding windows, and a multivariate evaluation framework significantly improved prediction accuracy. These findings demonstrate that ML models can enhance water level forecasting and disaster resilience in climate-vulnerable regions, even with limited data.