Enhancing monthly extreme water level predictions in a flood-prone river basin using regression-based machine learning
Md. Touhidul Islam, Sujan Chandra Roy, Nusrat Jahan, MN Islam, Asif Ahammed, Irfan Rizka Akbar, A. K. M. Adham
Abstract
ABSTRACT Reliable forecasting of extreme river water levels is crucial for flood mitigation, agricultural planning, and disaster preparedness, particularly in vulnerable regions like Bangladesh. Traditional hydrological models often struggle with the nonlinear dynamics of deltaic rivers influenced by monsoons, tides, and human activities. This study evaluates six regression-based machine learning (ML) models – linear regression (LR), random forest regression (RFR), XGBoost (XGBR), multilayer perceptron (MLPR), LightGBM (LGBMR), and polynomial regression (PR) – for predicting monthly maximum and minimum water levels in Bangladesh's Old Brahmaputra River. Using 34 years of data (1990–2024) from Islampur station, models were assessed via performance metrics and principal component analysis (PCA). Results demonstrated RFR's superior accuracy for maximum (R2 = 0.934, RMSE = 0.646 m) and minimum (R2 = 0.942, RMSE = 0.469 m) water levels, achieving the highest PCA-based composite scores. LGBMR and XGBR followed closely (R2 > 0.930, RMSE < 0.700 m), while MLPR performed poorly (RMSE = 0.978 m, R2 = 0.850). The significant performance gap between tree-based ensemble methods and other approaches highlights RFR's robustness in modeling nonlinear hydrological patterns. These findings underscore the potential of ML models for improving flood forecasting in data-scarce regions, aiding adaptive water management in the Brahmaputra Basin and similar flood-prone systems.