ABSTRACT Reliable river water level forecasting is crucial for flood management and sustainable water resource planning in climate-vulnerable deltaic regions like Bangladesh, where conventional hydrological models face challenges due to data scarcity and complex monsoon-driven dynamics. This study evaluates five machine learning and deep learning models – linear regression, random forest (RF), XGBoost, light gradient boosting machine (LGBM), and long short-term memory (LSTM) – to predict daily water levels in the Old Brahmaputra River under four scenarios. Using 26 years (1999–2024) of hydrological and meteorological data, we optimized hyperparameters via grid search and 5-fold cross-validation, incorporating lagged variables (1–5 days) for temporal dependencies. Performance was assessed using six metrics, with principal component analysis for model ranking. Results revealed exceptional accuracy in the water level-only scenario (RF: MAE 0.1445 m, R2 0.9916, NSE 0.9916). Climate-driven scenarios demonstrated LSTM's superiority, achieving R2 of 0.8145 (rainfall-temperature), 0.8064 (rainfall-only), while temperature-only scenarios showed limited predictive capability (R2 0.5346). Spatial transferability assessment at Sarishabari station validated robust cross-station performance without recalibration (LGBM: MAE 0.1205 m, R2 0.9917 for water level scenario). The study provides operational frameworks for climate-driven forecasting in data-scarce deltaic environments, highlighting LSTM's exceptional capability for ungauged catchments.
Islam et al. (Thu,) studied this question.