Rainfall and temperature-driven machine learning models for multi-station water level forecasting in the old Brahmaputra river, Bangladesh

ABSTRACT Reliable water level forecasting remains critical for flood management in data-scarce deltaic regions lacking adequate hydrological monitoring infrastructure. This study evaluates five machine learning algorithms – Linear Regression, Random Forest (RF), XGBoost, Support Vector Regression, and Light Gradient Boosting Machine (LGBM)—for water level prediction in Bangladesh's Old Brahmaputra River using exclusively meteorological inputs. Daily time series data spanning 1999–2024 from three hydrological stations and one meteorological station were systematically analyzed across four predictive scenarios: water level-only (WL), rainfall-only (RL), temperature-only (MT), and combined rainfall-temperature (RL-MT). Results demonstrated exceptional WL scenario performance, with RF achieving R2 values exceeding 0.98 and mean absolute errors below 0.15 m across all monitoring stations. The RL-MT scenario yielded substantial predictive accuracy (R2: 0.72–0.78, MAE: 0.64–0.98 m), significantly outperforming individual meteorological scenarios. Ensemble algorithms consistently outperformed linear approaches, with LGBM and XGBoost achieving optimal performance in multivariate scenarios. Temperature-only predictions showed limited effectiveness (R2: 0.11–0.46), confirming rainfall dominance in short-term hydrological responses. Cross-validation scores demonstrated robust generalization capabilities across temporal partitions. These findings establish the feasibility of reliable climate-driven forecasting systems for ungauged river basins, offering practical flood risk management solutions in data-constrained deltaic environments where comprehensive monitoring infrastructure remains economically challenging.