River water level forecast sensitivity to hydrometeorological input variable selection: a comparison of machine and deep learning models in Bangladesh

Conceptual summary showing LSTM and GRU outperform ML models under meteorological-only inputs, while RFR and LGBMR dominate when lagged water level data are available, across five scenarios in the Old Brahmaputra River, Bangladesh. Accurate river water level forecasting is critical for flood management in Bangladesh's monsoon-dominated deltaic environment, yet optimal model selection under varying data availability remains unexplored. This study systematically compared six machine learning (ML) and deep learning (DL) algorithms, specifically linear regression, support vector regression, random forest regression (RFR), light gradient boosting machine regression (LGBMR), long short-term memory (LSTM), and gated recurrent unit (GRU), for water level prediction at Islampur station using 25 years of daily hydrometeorological data (1999–2024) across five input scenarios employing 5-day lag configurations. DL architectures dominated meteorological-only scenarios, with LSTM achieving R2 = 0.8809, explained variance score (EVS) = 0.8874, and symmetric mean absolute percentage error (SMAPE) = 4.83%, confirming unbiased symmetrical predictions. ML algorithms excelled with autoregressive inputs, where LGBMR achieved R2 = 0.9960 and RFR attained R2 = 0.9984; the latter's EVS = 0.9984 and SMAPE = 0.40% confirmed purely stochastic residuals. Principal component analysis (PCA)-based composite scoring of nine complementary metrics provided objective model rankings. Zero-shot spatial transferability at Sarishabari confirmed reproducible hierarchies without retraining, with violin plots and PCA composite scoring corroborating near-zero prediction bias across all transferred configurations. Results demonstrate that LSTM and GRU enable reliable forecasting under input-limited, meteorological-only conditions while ensemble ML provides superior accuracy when water level observations are available, offering actionable guidance for flood forecasting in data-constrained environments.