Operational river stage forecasting in a data-scarce delta: prophet model performance across multiple lead times and hydrometeorological input scenarios

Flowchart showing the study workflow: hydrometeorological data acquisition, preprocessing, feature engineering into four input scenarios, temporal lag configurations, training of Prophet, ML, and DL models, performance evaluation, and key result of NSE approximately equal to 0.9984. River stage forecasting in data-scarce, monsoon-driven deltaic systems remains a persistent challenge for flood early warning. This study evaluates Facebook's extended Prophet framework against established machine learning (linear regression, LightGBM regression, random forest regression) and deep learning (bidirectional LSTM, gated recurrent unit) models for daily river stage prediction at multiple lead times (3, 7, 15, and 30 days) on Bangladesh's Old Brahmaputra River using 15 years of hydrometeorological data (2007–2021). Four input scenarios were assessed: (i) rainfall-only, (ii) temperature-only, (iii) combined rainfall–temperature, and (iv) river stage–rainfall–temperature integration. For meteorological-only inputs, Prophet achieved 30-day Nash–Sutcliffe efficiency (NSE) of 0.79, outperforming all alternatives – a counterintuitive result where accuracy improved with increasing lead time due to Prophet's additive seasonal decomposition. When the antecedent river stage was included, all models improved, with Prophet and linear regression achieving statistically indistinguishable accuracy (MAE ≈ 0.043 m, NSE ≈ 0.9984), demonstrating that autoregressive dependencies dominate when hydrological measurements are available. A novel finding is that temperature-only inputs sustained Prophet-based forecasting at 98.8–99.4% of rainfall-driven accuracy, providing a viable contingency for data-scarce networks. Principal component analysis–based ranking synthesized nine performance metrics, revealing systematic underprediction during extreme flood events – a critical limitation for operational deployment in vulnerable deltaic regions.