Benchmarking process-based SWAT and customized LSTM rainfall-runoff models for the climate change impact assessment on the streamflow of the Upper Meghna River Basin

ABSTRACT Graphical abstract showing climate projections, SWAT-based hydrological modeling, and LSTM machine learning used to predict future river flow changes from rainfall and temperature inputs. The transboundary Upper Meghna Basin (UMB), shared by Bangladesh and India, is very susceptible to climate change due to its complex region characterized by its mountainous origins and deltaic terrain. Cities in the northeastern districts of Bangladesh are particularly vulnerable to climate change, especially in the context of flash floods and monsoon floods. The lack of upstream data limits the performance and efficiency of traditional hydrological models in this basin. To overcome these constraints, this study integrates a deep learning-based Long Short-Term Memory (LSTM) model with CMIP6 models and compares results with a process-based Soil and Water Assessment Tool (SWAT) model. LSTM model outperformed SWAT in hydrological performances, achieving a higher R2 of 0.95 and 0.91 against 0.86 and 0.89, and lower errors (RMSE) for calibration and validation. Annual streamflow will be increased by 7.1% (3.0%–20.2%) and 5.3% (3.1%–14.3%) by the mid-century (2031–2060), and 10.9% (6.9%–20.5%) and 15.0% (6.6%–27.3%) by the end of the century (2071–2100) under SSP2-4.5 and SSP3-7.0, respectively. Monthly low flows in dry season and peak flows in pre-and post-monsoon may rise as much as 30.8% (−3.4%–60.8%), 14.5% (−2.5%–36.3%), and 17.0% (−5.2%–45.0%) by the mid-century under SSP3-7.0.