Shahariar et al. (2026) Enhanced Streamflow Prediction Using a Hybrid CNN-LSTM Model: Integrating Climate-Driven Spatio-Temporal Features in a Deep Learning Framework

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-16
• Authors: Shadman Shahariar, Hasan Zobeyer, Nasreen Jahan, Md. Mostafa Ali
• DOI: 10.1007/s11269-026-04491-9

Research Groups

• Bangladesh University of Engineering and Technology, Dhaka, Bangladesh

Short Summary

This study developed a climate-driven hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model for daily streamflow prediction in the Brahmaputra River Basin, relying solely on precipitation and temperature. The model demonstrated superior performance over standalone deep learning models and achieved comparable or superior skill to a calibrated SWAT model, particularly for low-flow and high-flow extremes.

Objective

• To develop and evaluate a climate-driven hybrid CNN-LSTM model for daily streamflow prediction in the Brahmaputra River Basin, relying exclusively on gridded climatic predictors (precipitation and temperature) without incorporating historical discharge as a model input.
• To benchmark the predictive performance of the proposed deep learning models against standalone deep learning architectures (CNN, LSTM, MLP) and a calibrated, physically-based SWAT hydrological model under identical climatic forcing.

Study Configuration

• Spatial Scale: Brahmaputra River Basin (approximately 580,000 km²), focusing on the Bahadurabad outlet. Gridded climate data at 0.5° resolution (180 cells).
• Temporal Scale: Daily streamflow prediction with a 1-day lead time. Data period: 1985–2014. Training period: 1985–2002 (60% of data). Testing period: 2003–2014 (40% of data). Optimal input lag length: 15 days.

Methodology and Data

• Models used: Hybrid CNN–LSTM, standalone Convolutional Neural Network (CNN), standalone Long Short-Term Memory (LSTM), Multilayer Perceptron (MLP), and a benchmark calibrated Soil and Water Assessment Tool (SWAT) model.
• Data sources:
  • Daily rainfall, maximum temperature, and minimum temperature: W5E5 reanalysis dataset (global data at 0.5° resolution, 1985–2014).
  • Observed daily discharge data: Bangladesh Water Development Board (BWDB) at Bahadurabad Transit (1985–2014), used solely as target outputs.

Main Results

• The hybrid CNN–LSTM model achieved the highest predictive skill, with Nash–Sutcliffe Efficiency (NSE) values of 0.92 during training and 0.86 during testing, and Root Mean Squared Error (RMSE) values of 5,401 cubic meters per second and 5,700 cubic meters per second, respectively.
• Rainfall was identified as the dominant driver of streamflow variability, while temperature, when combined with rainfall, enhanced predictive skill, particularly during low-flow and transitional periods.
• An optimal input lag length of 15 days was determined, effectively capturing the basin's delayed and nonlinear rainfall–runoff response.
• The 60:40 training-testing data split (1985–2002 for training, 2003–2014 for testing) yielded the best generalization performance.
• The CNN–LSTM model consistently outperformed standalone CNN, LSTM, and MLP models across all performance metrics.
• Benchmarking against the calibrated SWAT model showed that CNN–LSTM achieved comparable or superior performance. Specifically, CNN–LSTM demonstrated stronger skill in low-flow conditions (NSE = 0.92 vs. SWAT NSE = 0.89) and substantially outperformed SWAT for high-flow extremes (top 5% flows, CNN–LSTM NSE = 0.68 vs. SWAT NSE = 0.45), with lower peak-flow bias (-4.86%) and high-flow volume error (-9.52%).
• CNN–LSTM also exhibited superior accuracy in peak-flow timing, with errors tightly clustered around zero.

Contributions

• Development of a novel climate-driven hybrid CNN-LSTM framework for daily streamflow prediction that relies exclusively on universally available climatic variables (precipitation and temperature), without requiring historical discharge data as input.
• Successful application and rigorous evaluation of this deep learning framework in the Brahmaputra River Basin, one of the world’s largest, most climate-sensitive, and data-scarce transboundary river systems.
• Direct and transparent benchmarking of the deep learning models against a calibrated, physically-based hydrological model (SWAT) under identical climatic forcing, providing a robust comparison of data-driven and process-based approaches.
• Demonstrates the potential of climate-driven deep learning frameworks as robust and transferable tools for daily streamflow prediction, particularly for flood forecasting, water resources management, and climate adaptation in data-scarce basins facing increasing hydroclimatic uncertainty.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Citation

@article{Shahariar2026Enhanced,
  author = {Shahariar, Shadman and Zobeyer, Hasan and Jahan, Nasreen and Ali, Md. Mostafa},
  title = {Enhanced Streamflow Prediction Using a Hybrid CNN-LSTM Model: Integrating Climate-Driven Spatio-Temporal Features in a Deep Learning Framework},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04491-9},
  url = {https://doi.org/10.1007/s11269-026-04491-9}
}