Aryal et al. (2026) A novel approach for river discharge prediction in Lancang Mekong River Basin: Incorporation of multisource remote sensing and LSTM model

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-12
• Authors: Tilasmi Aryal, Sangam Shrestha, Salik Bhusal, Pragya Pradhan
• DOI: 10.1016/j.ejrh.2026.103326

Research Groups

• Department of Water Resources and Irrigation, Government of Nepal, Nepal
• Water Engineering and Management, School of Engineering and Technology, Asian Institute of Technology, Pathum Thani, Thailand
• Global Water and Sanitation Center, Asian Institute of Technology, Pathum Thani, Thailand

Short Summary

This study developed a Long Short-Term Memory (LSTM) model framework integrating multi-mission satellite altimetry and optical sensor data to predict daily river discharge in the Lancang Mekong River Basin (LMRB). The model achieved robust performance, particularly in downstream reaches, and demonstrated spatial transferability to ungauged locations, though predictive accuracy declined with increasing distance from training sites.

Objective

• To evaluate how effectively the integration of multi-mission satellite data within a deep learning framework can reconstruct daily river discharge compared to single-sensor approaches.
• To assess the extent to which the developed model can be spatially transferred to predict discharge at ungauged locations where training data is unavailable.
• To analyze how the model’s predictive performance varies across different river reaches characterized by distinct geomorphological conditions.

Study Configuration

• Spatial Scale: Lancang Mekong River Basin (LMRB), covering approximately 795,000 km², with a focus on river reaches and 54 virtual stations along the river.
• Temporal Scale: Daily time series data from 2003 to 2018 for satellite altimetry and 2003 to 2021 for MODIS optical data.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) deep learning model.
• Data sources:
  • Satellite Altimetry: Jason-1/2/3, ENVISAT, SARAL/ALTIKA missions for water level time series.
  • Satellite Optical: Moderate-Resolution Imaging Spectro-Radiometer (MODIS) AQUA (MOD09GQ) and TERRA (MYD09GQ) products for surface reflectance ratios (C/M).
  • In-situ Data: Daily river discharge data from hydrological stations obtained from the Mekong River Commission (MRC) data portal, used for model training and validation.
  • Ancillary Data: LANDSAT 7/8 images for river width and centerline extraction using the RivWidthCloud algorithm, EGM2008 geoid for altimetry corrections.

Main Results

• Water Level Retrieval: 54 virtual stations were established. Altimetry-derived water levels showed high accuracy when compared to the DAHITI database (coefficient of determination (R²) ranging from 0.93 to 0.99, Root Mean Squared Error (RMSE) from 0.136 m to 0.315 m). Comparison with in-situ data at Pakse and Kratie yielded R² values of 0.92 and 0.85, with RMSE values of 0.72 m and 0.67 m, respectively.
• Surface Reflectance Ratio: Exponential smoothing of the C/M ratio significantly improved its correlation with observed discharge. For AQUA products, the correlation coefficient improved from a range of 0.44–0.63 to 0.64–0.76, and for TERRA products, from 0.55–0.76 to 0.63–0.83. TERRA products generally showed higher correlations.
• Discharge Modeling Performance: The integrated model combining MODIS Aqua, MODIS Terra, and altimetry-derived water level (Aq + Te + WL) consistently achieved the highest performance. Nash-Sutcliffe Efficiency (NSE) values were consistently above 0.90 at downstream stations (e.g., Nong Khai: 0.91, Pakse: 0.92, Kratie: 0.90), with RMSE reductions of up to 45% compared to models using only optical data.
• Spatial Performance Gradient: Model performance was significantly better in broader, flatter downstream reaches (NSE: 0.86–0.92) compared to narrow, meandering upstream reaches (NSE: 0.54–0.66). This disparity is attributed to reduced land contamination in radar altimetry and fewer mixed pixel effects in optical data in downstream areas.
• Spatial Transferability: Predictive accuracy consistently declined with increasing longitudinal distance from the training location, particularly in the upstream direction. For example, the Kratie model maintained satisfactory performance (NSE > 0.5) up to 1100 km upstream, while the Nong Khai model's performance rapidly declined upstream beyond 100 km (NSE < 0.4).

Contributions

• Developed a novel LSTM framework that effectively integrates multi-mission satellite altimetry and optical sensor data to provide accurate daily river discharge predictions in the data-scarce Lancang Mekong River Basin.
• Demonstrated the synergistic value of combining altimetry-derived water levels with optical surface reflectance ratios, showing that optical data can compensate for altimetry's temporal gaps and retracking errors, especially during high flow events.
• Quantified the influence of river geomorphology on satellite-based discharge monitoring, identifying a clear downstream accuracy gradient where broader, flatter channels facilitate more reliable measurements.
• Assessed the spatial transferability of the LSTM model to ungauged locations, providing a viable framework for virtual gauging in transboundary river basins, while highlighting the need for segment-specific modeling due to localized flow dynamics.

Funding

Not explicitly stated in the paper.

Citation

@article{Aryal2026novel,
  author = {Aryal, Tilasmi and Shrestha, Sangam and Bhusal, Salik and Pradhan, Pragya},
  title = {A novel approach for river discharge prediction in Lancang Mekong River Basin: Incorporation of multisource remote sensing and LSTM model},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103326},
  url = {https://doi.org/10.1016/j.ejrh.2026.103326}
}