Qiu et al. (2025) Enhancing flood prediction in the Lower Mekong River Basin by a scale-independent interpretable deep learning model

Identification

• Journal: Environmental Impact Assessment Review
• Year: 2025
• Date: 2025-08-23
• Authors: Yangzi Qiu, Xiaogang Shi, Xiaogang He
• DOI: 10.1016/j.eiar.2025.108130

Research Groups

• School of Social & Environmental Sustainability, University of Glasgow, UK
• LEESU, École Nationale des Ponts et Chaussées, Institute Polytechnique de Paris, France
• Department of Civil and Environmental Engineering, National University of Singapore, Singapore

Short Summary

This study develops an interpretable Long Short-Term Memory (LSTM) model for flood prediction in the Lower Mekong River Basin, employing SHapley Additive exPlanation (SHAP) and Universal Multifractal (UM) analyses to identify key contributing variables and their scale-dependent and scale-independent impacts on river discharge. The model demonstrates high predictive power, with interpretations revealing the dynamic influence of soil, vegetation, and hydrometeorological variables on flood events across different temporal scales.

Objective

• Evaluate the performance of an LSTM model in predicting river discharge across the Lower Mekong River Basin (LMB).
• Identify the key variables contributing to flood prediction through SHAP (scale-dependent) and Universal Multifractal (UM) analyses (scale-independent).
• Understand complex feature interactions and non-linear scaling effects of input features on discharge prediction by the LSTM model in the LMB.

Study Configuration

• Spatial Scale: Lower Mekong River Basin (LMB), covering a watershed area of 795,000 km², divided into eight subbasins. Data resolutions include 0.25 degrees (ERA5), 0.1 degrees (GloFAS-ERA5), and 30 km (SRTM1).
• Temporal Scale: Data period from 1 January 1979 to 31 December 2020 (42 years). Data frequency is hourly (ERA5) and daily (GloFAS-ERA5, MRC). The LSTM model uses input features from the past 7 time steps. Universal Multifractal analysis identifies small-scale (1 to 16 days) and large-scale (16 to 2048 days) ranges.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) neural network for river discharge prediction.
  • SHapley Additive exPlanation (SHAP) for scale-dependent interpretability of feature contributions and interactions.
  • Universal Multifractal (UM) framework (including Trace Moment technique and maximum probable singularity (γs)) for scale-independent interpretability of temporal variability in extremes.
• Data sources:
  • ERA5 hourly data (ECMWF reanalysis): Total precipitation (m), air temperature at 2 m (K), evaporation (m), forecast albedo (dimensionless), forecast surface roughness (m), LAI high/low vegetation (m² m⁻²), relative humidity (%), specific humidity (kg kg⁻¹), skin reservoir content (m), runoff (m), volumetric soil water layers 1-4 (0-7 cm, 7-28 cm, 28-100 cm, 100-289 cm, in m³ m⁻³), U/V-wind at 10 m (m s⁻¹). Also includes static attributes: high/low vegetation cover, soil type, type of high/low vegetation, and catchment mean elevation (m).
  • GloFAS-ERA5 operational global river discharge reanalysis (Copernicus Climate Change Service): Gridded river discharge (m³ s⁻¹).
  • Shuttle Radar Topography Mission 1 (SRTM1): Digital Elevation Model.
  • Mekong River Commission (MRC): Daily discharge (m³ s⁻¹) from ten hydrological stations.

Main Results

• The LSTM model achieved satisfactory performance in discharge prediction across the LMB, with Pearson's correlation coefficient (Corr) values exceeding 0.95, Nash–Sutcliffe Efficiency (NSE) values higher than 0.9, and Kling-Gupta Efficiency (KGE) values ranging from 0.75 to 0.95 for all subbasins.
• The model tends to underestimate the largest peak flows, particularly in midstream subbasins (e.g., subbasin 5 with Percent Bias (PBIAS) ranging from -30% to -45%, subbasin 3 with approximately 28.9% underestimation, and subbasin 6 with approximately 36.0% underestimation).
• SHAP analysis revealed that soil-related variables (e.g., volumetric soil water layers 3 and 4) are consistently important contributors to discharge prediction, with their impacts partially manifested through interactions with precipitation and runoff.
• The dominant contributing variables influencing flood prediction shifted over time: soil-related and vegetation-related variables were more significant in earlier years (2013-2017), while hydrometeorological variables (e.g., runoff, air temperature at 2 m, evaporation) became more dominant after 2017.
• Universal Multifractal (UM) analysis showed that all selected variables exhibit a similar scaling behavior with a break at 16 days. In the small-scale range (1-16 days), hydrometeorological-related variables (total precipitation, runoff, skin reservoir content) have similar temporal variability in extremes (absolute difference in γs < 0.1) to discharge, suggesting their strong association with discharge extremes.
• The model's performance (NSE) generally improves as the extreme variability (indicated by γs) of the combined feature sets decreases within the 1-16 day range across most subbasins.

Contributions

• First study to integrate LSTM with SHAP and Universal Multifractal (UM) analyses for flood prediction in the Lower Mekong River Basin, providing both scale-dependent and scale-independent interpretations of model behavior.
• Offers a comprehensive assessment of LSTM model performance in discharge prediction, highlighting the impact of the variability in the extremes of combined features.
• Transforms the deep learning model from a "black-box" into an interpretable tool, yielding new insights into hydrological processes and underlying physical and scaling mechanisms.
• Identifies dynamic shifts in dominant flood drivers over time, showing a transition from land surface conditions (soil moisture, vegetation) to meteorological factors in recent years.
• Emphasizes the critical role of soil-related variables and their interactions with precipitation and runoff in discharge prediction.
• Provides valuable insights for stakeholders to enhance flood risk management strategies across the LMB.

Funding

• British Council UK-ASEAN Institutional Links Scheme (Grant No. 924846643)

Citation

@article{Qiu2025Enhancing,
  author = {Qiu, Yangzi and Shi, Xiaogang and He, Xiaogang},
  title = {Enhancing flood prediction in the Lower Mekong River Basin by a scale-independent interpretable deep learning model},
  journal = {Environmental Impact Assessment Review},
  year = {2025},
  doi = {10.1016/j.eiar.2025.108130},
  url = {https://doi.org/10.1016/j.eiar.2025.108130}
}