Shourabi et al. (2025) Enhancing Flood Frequency Predictions Under Climate Change and Uncertainty Using Machine Learning Model Fusion and Wavelet Transform

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-10-17
• Authors: Kiana Yahyazadeh Shourabi, Mohammad Hossein Niksokhan, Mohammad Reza Nikoo
• DOI: 10.1007/s41748-025-00859-w

Research Groups

• Faculty of Environment, University of Tehran, Tehran, Iran
• Department of Civil and Architectural Engineering, Sultan Qaboos University, Al-Khod, Muscat, Oman

Short Summary

This study developed an integrated framework to enhance flood frequency predictions under climate change and uncertainty in the Haraz River Basin, Iran. It combined machine learning model fusion, wavelet transform, and uncertainty quantification, projecting a general decrease in future flood frequency, particularly under high-emission scenarios.

Objective

• To develop an integrated, multi-step data-driven framework to assess the effects of climate change on river discharge and flood frequency, quantify model uncertainty, and perform sensitivity analysis in the Haraz River Basin, Iran.

Study Configuration

• Spatial Scale: Haraz River Basin, northern Iran. Area: 4013.02 square kilometers. Geographic coordinates: 35.742° N to 36.347° N latitude, 51.443° E to 52.610° E longitude. Elevation range: 300 meters to 5587 meters above sea level.
• Temporal Scale:
  • Historical/Baseline Period: 1999-2014 (for GCM comparison and model calibration), 1976-2016 (for observed hydrological and meteorological averages).
  • Future Projection Period: 2031-2050.

Methodology and Data

• Models used:
  • Statistical Downscaling: LARS-WG (weather generator), Random Forest (RF), Gradient Boosting Regression Tree (GBRT), Least Square Support Vector Regression-Particle Swarm Optimization (LSSVR-PSO).
  • Model Fusion: LSSVR-PSO (for downscaling outputs), Multi-Layer Perceptron (MLP) (for river flow prediction, combining MLP and LSSVR outputs).
  • Feature Selection: SelectKBest algorithm with F-statistic (f-regression).
  • Denoising: Wavelet Transform (Discrete Wavelet Transformation (DWT) with Daubechies (db4) wavelet, soft thresholding).
  • Uncertainty Analysis: Model Agnostic Prediction Interval Estimator (MAPIE) with CV-minmax method.
  • Flood Frequency Analysis: Three-parameter Weibull distribution.
  • Evaluation Criteria: Coefficient of determination (R²), Root Mean Square Error (RMSE), Normalized Root Mean Square Error (NRMSE), Mean Absolute Error (MAE), Nash–Sutcliffe Efficiency (NSE).
• Data sources:
  • Climate Projections: CMIP6 dataset (specifically ACCESS-CM2 Global Climate Model) under Shared Socio-economic Pathways (SSP1-2.6 and SSP5-8.5 scenarios).
  • Observational Climate Data: National Meteorological Organization (Siahbisheh synoptic station) for daily precipitation, minimum temperature, and maximum temperature (1999-2014).
  • Observational Hydrological Data: Iran Water Resources Management Company (Koresang hydrometric station) for daily runoff/discharge (1999-2014).
  • Topographic Data: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) files (N35E051, N35E052, N36E051, N36E052).

Main Results

• Downscaling Performance: The ML model fusion achieved a coefficient of determination (R²) of 0.76 for precipitation, 0.94 for maximum temperature, and 0.94 for minimum temperature.
• River Flow Prediction: The MLP model fusion predicted river flow with an R² of 0.84. Cross-validation demonstrated model stability with an average performance of 0.91.
• Climate Projections (2031-2050): Precipitation is projected to increase in most months but decrease in March, September, November, and December. Both maximum and minimum temperatures show an overall upward trend. The SSP5-8.5 scenario generally projects larger increases in precipitation and temperature compared to SSP1-2.6, especially for the ML model.
• Discharge Projections: Future discharge patterns generally align with precipitation, showing increases in late winter/spring (January–April) and decreases in summer/autumn (August–October or July–December).
• Uncertainty Analysis: The MAPIE method (specifically CV-minmax) successfully quantified prediction intervals for discharge at a 95% confidence level, confirming model robustness.
• Sensitivity Analysis: Precipitation was identified as the primary driver of river discharge. Discharge increased with increasing precipitation and minimum temperatures, but decreased when high maximum temperatures accompanied rainfall. Discharge was more sensitive to precipitation variations under low-to-medium maximum temperatures or medium-to-high minimum temperatures.
• Flood Frequency Analysis: A decrease in the frequency of maximum daily discharges is projected across all future scenarios compared to the base period. The ML model under SSP1-2.6 showed the smallest decline, while the LARS-WG model under SSP5-8.5 showed the most significant decrease. Future peak discharges for higher return periods (>200 years) were notably lower than historical observations.

Contributions

This research introduces a novel, integrated, multi-step modeling framework for hydrological impact assessment under climate change, specifically for flood frequency analysis. Its originality lies in systematically combining: 1. Machine learning (RF, GBRT, LSSVR-PSO) and LARS-WG for downscaling climatic variables, including model fusion. 2. Feature selection for optimizing predictors. 3. Wavelet-based noise reduction for improving input data quality. 4. Streamflow prediction and flood frequency estimation. 5. Model uncertainty quantification using the Model Agnostic Prediction Interval Estimator (MAPIE) with 95% confidence intervals. 6. Sensitivity analysis to determine the effect of input parameters on discharge. This comprehensive framework, applied with CMIP6 climate change scenarios, provides an improved and robust estimate of river discharge and flood frequency, addressing a gap in existing literature by integrating these components into a cohesive methodology.

Funding

Not applicable

Citation

@article{Shourabi2025Enhancing,
  author = {Shourabi, Kiana Yahyazadeh and Niksokhan, Mohammad Hossein and Nikoo, Mohammad Reza},
  title = {Enhancing Flood Frequency Predictions Under Climate Change and Uncertainty Using Machine Learning Model Fusion and Wavelet Transform},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00859-w},
  url = {https://doi.org/10.1007/s41748-025-00859-w}
}