Yeşilyurt et al. (2026) Integration of Satellite-Derived Meteorological Inputs into SWAT, XGBoost, WGAN, and Hybrid Modelling Frameworks for Climate Change-Driven Streamflow Simulation in a Data-Scarce Region

Identification

• Journal: Water
• Year: 2026
• Date: 2026-01-16
• Authors: Sefa Nur Yeşilyurt, Gülay Onuşluel Gül
• DOI: 10.3390/w18020239

Research Groups

• Graduate School of Natural and Applied Sciences, Dokuz Eylül University, Izmir, Türkiye
• Department of Civil Engineering, Dokuz Eylül University, Izmir, Türkiye

Short Summary

This study evaluates the transferability of satellite-derived meteorological inputs to process-based (SWAT), data-driven (XGBoost, WGAN), and hybrid hydrological models for streamflow simulation in a data-scarce region. It finds that hybrid SWAT+WGAN models achieve superior predictive accuracy, and satellite data can reliably substitute ground observations, projecting significant future streamflow reductions under climate change.

Objective

• Evaluate the accuracy of satellite-derived meteorological inputs as substitutes for observed data in hydrological modeling.
• Discuss the capability of machine learning techniques to improve the performance of physically based models, including SWAT.
• Establish the ability of independent machine learning models to simulate process-based hydrological characteristics.
• Assess the projected future hydro-meteorological regime of the Büyük Menderes Basin based on CMIP6 climate scenarios.

Study Configuration

• Spatial Scale: Büyük Menderes Basin, western Türkiye, covering approximately 25,987 square kilometers. A specific sub-basin was selected for detailed analysis.
• Temporal Scale: Observation period: 1994–2018 (70% for calibration, 30% for validation). Future projection periods: 2018–2040, 2041–2060, and 2061–2099. All meteorological inputs were processed at daily resolution, and model outputs were evaluated at a monthly time scale.

Methodology and Data

• Models used:
  • Process-based: Soil and Water Assessment Tool (SWAT)
  • Data-driven: eXtreme Gradient Boosting (XGBoost), Wasserstein Generative Adversarial Network (WGAN)
  • Hybrid: SWAT + XGBoost, SWAT + WGAN
  • Explainable AI: SHapley Additive exPlanations (SHAP) for interpretability.
• Data sources:
  • Ground-based observations: Precipitation, minimum/maximum/average temperature, relative humidity, wind speed from Turkish State Meteorological Service (MGM). Streamflow data from General Directorate of State Hydraulic Works (DSI).
  • Satellite-based meteorological data: NASA POWER (Prediction of Worldwide Energy Resources) for precipitation, temperature, humidity, wind speed, and solar radiation.
  • Reanalysis data: ERA5 for hourly direct solar radiation (aggregated to daily).
  • Topography: GLO-30 dataset from Copernicus programme (Digital Elevation Model - DEM).
  • Soil data: Global Hydrological Soil Groups (HYSOGs250m) dataset.
  • Climate projections: NEX-GDDP-CMIP6 (NASA Earth Exchange Global Daily Downscaled Projections—Coupled Model Intercomparison Project Phase 6) from three Global Climate Models (GCMs): MPI-ESM1-2-HR, GFDL-ESM4, and EC-Earth3, under two Shared Socioeconomic Pathways (SSPs): SSP2-4.5 and SSP5-8.5. Bias-corrected using Empirical Quantile Mapping (EQM).

Main Results

• Model Performance: All 11 model configurations performed well, with Nash-Sutcliffe Efficiency (NSE) values between 0.78 and 0.98 and Kling-Gupta Efficiency (KGE) values between 0.76 and 0.93.
• Hybrid Model Superiority: Hybrid models, particularly SWAT + WGAN, significantly enhanced streamflow simulation accuracy. The SWAT + WGAN framework achieved the best overall performance with a validation NSE of approximately 0.90 and KGE of approximately 0.89, outperforming the baseline SWAT model (NSE ≈ 0.86, KGE ≈ 0.80).
• Satellite Data Feasibility: Models forced with satellite-derived meteorological inputs performed comparably to those using ground-based observations, validating their feasibility as alternative data sources in data-scarce regions.
• AI Model Capability: Independent XGBoost models achieved slightly higher predictive capability (NSE ≈ 0.87) than the baseline SWAT model, demonstrating the potential of data-driven approaches. WGAN models performed comparably to SWAT but showed lower accuracy for satellite-based configurations.
• Explainable AI (SHAP): SHAP analysis confirmed that both data-driven and hybrid models captured physically consistent rainfall–runoff relationships, with precipitation persistence being a primary driver of streamflow variability. For hybrid models, SWAT-simulated discharge was the strongest predictor, with lagged values and seasonal terms refining the estimates.
• Climate Change Impacts: Future hydrological projections using the best-performing SWAT + WGAN model and CMIP6 scenarios indicate a clear drought signal. Mean streamflow is projected to decrease by up to 58% (e.g., MPI-ESM1-2 SSP245) compared to the reference period (1995–2017). Increased skewness and kurtosis suggest more extreme events, longer low-flow periods, and more rapidly increasing peaks.

Contributions

This study provides a comprehensive, integrated, and reproducible framework that: - Jointly assesses the hydrological usability of satellite-derived meteorological datasets. - Compares the accuracy of process-based, data-driven, and hybrid modeling paradigms in data-scarce environments. - Explores the future hydroclimatic sensitivity of a basin to CMIP6 scenarios. - Integrates explainable AI (SHAP) to enhance the interpretability and physical consistency of advanced machine learning and hybrid models, addressing the "black-box" limitation. - Demonstrates that AI-assisted hybridization can achieve both higher predictive performance and physically interpretable functions, offering a data-efficient and technology-compatible solution for resilient basin management.

Funding

• Scientific and Technological Research Council of Türkiye (TUBİTAK), Directorate for Scientist Support Programs (BİDEB), under the 2211-E Domestic Direct PhD Scholarship Program.
• Application number: 1649B032205678.

Citation

@article{Yeşilyurt2026Integration,
  author = {Yeşilyurt, Sefa Nur and Gül, Gülay Onuşluel},
  title = {Integration of Satellite-Derived Meteorological Inputs into SWAT, XGBoost, WGAN, and Hybrid Modelling Frameworks for Climate Change-Driven Streamflow Simulation in a Data-Scarce Region},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18020239},
  url = {https://doi.org/10.3390/w18020239}
}