Wei et al. (2025) Future projections of China runoff changes based on CMIP6 and deep learning

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-27
• Authors: Xikun Wei, Guojie Wang, Britta Schmalz
• DOI: 10.1016/j.ejrh.2025.102998

Research Groups

• Key Laboratory for Climate Risk and Urban-Rural Smart Governance, School of Geography, Jiangsu Second Normal University, Nanjing 211200, China
• Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science and Technology (NUIST), Nanjing 210044, China
• Chair of Engineering Hydrology and Water Management, Technical University of Darmstadt, Darmstadt 64287, Germany

Short Summary

This study projects future runoff changes across mainland China at 185 hydrological stations under CMIP6 Shared Socioeconomic Pathway scenarios using deep learning models (LSTM-SA, GRU-SA) with DL-downscaled climate inputs, revealing overall runoff increases, particularly in central transitional and southern humid regions, with pronounced summer increases and winter declines.

Objective

• To project future runoff changes across mainland China at 185 hydrological stations under SSP2–4.5, SSP3–7.0, and SSP5–8.5 scenarios using deep learning methods, providing a comprehensive understanding of regional hydrological responses to climate change.

Study Configuration

• Spatial Scale: Mainland China (approximately 9.6 million square kilometers), encompassing 185 hydrological stations across ten major river basins and diverse climatic zones.
• Temporal Scale: Monthly resolution. Historical period: 1961–2014 (CMIP6), 1961–2018 (observations). Future projections: 2015–2100 (CMIP6). Analysis periods: 1995–2014 (historical reference), 2025–2045 (near-term), 2045–2065 (mid-term), and 2080–2100 (long-term).

Methodology and Data

• Models used:
  • Deep Learning (DL) models for downscaling: YNET, Enhanced Deep Residual Network (EDSR), Residual Channel Attention Networks (RCAN). Compared with Random Forest (RF) and Ensemble Mean (EM).
  • Deep Learning (DL) models for runoff simulation: Long Short-Term Memory with Self-Attention (LSTM-SA), Gated Recurrent Unit with Self-Attention (GRU-SA).
  • Coupled Model Intercomparison Project Phase 6 (CMIP6) General Circulation Models (GCMs) (23 models).
• Data sources:
  • CMIP6 dataset: Air temperature (mean, maximum, minimum), precipitation, relative humidity, and runoff depth from 23 GCMs for historical (1961–2014) and future (2015–2100) periods under SSP2–4.5, SSP3–7.0, and SSP5–8.5 scenarios.
  • Observed climate gridded monthly dataset: From China Meteorological Administration (http://data.cma.cn/site/index.html), comprising data from over 2400 meteorological stations in China (1961–2018) at 0.25° × 0.25° spatial resolution.
  • China Natural Runoff Dataset (CNRD) v1.0: Gridded natural runoff dataset generated using the Variable Infiltration Capacity (VIC) model, based on 200 gauged stations.

Main Results

• DL-based downscaling methods significantly outperformed Random Forest and Ensemble Mean methods, with EDSR showing the best performance for most climate variables and RCAN for precipitation. DL-merged data achieved high correlation (e.g., R ≈ 0.85 for precipitation, 0.90 for relative humidity) and lower bias.
• The single-station GRU-SA model demonstrated the best overall runoff simulation performance, achieving a mean Nash–Sutcliffe Efficiency (NSE) of 0.656 and a median NSE of 0.727 across 185 stations, with most stations exceeding an NSE of 0.7.
• Reliable monthly rainfall–runoff models can be constructed with as few as 500 training samples using DL approaches.
• Future runoff projections indicate an overall increasing trend at most stations across China under all SSP scenarios (SSP2–4.5, SSP3–7.0, and SSP5–8.5).
• The strongest runoff increases are projected in the central transitional and southern humid regions of China.
• The magnitude of runoff increase intensifies under higher emission scenarios (SSP5–8.5) compared to lower ones (SSP2–4.5), particularly in the central-east transitional and southern humid regions.
• Seasonal contrasts are pronounced: summer runoff is projected to increase markedly across most regions, while winter runoff tends to decline in humid regions and exhibits spatial heterogeneity in transitional zones.
• These patterns suggest heightened risks of both floods (summer) and droughts (winter), with more frequent extreme events likely in humid areas.

Contributions

• Expanded the spatial coverage of future runoff projections to 185 hydrological stations across mainland China, providing a more detailed and comprehensive understanding of regional hydrological responses to climate change across diverse climatic zones and ten major river basins.
• Integrated deep learning (DL) methods for both downscaling CMIP6 climate data and simulating rainfall-runoff at a national scale, demonstrating their robustness and efficiency, even with limited training data (as few as 500 samples).
• Provided new insights into regional and seasonal hydrological responses to climate change across China, highlighting increased flood risks in summer and drought risks in winter, particularly in humid regions.
• Delivered valuable scientific evidence to support adaptive water resource allocation, flood control, and drought mitigation strategies across China.

Funding

• National Natural Science Foundation of China International Cooperation Research (W2412033)
• National Natural Science Foundation of China (42405176)
• Jiangsu Provincial Higher Education Basic Science (Natural Science) Research General Program (25KJD17002)

Citation

@article{Wei2025Future,
  author = {Wei, Xikun and Wang, Guojie and Schmalz, Britta},
  title = {Future projections of China runoff changes based on CMIP6 and deep learning},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102998},
  url = {https://doi.org/10.1016/j.ejrh.2025.102998}
}