Ji et al. (2025) Distinct hydrologic response patterns and trends worldwide revealed by physics-embedded learning

Identification

• Journal: Nature Communications
• Year: 2025
• Date: 2025-10-15
• Authors: H. Ji, Yalan Song, Tadd Bindas, Chaopeng Shen, Yuan Yang, Ming Pan, Jiangtao Liu, Farshid Rahmani, Ather Abbas, Hylke E. Beck, Kathryn Lawson, Yoshihide Wada
• DOI: 10.1038/s41467-025-64367-1

Research Groups

• Department of Civil and Environmental Engineering, The Pennsylvania State University, University Park, PA, USA
• Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, University of California, San Diego, La Jolla, CA, USA
• Physical Science and Engineering Department, King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia
• Biological and Environmental Science and Engineering Division, King Abdullah University of Science and Engineering (KAUST), Thuwal, Saudi Arabia

Short Summary

This study introduces a high-resolution, physics-embedded, big-data-trained hydrologic model to accurately capture global hydrologic response patterns and their shifts. The model reveals widespread and significant shifts in green-blue-water partitioning and baseflow ratios worldwide over the past two decades, with critical implications for flood risks, water supply, and aquatic ecosystems.

Objective

• To introduce a high-resolution, physics-embedded, big-data-trained model capable of reliably capturing characteristic hydrologic response patterns (signatures) and their shifts globally.
• To quantify the shifts in basic hydrologic partitioning (evapotranspiration-to-precipitation and baseflow-to-runoff ratios) worldwide over the past two decades and assess their implications for estuaries.
• To characterize how winter and summer runoff globally respond to accumulated precipitation in previous months, using seasonal runoff elasticity.
• To determine if big-data model training can significantly improve Global Water Model (GWM) performance, enhancing reliability for continental-scale impact assessments and increasing relevance for local stakeholders.

Study Configuration

• Spatial Scale: Global, with simulations at high resolution (median MERIT unit basin area of approximately 37 square kilometers, covering around 2.94 million MERIT unit basins worldwide). Evaluations conducted on 61 large global rivers (33 mixed-anthropogenic-impact, 28 natural) and over 5000 smaller basins (catchment areas less than 50,000 square kilometers).
• Temporal Scale: Analysis period primarily 2001–2020 for hydrologic shifts and elasticity. Model training data span 1980–2020. Evaluations for large rivers cover 1981–2000. Small basin evaluations cover 2001–2015 (temporal test), 1980–2010 (spatial test), and 1980–2015 (overall test). Simulations are performed at monthly and daily scales.

Methodology and Data

• Models used:
  • δHBV2δMC2-Globe2-hydroDL (δHBV2): A hybrid, multiscale, physics-embedded differentiable hydrologic model. It uses neural networks to generate physical parameters for a differentiable implementation of the conceptual Hydrologiska Byråns Vattenbalansavdelning (HBV) model, simulating rainfall-runoff processes at MERIT unit basin scale. Runoff is routed downstream using a differentiable Muskingum-Cunge (δMC) model.
  • δHBV1.0: A basin-lumped differentiable model.
  • LSTM: A purely data-driven Long Short-Term Memory neural network model.
  • GloFAS: A widely-used operational global-scale product.
  • Six established Global Water Models (GWMs) from ISIMIP2a: WaterGAP2, DBH, H08, LPJmL, MATSIRO, and PCR-GLOBWB (referred to as GWM0-GWM5).
• Data sources:
  • Streamflow observations: Daily streamflow from a global dataset (Abbas et al., 2025) comprising 4746 catchments (area < 50,000 square kilometers) from various global and national databases.
  • Meteorological forcings: Daily precipitation from Multi-Source Weighted-Ensemble Precipitation (MSWEP) V2.8, and maximum/minimum daily temperatures from Multi-Source Weather (MSWX) V1.
  • Static attributes: Topography, climate patterns, land cover, soil, and geological characteristics derived from diverse sources.
  • Hydrological simulation unit: MERIT-Basins, based on the 90-meter MERIT-Hydro digital elevation model (DEM) dataset.
  • Estuary database: Global Estuary Database from the Sea Around Us project.
  • GSWP3 atmospheric data: Used for forcing GWMs in comparison.

Main Results

• Superior Model Performance: δHBV2 significantly outperforms established GWMs and operational systems. It exhibits minimal bias for long-term mean annual runoff (32.4 millimeters per year for natural rivers, 42.0 millimeters per year for mixed-impact rivers), which is 20-55% lower than GWMs. It achieves high spatial R² values (0.92-0.97) for mean annual runoff and is the only model to capture spatial variability of streamflow-to-precipitation ratio trends with R² > 0.4. At monthly scales, it shows a median correlation of 0.89 for natural rivers. For daily streamflow in small-to-medium basins, δHBV2's median Nash-Sutcliffe Efficiency (NSE) is 0.63 (0.53 for ungauged basins), more than double GloFAS's ~0.26, and it edges out LSTM in Kling-Gupta Efficiency (KGE).
• Widespread Hydrologic Shifts (2001-2020): The model reveals significant shifts in green-blue-water partitioning (evapotranspiration-to-precipitation ratio) and local baseflow-to-runoff ratios, with some regions experiencing changes exceeding 20% over 20 years.
  • Green-Blue Water Partitioning: Mid-latitudes in North America and Asia, and tropical areas like Central America and Papua New Guinea show decreasing green water fluxes (increasing blue water), while Central Europe and subtropical/mid-latitude South America show increasing green water fluxes.
  • Baseflow Ratios: Shifts in baseflow-to-runoff ratios are widespread, implying pervasive changes in stream temperature and water quality.
  • Flood Risk and Water Scarcity: Increased blue-water fraction and decreased baseflow ratio in regions like northeastern China and mid-latitude North America correlate with higher flooding potential. Shifts towards green-water fluxes in Germany, central Siberia, southern Brazil, and other regions, often linked to declining precipitation, suggest disproportionate decreases in streamflow available for human use.
• Estuary Freshwater Inflows: Ten out of 55 analyzed estuaries, primarily along the North Sea coast of Germany and France, show statistically significant declining trends in mean annual freshwater inflows (some >1.5% per year, totaling 30% over 20 years), with ecological implications. δHBV2 accurately captures these trends (R² ~0.68), outperforming GWMs.
• Seasonal Runoff Elasticity: High summer runoff elasticity is prominent in arid and semi-arid regions (e.g., Sahel, central/south Africa, central Asia), indicating high vulnerability to precipitation changes but also potential for reliable seasonal streamflow prediction. Winter elasticity shows a smaller range, with high values in central Asia, northern Middle East, and parts of Africa and South America.
• Daily Streamflow Flashiness: Arid and semi-arid regions exhibit high flow flashiness (steep negative slopes of the flow duration curve), while tropical rainforests show low flashiness. Widespread, spatially mixed trends in flashiness are observed, with some regions becoming less variable (e.g., Mexico, northern India) and others more variable (e.g., central-western USA, central Asia), highlighting increasing challenges for water management.

Contributions

• Introduces a novel, high-resolution, physics-embedded, differentiable hydrologic model (δHBV2) that effectively integrates big-data learning with physical principles, overcoming limitations of traditional GWMs and purely data-driven models.
• Provides the first global-scale, high-resolution quantification and mapping of widespread shifts in fundamental hydrologic response patterns (green-blue water partitioning, baseflow ratios, seasonal elasticity, and flashiness) over the past two decades.
• Demonstrates a significant leap in global hydrologic modeling accuracy, achieving substantially more accurate simulations at monthly and daily scales than current operational systems and established GWMs, particularly in capturing spatial heterogeneity and temporal trends.
• Reveals previously unrecognized hydrologic shifts and their critical implications for increasing flood risks, heightening water supply stresses, and declining freshwater inputs to estuaries, with associated ecological impacts.
• Offers an advanced tool that enables global-scale models to deliver reliable and locally-relevant insights for water management and seasonal water availability forecasting, addressing a long-standing need for communities worldwide.

Funding

• National Science Foundation (award EAR-2221880)
• Cooperative Institute for Research to Operations in Hydrology (CIROH) (award NA22NWS4320003 from the NOAA Cooperative Institute Program)
• Federal Award Identification W912HZ-24-2-0001 (“Research, development and application of hydrometeorological, engineering and other capabilities in support of USACE FIRO objectives”)
• US Department of Energy, Office of Science (award DE-SC0021979)
• NASA (Award 80NSSC24K1646)

Citation

@article{Ji2025Distinct,
  author = {Ji, H. and Song, Yalan and Bindas, Tadd and Shen, Chaopeng and Yang, Yuan and Pan, Ming and Liu, Jiangtao and Rahmani, Farshid and Abbas, Ather and Beck, Hylke E. and Lawson, Kathryn and Wada, Yoshihide},
  title = {Distinct hydrologic response patterns and trends worldwide revealed by physics-embedded learning},
  journal = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-64367-1},
  url = {https://doi.org/10.1038/s41467-025-64367-1}
}