GoKawiilData
TWS data come from the GRACE and GRACE-FO mass concentration solutions (RL06) developed by the Center for Space Research at the University of Texas25. The dataset represents monthly total water mass variations (expressed as equivalent water height) at a gridded resolution of 0.25°. We omit data before August 2002 (due to sensor calibration at the outset of the GRACE mission) and for water-year 2017 (due to incomplete data during the mission gap between GRACE and GRACE-FO). We use GRACE TWS because it is an observational data product that holistically measures land water, capturing all surface and underground stocks.
To account for observational uncertainty, we use three daily gridded precipitation data products: the Global Precipitation Climatology Project (GPCP) Daily Precipitation Analysis v.1.3 (ref. 21), the Global Precipitation Climatology Center (GPCC) Daily Analysis v.2022 (ref. 23) and the National Oceanic and Atmospheric Administration Climate Prediction Center (CPC) Unified Gauge-Based Analysis22. GPCC is a station-based product (1° resolution) with coverage over 1982–2020, whereas CPC (0.5°) and GPCP (1°) integrate station and satellite observations over 1979–2022 and 1997–2022, respectively. We repeated our analysis for the more recent GPCP v.3.3 and find consistent main effects56 (Supplementary Fig. 4), but use the more complete earlier version for our main analysis (GPCP v.3.3 contains missing daily values at high latitudes after 2020).
We use two net all-sky surface shortwave radiation datasets: the Global Energy and Water Exchanges Surface Radiation Budget (GEWEX-SRB, limited here to 2002–2017)52 daily dataset and the NASA Clouds and the Earth’s Radiant Energy System Energy Balanced and Filled monthly dataset (NASA-EBAF, limited here to 2002–2022)53, each at 1° resolution. We use monthly evapotranspiration from the GLEAM v.3.8a, at 0.25° resolution for the overlapping GRACE period (2002–2020; ref. 54). Main river basin boundaries are from the HydroBASINS v.1.0 dataset, which is derived from Shuttle Radar Topography Mission elevation data collected in 2000 (ref. 51). All data URLs are enumerated in Supplementary Table 1.
Daily climate model data are from the land-hist experiment of the Coupled Model Intercomparison Project, Phase 6, Land-Surface, Snow and Soil Moisture Intercomparison Project (CMIP6-LS3MIP)55. In the land-hist experiment, the land-surface components of climate models are forced by atmospheric forcings from the Global Soil Wetness Project phase 3, whose precipitation data are derived from the NOAA-CIRES-DOE Twentieth Century Reanalysis (20CR). We include results from the two models that report all required variables: MIROC6 with land component MATSIRO and VISIT-e57, and CNRM-CM6-1 with land component ISBA-CTRIP58. We use the mrtws (total water storage), tas (mean surface air temperature) and pr (precipitation rate) variables. No participating models report these required variables plus evaporation, precluding extending our climate model analysis towards mechanisms.
Data aggregation and precipitation Gini coefficient
All data are temporally aggregated to the water-year scale, using October to September for the Northern Hemisphere water-year and July to June in the Southern Hemisphere. On this water-year scale, snow and ice storage changes are negligible. We further limit our spatial domain to regions without permanent snow and ice accumulation. We compute daily mean temperature data using the average of daily maximum and minimum temperatures from CPC. We also compute climatological mean precipitation as the average of the annual total over the full precipitation data record.
We linearly detrend GRACE data to isolate interannual variability, which is largely governed by hydroclimate variability, from long-term trends that include anthropogenic water use and management26,31. The rationale for this focus on interannual variability is to establish whether more concentrated daily precipitation distributions lead to cumulative water balance changes, averaged across the water-year, absent the influence of coincident economic, demographic and mean warming trends.
Correspondingly, and because GRACE only estimates TWS relative anomalies, we detrend and demean all other hydroclimate variables (except climatological precipitation, which is time-invariant). Finally, we interpolate all datasets to a common 0.5° grid as a compromise across the resolutions of the datasets. We repeated our analyses at alternative coarser common resolutions (1°, 2° and 3°) and found consistent results (Supplementary Fig. 5). We present long-term linear trends in terms of changes per 20 years to ease interpretation.
To quantify the annual concentration of daily precipitation, we apply the Gini coefficient, a common economic measure of income inequality59, to the three daily precipitation datasets. The Gini coefficient is based on the Lorenz curve, which plots the cumulative share of the total quantity (that is, cumulative share of annual precipitation) against the cumulative share of the population (that is, days per year), sorted from least to greatest (Fig. 1d). The Gini coefficient is defined as the area between the Lorenz curve and the 1:1 line of equality. For more concentrated distributions, such as all of the annual precipitation falling in 1 day, the Lorenz curve deviates further from the line of equality, giving a Gini coefficient closer to 1. If all days are equally rainy, the Gini yields a value of 0. We use a computationally efficient discrete formulation to calculate the annual Gini coefficient of sorted daily precipitation (G P ):
... continue reading