Evidence of human influence on Northern Hemisphere snow loss

Abstract

Documenting the rate, magnitude and causes of snow loss is essential to benchmark the pace of climate change and to manage the differential water security risks of snowpack declines^1-4. So far, however, observational uncertainties in snow mass^5,6 have made the detection and attribution of human-forced snow losses elusive, undermining societal preparedness. Here we show that human-caused warming has caused declines in Northern Hemisphere-scale March snowpack over the 1981-2020 period. Using an ensemble of snowpack reconstructions, we identify robust snow trends in 82 out of 169 major Northern Hemisphere river basins, 31 of which we can confidently attribute to human influence. Most crucially, we show a generalizable and highly nonlinear temperature sensitivity of snowpack, in which snow becomes marginally more sensitive to one degree Celsius of warming as climatological winter temperatures exceed minus eight degrees Celsius. Such nonlinearity explains the lack of widespread snow loss so far and augurs much sharper declines and water security risks in the most populous basins. Together, our results emphasize that human-forced snow losses and their water consequences are attributable-even absent their clear detection in individual snow products-and will accelerate and homogenize with near-term warming, posing risks to water resources in the absence of substantial climate mitigation.

Fig. 1. Observed long-term warming trends are robust throughout the Northern Hemisphere, but snowpack trends are not.

a,b, Agreement across observational products (Supplementary Table 1) on the sign of trends in November–March average temperature (winter T, a) and March SWE (b) from 1981 to 2020. Numbers in bottom left show the percentage of basins with each category of agreement indicated on the colour bar. Insets: the hemispheric trends for each individual product. c–e, The trends for the four most populous river basins in North America (c), Europe (d) and Asia (e) that are generally considered snow dominated, as well as each continent (Methods). The locations of the basins are indicated on the map in a, corresponding to the number in parentheses. Temperature (red triangles) is referenced to the top x axis and SWE (blue squares) is referenced to the bottom x axis. The 2020 basin population is indicated in the top-right corner. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Fig. 2. Climate model experiments reveal that human-caused warming has influenced Northern Hemisphere snowpack trends.

a–d, Trend in March SWE from 1981 to 2020 in in situ observations (a), the ensemble mean of five long-term gridded SWE products (b), and the multimodel mean of CMIP6 historical simulations with (c) and without (d) anthropogenic emissions. e, Spatial pattern correlation (ρ) of 1981–2020 March SWE trends between the CMIP6 multimodel mean HIST (red symbols) and HIST-NAT (blue symbols) simulations and each observational (OBS) SWE product (see legend). The grey histogram indicates the empirical probability density function of spatial correlations between trends from the historical simulations and all possible 40-year trends from unforced pre-industrial control (PIC) simulations (N = 78,601). The red (orange) vertical dashed line indicates the 99th (95th) percentile of this empirical distribution. Maps were generated using cartopy v.0.18.0. Source Data

Fig. 3. Empirical snowpack reconstructions reveal the countervailing effects of human-forced temperature and precipitation trends on basin-scale snow changes.

a, Average observed 1981–2020 March SWE trends from 5 long-term SWE data products in 169 major Northern Hemisphere river basins. b, As in a but for our observation-based reconstructions. c, Effect of anthropogenically forced temperature changes on March SWE trends, given by the ensemble mean difference between the statistically reconstructed historical trend and the reconstructed trend with forced changes to temperature removed. d, As in c but for forced precipitation changes. e, As in c and d but for forced changes to both temperature and precipitation. The hatching indicates basins where fewer than 80% of observations or reconstructed estimates agree on the sign of the trend or forced effect. Maps were generated using cartopy v.0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Fig. 4. The nonlinear sensitivity of snowpack to warming augers accelerating water security risks for highly populous snow-dependent basins.

a, Temperature sensitivity of March SWE across a range of climatological winter temperatures in in situ observations (green), gridded data products (blue), climate models (red) and our basin-scale statistical reconstructions (orange). The solid line (shading) indicates the average sensitivity (±1 s.d.) in a rolling 5 °C temperature window across all in situ locations, grid cells or river basins. The red vertical line indicates the change point at which the temperature sensitivity of snowpack becomes nonlinear (based on a change-point analysis using the basin-scale reconstructions). The bottom histograms show the distribution of climatological Northern Hemisphere March SWE and human population in 2° temperature bins, with the values indicating how much of each distribution falls on each side of the change point. Temperatures on the x axis are the average November–March temperature over the 1981–2020 period from each in situ location or grid cell. Only climatologically snow-covered grid cells are used to calculate the basin-average temperature. b, Percentage change in basin-scale March SWE-driven April–June runoff in 2070–2099 under SSP2-4.5 relative to 1981–2020 (Methods) versus basin population. The dots are coloured by the percentage change in March SWE in 2070–2099 relative to 1981–2020 and sized by the CMIP6 ensemble mean projected end-of-century temperature change. Source Data

Extended Data Fig. 1. Heterogenous long-term trends in observed March SWE make claims about snow responses to warming a challenge.

a-e, Trend in March SWE from 1981 to 2020 from individual gridded SWE data products. f, Average trend across all 5 products. Grid cells where fewer than 4 products agree on the sign of the trend are hatched. Maps were generated using cartopy v0.18.0. Source Data

Extended Data Fig. 2. Historical trends in March SWE from CMIP6 models exhibit uncertainty outside of the Western United States, Europe, and Northern Eurasia.

a-k, Trend in March SWE from 1981 to 2020 from historical climate model simulations. Details of models can be found in Extended Data Table 2. l, Ensemble mean trend. Grid cells where fewer than 80% of models agree on the sign of the trend are hatched. Maps were generated using cartopy v0.18.0. Source Data

Extended Data Fig. 3. Ensemble reconstructions based on the Random Forest model skillfully reproduce the pattern and magnitude of long-term SWE trends in each snow product.

Observed (a-f) and reconstructed (g-l) 1981–2020 March SWE trends for 5 gridded SWE data products and their mean. m-r, Scatterplot of reconstructed versus observed trends, where each dot represents a river basin. Dashed line denotes perfect reconstruction. Pearson’s correlation is shown in bottom right corner. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Extended Data Fig. 4. The Random Forest model exhibits high snowpack reconstruction skill based on temperature and precipitation data.

Basin-scale R² (a-e) and root-mean-square error (RMSE; f-j) for 5 gridded SWE data products over the period 1981–2020. Each metric shows the skill of the mean of all reconstructions for a single SWE product versus the observed values from that product. Insets show the distribution of skill across basins, with the red line and value indicating the median. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Extended Data Fig. 5. The ensemble reconstruction based on the Random Forest model skillfully predicts the variability and trends in out-of-sample in situ snowpack data.

R² (a) and RMSE (b) of Random Forest model predictions of in situ March SWE at 2,961 locations over the period 1981–2020. Insets show the distribution of skill across sites, with the red line and value indicating the median. Observed (c) and reconstructed (d) 1981–2020 March SWE trends. c, Scatterplot of reconstructed versus observed trends, where each dot represents an in situ location. Points are colored by their density. Dashed line denotes perfect agreement between reconstructed and observed trends. Pearson’s correlation is shown in bottom right corner. Maps were generated using cartopy v0.18.0. Source Data

Extended Data Fig. 6. CMIP6 model bias in winter temperature and precipitation trends largely within range of natural variability.

Observed trends in November-March average temperature (a) and total precipitation (d) from 1981 to 2020. b, e. Ensemble mean of historical CMIP6 simulations. c, f. Average bias in trends across all observation-model combinations. Hatching indicates regions where the observed trend falls outside the 2.5–97.5th percentile range of the CMIP6 trends. Maps were generated using cartopy v0.18.0. Source Data

Extended Data Fig. 7. Uncertainty in the attribution of human-caused snowpack trends resides with climate model structure and modeled internal variability, not observations.

a, Dominant source of uncertainty in reconstruction-based estimates of forced March SWE trends from 1981 to 2020. b-e, Percentage of total uncertainty in forced SWE trends attributable to (b) observational uncertainty in gridded SWE products, (c) observational uncertainty in temperature and precipitation data products, (d) uncertainty in the forced response of temperature and precipitation across different climate models, and (e) uncertainty in the forced response of temperature and precipitation arising from internal variability (Methods). Hatching indicates basins where the uncertainty attributable to a given source is insufficient to change the sign of the ensemble mean estimate of the forced SWE trend. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Extended Data Fig. 8. Historical associations among climate, snowpack, and snow-driven runoff portend accelerating changes to snow hydrology.

Column 1: Historical change in November-March average temperature (a), total precipitation (c), March average SWE (e), and snowpack-driven April-July runoff (g) over the period 1981–2020. Values represent averages across all data products and hatching indicated basins where fewer than 80% of products agree on the sign of the change. Column 2: 2070–2099 changes under the SSP2-4.5 forcing scenario relative 1981–2020. Temperature (b) and precipitation (d) are calculated as the difference within each model realization between the end-of-century and climatological periods and future SWE (f) and runoff (h) changes are calculated according to Equations 1 and 2, respectively. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data

Extended Data Fig. 9. The Random Forest snowpack reconstruction methodology exhibits high skill based on a perfect model framework.

CMIP6 ensemble mean forced (HIST minus HIST-NAT) trends in March SWE from 1981–2020 based on (a) climate model SWE output and (b) SWE estimated using climate model temperature and precipitation and Random Forest model. c, Scatterplot of reconstructed versus original trends, where each dot represents a grid cell. Points are colored by their density. Dashed line denotes perfect agreement between reconstructed and original trends. Spatial correlation is shown in the bottom right corner. Source Data

Extended Data Fig. 10. The Random Forest model is extended to predict runoff from snowmelt skillfully.

R² (a) and RMSE (b) of Random Forest model predictions of April-July basin-scale runoff from 1981 to 2020. Insets show the distribution of skill across sites, with the red line and value indicating the median. Observed (c) and reconstructed ensemble mean (d) 1981–2020 April-July runoff trends. e, Scatterplot of reconstructed versus observed trends, where each dot represents a basin. Dashed line denotes perfect agreement between reconstructed and observed trends. Spatial correlation is shown in center left. Maps were generated using cartopy v0.18.0. River basin boundaries come from the Global Runoff Data Centre’s Major River Basins of the World database. Source Data