Evidence of human influence on Northern Hemisphere snow loss

The study addresses whether and to what extent anthropogenic climate change has already influenced Northern Hemisphere spring snowpack (SWE) and how snow is expected to respond to further warming. Despite robust hemispheric warming since 1981, observational SWE products show inconsistent snowpack trends, complicating detection and attribution. The paper highlights three challenges: (1) large observational uncertainties in SWE estimates leading to disagreement on trend direction in many major basins; (2) substantial internal climate variability across timescales (for example, Pacific Decadal Oscillation and Atlantic Multidecadal Variability) that can mask or mimic forced signals; and (3) opposing effects of warming that can increase cold-season precipitation and snowfall extremes in cold regions, partially offsetting melt-driven losses. The authors aim to overcome these issues by integrating multiple observations with empirical and climate models to attribute snowpack changes at hemispheric and basin scales and to generalize snowpack sensitivity to warming.

The paper situates its work within prior assessments and studies: IPCC assessments state Northern Hemisphere spring SWE has generally declined since 1981, but the magnitude, timing, and spatial pattern attributable to anthropogenic forcing remain unclear at decision-relevant scales. Previous attribution studies sometimes relied on limited models/realizations, risking conflation of internal variability with model structural uncertainty, which is high for snowpack. Regional studies have normalized SWE by precipitation to isolate temperature signals, but this can omit the full climate change effect on snow by excluding precipitation changes. The role of internal variability (e.g., PDO/AMV) and observed increases in winter precipitation in some regions further complicate detection. The study builds on literature evaluating SWE products and CMIP6 performance, and on attribution methods using pattern correlation and large ensembles to separate forced from internal variability signals.

• Data integration: Combined multiple gridded datasets of SWE, temperature, precipitation, and runoff with in situ observations to form an observations-based ensemble. Five long-term gridded SWE products were used and compared with in situ measurements.
• Detection via spatial pattern correlation: Computed spatial patterns of 40-year (1981–2020) March SWE trends and compared them with CMIP6 ensembles: HIST (with anthropogenic and natural forcings) and HIST-NAT (natural-only). Generated a null distribution of pattern correlations using 40-year segments from pre-industrial control (PIC) simulations (N = 78,601) to evaluate whether observed patterns are distinguishable from natural variability.
• Empirical reconstruction: Built basin-scale SWE reconstructions using a random forest machine-learning approach relating March SWE to cold-season temperature and precipitation. Created a large ensemble by factorially combining observational products and climate model forcings to sample observational, empirical, and model uncertainties. Reconstructions skill: spatial pattern correlations of trends 0.90–0.97 across products; median RMSE <8% across products/basins; validated with ~3,000 in situ sites (trend pattern correlation 0.72; median RMSE 22%).
• Attribution at basin scale: Estimated anthropogenically forced changes in temperature and precipitation by differencing CMIP6 HIST and HIST-NAT. Removed forced components from observed temperature/precipitation to construct counterfactual no-anthropogenic-change reconstructions, enabling isolation of forced effects on SWE trends per basin.
• Bias assessment: Compared CMIP6 simulated historical temperature and precipitation trends with observations; quantified whether apparent biases exceed internal variability envelopes (<1% for temperature; <3% for precipitation exceedances), supporting model suitability for attribution.
• Uncertainty partitioning: Quantified contributions to uncertainty in forced SWE trends from internal variability, climate model structural differences, and observational SWE/meteorology uncertainties (Extended Data Fig. 7), identifying dominant sources by basin.
• Nonlinear sensitivity analysis: Estimated marginal sensitivity of March SWE to 1°C warming as a function of climatological winter (Nov–Mar) temperature using in situ, gridded products, climate models, and basin reconstructions. Change-point analysis identified a nonlinear inflection near −8°C.
• Future impacts: Assessed end-century (2070–2099) changes in March SWE and SWE-driven April–June runoff under SSP2-4.5 across basins, relating changes to basin populations and projected warming.

• Detection and attribution at hemispheric scale: Spatial patterns of observed March SWE trends (1981–2020) are highly similar to CMIP6 HIST but not to HIST-NAT, indicating anthropogenic influence. Probabilities that observed pattern similarity arises from natural variability alone: virtually certain (>99%) for in situ observations, ensemble mean of gridded SWE, TerraClimate, and JRA-55; extremely likely (~97%) for ERA5-Land and Snow-CCI; not detectable for MERRA-2 (~78%).
• Basin-scale trends and attribution: Empirical reconstructions show consistent direction of March SWE trends in ~48.5% of major basins (82/169). Anthropogenically forced changes in temperature and precipitation altered SWE trends in 31/169 basins. Forced SWE trends: mid-latitudes (south of 60° N) declined by 4.1 ± 3.4% per decade; high-latitude Arctic-draining basins increased by 2.5 ± 1.8% per decade.
• Internal variability masks forced signals: Some basins (e.g., Columbia: forced decline ~4.8% per decade; Saint Lawrence: ~6.9% per decade) exhibit significant forced declines despite small or non-significant observed historical trends, implying masking by internal variability. Conversely, basins like the Rio Grande show large historical declines (>10% per decade) but limited agreement that forcing caused those declines; counterfactual reconstructions indicate significant declines even without anthropogenic forcing in some basins (Rio Grande ~6.3% per decade).
• Countervailing drivers: Forced warming generally reduces March SWE except in the coldest basins, while forced precipitation increases partly offset warming-driven losses in cold regions. Outside cold continental interiors, forced precipitation-driven SWE increases are generally insignificant.
• Nonlinear temperature sensitivity: A robust inflection in SWE sensitivity occurs near climatological winter temperature of about −8°C. Below −8°C, SWE is relatively insensitive to warming; above −8°C, each additional 1°C yields accelerating SWE losses. This inflection is consistent across in situ, gridded products, CMIP6 models, and basin reconstructions.
• Mass and population distributions: ~80% of Northern Hemisphere March snow mass lies in regions colder than −8°C (historically insensitive), but ~80% of the hemisphere’s population resides in snow-influenced regions warmer than −8°C (sensitive), implying rapidly emerging water risks with modest additional warming.
• Future runoff impacts (SSP2-4.5, 2070–2099 vs. 1981–2020): Strong projected spring runoff declines in highly populated basins due to nonlinear snow loss despite modest regional warming. Examples: Upper Mississippi (84 million people): −30.2% spring runoff; Colorado (14 million): −42.2%; Columbia (8.8 million): −32.7%; San Joaquin (6.8 million): −40.9%. Europe: Danube (92 million): −41.0%; Volga (60 million): −39.5%; Rhine (51 million): −33.0%; Po (18 million): −40.5%. Cold, sparsely populated Arctic-draining basins may see increased SWE and >10% spring runoff increases on average.
• Model-data performance and biases: CMIP6 models slightly overestimate historical warming in some regions and underestimate drying in the U.S. Southwest and wetting on the Tibetan Plateau, but <1% (temperature) and <3% (precipitation) of biases fall outside internal variability ranges, supporting attribution robustness.
• Uncertainty attribution: Climate model structural differences in forced temperature/precipitation responses dominate uncertainty in forced March SWE trend magnitude in over half of basins (95/169). Observational SWE uncertainties limit constraints in data-sparse regions (e.g., High Mountain Asia). Internal variability is the dominant uncertainty in roughly one in eight basins.

The findings demonstrate that anthropogenic forcing has already shaped the spatial pattern of Northern Hemisphere spring snowpack trends and that internal variability can obscure or amplify these changes at basin scales. The identified −8°C inflection explains the historical elusiveness of widespread snow loss despite substantial warming: most snow mass resides in cold regions with low marginal sensitivity. However, because the majority of people live in snow-influenced regions warmer than −8°C, modest additional warming will push many basins into a highly sensitive regime, accelerating SWE losses and reducing spring runoff. Forced precipitation increases provide some offset in cold interiors, but are generally insufficient in warmer basins where warming-driven rain–snow partitioning and melt dominate. The data-model fusion and counterfactual analyses clarify where observed declines are attributable to forcing versus internal variability, enabling more targeted risk assessment for water resources. The results underscore the urgency of adaptation planning for winter flood management and reduced warm-season streamflow, as clear observational detection may lag behind the onset of impactful changes due to nonlinearity and variability.

This work provides robust attribution of human influence on Northern Hemisphere spring snowpack since 1981, resolves why snow has been a poor sentinel of warming to date, and identifies a simple, generalizable nonlinearity—an inflection near −8°C—that governs snowpack sensitivity to warming. It quantifies forced SWE trends across basins, reveals regions where internal variability has masked large forced declines, and projects substantial spring runoff reductions in many populous basins under moderate emissions. The study contributes a scalable data-model fusion framework that leverages observational and model uncertainties to isolate forced signals. Future research should: (1) improve regional climate model skill for cold-season temperature and precipitation; (2) better constrain SWE observations, especially in data-sparse regions like High Mountain Asia; (3) further quantify and communicate irreducible internal variability to guide robust water management; and (4) refine hydrologic impact assessments linking nonlinear snow loss to downstream water supply, ecosystems, and flood risks under different emissions pathways.

• Observational uncertainties: Gridded SWE products disagree on trend magnitudes and sometimes on trend direction, especially in regions lacking in situ measurements (e.g., High Mountain Asia, parts of central Asia and Siberia).
• Model biases and uncertainty: CMIP6 models slightly overestimate warming in some regions and misrepresent regional precipitation trends (e.g., U.S. Southwest drying, Tibetan Plateau wetting), though most biases lie within model internal variability ranges. Structural model differences are the dominant source of uncertainty in forced SWE trends in many basins.
• Internal variability: Large, irreducible internal climate variability can mask or amplify forced signals and remains the dominant uncertainty source in roughly one in eight basins, limiting attribution at regional scales over multidecadal periods.
• Dataset dependence: Detection results vary by dataset (e.g., no detectable anthropogenic signal in MERRA-2), indicating sensitivity to observational product choice.
• Precipitation uncertainty: Greater uncertainty in precipitation responses and larger contribution of internal variability to hydroclimate trends reduce confidence in forced precipitation effects outside cold interiors.