The rate, magnitude, and causes of snow loss are crucial for benchmarking climate change and managing water security risks. While warming trends are consistent across scales, observational data on snowpack loss is inconsistent, hindering preparedness. This inconsistency is puzzling given observed warming trends at hemispheric, continental, and river-basin scales. Although the IPCC acknowledges a general decline in Northern Hemisphere springtime snow water equivalent (SWE) since 1981, the extent of anthropogenic influence remains unclear, particularly at decision-relevant scales. The lack of robust attribution makes it difficult to identify vulnerable regions and develop effective water security strategies. Three factors contribute to inconsistent snowpack responses to warming: observational uncertainties in SWE estimates; high snowpack variability from low-frequency climate variability (Pacific Decadal Oscillation, Atlantic Multidecadal Variability); and the non-unidirectional relationship between forcing and snowpack (warming can enhance cold-season precipitation, offsetting losses). This study aims to address these uncertainties by combining observations-based ensembles with empirical and climate models to attribute snowpack changes to anthropogenic warming at various scales, assessing the effects of temperature and precipitation changes on snow water storage, and generalizing snowpack and runoff responses to future warming.

Previous research has highlighted the potential impacts of a warming climate on water availability in snow-dominated regions (Barnett et al., 2005; Immerzeel et al., 2020; Mankin et al., 2015; Qin et al., 2020), and the importance of understanding snow drought (Gottlieb & Mankin, 2022). However, inconsistencies in observational data on snowpack trends (Mortimer et al., 2020; Fox-Kemper et al., 2021) have made it difficult to definitively attribute snow loss to human activities. Studies have attempted to attribute snowpack declines to anthropogenic climate change at regional scales (Barnett et al., 2008; Pierce et al., 2008; Najafi et al., 2017; Jeong et al., 2017), but these have often been limited by the use of a small number of climate models and/or model realizations, leading to challenges in disentangling internal variability and model structural uncertainties (Mankin & Diffenbaugh, 2015; Mankin et al., 2020; Lehner et al., 2020). The complexities of the snowpack response to climate change, including the influence of precipitation variability (Mankin & Diffenbaugh, 2015) and the potential for enhanced snowfall extremes to offset warming-driven losses (O'Gorman, 2014; Brown & Mote, 2009), have further complicated attribution efforts.

This study combines an ensemble of observations-based snowpack, temperature, precipitation, and runoff data products with empirical and climate models to attribute snowpack changes to anthropogenic warming. The analysis uses a data-model fusion approach employing a random forest machine-learning algorithm. This approach combines empirical models of SWE with climate model simulations to estimate the effects of anthropogenic emissions on temperature and precipitation at finer scales. Multiple gridded snowpack, temperature, and precipitation datasets were used to create an ensemble of empirical reconstructions of historical March SWE at the basin scale. This strategy aimed to effectively sample observational uncertainty, reconstruct snowpack as a function of temperature and precipitation to isolate the influence of forced and unforced changes, and assess whether signals of forced snowpack changes emerged above the noise of observational, internal variability, and climate model uncertainties. The Coupled Model Intercomparison Project Phase 6 (CMIP6) HIST and HIST-NAT experiments were used to estimate the forced response of temperature and precipitation. A counterfactual no-anthropogenic-climate-change snowpack was estimated by removing the forced response from observed temperature and precipitation time series and re-estimating snowpack reconstructions. The nonlinear sensitivity of snowpack to warming was examined by analyzing the relationship between average winter temperatures and the marginal sensitivity of snow change to temperature change. Finally, population exposure to projected snow loss and runoff change was assessed using the Shared Socioeconomic Pathway (SSP) 2-4.5 scenario.

Despite substantial uncertainty in spatially distributed estimates of snowpack, gridded snow products shared a distinct spatial pattern of historical trends agreeing with in situ observations. March SWE declined sharply in the southwestern USA and much of western, central, and northern Europe (10-20% per decade). Coupled climate model simulations forced with historical human and natural forcing captured some features of the observed spatial pattern of snow change, particularly large snow loss over Europe and modest gains over Northern Eurasia. Simulations excluding anthropogenic emissions failed to capture the observed pattern. Attribution analysis showed that it is virtually certain (>99% probability) that human emissions contributed to the observed pattern of March snowpack trends in in situ observations and in the average of the gridded ensemble. Basin-scale SWE reconstructions revealed that spring snowpack declined in many mid-latitude basins (around 10% per decade in the southwestern USA and Europe) and increased in cold, high-latitude basins. Human-forced changes to temperature and precipitation altered spring snowpack trends in 31 major river basins. Anthropogenic temperature changes generally reduced March SWE except in the coldest basins. Anthropogenically forced precipitation increases offset some warming-driven losses, but were generally insignificant outside cold continental interiors. Analysis of the relationship between temperature and snowpack sensitivity revealed a highly nonlinear response. Below -8°C, snowpack is little affected by warming; however, each additional 1°C above this threshold results in accelerating losses. This explains the lack of widespread snow loss despite warming. Under SSP2-4.5, highly populated basins are projected to see strong declines in spring runoff due to nonlinear snow loss, even with modest warming. The western USA and Europe are particularly vulnerable.

This study demonstrates that human-caused warming has significantly impacted Northern Hemisphere snowpack, despite the challenges posed by observational uncertainties and the complex interplay of temperature and precipitation. The finding of a highly nonlinear temperature sensitivity of snowpack provides a critical explanation for why snow loss has not been more widely detected to date and underscores the accelerating risks to water security in populous regions as warming continues. The results highlight the importance of considering both temperature and precipitation changes when assessing snowpack trends. Furthermore, the study emphasizes the need to incorporate multiple data sources and modeling approaches to account for uncertainties and improve the accuracy of snowpack projections. The findings have important implications for water resource management and adaptation planning in snow-dependent regions.

This study provides robust evidence of human influence on Northern Hemisphere snow loss, revealing a nonlinear relationship between temperature and snowpack decline. The findings highlight the vulnerability of populous, mid-latitude basins to accelerating snow loss and reduced spring runoff. Future research should focus on improving climate model representations of regional climate, refining observational estimates of SWE, and further investigating the role of internal climate variability in shaping snowpack trends.

While the study uses a comprehensive approach to address uncertainties, limitations remain. The accuracy of the snowpack reconstructions depends on the quality and availability of input data, and there are inherent uncertainties in climate model projections. The focus on March SWE might not fully capture the complexities of snowpack dynamics throughout the winter season. The analysis relies on a specific emissions scenario (SSP2-4.5), and results may differ under alternative scenarios.