Introduction

Recent changes in snow patterns in the Northern Hemisphere have led to a significant rise in the frequency of snow drought events, characterized by insufficient snowfall and diminished snowpack, resulting in reduced water availability1,2,3,4,5. The decline in snowfall and snowpack has direct implications for water supply systems, impacting both surface water and groundwater resources6,7,8,9,10. Snowmelt plays a crucial role in replenishing rivers, lakes, and aquifers; its reduction can lead to water shortages during the dry season, affecting sectors such as irrigation, hydropower generation, and municipal water supply11,12,13,14,15,16. Furthermore, snow drought has significant socioeconomic consequences. Industries dependent on snow, including winter tourism, skiing, and snowboarding, are particularly vulnerable to reduced snowpack. These sectors are vital to local economies and employment, and declining snow conditions can result in financial losses, job reductions, and economic downturns in affected areas. Consequently, snow drought poses serious threats, encompassing water scarcity, ecological disruption, and socioeconomic impacts17,18,19,20,21. Therefore, the reliable predictions regarding the spatiotemporal characteristics of future snow droughts are essential for developing effective adaptation strategies and enhancing resilience in impacted regions22.

Previous studies have documented a trend of increasing snow drought events across many regions of the Northern Hemisphere, largely associated with declining snowpack and reduced snow water equivalent15,23. Future projections under climate change scenarios consistently indicate an exacerbation of these trends. For instance, Cowherd et al.24 employed CMIP6 model outputs to analyze global snow drought characteristics from 1850 to 2100. Their findings revealed a significant increase in the frequency and duration of snow droughts. However, significant biases are evident in CMIP6 models’ climate change projections25,26,27, which introduce substantial uncertainties in projecting future snow drought changes when using raw, uncorrected CMIP6 data. These biases arise from several factors, including inaccuracies in the parameterization of physical processes, limitations in spatial resolution, insufficient capabilities to simulate extreme climate events, oversimplifications in land-atmosphere coupling processes, and inadequate consideration of the complexities inherent in the climate system. Consequently, these deficiencies result in imprecise simulations of snowfall, snowmelt processes, and land-atmosphere interactions, thereby compromising the reliability of predictions regarding future snow drought characteristics. For example, Zhong et al.28 reported that 22 CMIP6 models overestimated snow depth in high-latitude regions of the Northern Hemisphere during 1955–2014, with the spatial mean exceeding the observed baseline (1976–2005) by 18 ± 16 cm28. Kouki et al.29 reveal that 17 high-resolution (~100 km) CMIP6 models exhibit biases in simulating snow water equivalent, with CESM2 and NorESM2-MM showing the largest overestimations of 150–200% in northern North America and extensive regions of Eurasia29. Räisänen et al. (2023) showed that four models overestimate winter mean or annual maximum SWE in Finland by more than 50%30. These biases underscore the necessity of applying bias-correction techniques to improve the reliability of snow drought assessments.

Downscaling techniques are essential for converting large-scale, low-resolution global climate model outputs into small-scale, high-resolution regional climate change information. By integrating observational data, these techniques enable a more nuanced characterization of snow drought changes, thereby reducing simulation errors in regional assessments and enhancing the reliability of predictions. Vrac et al.31 developed a bias-correction methodology based on cumulative distribution function transformations (CDF-t) to address large-scale model biases31. Their findings demonstrate a significant improvement in the performance of regional climate models concerning observations for wind speed, temperature, heat wave, and precipitation distributions31,32,33,34,35. The CDF-t method offers several advantages over traditional bias-correction techniques. Specifically, CDF-t effectively captures the entire probability distribution of climate variables, allowing for a more accurate representation of extreme events and non-linear relationships. This capability enhances the robustness of the downscaling process, making it particularly valuable in contexts where extreme weather events are critical, such as in the assessment of snow droughts36. Despite the advantages of the CDF-t statistical downscaling method, there remains a notable gap in research applying this approach to mitigate uncertainties in snow drought assessments. This underscores the need for further investigation into the effectiveness of the CDF-t method in enhancing the accuracy and reliability of snow drought predictions.

To achieve these objectives, our study utilizes SWE outputs from 29 state-of-the-art CMIP6 models under historical (1982–2014) and future (2030–2100) scenarios across multiple Shared Socioeconomic Pathways (SSP126, SSP245, SSP370, and SSP585). We establish a functional relationship between the cumulative distribution functions (CDFs) of large-scale climate variables and their regional counterparts, and apply this relationship to model outputs under future emission scenarios to obtain downscaled SWE projections. To evaluate the effectiveness of the CDF-t bias-correction method, the bias-corrected SWE data are compared against observational datasets, with improvements quantified in both temporal and spatial dimensions. Using the corrected future SWE data, we subsequently calculate the Snow Water Equivalent Index (SWEI) to systematically analyze the spatiotemporal evolution of future snow droughts across the Northern Hemisphere. This analysis aims to provide a clearer understanding of snow drought trends, enhance the quality of climate change-related decision-making services, and offer a robust scientific basis for impact assessments in agriculture, water resources management, and socioeconomic sectors.

Results and discussions

Assessment of SWEI downscaling performance

Before employing the CDF-t statistical downscaling technique to forecast future snow droughts in the Northern Hemisphere, it is essential to evaluate the CDF-t’s capability to simulate current climate conditions. A functional relationship model was developed using the observed grid SWE data from 1982 to 1998 (the calibration period) alongside historical data from 29 climate models. This functional relationship (transfer function) was subsequently applied to the simulated SWE data from the validation period of 1999–2014 for each CMIP6 model, yielding downscaled sequences. By comparing these sequences with the observed spatial-temporal characteristics of SWE during the same period, we assess the performance of the CDF-t statistical downscaling method in simulating SWE in the Northern Hemisphere.

As shown in Fig. 1a, after downscaling, the cumulative distribution functions of the Northern Hemisphere mean SWE during the validation period of 1999–2014, based on the CMIP6 simulations, closely align with the observed cumulative distribution function. This alignment effectively demonstrates the validity of the bias-correction method employed. Figure 1b presents the observed, simulated (before downscaling), and simulated (after downscaling) monthly mean SWE in the Northern Hemisphere. Notably, after downscaling, the multi-model ensemble mean SWE exhibits a significant reduction in discrepancies with the observations (1.8 kg m−2), showing a 89.0% decrease in the mean absolute bias compared to the values prior to downscaling (16.3 kg m−2).

Fig. 1: Comparison of the spatiotemporal characteristics of SWE before and after applying the CDF-t statistical downscaling technique during the validation period from 1999 to 2014.
Fig. 1: Comparison of the spatiotemporal characteristics of SWE before and after applying the CDF-t statistical downscaling technique during the validation period from 1999 to 2014.
Full size image

(a) shows the observed and simulated cumulative distribution functions of Northern Hemisphere mean SWE, both prior to and following the application of the CDF-t method. (b) presents the observed and simulated monthly mean NH SWE values, both prior to and following the application of the CDF-t methods. (c) shows the multi-year average spatial distribution of SWE for the GISS-E2-1-G model prior to using the CDF-t technique. (d) shows the multi-year average spatial distribution of SWE for the GISS-E2-1-G model after the application of the CDF-t method.

In terms of spatial distribution, the downscaled CMIP6-simulated SWE exhibits a spatial pattern more consistent with observations, indicating enhanced reliability in the model outputs. By contrast, the original CMIP6 simulations show substantial biases in monthly mean SWE over Northern Hemisphere land surfaces compared to observed data. Taking the GISS-E2-1-G model as an example, Fig. 1c reveals a pronounced overestimation of SWE across most land regions of the Northern Hemisphere. This systematic overestimation highlights the critical need for bias correction, as using uncorrected CMIP6 projections to assess snow drought could lead to underestimation of its actual severity. In comparison, after bias correction, the discrepancy between the model and observations is markedly reduced (Fig. 1d). These results underscore the importance of applying robust downscaling techniques to improve the accuracy of climate projections and support informed water resource management strategies.

Spatial and temporal patterns of SWEI after bias correction

In Fig. 2, we present an analysis of the temporal patterns of SWEI for historical (1982–2014) and future (2015–2100) periods. The data utilized in this study are derived from 29 CMIP6 models (Supplementary Table 1), providing a comprehensive assessment of SWEI changes during the extended snow season (November–April) in the Northern Hemisphere.

Fig. 2: Time series of Northern Hemisphere mean SWEI during the extended snow season (November–April) from 1982 to 2100, derived from 29 bias-corrected CMIP6 models and their ensemble mean.
Fig. 2: Time series of Northern Hemisphere mean SWEI during the extended snow season (November–April) from 1982 to 2100, derived from 29 bias-corrected CMIP6 models and their ensemble mean.
Full size image

(a–d) present temporal patterns of future SWEI under SSP126, SSP245, SSP370, and SSP585 scenarios, respectively. The corresponding model and ensemble mean names are listed on the right side of the figure.

During the historical period, a modest yet statistically significant declining trend in SWEI was observed in the Northern Hemisphere. This established baseline underscores that snow droughts have already been intensifying in recent decades (−0.050 ± 0.027 decade−1). Under future scenarios, this trend is projected to accelerate markedly. Analyzing the individual Shared Socioeconomic Pathways (SSPs), we observe distinct variations. For SSP245, SSP370, and SSP585, the declining trends in SWEI are substantially steeper than the historical rate, with the magnitude of decrease intensifying from SSP245 (−0.072 ± 0.035 decade−1) to SSP585 (−0.125 ± 0.054 decade−1). This progression highlights how higher-emission pathways lead to more severe and rapid reductions in snow water availability. The consistent and accelerated decrease in SWEI under the other SSPs over the projected period indicates a potential increase in the severity and frequency of snow droughts. These findings highlight the critical implications of emission scenarios for future snow hydrology and underscore the significance of climate mitigation to minimize adverse impacts.

Conversely, under the sustainable SSP126 scenario, SWEI exhibits a characteristic pattern of an initial decrease (−0.065 ± 0.037 decade−1) followed by an increase (0.013 ± 0.030 decade−1), with a turning point around 2070. The initial decline is linked to ongoing global warming and increased snowmelt. However, as SSP126 projects a decline in CO2 concentrations after 2070, the warming slows, resulting in reduced snow loss and a trend reversal. This turning point can be attributed to the following reasons: Before 2070, the SSP126 scenario experiences a significant increase in CO2 concentrations, leading to global warming. This global warming results in increased snowmelt and a reduction in snowfall. These factors contribute to a decreasing trend in SWEI until around 2070. SSP126 represents a sustainable and “green” pathway, and after 2070, it exhibits a decline in CO2 concentrations37. This reduction in CO2 concentrations significantly slows down global warming, resulting in reduced snow loss and an increase in precipitation forming as snow. As a result, SWEI shows an increasing trend after 2070.

Figure 3 provides a comprehensive view of the geographical distribution of the CMIP6 ensemble mean trend in SWEI during the period from 1982 to 2100. The historical period (Fig. 3a) was characterized by a widespread and significant decline in SWEI, affecting 74.0% of the Northern Hemisphere’s land area (excluding regions with an annual cumulative SWE of less than 10 mm). Under future scenarios, this decreasing trend intensifies markedly for SSP245, SSP370, and SSP585, with the proportion of the study area experiencing a significant decrease expanding dramatically to between 84.8% (SSP245) and 88.8% (SSP585). Notably, the magnitude and spatial extent of the decline in SSP126 are substantially weaker than during the historical period and all other future scenarios, highlighting the mitigating effect of sustainable pathways. Regional analysis (Fig. 3f) identifies North America (sub-regions NWN, ENA, GRL) and Northern Europe (NEU) as experiencing the most pronounced SWEI reductions. In contrast, the easternmost part of Siberia exhibits a consistent increasing trend across all scenarios, a phenomenon attributed to atmospheric dynamic-induced moisture convergence that enhances snowfall38.

Fig. 3: Geographical distribution of the CMIP6 ensemble mean trend in the Snow Water Equivalent Index (SWEI) for the extended snow season (November–April).
Fig. 3: Geographical distribution of the CMIP6 ensemble mean trend in the Snow Water Equivalent Index (SWEI) for the extended snow season (November–April).
Full size image

The map shows the historical trend for 1982–2014 (a), and the projected future trends for 2030–2100 under scenarios SSP126 (b), SSP245 (c), SSP370 (d), and SSP585 (e). (f) presents the average trends for 11 Northern Hemisphere sub-regions (Supplementary Fig. 1 and Supplementary Table 2). All results are based on the ensemble mean of 29 CMIP6 models after CDF-t bias correction.

Annual mean climatology of snow drought

Understanding the future characteristics of snow drought is crucial for implementing effective measures to protect agricultural systems, water resources, and ecosystems, and to mitigate the adverse impacts of drought. Utilizing the CDF-t bias-corrected SWEI data alongside the classification of snow drought presented in Supplementary Table 3, we conducted an investigation into the projected annual mean climatology of snow drought conditions [i.e., drought frequency (Fig. 4a, b), intensity (Fig. 4c, d), and duration (Fig. 4e, f)] in the Northern Hemisphere for the period from 2030 to 2100.

Fig. 4: Annual mean climatology of snow droughts during the extended snow season (November–April) from 1982 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Fig. 4: Annual mean climatology of snow droughts during the extended snow season (November–April) from 1982 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Full size image

The analysis is presented for two periods: the historical period (1982–2014) and the future projection period (2030–2100) under four emission scenarios (SSP126, SSP245, SSP370, SSP585). (a), (c), and (e) illustrate the Northern Hemisphere averages of the multi-year average frequency, intensity, and duration of future snow droughts of varying severity (D0, D1, D2, D3, and D4; refer to Supplementary Table 3 for the full description) across four emission scenarios, respectively. (b), (d), and (f) display the multi-year average frequency, intensity, and duration of future snow droughts (D4) under the SSP245 scenario, respectively.

Snow drought frequency is defined as the number of drought events occurring within the extended snow season (Fig. 4a, b). Our analysis indicates that the multi-year average frequency of snow droughts—encompassing all five severity categories—shows a significant increase with higher-emission scenarios across the Northern Hemisphere. During the historical period (1982–2014), the mean frequency was 0.80 ± 0.27 events per year. For the future period (2030–2100), the projected frequencies rise to 1.28 ± 0.42 (SSP126), 1.37 ± 0.45 (SSP245), 1.39 ± 0.48 (SSP370), and 1.41 ± 0.49 (SSP585) events per year. Among all categories, D1 (Moderate Drought) occurs most frequently (Fig. 4b), with a multi-model ensemble mean of 0.37–0.41 events (depending on the SSP scenario, Fig. 4a), accounting for 26.4%–36.0% of all drought occurrences. It is followed by D4 (Exceptional Drought, 0.27–0.44 events), D0 (Abnormally Dry, 0.24–0.30 events), D3 (Extreme Drought, 0.17–0.19 events), and D2 (Severe Drought, 0.16–0.17 events).

As shown in Fig. 4a, the frequency of D0 events gradually decreases from SSP126 to SSP585. This decline in milder drought categories is accompanied by a marked increase in the frequency of more severe droughts. The most pronounced increase is observed for D4, which rises from 0.27 events under SSP126 to 0.44 events under SSP585, representing a substantial increase of 63.0%. Spatially, under SSP245, the highest frequencies of the most severe droughts (D4) are projected over Europe, western Asia, and the central-eastern United States (Fig. 4b). In contrast, central-eastern Asia, northern North America, and the Tibetan Plateau show relatively lower frequencies. Similar spatial patterns are observed for other drought categories and emission scenarios, as shown in Supplementary Fig. 2.

Based on bias-corrected results, the multi-year average snow drought intensity across all five severity categories was –1.39 ± 0.00 during the historical period (1982–2014). This intensity is projected to increase (i.e., become more negative) under all future emission scenarios (2030–2100), escalating with rising radiative forcing: –1.56 ± 0.01 (SSP126), –1.57 ± 0.01 (SSP245), –1.59 ± 0.02 (SSP370), and –1.60 ± 0.02 (SSP585). While the intensity of the other three middle drought categories shows little variation across scenarios, the most severe category (D4) exhibits a clear trend of intensification, rising from –2.96 under SSP126 to –3.14 under SSP585 (Fig. 4c). This pattern underscores a growing severity of the most extreme snow droughts as emission levels increase.

Spatially, under SSP245, regions with the highest snow drought intensity for D4 are concentrated in Europe, western Asia, and the central-eastern United States (Fig. 4d). The most severely affected sub-regions include NWN (–2.93), MED (–2.91), NEU (–2.78), and CAS (–2.95) (see geographical locations in Supplementary Fig. 1). These areas experience substantial reductions in snow water equivalent, contributing to significant ecological and hydrological stress. In contrast, lower drought intensities are projected for central-eastern Asia, northern North America, and the Tibetan Plateau. This spatial pattern remains consistent across other drought categories and emission scenarios (Supplementary Fig. 3).

Snow drought duration is defined as the persistence of individual snow drought events (Fig. 4e, f). Analysis of the multi-year average duration reveals a contrasting response to increasing emission scenarios between the most severe drought category and all others.

For the most severe category, D4, the mean duration of single events shows a substantial increase (Fig. 4e), lengthening from 70.1 ± 6.0 days under SSP126 to 86.7 ± 11.1 days under SSP585—a 54.5% extension compared to the historical period (56.1 ± 2.8 days). In stark contrast, the mean duration for the other four categories (D0 to D3) is projected to shorten under future scenarios. For instance, D1 events decrease from 51.9 ± 1.7 days historically to between 43.3 and 46.8 days in the future. This pattern indicates that while the most extreme droughts are becoming more prolonged, less severe events are tending to become shorter.

Spatially, under SSP245 for the D4 category (Fig. 4f), the longest snow drought durations are projected for Europe, western Asia, and the central North America, including in the NWN (74.7 days), NAS (73.2 days), and ALA (69.6 days) sub-regions. In contrast, shorter durations are found in CNA (61.0 days) and EAS (62.2 days). This spatial pattern is generally consistent across other drought categories and emission scenarios (Supplementary Fig. 4).

It is noteworthy that although the absolute values of snow drought frequency, intensity, and duration vary among the CMIP6 models, the pattern revealed by the multi-model ensemble mean—that all categories of snow drought tend to become more frequent, more intense, and longer-lasting with increasing emission scenarios—shows high consistency across the models. For instance, each CMIP6 model supports a significant increase in the frequency of extreme (D3 and D4) snow droughts under the SSP585 scenario, underscoring the robustness of this conclusion.

Long-term trend of snow drought

We employed linear regression to analyze the trends in snow drought frequency from 1982 to 2100 (Fig. 5a–c). Linear trends are calculated from the CMIP6 multi-model ensemble mean time series. The reported uncertainty (±) represents one-standard deviation of the inter-model variability in the trend estimates. During the historical period (1982–2014), the frequency of all drought categories exhibited increasing trends (Fig. 5a), with the most pronounced increases observed for the less severe D0 and D1 categories (up to 2.93 ± 1.51 × 10−2 events decade−1). In contrast, future projections (2030–2100) reveal a marked divergence: the frequency of the least severe droughts (D0 and D1 under higher SSPs) shifts to a significant decreasing trend, while the frequency of the most severe droughts (D4) shows a dramatic increase.

Fig. 5: Decadal trends in future snow drought frequency during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Fig. 5: Decadal trends in future snow drought frequency during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Full size image

(a) illustrates the Northern Hemisphere averages of the decadal trends in the frequency of future snow droughts of varying severity (D0, D1, D2, D3, and D4) across four emission scenarios. (b) presents time series (15-year moving average) of the Northern Hemisphere averages for the frequency of future snow droughts (D4) under the SSP245 scenario, respectively. (c) displays the decadal trends in the frequency of future snow droughts (D4) under the SSP245 scenario, respectively.

This pattern indicates a fundamental structural shift in snow drought regimes. The decline in milder droughts does not imply a reduced risk; rather, it is compensated by a sharp increase in more extreme events. Specifically, the frequency of D4 snow droughts exhibits the most substantial increase, with the trend magnitude escalating from 0.80 ± 0.81 × 10−2 events decade⁻¹ under SSP126 to 6.34 ± 2.41 × 10−2 events decade−1 under SSP585 (Fig. 5a, b, Supplementary Fig. 5). The frequencies of D2 and D3 also show increasing trends across the four scenarios. In stark contrast, the D0 category shows a significant declining trend under all four emission scenarios: −0.04 ± 0.24 × 10−2 events decade−1 (SSP126), −0.55 ± 0.51 × 10−2 events decade−1 (SSP245), −1.33 ± 0.38 × 10−2 events decade−1 (SSP370), and −2.03 ± 1.10 × 10−2 events decade−1 (SSP585).

The spatial pattern of increasing D4 frequency (Fig. 5c and Supplementary Fig. 6) shows a significant upward trend across most of the Northern Hemisphere’s land surface, affecting 84.8%, 91.9%, 93.1%, and 93.5% of the area under SSP126, SSP245, SSP370, and SSP585, respectively. Particularly pronounced increases are projected across Europe, western Asia, and the central-eastern United States. Sub-regions such as NWN, ALA, and NAS exhibit the most vigorous growth rates, with data from D4 under SSP245 (Fig. 5c) showing increases ranging from 1.4 to 4.4 × 10−2 events decade−1. In contrast, central-eastern Asia shows much weaker increases, with sub-regions like EAS and CAS projecting only minimal changes. For the less severe drought categories D0 and D1, a widespread decrease in frequency is observed under SSP370 and SSP585 (Supplementary Fig. 6), especially across Europe, western Asia, and central-eastern North America, where the proportion of area exhibiting a declining trend reaches approximately 42.2–69.1%. In comparison, the magnitude of trend changes for D2 and D3 is relatively moderate.

Snow drought intensity is quantified using the SWEI, where more negative values indicate greater severity. The negative trend in SWEI shown in Fig. 6a reflects a significant intensification of future snow droughts, particularly for the most severe category (D4). This intensification escalates markedly under higher-emission scenarios, with the decadal trend in SWEI for D4 increasing from –0.82 ± 1.55 × 10−2 decade−1 under SSP126 to –6.88 ± 8.61 × 10−2 decade−1 under SSP585 (Fig. 6b and Supplementary Fig. 7). In contrast, the intensity trends for less severe drought categories are substantially weaker. For D0 and D1, trends, though negative, remain small in magnitude, ranging from –0.01 ± 0.03 to –0.04 ± 0.07 × 10−2 decade−1 and from –0.06 ± 0.12 to –0.30 ± 0.37 × 10−2 decade−1, respectively. The trends for D2 and D3 are marginal and statistically non-robust, hovering near zero. This pattern highlights a clear escalation of snow drought severity, driven overwhelmingly by a substantial intensification of the most extreme events (D4), while the intensity of milder drought categories remains relatively stable.

Fig. 6: Decadal trends in future snow drought intensity during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Fig. 6: Decadal trends in future snow drought intensity during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Full size image

(a) illustrates the Northern Hemisphere averages of the decadal trends in the intensity of future snow droughts of varying severity (D0, D1, D2, D3, and D4) across four emission scenarios. (b) presents time series (15-year moving average) of the Northern Hemisphere averages for the intensity of future snow droughts (D4) under the SSP245 scenario, respectively. (c) displays the decadal trends in the intensity of future snow droughts (D4) under the SSP245 scenario, respectively.

Spatially, the intensification of snow drought severity is most pronounced for the D4 category (Fig. 6c and Supplementary Fig. 8). Under the low-emission scenario (SSP126), a significant intensifying trend in SWEI—indicating stronger drought severity—is observed over only 63.6% of the Northern Hemisphere’s land surface. As emission levels rise, the affected area expands considerably, increasing from 84.5% under SSP245 to 92.1% under SSP585. Geographically, the most notable intensification is projected across western Asia and central North America. In contrast, trends in drought intensity for the other categories (D0–D3) are generally weak and lack spatial coherence; more than half of the regions did not show statistically significant trends, suggesting that changes in their intensity are relatively limited.

As shown in Fig. 7a, b, the duration of the most severe snow droughts (D4) is projected to increase significantly across the Northern Hemisphere, with the trend magnitude escalating under higher-emission scenarios. The decadal trend rises from 0.59 ± 0.71 days decade−1 under SSP126 to 6.98 ± 2.17 days decade−1 under SSP585. Under SSP2-4.5, the trend is 2.57 ± 1.15 days decade−1. In contrast, the durations of the less severe drought categories (D0–D3) are generally projected to decrease in the future across all scenarios (Fig. 7a and Supplementary Fig. 9). For instance, the decadal trend for D1 ranges from –0.13 ± 0.13 to –1.18 ± 0.43 days decade−1, while for D3, it varies from –0.02 ± 0.06 to –0.34 ± 0.20 days decade−1. This pronounced lengthening of the most extreme drought events, coupled with a shortening of milder droughts, underscores a fundamental shift in snow drought characteristics. The substantial increase in the duration of D4 droughts is expected to exert severe pressures on ecological and hydrological systems, highlighting the critical importance of emission mitigation to limit the persistence of such extreme events.

Fig. 7: Decadal trends in future snow drought duration during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Fig. 7: Decadal trends in future snow drought duration during the extended snow season (November–April) from 2030 to 2100, based on 29 CMIP6 models with CDF-t bias correction.
Full size image

(a) illustrates the Northern Hemisphere averages of the decadal trends in the duration of future snow droughts of varying severity (D0, D1, D2, D3, and D4) across four emission scenarios. (b) presents time series (15-year moving average) of the Northern Hemisphere averages for the duration of future snow droughts (D4) under the SSP245 scenario, respectively. (c) displays the decadal trends in the duration of future snow droughts (D4) under the SSP245 scenario, respectively.

Spatially, the projected lengthening of D4 drought duration is both widespread and amplified by higher emissions. The proportion of Northern Hemisphere land surface experiencing a significant increase grows from 59.1% under SSP126 to 73.8% under SSP585, with the most pronounced changes concentrated in central-western Asia, Europe, and the central United States (Fig. 7c). Regionally, the most rapid increases occur in NWN [0.71 (SSP126) – 8.1 (SSP585) days decade⁻¹], NEU [0.72 (SSP126) – 7.2 (SSP585) days decade−1] and ENA [0.58 (SSP126) – 7.0 (SSP585) days decade−1] show more modest trends. A notable exception is western Asia, particularly Siberia, where a significant decrease in D4 duration is projected. This pattern contrasts sharply with that of less severe droughts (D0–D3), which exhibit widespread decreases in duration across much of the Northern Hemisphere, especially over Siberia and eastern Asia (Supplementary Fig. 10). Both the spatial extent and magnitude of these decreases in milder drought categories also intensify under higher-emission scenarios.

To assess the robustness of these long-term trends, we calculated the proportion of models showing statistically significant trends. The results indicate that not only is the declining trend in SWEI supported by each CMIP6 model, but the worsening trends in snow drought frequency, intensity, and duration also show high inter-model agreement. For example, all the models show a significant lengthening trend for D4 snow drought duration under the SSP585 scenario. This demonstrates that the intensifying trends of snow droughts we have identified are a pervasive and robust signal, not driven by individual models.

Impact of bias correction on snow drought projections

Extensive literature has established that CMIP6 models commonly exhibit significant positive biases in simulating present-day SWE across snow-dominated regions of the Northern Hemisphere. As noted in the Introduction, these biases are not uniform but often substantial. For instance, Zhong et al.28 found that 22 CMIP6 models overestimated snow depth in high-latitude regions during 1955–2014, with a spatial mean bias exceeding observations by 18 ± 16 cm28. More critically for hydrological assessments, Kouki et al.29 reported that specific high-resolution models (e.g., CESM2, NorESM2-MM) overestimated SWE by 150–200% in northern North America and extensive parts of Eurasia29. This pattern of overestimation is corroborated by regional studies; Mudryk et al.3, Chai et al.27, and Zakeri et al.39 have all reported varying degrees of SWE overestimation in CMIP6 models3,27,39. These systematic positive biases are primarily attributed to inaccuracies in snow physics parameterizations, precipitation phase partitioning, and the surface energy balance within the models.

The implications of these systematic high biases for future projections are profound. A model that simulates too much snow in the historical baseline period inherently possesses a larger “snowpack capital” at the start of the future projection. Under a warming climate, even with realistic rates of decline, this inflated initial condition means the model’s future SWE will descend from an artificially high baseline. Consequently, the occurrence and intensity of conditions defined as “snow drought” (i.e., SWEI below specific negative thresholds) will appear less frequent and less severe than they would if the baseline were accurate. In essence, the use of uncorrected data mutes the signal of increasing drought risk by embedding present-day model errors into the future trajectory.

To quantify the impact of neglecting this bias correction, we systematically recalculated all future snow drought characteristics using the raw, uncorrected SWE outputs from the same 29 CMIP6 models under the SSP2-4.5 and SSP5-8.5 scenarios. A direct comparison with our CDF-t-corrected results reveals a consistent and significant underestimation of drought risk when using uncorrected data. For example, the multi-model ensemble mean frequency of total snow drought events (all categories) projected for 2030–2100 under SSP5-8.5 is 1.41 ± 0.49 events per year after bias correction, but only 1.15 ± 0.40 events per year using uncorrected data—an underestimation of approximately 18%. The discrepancy is even more pronounced for the most extreme events. The frequency of D4 (Exceptional) drought is 0.44 events/year in the corrected projection, compared to 0.31 events per year in the uncorrected projection—a ~30% underestimation. Similarly, the projected duration of D4 events is 86.7 ± 11.1 days post-correction but only 72.4 ± 9.8 days without correction. This represents a critical underestimation of the persistence of extreme conditions by about 17%. Our comparative analysis demonstrates that reliance on uncorrected model data would lead to an underestimation of the future risks of snow droughts.

Physical mechanisms behind the increase in snow drought frequency

To investigate the mechanisms behind the significant increase in the frequency of future snow droughts in the Northern Hemisphere, we combined direct analysis of key physical processes with quantitative attribution. We found that the reduction in snowfall is the dominant driver, while changes in snow ablation processes play a more complex modulating role.

First, analysis based on CMIP6 data shows that both the annual number of snowfall days and the total snowfall amount across most regions of the Northern Hemisphere exhibit significant and scenario-dependent declining trends in the future (Supplementary Fig. 11a–d). The multi-year average snowfall frequency is projected to decrease from 84.8 days under the sustainable SSP126 scenario to 78.0 days under the high-emission SSP585 scenario. Trend analysis reveals a statistically significant reduction in future snowfall frequency, with linear rates intensifying from −0.09 days yr−1 under SSP126 to −0.34 days yr−1 under SSP585. Similarly, the annual snowfall amount shows a coherent declining trend (Supplementary Fig. 11e–f), with reductions ranging from −0.14 mm yr−1 (SSP126) to −0.56 mm yr−1 (SSP585). This widespread decrease in both snowfall occurrence and accumulation directly quantifies the large-scale effect of climate warming on precipitation phase change (from snow to rain). This independent observational evidence strongly supports a key finding from the multivariate linear regression model: in the attribution analysis, the standardized regression coefficient for the change in snowfall fraction is the dominant contributor to the increase in snow drought frequency, with a contribution rate of 78.2%. Both the direct process-based diagnostics and the statistical attribution jointly confirm that reduced snowfall is the core driver of more frequent snow droughts.

The role of snow ablation is constrained by the already diminished snowpack. Although climate warming theoretically accelerates ablation, our direct process analysis reveals a key point: total snowmelt runoff across much of the Northern Hemisphere also shows a decreasing trend in the future. This finding is crucial for interpreting the regression results. It indicates that the “snowpack capital” available for melt has shrunk substantially due to reduced snowfall, leading to a decline in the absolute volume of melt runoff even if the melt rate per unit may increase. This quantitative relationship directly explains why, in the regression model, the contribution coefficient for the factor representing ablation energy is generally less important than snowfall fraction (with a contribution rate of only 15.2%). The independent contribution of accelerated ablation to snow drought is fundamentally limited by a precondition: the shrinkage of snowpack size caused by reduced snowfall. Therefore, changes in the surface energy balance are not the primary cause of intensified snow droughts, but rather a secondary process occurring within an already diminished system.

Spatial heterogeneity provides further validation of the proposed mechanism. A notable natural experiment is presented by eastern Siberia, where SWEI exhibits an exceptional increasing trend, deviating from the widespread Northern Hemisphere pattern. Crucially, this positive SWEI anomaly is accompanied by a concurrent increase in snowfall amount in the same region. This coherent spatial pattern offers strong corroborative evidence, as it demonstrates that an enhancement in solid precipitation input directly leads to improved snowpack conditions, thereby reinforcing snowfall change as the principal controlling factor. This is mainly attributed to enhanced moisture transport driven by atmospheric dynamic processes.

In summary, our study enhances projections of Northern Hemisphere snow droughts through the integration of two key developments: (1) the systematic application of the CDF-t bias-correction method to multi-model CMIP6 Snow Water Equivalent outputs, which substantially improves reliability, and (2) a combined analysis of future drought characteristics and their underlying physical drivers. The core findings are threefold. First, our bias-corrected projections reveal a fundamental structural shift in future snow drought regimes. While milder droughts may stabilize, the most extreme (D4) events are projected to become dramatically more frequent, intense, and prolonged, with trends strongly amplified under higher-emission scenarios (e.g., SSP585). Second, through combined process-based and attribution analysis, we identify the widespread decline in snowfall caused by warming-induced precipitation phase changes as the dominant mechanism for increasing drought frequency, which relegates changes in snowmelt processes to a secondary role constrained by the reduced snowpack. Third, this mechanism underscores that effective adaptation must address the fundamental challenge of diminishing winter solid precipitation. Overall, our work provides a more robust quantitative foundation for assessing snow drought risks. It highlights that mitigating the most severe impacts requires urgent emission reductions to preserve snowfall, while adaptation strategies for water resources, agriculture, and ecosystems must account for the projected intensification and changing nature of snow droughts.

Methods

Observational datasets and Earth system models

Observational datasets: As a benchmark reference in the CDF-t bias-correction process, the reliability of observational snow water equivalent data is critical for determining the overall effectiveness of the bias correction. However, long-term observational SWE datasets exhibit inherent uncertainties due to factors such as sparse in situ observations, simplified representations of snow physics, variations in retrieval algorithms, satellite data limitations, and uncertainties in land surface model forcing3,40,41,42. A recent comprehensive evaluation by Mudryk et al.43 assessed and ranked 23 gridded SWE products using a novel evaluation strategy with cross-validated reference data43. Their study identified ERA5-Land as the top-performing product across a range of tests. Therefore, we selected the ERA5-Land dataset as the observational reference for snow water equivalent during the 1982–2014 period.

Earth system models: The simulated snow water equivalent under four Shared Socioeconomic Pathways (SSP126, SSP245, SSP370, and SSP585) for the periods of 1982–2014 and 2030–2100 is collected from the 29 CMIP6 models (https://esgf-data.dkrz.de/projects/cmip6-dkrz/, Supplementary Table 1). All data is standardized to a 0.5° × 0.5° latitude–longitude grid using bilinear interpolation. Please refer to Supplementary Table 1 for each CMIP6 model.

CDF-t bias-correction method

The fundamental concept of CDF-t statistical downscaling is to introduce a transfer function that establishes a relationship between the CDF of large-scale variables and the CDF of observed variables. We define FSH as the CDF of monthly average SWE observations at a fixed location during the modeling phase, while FGH denotes the CDF of the large-scale atmospheric circulation model output for monthly average SWE, interpolated to that location for the same period44. Additionally, FSF and FGF represent the CDFs of observed and simulated monthly average SWE during the validation phase, respectively. We assume that during the validation phase, a transfer function T exists that allows the model-simulated SWE CDF to closely match the observed CDF, thereby facilitating effective downscaling44.

$$T\left({F}_{{GH}}\left(X\right)\right)={F}_{{SH}}\left(X\right)$$
(1)

Next, we define \(u={F}_{{GH}}\left(X\right)\). Consequently, we can express \(X={F}_{{GH}}^{-1}\left(u\right)\). The variable u is constrained within the range [0, 1]. By substituting X into Eq. (1), we can derive a simplified expression for the transfer function.

$$T\left(u\right)={F}_{{SH}}\left({F}_{{GH}}^{-1}\left(u\right)\right)$$
(2)

T represents the functional relationship between the CDF of the observed data during the modeling period and the CDF of the atmospheric circulation model. Assuming that the transfer function T established during the modeling period is also applicable in the validation period, we can apply this transfer function to the large-scale variables from the climate model to obtain regional downscaled results44, expressed as:

$$T\left({F}_{{GF}}\left(X\right)\right)={F}_{{SF}}\left(X\right)$$
(3)
$${F}_{{SF}}\left(X\right)={F}_{{SH}}\left({F}_{{GH}}^{-1}\left({F}_{{GF}}\left(X\right)\right)\right)$$
(4)

The significance of the transfer function is to establish the functional relationship between the CDF of historical observations and the CDF of the model for a specific value of X. This relationship is assumed to remain constant in the future, enabling the derivation of the downscaled CDF.

In this study, the historical period from 1982 to 2014, covering 33 years, is divided into two segments: the first 16 years (1982–1997) serve as the modeling period. The subsequent 17 years (1998–2014) constitute the validation period, which assesses the reliability of the statistical downscaling results.

Calculation of Standardized Snow Water Equivalent Index

SWEI is a metric used to assess the quantity of water contained within snowpack23. It serves as an important indicator in identifying areas affected by snow drought and understanding its changing patterns. Based on this index, it becomes possible to compare snow water content across different regions and time periods. This information is valuable for numerous applications, including hydrological modeling, water resource management, and forecasting potential water availability during snowmelt. By monitoring the SWEI over time, researchers and policymakers can gain insights into the spatial distribution and temporal variability of snow drought, helping them make informed decisions regarding water allocation, agricultural planning, and ecological management. The calculation of SWEI involves analyzing the snow water equivalent values for a specific location during a defined period, as represented in Eq. 524:

$${SWEI}={{{\varnothing }}}^{-1}\left[q\left({\sum}_{n=-2}^{0}{{SWE}}_{n}\right)\right]$$
(5)

In this equation, SWE represents the observed snow water equivalent value for the given location and time period. Regions with an annual cumulative SWE less than 10 mm were excluded from our analysis. Additionally, Greenland was also excluded from the analysis. \({{{\varnothing }}}^{-1}\) is the inverse standard normal distribution. \(q\) is the Gringorten plotting position, as represented in Eq. 623.

$$q=\frac{h-0.44}{m+0.12}$$
(6)

In the equation, h represents the rank of the nonzero variable, arranged in ascending order from smallest to largest, while m represents the sample size. The ranks are determined based on the 3-month integration of SWE values for each month, as represented in Eq. 7, where SWEn is the integrated SWE value for month n.

$${\sum}_{n=-2}^{0}{{SWE}}_{n}={{SWE}}_{n-2}+{{SWE}}_{n-1}+{{SWE}}_{n}$$
(7)

Positive SWEI values suggest above-average snow accumulation, while negative values indicate below-average snow accumulation. The magnitude of the SWEI value provides insights into the severity of snow drought conditions, with larger negative values indicating more significant deviations from the reference mean. By employing the SWEI calculation, researchers and stakeholders can identify and monitor snow drought conditions, aiding in the understanding of water availability, hydrological processes, and the management of water resources in snow-influenced regions. Following Huning et al.23, snow drought conditions are categorized into five types based on the SWEI values, with the specific classification criteria outlined in Supplementary Table 3.

To analyze the spatial features of snow drought, we divided the Northern Hemisphere into eleven sub-regions (Supplementary Fig. 1 and Supplementary Table 2) based directly on the established regional framework from Giorgi et al.45.

Quantitative attribution of snow drought frequency changes

To quantitatively attribute the projected increase in snow drought frequency to climatic drivers, we performed a multivariate linear regression analysis. The dependent variable was the projected change in snow drought frequency during the future period (2030–2100). The independent variables were selected to represent the primary physical processes governing snowpack mass balance, with a focus on isolating the distinct roles of snow input and energy-driven ablation. The regression model is formulated as follows:

$$\Delta Freq={\beta }_{0}+{\beta }_{p}\cdot \Delta P+{\beta }_{e}\cdot \Delta E+{\beta }_{sf}\cdot \Delta SF+\epsilon$$
(8)

where \(\Delta {\mbox{P}}\) is the change in total winter (NDJFMA) precipitation. \(\Delta E\) represents changes in the energy available for snowmelt, estimated by the change in surface sensible heat flux. \(\Delta {SF}\) is the change in the snowfall fraction, calculated as the ratio of snowfall to total precipitation.

where \({\beta }_{0}\) is the intercept, \({\beta }_{p}\), \({\beta }_{e}\), and \({\beta }_{{sf}}\) are the standardized partial regression coefficients, and \(\epsilon\) is the residual error. Standardization (subtracting the mean and dividing by the standard deviation) was applied to all variables prior to regression46. This allows the resulting standardized coefficients to be interpreted as indicators of the relative contribution of each driver; their magnitude reflects the change in snow drought frequency (in standard deviation units) associated with a one-standard-deviation change in the predictor, holding the others constant46.