Creeping snow drought threatens Canada’s water supply

Abstract

Snow water is key to water supply in cold regions and beyond. Here we introduce Snow Water Availability that quantifies water stored in the snow-covered portion of an area. By integrating a plausible set of available gridded datasets for snow depth, density, and cover fraction, we form four estimates of Snow Water Availability at the 25 × 25 km² across Canada and Alaska. We show that annual long-term mean of Snow Water Availability over the domain was 996 ± 170 km³ during 2000–2019. While annual Snow Water Availability increased from 799 ± 121 km³ in 2000-2009 to 1208 ± 231 km³ in 2010-2019, significant losses (p-value ≤ 0.05) were observed in ~3% of the domain, mainly in North American Cordillera, headwaters to major rivers in western Canada. These losses alongside insignificant decreases across southern Canada can threaten water supply in a quarter of the country, where ~86% of its population reside.

Introduction

Seasonal snow plays a critical role in forming local to regional climate^1,2,3, hydrology^4,5,6, and ecosystem functions^7,8, and supports diverse socio-economic activities^9,10. High reflectivity of snow cover reduces energy absorption, helping to moderate surface temperatures; and, its insulation properties protect underlying soil from extreme temperature fluctuations^11,12,13,14. Most importantly for socio-economic systems, snow acts as a natural water reserve, storing cold-season precipitation. In warmer months, when sufficient energy is available, accumulated snow melts and replenishes soil moisture, recharges groundwater, flows overland, and ultimately joins rivers that transport melted snow water to downstream^15,16,17. The recurring pattern of snow accumulation and melt is crucial to water resource management in cold regions and beyond. As the snowmelt generated streamflow passes through lower and drier lands, it becomes a vital source for food and energy productions in places with greater water demands^18,19,20,21.

Climate change, however, is increasingly disrupting snow-related processes^22,23,24. Having said that, variations in impacts of global warming on regional snow processes differ widely, as does the response of the regional hydrology and ecosystem to changes in snow processes^25,26. While many areas report gradual declines in snow cover and depth²⁷, some also experience record-breaking snow accumulation events²⁸. Nevertheless, there is a broad consensus that mountainous regions are among the most impacted areas globally^29,30,31,32. Rising temperatures in highlands have accelerated melt rates, shifted snowmelt timing, and increased the likelihood of rain in cold season^{25,33,34,35,36}, all contributing to a reduced snow accumulation by the start of the warm season, a phenomenon termed as snow drought^{34,37,38,39,40}. The cascading impacts of snow drought on natural systems are intricate⁴¹; and difficult to predict⁴², with new insights continually emerging^43,44,45. Various human systems are also impacted by snow drought ^{46,47,48,49,50}, with far-reaching consequences beyond snow-dominated regions^{9,20,21,51,52}.

To assess the outbreak of snow droughts and to advise adaptation strategies, monitoring snow variables is critical. The water resources community often uses Snow Water Equivalent (SWE), the average depth of snow water stored over an area, as a diagnostic measure to monitor the onset and severity of snow drought. However, monitoring SWE presents significant technological and conceptual challenges. Technically, in-situ data provide the most direct and presumably high-quality measurements of snow variables^53,54. While the number of emerging data archives for in-situ snow depth and SWE is growing, e.g.⁵⁵, there are major limitations in the spatial and temporal coverage of in-situ measurements due to their high cost and labor requirmenents^56,57. Furthermore, most in-situ observations are distributed at lower elevations and latitudes; therefore, highlands and northern regions with limited accessibility do not have sufficient monitoring stations.

Remote sensing offers a more expansive and global perspective on snowpack conditions^58,59,60, but challenges still remain⁶¹. Optical sensors, e.g., the Moderate Resolution Imaging Spectroradiometer (MODIS)⁶², Visible Infrared Imaging Radiometer Suite (VIIRS)⁶³, Landsat⁶⁴, and Sentinel-2⁶⁵ can detect snow cover effectively^66,67, yet cloud cover, vegetation, and polar nights obscure optical observations^68,69,70. Monitoring snow depth and SWE presents even greater complexity^71,72. Passive microwave sensors, e.g., the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E⁷³ and AMSR2⁷⁴) and the Special Sensor Microwave/Imager (SSM/I⁷⁵), are available; however, they have coarse spatial resolutions; and, their accuracy declines in complex topographies and areas with deep and/or wet snow^76,77.Active sensing through Synthetic Aperture Radar (SAR⁷⁸), e.g., C-band Sentinel-1⁷⁹ and L-band Advanced Land Observing Satellite-2 (ALOS-2⁸⁰), can significantly improve spatial resolution and can detect wet snow. Nonetheless, they present major limitations in dry snow and in both shallow and deep snowpacks⁸¹. Light Detection and Ranging (LiDAR⁸²) sensors, e.g., Ice, Cloud, and Elevation Satellites (ICESat and ICESat-2⁸³) can provide snow depth at high resolution globally; however, they present data gaps and coarse temporal resolutions. It is envisioned that some of the noted limitations can be addressed by combining both active and passive sensing across a range of frequencies and polarizations⁸¹. Even so, currently available remote sensing products cannot provide all snow variables with high accuracy, enough data length, and/or spatial and temporal resolutions needed for impact assessments and decision making applications⁸¹.

One way to bypass current limitations in observational data, whether directly measured or remotely sensed, is to complement them with model-based datasets. Climate reanalysis products, i.e., climate models that update their forecasts with observations using data assimilation techniques, have provided spatially continuous and long-term records of climate variables^84,85. The idea of reanalysis has been also extended to land models^86,87, including snow variables^88,89,90. Reanalysis products enable broad-scale assessments of trends and variability in snow across space and time. Having said that, snow reanalysis products, particularly if snow variables remain unassimilated, may exhibit large biases⁸⁸. This is particularly the case for snow depth and SWE^57,91,92, though snow cover and snow density tend to align more closely with observational data⁹³. As a result, some level of data integration for accurate snow monitoring is often necessary⁹⁴.

Beyond technical challenges, the impact community often uses SWE during spring-freshet to approximate meltwater availability⁹⁵. Yet focusing only on SWE at the end of the snow season ignores the dynamic of snow accumulation and melt during the snow season, including critical losses due to mid-season melts and/or rain-over-snow events, two phenomena that become more common in a warmer climate^96,97. Gridded SWE products also add another blind spot as they report grid-averaged values, ignoring the snow cover distribution within each grid. Past findings indicate that changes in snow cover area cannot be simply translated into changes in SWE or snow depth⁹⁸. This means that grid-averaged SWE may appear stable even though snow cover fraction shrinks or expands, which causes water stored in remaining patches inflates or drops, respectively. Given the sub-grid heterogeneity in hydrological response well-noted in the literature^99,100,101, accounting for snow water where the snow cover exists can better reveal how it contributes to water supply upon melting. This is particularly relevant to patchy snow covers, typical in complex terrains and/or during the onset and offset of the snow season.

To address this conceptual challenge, we propose Snow Water Availability (SWA), defined as the grid-averaged SWE normalized by snow cover fraction (see Eq. 1 in “Methods”). As a result, SWA reports the average snow water content over the snow-covered fraction of a grid. This scaling makes SWA more sensitive to patchy snow covers: By revealing how much water remains where snow exists, SWA offers a sharper diagnostic of when and where snow water availability can lead to extreme thresholds, both in terms of droughts and floods. As a result, SWA can be aligned better with impact assessments, flood and drought forecasting, and water management. The definition of SWA also lends itself to data fusion: for instance, multiple hypotheses for SWE data (or its components, snow depth and density) can be paired with satellite- and model-derived snow cover fraction to form an ensemble estimation of SWA over a region. This can lead to sensitivity analysis and quantification of uncertainty bounds in SWA estimates in light of the information available about constituting snow variables.

Using SWA, we track monthly, seasonal, and annual changes in water stored in snow-covered areas of the northern North America, i.e., Canada and Alaska. Our domain represents ~9% of the global land area and includes 25 drainage regions that release into three oceans (Fig. 1a). Snow water has underpinned water resources management, food production, and hydropower throughout this region^{102,103,104,105}. To compute SWA, we pair gridded snow depth data from the Canadian Meteorological Center’s snow depth reanalysis (CMC⁹⁰), a widely-used assimilated snow depth reanalysis data^{106,107,108,109}, with two alternative fields for snow density and snow cover fraction, yielding four distinct SWA estimates (see Methods). Snow density estimates are based on CMC’s SWE reanalysis¹¹⁰ and the fifth generation of the European Centre for Medium-Range Weather Forecasts (ECMWF) Land Reanalysis (ERA5-Land⁸⁷) products. Estimates of snow cover fractions come from ERA5-Land and MODIS. These gridded products are selected after a rigorous benchmarking against in-situ measurements and cross-product intercomparisons (see Methods and Supplementary Method in the Supplementary Information). Accordingly, we map hotspots of SWA changes, examine specific conditions in constituting snow variables that led to significant SWA changes, and relate these changes to various climatic drivers. We further contrast SWA with SWE, diagnose regions prone to snow drought, and explore the uncertainty of SWA estimates.

Fig. 1: Estimates of Snow Water Availability (SWA) over northern North America.

a The northern North America, compromising Canada and Alaska, and its 25 drainage regions¹¹¹, mapped over the 25 × 25 km² grid system of National Aeronautics and Space Administration’s (NASA’s) Making Earth System Data Records for Use in Research Environments (MEaSUREs) for the global landscape Freeze-Thaw Earth System Data Record (FT-ESDR¹⁶⁵). This discretization divides the total domain into 18,060 grid cells; b Categorizations of the area-averaged 25 × 25 km² elevation across 0–500 m, 500–1000 m, 1000–1500 m, and 1500–3000 m elevation bands. The elevation data is based on the 1 km Global Multi-resolution Terrain Elevation Data 2010 (GMTED2010¹⁶³); c Clustering the 25 drainage regions under the five major ocean-drained basins in Canada and Alaska; d Geographic extents of five SWA classes, based on annual long-term mean during 2000–2019. Five SWA classes considered include low (50 mm and below; dark blue), below-average (50–70 mm; light blue), average (70–100 mm; green), above-average (100–120 mm; pink) and high (120 mm and more; red). Each panel corresponds to one of the four estimates of SWA across Canada and Alaska. SWA1 and SWA2 have a cutoff at 64 °N due to missing MODIS data during polar nights.

Results and discussion

Analysis of SWA over the northern North America

Based on the four SWA estimates (see “Methods”), the long-term average of annual SWA over Canada and Alaska is estimated as 996 ± 170 km³ during 2000–2019 (i.e., mean of the four SWA estimates ± their standard deviation; see Fig. 1d). We also quantify long-term average of annual SWA across the 25 drainage regions, calculate the contributions of each region to the domain-wide SWA, and estimate the portion of SWA in the regional water availability based on the water yield data provided by Statistics Canada¹¹¹ (see Supplementary Table 1). To ensure the robustness in our estimations in each drainage region, we only consider those SWA estimates that have missing data in less than 50% of their respective area (see Supplementary Figs. 1a and 2a). Using these SWA estimates, we present spatial variability in annual long-term SWA (Supplementary Fig. 1b) and extract its expected monthly variations in each region (Supplementary Fig. 2b).

Five Pacific drainage regions together account for ~41% of the total annual SWA in the domain. While they collect the largest amount of SWA, they also include the smallest drainage region, Okanagan–Similkameen (DR3), providing just 0.12 ± 0.02% of the domain-wide SWA. This tiny fraction of SWA, however, supplies 27 ± 4% of the water availability in this region. Arctic drainage regions contribute ~38% of the domain-wide annual SWA, with Arctic Coast–Islands (DR8), the largest drainage region, receiving the highest regional SWA of 292 ± 87 km³, accounting for 29 ± 9% of the domain’s annual SWA during the study period. The nine regions draining into Hudson Bay account for $\sim$22% of the total SWA, including Assiniboine–Red (DR12), where snow water provides 43 ± 9% of the total regional water availability, the largest in the populous southern parts of the domain. Regions draining to the Atlantic Ocean receive slightly more than 10% of the total annual SWA, including Great Lakes (DR19), home to over 11 million as of 2011, intaking $\sim$70% of the Canada’s total water use¹¹¹.

Despite mainstream expectations under climate change, domain-wide annual average SWA increased from 799 ± 121 km³ in 2000–2009 to 1208 ± 231 km³ in 2010–2019, indicating an overall gain of 91 ± 44% during the 19-year period, implying a domain-wide Sen’s slope of 42 ± 14 km³/year. However, this total gain includes a significant spatial heterogeneity. Although 80 ± 3% of the domain show an increase in SWA, only 34 ± 4% of the area demonstrate a statistically significant increase (p-value ≤ 0.05), primarily in the eastern and northern parts of the domain (Fig. 2a). Regions with SWA loss are dispersed across the southern and western parts, where a small portion (3 ± 1% of the domain) shows significantly decreasing trends in annual SWA (p-value ≤ 0.05). This area is mainly located between elevations 800–2200 m particularly within North American Cordillera (Fig. 2b). Compared to gridded average annual SWA, i.e., 88 ± 15 mm, this region contains above average and high SWA (expected value of 104 ± 11 mm; see Fig. 1d), where SWA decreased sharply from 38 ± 5 km³ (about 5 ± 1% of total SWA in 2000–2001) to 11 ± 2 km³ (approximately 0.8 ± 0.2% of total SWA in 2018–2019), a major 71 ± 13% loss during the 19-year period. Examining trend categories of SWA, i.e., positive significant, positive insignificant, negative insignificant, and negative significant, along elevation gradients (Fig. 2b, right panels) reveals a consistent increase in the likelihood of significant SWA losses with elevation, reaching 14 ± 2% in elevations above 1500 m. Previous studies have also revealed the sensitivity of mountains’ snowpack to warming^112,113. In contrast, lowland areas (elevation between 0 and 500 m), concentrated in northern and eastern regions, consistently show significant SWA gains, reaching 40 ± 5% in likelihood.

Fig. 2: Changes in annual SWA in the northern North America during 2000-2019.

a Geographic extents of four annual trend classes, i.e., positive significant (dark blue), positive insignificant (light blue), negative insignificant (pink) and negative significant (red), based on the four estimates of SWA across Canada and Alaska; b Gradients of normalized annual trends across longitudes (Left column), latitudes (Second column from left) and elevation (first and second columns from right). Dot sizes and colors correspond to estimated annual long-term SWA and trend classes respectively. Right column summarizes the class frequencies of gridded annual trends of SWA across four elevation bands, i.e., 0–500 m, 500–1000 m, 1000–1500 m, and 1500–3000 m. SWA1 and SWA2 have a cutoff at 64^°N due to missing MODIS data during polar nights.

Similar to annual trends, SWA increased in majority of the grids during fall (October–December), winter (January–March), and spring (April–June), with a spatial distribution similar to that of the annual SWA trends (Supplementary Figs. 3–6). Nevertheless, decreasing SWA trends are still observed over 23 ± 3%, 20 ± 2%, and 16 ± 1% of the domain during fall, winter, and spring, respectively. While the extent of areas with decreasing SWA is relatively consistent across seasons and months, statistically significant SWA losses are more pronounced during winter, where 3 ± 0.2% of the domain shows a significant loss in SWA, compared to 2 ± 0.6% in fall and 2 ± 0.2% in spring. Significant winter SWA losses mainly occurred in snow-rich areas (expected SWA of 200–500 mm; Supplementary Figs. 7a and 8), where the likelihood of significant losses reached to 14 ± 3%. As with the annual pattern, the likelihood of significant seasonal and monthly losses consistently increases with elevation across all seasons and months, with the exception of October and June, reaching up to 17 ± 2% in elevations above 1500 m (Supplementary Figs. 7b and 9). This finding confirms the susceptibility of mountainous areas, particularly within North American Cordillera, which is also noted in the literature^30,34,112.

Hotspots of SWA change

Based on statistically significant positive and negative trends identified by the four SWA estimates, we delineate hotspots for SWA change across the domain (Supplementary Fig. 10, top row). Grids exhibiting significant increase in SWA (positive hotspots) show an elevational distribution that closely mirrors the domain-wide elevation profile, suggesting a relatively homogeneous distribution of SWA gains across the elevation bands. In contrast, grids with significant decreasing SWA (negative hotspots) exhibit a distinctly different elevational signature, with a pronounced peak consistently occurring between 1000 and 1700 meters (Supplementary Fig. 10, bottom row). This again suggests that mid-elevation mountainous zones are disproportionately affected by disappearing water reserve over their snowpack.

We accordingly attribute significant changes in SWA in hotspot zones to the corresponding changes in the constituting snow variables, i.e., snow depth, snow density and snow cover fraction (see Eq. 1). By clustering significant annual changes in SWA based on the associated trends in constituting variables, we extract joint probabilities for specific conditions in snow variables that led to significant changes in SWA (Fig. 3). We note that significant changes in SWA almost always correspond to changes in snow depth in the same direction, whereas similar associations with snow density and snow cover fraction cannot be taken for granted (Supplementary Fig. 11). The probability for alignments of trends in snow depth and SWA is 99 ± 0.4% in negative hotspots, and 99 ± 1% in positive hotspots. This finding determines snow depth as the key driver of significant changes in annual SWA. The likelihood of significant SWA trends occuring with the same direction of change in snow density drops to 74 ± 1% in negative and to 67 ± 13% in positive hotspots. The likelihood of concurrent changes in SWA and snow cover is 75 ± 2% in positive hotspots and only 51 ± 8% in negative hotspots. While the probability of alignment in the direction of change for SWA and snow depth decreases across monthly and seasonal timescales, particularly in negative hotspots (Supplementary Fig. 12), it stills prevails the impacts of snow density and cover fraction. However, there are exceptions during the onset and offset of the snow season, when significant negative trends in SWA are more frequently linked to decreasing snow cover fraction. This observation highlights the importance of sub-grid snow cover in determining water stored in the snowpack during transitions from and to snow-free season.

Fig. 3: Association between significant annual SWA trends with corresponding trends in snow depth (${T}_{{SD}}$), density (${T}_{{Rho}}$), and cover fraction (${T}_{{SCF}}$).

a correspondence in positive hotspots b  correspondence in negative hotspots. The measure reported in each box is the likelihood of significant trends in SWA coinciding with increasing (blue boxes) and decreasing (red boxes) trends in constituting snow variables based on the four SWA estimates (mean of the four SWA estimates ± their standard deviation).

Analyzing joint probabilities of positive and negative trends in constituting snow variables in hotspot areas reveals how compounding patterns of change in snow variables lead to significant SWA shifts. At the annual scale, the most frequent pattern in positive hotspots (Fig. 3a) involves simultaneous increases in snow variables, which is the case in 38 ± 7% of grids. The second common pattern includes increasing snow depth and density with decreasing snow cover fraction, accounting for 21 ± 11% of grids. The third, and comparably frequent pattern, includes increasing snow depth alongside decreasing snow density and cover fraction, the case for 18 ± 16% of the grids. The least frequent pattern involves increasing snow depth and cover fraction paired with a decrease in snow density, the case of 12 ± 7% of grids. These patterns are shifted in negative hotspots (Fig. 3b), where the dominant pattern is the concurrence of decreasing trend in snow variables, the condition led to significant SWA losses in 58 ± 6% of the grids. This is followed by a combination of decreasing snow depth and cover fraction with increasing snow density, the case for 17 ± 4% of grids within negative hotspots. A comparable pattern includes the coincidence of decreasing snow depth and density with increasing snow cover, which holds in 16 ± 7% of grids. The least frequent pattern in negative hotspots includes the decreasing snow depth with increasing snow density and snow cover fraction, the case of 9 ± 5% of grids over negative hotspots.

Climatic drivers of SWA change

To investigate the climate links of significant SWA changes across hotspot areas, we examined the dependency between the gridded annual SWA and twelve Large Scale Climate Indices (LSCIs) using Kendall’s tau rank dependence¹³ (p-value ≤ 0.05; see “Methods”). The selected LSCIs cover both oceanic-atmospheric teleconnections and radiative forcing indicators, and include the Pacific Decadal Oscillation (PDO¹¹⁴), Northern Oscillation Index (NOI¹¹⁵), Pacific North American pattern (PNA¹¹⁶), North Atlantic Oscillation (NAO¹¹⁷), North Pacific Pattern (NPP¹¹⁸), Western Hemisphere Warm Pool (WHWP¹¹⁹), Atlantic Multidecadal Oscillation (AMO¹²⁰), Solar Flux^121,122, Western Pacific Index (WPI¹¹⁵), Global Mean Surface Temperature (GMST¹²³), Pacific Warm Pool Area Average (PWPAA¹²⁴), and Arctic Sea Ice Area Average (ASIAA¹²⁵); see Data Availability. Our analysis, summarized in Fig. 4, reveals that 91 ± 1% and 94 ± 0.4% of the areas identified as positive and negative hotspots (blue and red regions in Supplementary Fig. 10) have a significant dependence with at least one LSCI. The six most influential LSCIs in positive hotspots are ASIAA, GMST, NAO, WHWP, PWPAA, and Solar Flux, affecting 15–58% of grids. For negative hotspots, the dominant drivers shift slightly to GMST, PWPAA, WHWP, ASIAA, PDO, and NOI, accounting for 25–80% of hotspot area. GMST, which proxies global warming, emerges as the most influential LSCI, showing significant dependence in ~50% and ~65% of grids across positive and negative hotspots. The correspondence of GMST to both increasing and decreasing SWA underscores complex and regionally specific impacts of warming, which can be sometimes counterintuitive to the mainstream expectation. For instance, significant SWA gains along Arctic coasts coincide with increasing GMST and declining ASIAA. As sea ice retreats, it expands open water areas, and amplifies surface-atmosphere heat and moisture exchange, which also enriched by higher moisture capacity in the warmer atmosphere over the Arctic Sea. When this moist air parcel penetrates into neighboring lands, it can enhance snowfall¹²⁵, leading to increasing SWA near costal regions.

Fig. 4: Correspondence between gridded annual SWA and Large Scale Climate Indices across positive and negative hotspots.

In all panels blue and red bars show the percentage of the areas within positive and negative hotspots that have significant dependence (p-value ≤ 0.05) with Large Scale Climate Indices (LSCI) – see Supplementary Fig. 10 for the location of hotspot areas. LSCIs considered include (from left to right in each panel), the Pacific Decadal Oscillation (PDO), Northern Oscillation Index (NOI), Pacific North American index (PNA), North Atlantic Oscillation (NAO), North Pacific Pattern (NPP), Western Hemisphere Warm Pool (WHWP), Atlantic Multidecadal Oscillation (AMO), Solar flux, Western Pacific Index (WPI), Global Mean Surface Temperature (GMST), Pacific Warm Pool Area Average (PWPAA), and Artic Sea Ice Area Average (ASIAA).

Comparison between SWE and SWA estimates

Gridded SWE quantifies area-averaged snow water over a grid cell regardless of its snow cover extent. SWA, however, represents water stored over the snow-covered portion of the grid. This enables understanding the impact of snow cover, and offers a more nuanced view of the available water stored in the snowpack when and where the snow cover fraction rapidly changes.

By definition (see Eq. 1), SWA is expected to exceed SWE if the grid is not fully covered by snow. The comparison of long-term means of SWE and SWA across annual, seasonal, and monthly scales confirms this (Supplementary Figs. 13 and 14). The spatial comparison is made using three classes: class 1, where long-term mean of SWA equals that of SWE; class 2, where SWE exceeds SWA; and class 3, where SWA exceeds SWE. Note that CMC’s SWE aligns with SWA1 and SWA3 (Supplementary Fig. 13), while a synthesized SWE, produced by combining CMC snow depth with ERA5-Land snow density, corresponds with SWA2 and SWA4 (Supplementary Fig. 14). Across all SWA estimates, class 3 (SWA > SWE) dominates during the core snow season prior its end in May and June. Grids where long-term means of SWE and SWA are equal (class 1) correspond to locations where both SWE and SWA are zero. In contrast, class 2 areas, where SWE exceeds SWA reflect conditions where zero snow cover fractions imply zero SWA, yet SWE may still report non-zero values. This results in a false signal of available snow water in snow-free locations, the case for up to 15% of the domain during May and June.

We similarly compare the differences in annual, seasonal, and monthly trends derived from SWE and SWA (Supplementary Figs. 15 and 16). The comparison is based on four trend classes, representing conditions in which SWE and SWA show (1) the same trend direction and significance level; (2) the same direction but differing significance levels; (3) opposite trend directions but the same significance level; and (4) both opposite directions and significance levels. Unlike the comparison made between long-term means, SWE and SWA trends tend to show strong agreement during the core snow season (December to February). However, at the onset and offset of the snow season, large spatial mismatches emerge. For example, in May, over 50% of snow-covered grids exhibit opposite trend directions between SWE and SWA. Even for annual trends that show agreements between SWE and SWA over most of the domain, the existing discrepancies can shift the frequency of trend classes across elevation bands and/or long-term means (Supplementary Fig. 17). These divergences reflect the importance of considering snow cover fraction when and where sub-gird heterogeneity in snow cover prevails. Handling patchy snow covers becomes increasingly important in a warming climate, when snow cover is more intermittent. SWA provides a new diagnostic measure for handling such conditions, which can provide in some cases contradictory results to those of SWE.

To understand the conditions in which trends in SWE and SWA diverge, we analyze corresponding joint probabilities of trends in constituting snow variables (Supplementary Fig. 18). In grids where SWE increases while SWA decreases (Supplementary Fig. 18a), three patterns account for ~90% of cases across the four SWA estimates: (1) coincidence of positive trends in snow depth, density, and cover; (2) increasing depth and cover, with decreasing density; and (3) increasing depth, with decreasing density and cover. Similarly, where SWE decreases while SWA increases (Supplementary Fig. 18b), three dominant patterns emerge: (1) losses in snow depth, density, and cover; (2) losses in snow depth and cover but gains in density; and (3) losses in snow depth with gains in density and cover. These patterns are also observed in positive and negative hotspots, respectively (Fig. 3), highlighting complex interactions between snow variables in constituting SWA trends.

Implications of SWA losses

While the geographic extent of significant SWA gains is much larger than significant SWA losses (Supplementary Fig. 10), we note that our estimates in northern areas may include a large and unquantifiable uncertainty. Although we choose CMC snow depth as the sole snow depth product due to its superiority to ERA5-Land (see Supplementary Method), we recognize that CMC snow depth remains largely unassimilated in northern latitudes⁹⁰. In addition, limitations in available in-situ stations, particularly above 64 °N (see Supplementary Method) prevents quantifying both accuracy and uncertainty of CMC’s snow depth product, where most significant SWA gains took place. In contrast negative hotspots are located in relatively data-rich areas (see Supplementary Method), where our benchmarking results can provide a notion for the accuracy and uncertainty of gridded snow products for constituting snow variables. In addition, negative hotspots are concentrated mainly in headwaters of six drainage regions that support major water demands downstream. The existence of significant SWA losses in headwaters are relevant to the analysis of snow drought; and become more concerning by their alignments with insignificant losses in downstream and nearby areas (Supplementary Fig. 19), some with rates of losses similar to those in hotspots (Supplementary Fig. 20).

Historical evidence shows that compounding impacts of significant snow droughts in headwaters with slightly below- or near-normal snowpack downstream can be major. The 2015 snow drought in western Canada stands as an example with cascading socio-economic and ecological impacts that extended well beyond mountainous regions. During winter of 2015, snowpack was observed very low across Rocky Mountains, with some basins, e.g., Okanagan and Lower Fraser, registering near zero SWE¹²⁶. This caused significant decline in streamflow, stressing municipal water supplies in a way that required emergency actions¹²⁷. The ecological impacts were equally severe. The lack of meltwater led to stream temperatures rising beyond thermal thresholds for aquatic life, resulting in mass die-offs of sockeye salmon^128,129. Compounded by a warm spring and dry summer, the drought reached to the northern Peace River (DR6) and Saskatchewan River system (DR10, DR11, DR14), reducing inflows to Lake Diefenbaker and threatening downstream agricultural and hydroelectric productions¹³⁰. Although managed, the 2015 snow drought exemplifies the far-reaching vulnerabilities of western Canada to snowpack declines.

We also note that majority of highly populated areas in western and southern Ontario and Quebec are experiencing SWA losses, although these changes are not statistically significant (Supplementary Figs. 19). There are historical examples that these areas are also vulnerable to snow drought. One example during the study period is the snow drought unfolded over southern Ontario and Quebec during the winter of 2011–2012¹³¹, when late winter snowpack in Great Lakes basin, southern Ontario, and the St Lawrence corridor were low¹³² due to a mix of atmospheric processes that kept the Atlantic and Arctic air away, bringing warm advection from the Gulf of Mexico¹³³. With little meltwater to replenish the lakes, water levels on Lakes Superior, Michigan–Huron, and Erie slumped toward historic lows, triggering emergency dredging of harbours and lightening cargo¹³⁴. Downstream water levels near Montreal ran also well below average, requiring major water release from upstream reservoirs to fend off navigation constraints¹³⁵. Snow drought of 2011-2012 shows how in places with insignificant SWA losses, a single warm and low-snow winter propagates through the whole water system and can even affect an international shipping lane.

Droughts are creeping phenomena^136,137,138, so are snow droughts. Given the context of basin-scale water management, the snow drought in one location can extend to other locations through streamflow. As a result, there is a need to put the SWA trends in perspective across each drainage basin (see Fig. 5 and Supplementary Fig. 21a, b). Depending on SWA estimate, 2–6 drainage regions experiencing SWA loss due to having a negative expected gridded trend. To further assess the potential vulnerabilities, we classify the 25 drainage regions into three categories of highly vulnerable, vulnerable, and at risk, depending on the nature of the expected basin-wide SWA loss. Highly vulnerable regions are those in which all SWA estimates consistently indicate an expected decreasing trend across all time scales. Vulnerable regions are those that show a decreasing basin-wide trend by at least two SWA estimates, either at the annual scale, winter, spring, or during March, when SWA historically peaks in most drainage regions (see Supplementary Table 1 and Supplementary Fig. 2). At risk regions are those in which SWA decreases based on at least one estimate, during at least one timescale.

Fig. 5: Characteristics of annual changes in SWA across 25 drainage regions in Canada and Alaska based on four estimates of SWA.

(top row) normalized gridded trends and their expected values across 25 drainage regions, shown by boxplots and solid lines, respectively; (second row from the top) expected changes in SWA during the study period (solid black lines in each panel quantified by the left y axis), as well as percentages of trend categories (colored bars in each panel quantified by the right y axis) in each drainage region; (third row from the top) changes in gridded long-term SWA per unit of change in elevation. Colored bars denote drainage regions in which long-term SWA and elevation have a significant dependence; (bottom row) changes in gridded trends in SWA per unit of change in elevation. Colored bars denote drainage regions in which gridded trends in SWA and elevation have a significant dependence.

Based on this classification, Okanagan–Similkameen (DR3) and Assiniboine–Red (DR12) are identified as highly vulnerable regions to SWA loss. Regional water availability in both regions is heavily dominated by spring snowmelt (Supplementary Table 1). During the study period, nearly 50% of the area of Okanagan–Similkameen, Canada’s third-highest population density¹¹¹, experienced a significant loss in SWA (see Fig. 5 and Supplementary Fig. 21a, b), leading to total basin-wide SWA loss of −1.2 ± 0.2 km³ during the 19-year period. This rate of loss is slightly lower than basin-wide expected annual SWA and marks the steepest SWA loss across the domain. Given the fact that Okanagan–Similkameen has the second lowest water yield in Canada¹¹¹, the continuation of the SWA loss can be devastating for its agricultural activity, which includes third highest irrigation demand in Canada¹¹¹. The Assiniboine–Red basin is the other highly vulnerable region, experiencing a SWA loss of −0.9 ± 0.2 km³. This loss corresponds mainly with insignificant negative trends in SWA over nearly 70% of the region’s area, while only around 5% of the region undergo significant SWA losses (see Fig. 5 and Supplementary Fig. 21a, b). Given extensive agricultural activity (73.2% of its total area as of 2011¹¹¹), and  being Canada’s second-lowest water yield per area¹¹¹, the continuation of SWA losses severely impacts food production in this region.

Five additional regions, i.e., North, South and Lower Saskatchewan–Nelson (DR10, DR11 and DR14, respectively), Great Lakes (DR19) and Saint John–St. Croix (DR23) is categorized as vulnerable region. North Saskatchewan (DR10) shows decreasing trends in accumulated snow water in spring based on SWA1 and SWA3. The situation is much more severe in the case of South Saskatchewan (DR11), in which all estimates show spring SWA losses. More downstream and across Lower Saskatchewan–Nelson, three estimates show SWA loss in March. Decreasing SWA in the North, South, and Lower Saskatchewan has major regional water supply implications. The three regions together cover an area of 688,661 km² (7% larger than France, the largest EU country) that home nearly 4 million people, and 5500 km² irrigated lands with the demand of 1.3 km³/year as of 2014¹¹¹. Decreasing spring SWA in South Saskatchewan can be particularly pressing on hydropower production, manufacturing, mining, and the expansion of irrigated agriculture, which is envisioned in the province of Saskatchewan¹³⁹. Increasing irrigation without enough water supply can lead to major socioeconomic consequences and endangers ecosystem integrity of Saskatchewan River Delta, the largest inland delta in North America, and home to Indigenous communities and precious ecosystems¹³⁹. In addition, reduced streamflow as a result of SWA loss can disrupt water diversion to Assiniboine–Red to manage water quantity and quality in Qu’Appelle River^{140,141,142,143}. More importantly, there is a Master Agreement on Apportionment between governments of Alberta, Saskatchewan, Manitoba, and Canada¹⁴⁴. It has been shown that streamflow losses at the headwaters, particularly when combined with earlier peak timing, can lead to failure in maintaining the Master Agreement without adaptations¹⁸.

Great Lakes (DR19) is another vulnerable region to winter SWA loss, particularly in March, when all estimates suggest a basin-wide loss. Great Lakes basin supports major freshwater intake, exceeding 23 km³/year as of 2014¹¹¹. Decreasing SWA in March can severely affect flow peak in Great Lakes occurring in April, with major consequences for shipping industry during snow-free months^134,135. Saint John–St. Croix (DR23) is also classified as a vulnerable region, as two estimates suggest decreasing annual and winter SWA. Based on SWA2 and SWA4, this relatively small region ($\sim$42,000 km²) has experienced more than 73 million m³ of SWA loss during the study period, exceeding one third of its surface water intake¹¹¹.

Seven regions are identified at risk. The most highlighted region is Fraser–Lower Mainland (DR2), Canada’s third-largest population center and home to the city of Vancouver¹¹¹, showing SWA loss in the fall season (all estimates) and months (two estimates in October and November and all estimates in December) as well as January (two estimates), and spring (one estimate), particularly in May (all estimates). Columbia basin (DR4), another drainage region in British Columbia, shows SWA loss during spring (SWA3) as well as in months October and May (SWA4). Peace–Athabasca (DR6) shows SWA losses in spring (SWA3), October (all estimates), and April (SWA1). The Missouri region (DR9), a small and heavily agricultural region in southern Alberta and Saskatchewan, shows spring SWA loss based on one estimate (SWA2). Winnipeg (DR13) shows SWA loss in fall (SWA4), particularly in December (SWA1 and SWA3). Ottawa (DR20) also shows SWA loss in fall (SWA1 and SWA2), particularly in December (all estimates). This region includes Canada’s capital city and is home to nearly 2 million as of 2011¹¹¹. St. Lawrence (DR21), home to Montreal and Quebec city, and Canada’s densest population zone with the second highest surface water intake¹¹¹, consistently shows SWA loss in November across all estimates.

We also examine the changes in SWA and its trends across elevation gradient in annual (Fig. 5) and seasonal scales (Supplementary Fig. 21c, d), investigating whether elevation influences expected values and trends of SWA in each drainage region. Our analysis reveals that SWA tends to increase with elevation in the majority of regions. There is a clear regional variation in the manifestation of SWA loss across elevation gradients. All SWA estimates indicate that in the Pacific regions SWA decreases more rapidly at higher elevations, while in Atlantic regions, the decreasing trend is steeper in lowland areas. Arctic and Hudson Bay regions exhibit a mixed pattern, with the direction and magnitude of SWA change varying by the drainage region and SWA estimate.

Quantifying uncertainty in SWA estimates

Our 4-member ensemble of gridded SWA estimates provides an opportunity for accounting for uncertainty in SWA due to the choice of datasets for snow cover and density (see Table 1 in Methods) that are only marginally different from one another. To quantify this uncertainty, we calculate the dispersion of the gridded long-term mean and trend in SWA, relative to their expected values based on the four estimates. In the annual scale (Fig. 6a), relative uncertainties can be substantial, particularly for the estimated trends, where the difference between maximum and minimum estimates can exceed six times the expected values. We note that such extreme cases of volatility in SWA estimates mainly occur in grids where trends and/or long-term means are near-zero; and therefore, relative uncertainty estimates may inflate. Nonetheless, at the regional scale, expected values of dispersion do not exceed the expected basin-wide values. Monthly and seasonal SWA estimates are less sensitive to the choice of dataset chosen for snow density and cover fraction in winter; however, relative uncertainty increases in fall and spring, particularly at the onset and offset of the snow season (Supplementary Fig. 22). While relative uncertainties are notably higher for trends in October and November, larger uncertainties for the long-term mean are observed in May and June.

Fig. 6: Uncertainties in annual estimates of long-term mean and trend of SWA and their spatial dependence structure.

a relative uncertainties in gridded (boxplots) and basin-wide (solid red line) estimations of the long-term mean (top panel) and the trend (bottom panel) of SWA across 25 drainage regions in Canada and Alaska; spatial structure of relative uncertainties in estimations of the gridded b long-term mean and c trend of SWA, quantified through dependencies with longitude, latitude and elevation. Drainage regions, where significant dependencies exist between SWA estimates and spatial elements (p-value ≤ 0.05), are colored in both panels. Color saturations correspond with the absolute value of Kendall’s tau. Drainage regions with negative dependencies between relative uncertainty and spatial elements are marked with asterisks (^*). Drainage regions where dependencies are not significant are sliced out and remained uncolored.

We further examine how uncertainty in annual long-term mean and trend varies across longitudinal, latitudinal, and altitudinal gradients (Fig. 6b, c). In the Pacific region, uncertainty in the long-term mean significantly increases eastward (p-value ≤ 0.05). In contrast, the Arctic, Missouri, and Hudson Bay regions, except Lower Mackenzie and Assiniboine–Red, show significantly higher uncertainties in the western parts. This pattern is reversed again in the Atlantic region, where uncertainty increases eastward, except in North Shore–Gaspé and Ottawa. Dispersion in the long-term mean reduces at higher latitudes within the Pacific Coastal, Fraser–Lower Mainland, along with Arctic and Atlantic regions, and most of the Hudson Bay. However, exceptions include Assiniboine–Red and Northern Ontario, where, similar to Columbia and Yukon, uncertainty increases at higher latitude. In Missouri and Keewatin–Southern, the changes in uncertainty with respect to latitude are statistically insignificant (Fig. 6b, middle pie chart). Relative uncertainty in the long-term mean generally increases at higher elevations in Pacific, Arctic, Missouri, and Hudson Bay, except in Okanagan–Similkameen and Winnipeg, as well as Northern Ontario and Northern Quebec. In contrast, in Atlantic regions, the uncertainty in long-term mean consistently reduces at higher elevations. In Assiniboine–Red, uncertainty in annual long-term mean does not show a significant dependency with elevation, given its flat terrain with negligible elevation gradient across the region.

Spatial dependencies are weaker and less consistent for uncertainty in trend estimates, with several cases showing insignificant associations. Specifically, dependencies between uncertainty in annual trend estimates and spatial characteristics are insignificant in 9 out of 25 regions. In the Pacific region, uncertainty in trend estimates increases eastward in Fraser–Lower Mainland but westward in Yukon. A similar divergent pattern appears in the Lower Mackenzie and Arctic Coast–Islands, where uncertainty in trend increases eastward and westward, respectively. In the Hudson Bay region, uncertainty in annual trend generally increases westward, with the exception of Assiniboine–Red and Winnipeg. Uncertainty in annual trend generally increases northward across Pacific basins. The pattern is reversed in Arctic regions, where uncertainty in trend estimates significantly decreases northward. This pattern remains dominant in Hudson Bay, except in Lower Saskatchewan–Nelson and Northern Quebec, where uncertainty in annual trend estimates increases northward. In Atlantic regions, a mixed pattern is seen where uncertainty in annual trend increases northward in Great Lakes but southward in St. Lawrence. The dispersion in trend estimates generally increases with elevation, with the exception of Yukon, Assiniboine–Red, Lower Saskatchewan–Nelson, and Northern Quebec.

Conclusion

By introducing SWA, we provide a four-member ensemble for quantifying snow water supply from 2000 to 2019 across Canada and Alaska. This ensemble estimate enables us to quantify trends and long-term means of snow water availability, and their associated uncertainties at the grid scale and across the 25 drainage regions in annual, seasonal and monthly timescales. While SWA has increased over the domain, we note both significant and insignificant SWA losses mainly in the western and southern parts of Canada. These SWA reductions can stress 14 out of 25 drainage regions, from highly-vulnerable to vulnerable to at-risk, and jeopardize water management in nearly 2.5 million km², 25% of Canada’s total area, equal to ~60% of the EU. This can engage ~86% of Canadian population and important ecosystems, e.g., Saskatchewan River Delta. Most notably Okanagan–Similkameen, Assiniboine–Red, and Saskatchewan River system (North and South Saskatchewan along Lower Saskatchewan–Nelson) are most impacted by SWA loss. Except Okanagan–Similkameen, where around half of the basin area undergoes significant SWA losses, other regions are mainly impacted by insignificant SWA losses across a large proportion of their land. This shows that compound impacts of insignificant SWA losses over a large area can lead to major consequences, which highlighs the creeping nature of snow droughts in these regions.

We also note the role of LSCIs in shaping the interannual variability of SWA in hotspot regions and identified GMST as a key driver of interannual change in SWA, corresponding with ~50% and ~65% of grids with significant SWA gains and losses. However, critical questions remain regarding the spatial heterogeneity in impacts of LSCIs and their interactions, before a robust causal attribution between LSCIs and SWA trends can be established. Another point is the fact that we only used CMC’s assimilated product as plausible data support for snow depth. This decision was supported by CMC’s strict superiority to ERA5-Land, which was recently cited as the top performing gridded SWE data over the Northern Hemisphere¹⁴⁵. CMC’s superiority was revealed through a benchmarking exercise in 1000+ in-situ stations that are mainly located in the areas with SWA losses (see Supplementary Method). Having said that, we do recognize that other gridded products of snow depth or SWE, particularly those that fuse multiple products, e.g.¹⁴⁶, are available and can be used as basis for SWA calculation. Examining all available data supports is indeed beyond the scope of this study; however, our analysis using two hypotheses for snow density and snow cover fraction underscored a heightened uncertainty in SWA estimates during the onset and offset of the snow season, where calculating SWA actually matters. This highlights the need for improving the quality of gridded snow products, particularly during transitional periods. This goal cannot be achieved without investing on in-situ and remote sensing observations, as well as improving our modeling and assimilation capabilities.

As a final remark, we should emphasize that SWA does not compete with SWE, but offers a complementary perspective, valuable during periods and in regions that snow cover evolves rapidly. In addition, while our SWA estimates are presented at a spatial resolution of 25 × 25 km² to maintain consistency across the four estimates, the methodology allows for downscaling using higher-resolution inputs. Specifically given the datasets used here, ERA5-Land snow density and snow cover fraction (0.1° × 0.1°), and/or MODIS-derived snow cover fraction (5 × 5 km²), can be used to refine SWE into finer-scale SWA estimates. This represents a promising direction, though it should be first validated through a benchmarking study.

We hope our findings can contribute to a deeper understanding of changing snow water supply in Canada, the Great White North, and can inform policy makers and the public about the state of Canadian freshwaters, our most precious natural resources in the era of heightened climatic and anthropogenic changes.

Methods

Rationale

SWA is defined as the water available within the sub-grid snowpack and can be calculated as:

$${{\rm{SWA}}}=\left\{\begin{array}{cc}0 & {r}_{s}=0\\ \frac{{{\rm{SWE}}}}{{r}_{s}}=\frac{\bar{{h}_{s}}}{{r}_{s}}\times \frac{\bar{{\rho }_{s}^{* }}}{{p}_{w}} & {r}_{s} > 0\end{array}\right.$$

(1)

Where SWE is the gridded snow water equivalent, $\bar{{h}_{s}}$ is the grid-averaged snow depth, ${r}_{s}$ is the fraction of the grid cell covered by snow, $\bar{{\rho }_{s}^{* }}$ is the estimated area-averaged snow density and ${p}_{w}$ is the density of water.

SWA recognizes that snow water is only available over snow covered areas of the grid, a subtle yet a fundamental requirement for the existence of snow water. SWA, therefore, can better represent sub-grid heterogeneity in the snow cover and water availability. Figure 7 shows how by blending SWE (or its constituents, snow depth and snow density; see Eq. 1) with snow cover fraction, SWA can quantify snow water availability over sub-grid snow cover patches (see also Supplementary Fig. 23 for an example). Given the widely-noted heterogeneity in hydrologic response^99,100,101, accounting for snow cover distribution is critical for understanding the fate of meltwater, how it is partitioned into various hydrologic processes (e.g., surface runoff vs. infiltration), and ultimately contributes to streamflow. This gives SWA an advantage for water resource management application.

Fig. 7: From Snow Water Equivalent (SWE) to Snow Water Availability (SWA).

a SWE as the area-averaged depth of snow water over a grid; b identification of sub-grid snow patches and estimation of snow cover fraction (${r}_{s}$); c scaling SWE using ${r}_{s}$ to SWA over snow patches.

Data support

We choose our data support for estimating SWA based on a rigorous benchmarking study (see Supplementary Method). In a nutshell, our benchmarking looks at (i) how gridded data are able to reproduce the empirical dependencies of snow variables on temperature, observed in 1000+ in-situ stations, (ii) how the timeseries of gridded data match those of observed in-situ data, and (iii) how the estimates of long-term mean, variability and trend obtained from competing gridded data differ from one another in both space (drainage regions) and time (monthly, seasonal and annual scales). The in-situ data used are obtained from CanSWEv.7 dataset⁵⁵ (see Data Availability). We used temperature data from Multi-Source Weather reanalysis data (MSWX¹⁴⁷; see “Data Availability”) that were not used for producing and/or post-processing any of the data supports used in our study. We consider three competing hypotheses for SWE, i.e., CMC, ERA5-Land, and AMSR (merged AMSR-E/AMSR2), two competing hypotheses for snow depth and density, i.e., CMC and ERA5-Land, and two competing hypotheses for snow cover fraction, i.e., MODIS and ERA5-Land—see Data Availability. Our benchmarking study revealed that temperature sensitivities witnessed in in-situ stations can be reconstructed to a great extent by gridded data, particularly by CMC. Having said that, ERA5-Land’s snow depth can be falsified in favour of CMC as (1) CMC provides a superior performance compared to ERA5-Land in reproducing the in-situ observations of both snow depth and SWE; and (2) there are systematic differences between estimates of CMC and ERA5-Land across drainage regions. We consider both CMC and ERA5-Land as plausible data supports for snow density as they provide comparable performance in reproducing the in-situ observations, and, no systematic difference between estimates across different drainage regions and/or temporal scales can be observed. For the same reason, we consider both MODIS and ERA5-Land as plausible data supports for snow cover fraction as there are only marginal differences between the two products in terms of reproducing the long-term mean and trend estimates of snow cover fraction. For more details, see Supplementary Method in Supplementary Information file. Integrating these data sources yields four SWA estimates (see Table 1). In the lack of any posterior information, we assign equal epistemic weight to these four SWA estimates, reflecting their empirical comparability and therefore the ability to quantify uncertainty in estimations through mean ± standard deviation of the four estimates. Table 1 introduces these SWA estimates and the selected data support for constituting snow variables.

Table 1 Four estimates of Snow Water Availability (SWA) and their constituting data support

CMC’s snow depth, our only data support for ${\bar{h}}_{s}$, is widely recognized for its accuracy over North America and has been a frequently-used reference data in snow assessments and benchmarking studies^{106,107,108,109}. CMC provides daily snow depth analysis and monthly SWE estimates from 1998 to 2020 across the Northern Hemisphere at a 24 km grid resolution¹⁴⁸ and is publicly available via NASA’s Snow and Ice Data Center (NSIDC; see Data availability). The CMC snow depth product assimilates in-situ snow depth observations from the World Meteorological Organization (WMO) into simulated snow depth from a conceptual snowpack model. The model simulates SWE using the mass balance and is forced with analyzed temperature and forecast precipitation fields of the Canadian meteorological forecast system^90,105. Within this snow model, fresh snow density is calculated empirically based on the air temperature¹⁴⁹, which evolves according to different aging processes for cold¹⁵⁰ and warm snowpacks¹⁵¹. The model estimates snowmelt using a variable degree-day melt factor, adjusted for snow density and vegetation cover¹⁵². Initial simulated snow depth, named “first-guess”, is subsequently corrected by assimilating in-situ observations through an Optimum Interpolation scheme⁹⁰. While comparison with observed data shows a good accuracy of CMC data across Northern Hemisphere, it has been noted by the developers that in certain regions snow depth is largely overestimated¹⁴⁸. These regions are masked out across the study domain.

CMC does not compute monthly SWE dynamically but rather calculate it post-hoc by multiplying the assimilated monthly-mean snow depth by a prescribed monthly snow density¹⁵³. As CMC does not readily provide the data of snow density, we extracted monthly area-averaged snow density in each grid cell by dividing the monthly area-averaged SWE to the monthly area-averaged snow depth:

$$\bar{{\rho }_{s,j}^{* }}=\frac{{\overline{{{\rm{SWE}}}}}_{j}}{\left(\mathop{\sum }_{i\in I(j)}\bar{{h}_{s,i}}\right)/\left|I(j)\right|}$$

(2)

where $\bar{{\rho }_{s,j}^{* }}$ is the estimated area-averaged snow density in month j; ${\overline{{SWE}}}_{j}$ is the CMC’s estimated area-averaged SWE in month j; i indexes the days in month j, i.e., $I(j)$; $\bar{{h}_{s,i}}$ is the CMC’s area-averaged assimilated snow depth on the day i of month j and $\left|I(j)\right|$ is the total number of days in month j. Snow density cannot be estimated during July, August, and September as SWE is missing.

ERA5-Land⁸⁷ is an unassimilated land reanalysis data product. It uses spatially disaggregated climatic fields of ERA5¹⁵⁴ to force a higher resolution land-hydrology model^155,156, offering hourly unassimilated simulations of land variables at a 0.1° resolution from 1950 to present. Similar to CMC, ERA5-Land calculates the SWE using the mass-balance, where snowmelt and sublimation are determined through a energy balance scheme. Snow density evolution incorporates metamorphism, compaction, and melt¹⁵⁷. Knowing SWE and snow density, snow depth can be calculated, from which snow cover fraction is estimated:

$${r}_{s}=\min \left(1,\frac{\bar{{h}_{s}}}{0.1}\right)$$

(3)

ERA5-Land’s snow density and snow cover fields are publicly available and are obtained using the ECMWF Climate Data Store API – see Data availability.

MODIS global monthly Level-3 (L3) snow cover product provide a monthly composite data by averaging daily observations from the MODIS/Terra Daily L3 (MOD10C1) product at a 5 km resolution from March 2000 onward¹⁵⁸. The data is available at NASA’s NSIDC—see Data availability. MODIS determines snow cover using the Normalized Difference Snow Index (NDSI⁶⁶), from Terra satellite observations^159,160. Clouds are resolved using the MODIS Cloud Mask Product (MOD35¹⁶¹), which assesses the presence of clouds using a combination of visible, infrared, and near-infrared bands¹⁶². To upscale from daily to monthly estimates, grids with low snow cover (less than 10%) or high cloud cover are eliminated. If there is no day within a month that has upscaled snow cover, then the monthly averaging is not done, and the data is tagged as missing.

We monitor the changes of SWA estimates across elevation using the Global Multi-resolution Terrain Elevation Data 2010 (GMTED2010¹⁶³), a digital elevation model produced jointly by the U.S. Geological Survey and the National Geospatial-Intelligence Agency. Here we consider mean elevation data at 1 km resolution. The data is publicly available—see “Data availability”.

Data harmonization and SWA assessments

To estimate SWA, we integrate the CMC’s snow depth with two sources for snow density data, i.e., CMC and ERA5-Land, and two sources for snow cover fraction, i.e., ERA5-Land and MODIS. All datasets are harmonized to a common spatiotemporal scale to allow consistency between SWA estimates. We focus on the common data period spanned from 1 March 2000 to 31 December 2019 and clip the individual data over Canada and Alaska using shapefiles available at Global Administrative Areas (GADM)—see Data availability. We temporally upscale the ERA5-Land’s hourly snow cover and density data, as well as CMC’s daily snow depth data to monthly by simple averaging. We then spatially upscale ERA5-Land’s temporally aggregated snow cover and density data, along with MODIS’s snow cover, and GMTED2010’s elevation data to $\sim$25 km grid cells by weighted averaging. The weight for each grid is assigned based on the percentage of the grid area that is located inside the larger$\, \sim$25 km grids. Any missing value in smaller grids results in a missing value for the larger grid cell. To precisely align the grids in individual data before calculating SWA, we re-grid all aggregated data using the k-nearest neighbor interpolation¹⁶⁴ to the 25 km grid of NASA’s Making Earth System Data Records for Use in Research Environments (MEaSUREs) for the global landscape FT Earth System Data Record (FT-ESDR¹⁶⁵).

Four SWA estimates are calculated at the monthly scale during the study period (see Supplementary Fig. 23 for an example) and then upscaled from monthly to seasonal and annual scales. Seasonal averages are defined for fall (October to December), winter (January to March), and spring (April to June). If the data in one of the months is missing, the seasonal data is tagged as missing too. Annual averaging is done during October to June, as CMC’s snow density from July through September is missing. If a SWA estimate is missing in one of the seasons in a given grid cell, it is also considered missing at the annual scale as well.

For each grid cell, we calculate long-term means and trends across 13 timescales (9 monthly, 3 seasonal, 1 annual). At each timescale, we only consider those grids for which the SWA estimates have no missing data. Trend is quantified using the Sen’s slope, and the significance of the trend is quantified using the Mann–Kendall’s trend test¹⁶⁶. The trend is identified as significant when p-value$\,\le 0.05$. Normalized gridded SWA trends are calculated by diving the gridded Sen’s slope by gridded long-term mean. The uncertainty in long-term mean and trend of SWA at each grid is quantified by the ratio of difference between maximum and minimum estimates divided to the expected SWA, providing a non-parametric measure of spread. Note that for long-term mean, the uncertainty estimate is always positive; however, the uncertainty estimate for trend can have negative values when average trend estimate at a grid is negative. We use Kendall’s tau, a non-parametric dependence measure^13,164,167, with significance threshold of p-value$\,\le 0.05$ to explore the dependencies between snow variables and temperature, to understand the link between SWA and LSCIs, and to investigate spatial structures of trends, long-term means and their uncertainties.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.