Our field surveys are designed to sample hypothesized endmembers of variability that are missing from the regional in-situ network: deep vs. shallow snow, low vs. high elevations, forested vs. open areas, different slope aspects, and unique types of snow such as avalanche debris. This is similar to the approach of Broxton et al. (2019b), who also used snow pits to constrain anticipated endmembers of variability. Unlike many snow field surveys, we prioritize snow density instead of depth measurements since high resolution depth data are available from contemporaneous airborne lidar surveys. Additionally, we only collect snow pit measurements instead of using a SWE coring tube (i.e., Federal sampler) because deep, high-density snow with thick ice layers is not conducive to inserting or retrieving a SWE coring tube by hand (Federal Wilderness regulations prohibit power tools), and in shallower snow, the rocky alpine ground is not conducive to obtaining a dirt plug to confirm that the ground was reached. Moreover, snow pits are widely regarded as the standard validation measurement technique against which other measurements are compared (e.g., López‐Moreno et al., 2020; Kaasik et al., 2023).

We conducted three fieldwork campaigns during the 2025 spring season. The first survey (24–26 May) involved six participants and yielded 13 snow pits, with a primary goal of digging and sampling one very large pit (5.9 m depth). We dedicated nearly a full day to digging this large pit and began measurements on the second day of the survey (25 May). The second survey (30 May–1 June) was conducted by a single individual and involved a traverse across the WRR to measure high-elevation snow on both sides of the mountain crest, yielding 18 snow pits. The third survey (1 June) involved two participants and yielded 5 snow pits. Combined, our three backcountry survey routes cover a distance of roughly 90 km with about 5400 m of vertical elevation gain. Although we could sample more snow pits within a smaller area, traversing these substantial distances ensures that our measurements span a wide variety of representative terrain across the crest of the WRR (Fig. 1). Additional surveys were also conducted in late May to early June of 2023, 2024, and 2026, including both repeated measurements from sites sampled in 2025 and additional sites approximately 50 km to the south on the Wind River Indian Reservation (North and South Fork Little Wind River). Eight profiles were collected in 2023 (snow depth range 0.6–5.8 m), another 20 profiles were collected in 2024 (snow depth range 0.25–3.2 m), and another 14 profiles were collected in 2026 (snow depth range 0.35–5.6 m). We focus on the 2025 surveys, because this was the most comprehensive year of measurements spanning different environments, providing a snapshot of spatial heterogeneity while controlling for confounding variability caused by interannual weather patterns.

At each snow pit, we measure vertical density profiles following standard protocols (Kinar and Pomeroy, 2015). Most snow pits extend to the ground, identified by visual presence of rock/dirt. In 2025, two deep pits did not reach the ground due to time constraints, and we offset our height measurements for these pits using a combination of probe measurements and the 2025 lidar-based snow depths at the pit locations. Similarly, three of the deepest drifts measured in 2023 and 2026 (5.8, 5.6, and 4.0 m) were partially sampled to depths of 4.8, 2.6, and 2.0 m, respectively, and all other profiles were sampled to the ground.

We measure density using several 1000 cc triangular wedge cutters (Snowmetrics Inc.), with care to ensure that the cutter is fully filled but not over-pressed into the snow face. The cutter is used vertically, so that a representative section of the vertical layering structure is obtained. Snow density samples are measured using A&D HT-3000 and HT-5000 digital scales with 1 g precision. To maximize accuracy, we re-tare the density cutter at each snow pit location to account for varying amounts of residual liquid water adhering to the cutter. Additionally, we carefully wipe off excess snow and liquid water from the outside of the cutter after each measurement, and we wipe out any snow adhering to the inside corners of the cutter before each new measurement. Despite our attention to detail, the precision of our density measurements is affected by the challenges of sampling dense, icy snowy in remote alpine locations, in addition to the baseline uncertainty associated with all field density measurements (Proksch et al., 2016). In sections with density exceeding approximately 450 kg m⁻³, we typically find it necessary to hammer in the density cutter. Brittle, icy snow can chip out of the cutter, and in some cases it is necessary to replace these chips manually to approximate a full cutter sample. This “messy” quality of the data is inherent to the challenge of sampling vertical density profiles in extreme conditions, i.e., thick ice layers and brittle sub-zero snow in remote wilderness settings. All cutters remain visually free of deformation during our measurements. As a precaution, since hammering the cutter into hard snow could potentially degrade its precise geometry, we use a separate dedicated cutter for softer snow measurements. Calibration by weighing the cutters when filled with liquid water indicates uncertainty of < 1 % in the nominal cutter volume.

We also record supplemental data for each snow pit and measure snow temperature profiles using an analog thermometer. Calibration of the thermometer in ice water shows a nominal 0 °C. For each snow pit, we also take a variety of photos showing the regional setting and snow pit interior for archival reference and cross-checking with geospatial attributes such as forest cover. We record the location of each snow pit using a smartphone GPS, which has sufficient accuracy for our objective of estimating general geospatial attributes such as elevation and slope. Anecdotally, residual excavations after re-filling our deepest snow pit remain clearly discernible in the airborne lidar data collected several days later, and our GPS point differs by less than 1 pixel (3 m).

To explore potential drivers of spatial variability in alpine snow density, we construct a statistical regression model fine-tuned to our snow pit datasets from each year, and we compare our results with several generalized empirical models from the literature to test whether these other models predict similar degrees of variability.

Repeat airborne lidar surveys provide high resolution topography and snow depth data that inform our analysis of alpine snow variability. The snow-free topography of the WRR is captured by lidar acquired on 18–19 August 2019, distributed as a digital terrain model (DTM) by the U.S. Geological Survey (USGS) as part of the 3DEP program. Since glaciers and perennial snow/ice features are rapidly melting in the WRR, the snow-free DTM changes continually in many areas of the northern WRR. To account for annual ablation, we update the snow-free DTM using a dedicated glacier lidar survey conducted by Airborne Snow Observatories, Inc. (ASO) on 6 October 2024. Another lidar survey conducted by ASO on 1–2 June 2025, provides snow depth data at 3 m resolution across much of the WRR, including the full extent of our field surveys. Lidar surveys used for the 2024 and 2026 SWE maps were likewise conducted by ASO on 31 May 2024, and 28 May 2026, relative to end-of-ablation-season glacier lidar updates on 7 October 2023, and 26 September 2025, respectively. Processing procedures used to retrieve snow depth from the fusion of lidar and imaging spectrometer data are presented in Painter et al. (2016) and Boardman et al. (2025). In 2025, our final SWE map (using ASO lidar depths with densities estimated from Eq. 1) was delivered to water managers and stakeholders on 7 June, 5 d after the conclusion of the airborne surveys and 2 d after ASO's delivery of the snow depth map. Similarly, the 2024 and 2026 SWE maps were delivered in near-real-time, making our study uniquely relevant for operational snowpack guidance and water supply forecasting compared to retrospective analyses of snowpack patterns that are not available for management in real-time.

Finally, we compare several other near-real-time gridded SWE datasets with our lidar-based dataset. Snow datasets selected for this comparison must be available in near-real-time, i.e., with a latency of days, since water supply forecasting (and related management guidance) is one of the primary motivations for annual snow surveys in the western USA. Additionally, we disregard very coarse datasets (larger than roughly 1 km resolution) that are unlikely to capture the relevant sub-watershed scales in the WRR region. We identify three qualifying datasets for this comparison: the SNOw Data Assimilation System (SNODAS) daily SWE product distributed by the National Oceanic and Atmospheric Administration (Barrett, 2003), a regression-based SWE product (CU-SWE) distributed by the University of Colorado (Yang et al., 2022), and a neural network-based SWE product (SWANN) distributed by the University of Arizona (Broxton et al., 2016, 2019b, 2024). For consistency and clarity, we refer to the two university experimental SWE products by their respective university names, i.e., U. Colorado SWE and U. Arizona SWE. The native resolutions of these products vary between 500 m for the U. Colorado dataset and 1/120° (approximately 800–1000 m) for SNODAS and the U. Arizona dataset. For each of these products, in addition to the lidar-based SWE maps, we interpolate missing grid cell values using the average within a 3 × 3 kernel and reproject to a common 500 m resolution using the weighted average of overlapping cells. Using both the native resolution and the 500 m resolution, we calculate and compare cumulative distribution functions (CDF) for SWE within each of three watershed boundaries encompassing our field survey area.

We also use the ASO snow-on and snow-off lidar topography data to explore the relationship between spatial snow density heterogeneity and wind speed to test the hypothesis that wind-topography interactions can explain observed snowpack heterogeneity in our deepest sampled location. Slower wind speeds in sheltered locations can contribute to snow deposition and deep drift formation, which may also impact density due to wind-packing. We simulate wind speed for a 2.2 km² domain surrounding a major drift where our deepest snow pit is located. The WindNinja software (Wagenbrenner et al., 2019) leverages a computational fluid dynamics module based on a Reynolds-Average-Navier-Stokes (RANS) solver implemented in OpenFOAM (Weller et al., 1998). We use WindNinja to implement a RANS simulation of the wind field through a 3 m mesh restructured from the 1 m lidar DTM. We initialize WindNinja with a single 20 m s⁻¹ west-to-east wind speed at a height of 10 m with a ground roughness length of 0.1 m (grass setting), and we sample the simulated wind field at 1 m height after 1000 RANS solver iterations. Since our goal is merely to obtain an indicative map of faster/slower wind speeds, the precise numeric values are less important than the spatial pattern. This combination of settings proves satisfactory for illustrating the general slowdown and speedup of wind in the simulation domain, though it is merely a snapshot of hypothetical conditions and does not account for variability and uncertainty in the actual wind speed and direction. Secondly, we repeat this simulation with an altered DTM representing the snow surface, calculated by adding the 2025 ASO lidar snow depth to the 2019 snow-off topography.

Our measurements reveal systematic heterogeneity in ablation-season alpine snow density, both spatially and within individual pits (Fig. 2). Similar patterns of snow density heterogeneity are also apparent in the data from other years (Fig. 2), but we focus our analysis on the 36 density profiles from 2025 to control for confounding interannual variability in accumulation and densification processes (Fig. S4). The coefficient of variation (CV) among all 36 locations sampled in 2025 is 13 % relative to a mean of 450 kg m⁻³. The mean vertically integrated density (entire snow pits) varies between 339 and 585 kg m⁻³, and individual density measurements (within pits) vary between 247 and 688 kg m⁻³. The snow in forested areas is significantly (p< 0.001) less dense compared to snow in open areas (mean 405 vs. 471 kg m⁻³). There is a significant correlation between snow depth and density (r= 0.59, p< 0.001), and snow at least 2 m deep is significantly (p= 0.02) more dense than snow that is less than 2 m deep (mean 497 vs. 438 kg m⁻³). Additionally, snow at higher elevations is generally less dense. Although there is no significant correlation directly between elevation and density due to confounding variability, snow above 3500 m is significantly less dense (p< 0.01) than snow below 3500 m considering only non-forested locations (437 vs. 490 kg m⁻³). However, we emphasize that these relationships are subject to multicollinearity, which prevents any straightforward interpretation. For example, there are significant correlations between elevation and forest cover (r=-0.63, p< 0.001) and between snow depth and forest cover (r=-0.41, p= 0.01) simply due to the existence of a treeline elevation and the lack of deep snow (> 2 m) in forested areas.

There is less vertical variability in snow density compared to the spatial variability between pits, but vertical variability is still substantial. Considering only the 2025 snow pits, the standard deviation across all 485 individual density measurements is 75 kg m⁻³, which reduces to 46 kg m⁻³ after subtracting the mean of each snow pit from the constituent intra-pit measurements. For comparison, the standard deviation of bulk density across all 36 pits is 57 kg m⁻³, which is 22 % more variable than the intra-pit vertical variability. Nevertheless, vertical trends in snow density can reveal salient information about physical processes operating in the snowpack that are missed by vertically integrated average measurements (such as those obtained from snow pillows or SWE coring tubes).

Isolating patterns of vertical density variability reveals several distinct categories of depth-density trends in the 2025 dataset (Fig. 3). Most of the snow pits in forested areas have significant (p< 0.05) trends toward lower density near the ground (N= 8, or 73 % of forested locations), and the other three have insignificant trends. Two of the three forested snow pits with insignificant depth-density trends nevertheless have lower-density snow near the ground, but the 10 cm vertical sampling resolution is too sparse for the trend to rise to statistical significance in shallow snow. Among the 25 snow pits in open areas (non-forested), 13 have significant trends toward lower density near the ground and 11 have no significant trend. Interestingly, the snow pits with significant trends toward lower density near the ground have widely varying snow depths, from 0.4 to 3.3 m. From field observations, we note that lower snow densities near the ground are commonly associated with depth hoar, i.e., faceted crystals caused by vapor gradients in shallow snow, though similar trends also emerge in deeper snow due to overlying avalanche debris and potentially other controls on densification.

Deep wind drifts exhibit different depth-density relationships compared to shallower snow. The deepest snow pit (5.9 m) has no significant trend with depth, a finding that we explore further in Sect. 3.2. The second-deepest (4.0 m) snow pit provides the sole example of a significant trend toward increasing density closer to the ground (r=-0.53, p< 0.001), a pattern that was also observed in a repeat measurement from another year (r=-0.57, p< 0.001). Notably, this deep pit location is at a relatively high elevation (3820 m) and had a mean temperature below freezing, which we explore further in Sect. 3.3. Thus, the lower density near the surface at this location could be associated with relatively fresh snow that has not undergone as much metamorphosis as the deeper, older snow.

Measuring the density of a very deep snow drift was a key goal of our 2025 field surveys, since these deep drifts are salient features of the WRR landscape (Boardman et al., 2025; Boardman, 2025b). In this section, we focus our analysis on a wind drift occupying a nivation hollow at roughly 3280 m elevation near Burrow Flat in the Dinwoody Creek watershed (Figs. 1 and S5). Figure 4 provides an overview of our nivation hollow study site, including photographs of the drift surface and the interior of our deepest snow pit. A longitudinal profile across the nivation hollow shows how a nearly planar snow surface overlies a topographic concavity, which results in exceptionally deep snow near the middle of the drift (4–6 m).

A 5.9 m deep pit near the center of the nivation hollow reveals a remarkably high snow density, with a range of 515–688 kg m⁻³ and a mean of 585 kg m⁻³. There is no significant trend in density with depth for this deep pit (r= 0.15, p= 0.24). Nevertheless, the upper 2 m of the profile has a slightly higher density compared to the lowest 2 m of the profile (mean 595 vs. 582 kg m⁻³), though this difference is not statistically significant. The deep drift remains quite dense even near the surface: the top meter has a bulk density of 588 kg m⁻³, and the top 10 cm has a density of 597 kg m⁻³, both of which are higher than the pit average. These measurements support our qualitative observation that it is quite easy to walk on the surface of the nivation hollow drift without sinking in or “postholing,” in contrast to nearby shallower snow.

We note several substantial ice layers (1–3 cm thick) beginning at a height of 2.2 m above the ground in the Burrow Flat nivation hollow drift. No horizontal ice layers are apparent below this height. Ice layers are more frequent closer to the surface, which may contribute to the higher bulk density of the upper 2 m of the pit. Notwithstanding, we still observe exceptionally high densities in deeper layers with no visible ice (e.g., 605, 628, and 684 kg m⁻³ at 0.4–0.5, 1.3–1.4, and 1.4–1.5 m heights, respectively). A re-frozen preferential “flow finger” (Marsh and Woo, 1984) was encountered at approximately 1.0–1.4 m height, but we chose to avoid sampling this anomalous feature and instead sampled adjacent snow without visible ice. Surprisingly, our lowest-density measurement (515 kg m⁻³ at 2.2–2.3 m height) is associated with the deepest and thickest ice layer (approximately 3 cm thick). Our highest-density measurement (688 kg m⁻³ at 4.4–4.5 m height) is also associated with several thinner ice layers. One explanation for this discrepancy could be different amounts of liquid water pooling/refreezing above and below the different ice layers, especially if the ice layers create preferential lateral flow pathways. However, these ice-layer measurements are also relatively uncertain due to the inherent challenge of sampling brittle ice with a metal density cutter, and the anomalous weights could be the result of measurement error.

Additional snow pit measurements reveal that density remains unexpectedly high near the lower edge of the nivation hollow, but snow at the upper edge is more comparable to the density of nearby locations. Based on prior fieldwork, we anticipated that shallower snow (< 2 m) would have a lower density more in line with the 350–450 kg m⁻³ range of typical ablation-season alpine snow. While this is largely true at the landscape scale (Fig. 2), a 1.0 m snow pit near the lower edge of the nivation hollow has a bulk density of 552 kg m⁻³, only 6 % lower density than the deep pit near the middle of the drift. Interestingly, the 10 samples from this shallower snow pit have a high correlation with the lowest 10 samples from the deep pit (r= 0.76, p= 0.01), indicating that both locations have similar density stratigraphy near the ground. This increases our confidence that vertical density variations are physically meaningful (not just the result of random sampling error), and seemingly implies that the processes controlling densification are independent of snow depth within parts of the nivation hollow drift. In contrast, a 1.5 m snow pit near the upper edge of the nivation hollow shows a different signature: its bulk density of 484 kg m⁻³ is 17 % lower than the center of the drift, and there is no significant correlation between the vertical density variation of the 1.5 m pit and the bottom 1.5 m of the deep pit (r= 0.23, p= 0.41). The upper edge of the nivation hollow drift fits within the density range of surrounding snow (bulk density range 444–488 kg m⁻³ across five other pits in non-forested locations). Due to the large amount of work required to sample a single ∼ 6 m snow pit profile (approximately nine person-days, or 1.5 d with six people), it was infeasible to sample additional lateral variability within the deepest part of the drift. However, future studies may consider measuring multiple profiles within a single snow pit to reduce potential measurement errors.

Wind simulations and repeat lidar observations indicate that the nivation hollow reduces wind speed until it fills with snow, resulting in an apparent self-limiting drift behavior (Fig. 5). The mid-July imagery in Fig. 5A shows how the deep drift persists through much of the summer after most other snow has melted, though field observations indicate that it melts out completely by September, at least in most years. Simulating the wind speed 1 m above the underlying ground surface results in a wind speed pattern that closely mirrors the snow depth pattern (Fig. 5B, D–G). Wind speeds are greatly reduced where the concave nivation hollow creates a sheltered zone, similar to the functioning of artificial snow fences. Snow transport is reduced in areas with lower wind speeds, which reduces the suspension and saltation fluxes, causing drift formation. Consequentially, there is a strong negative correlation between ground wind speed and snow depth (r=-0.71, p< 0.001). However, as the nivation hollow begins to fill with drifted snow, the topographic concavity has less effect on the wind speed. The wind speed 1 m above the 2025 snow surface is much faster than the wind speed 1 m above the ground surface (Fig. 5C). In areas with > 4 m snow depth, the mean wind speed is 143 % faster above the snow surface compared to the ground surface.

The nivation hollow wind drift stores nearly the same amount of snow each spring (Figs. 4–5) despite widely varying seasonal snowfall. For example, based on spatially complete lidar depth data with densities constrained by our fieldwork, we estimate that the surrounding Dinwoody Creek watershed has a mean snowpack storage of 0.21 m SWE on 2 June 2025, compared to 0.34 m on 31 May 2024, which is 37 % less in 2025 at nearly the same point in the season. However, areas of the nivation hollow with > 4 m snow depth in 2025 (0.015 km² area) had a mean snow depth decrease of 0.36 m relative to 2024, representing an estimated SWE volume storage difference of only 7 % in those locations. Similarly, in 2026, the watershed-mean SWE was 20 % lower than 2025, but the nivation hollow held almost exactly the same amount of SWE (0.5 % less). As illustrated in the Fig. 4 profile and Fig. 5 D–G, nearly identical snow surface elevations are observed by four different airborne lidar surveys: 11 June 2022; 31 May 2024; 1–2 June 2025, and 28 May 2026. The snow depth correlations between all six pairs of lidar surveys (2022–2026) range between r= 0.984 and r= 0.996 (p< 0.001) in the drift vicinity at the 3 m grid scale (Fig. 5D–F). Hence, the nivation hollow drift is largely insensitive to interannual variability.

Another nivation hollow at a similar elevation of 3140 m on nearby Whiskey Mountain (Fig. 1) exhibits a similarly high bulk snow density. A partial snow pit in this nivation hollow drift (top 1.3 m sampled of 3.5 m total depth) has a bulk density of 567 kg m⁻³, again with no significant depth-density trend (r=-0.02). This represents the second-highest mean snow density recorded during our 2025 field surveys, and all three snow pits with mean densities above 500 kg m⁻³ are associated with wind drifts in nivation hollows (Burrow Flat drift center, Whiskey Mountain drift, Burrow Flat drift lower edge).

Similar patterns repeat in the nivation hollows across years (Fig. 2). Partial snow pit profiles from deep nivation hollow drifts collected in 2023 and 2026 show similarly high bulk densities: means of 568, 526, and 564 kg m⁻³ for drifts with depths of 5.8, 5.6, and 4.0 m depth, respectively. Although these profiles only extend through the upper 2.0 to 4.9 m of the drifts, we similarly observe that there is no statistically significant trend of density with respect to depth within these additional three deep nivation hollow drifts.

We now analyze snow density at a second deep drift location, this time at a higher elevation with a different topographic setting, and we compare our observations to surrounding snow pits that illustrate a variety of interacting controls on density (Fig. 6). Downs Mountain (4071 m elevation) is the northernmost 4000 m peak on the Rocky Mountain Continental Divide, providing a unique combination of extreme alpine conditions. A 1.6 m snow pit in a drift near the summit (4040 m) has a much lower density of 401 kg m⁻³ compared to lower-elevation snow of a similar depth, e.g., 471 kg m⁻³ in a 1.1 m snow pit at 3630 m elevation on the nearby Goat Flat plateau (Fig. 1). A 0.8 m snow pit on a steep (43°) north-facing slope also has a relatively low density of 393 kg m⁻³ despite being at a similarly low elevation as the relatively dense Goat Flat snow pit (471 kg m⁻³). Compared to the nearby summit and north-facing snow pits, 4.0 m deep snow pit in an extensive drifted snowfield at 3820 m elevation has a substantially higher density of 457 kg m⁻³, though this is still much less dense than the nivation hollow drifts. Part of this difference in snow density between the deep nivation hollow drift and the deep Downs Mountain snowfield is associated with the latter's 540 m higher elevation, as we show in Sect. 3.4. Finally, a partial snow pit in an avalanche runout zone has the highest bulk density observed in the Downs Mountain vicinity (483 kg m⁻³), and individual blocks of recent wet slide debris have even higher densities in the range of 558–638 kg m⁻³.

The deeply drifted snowfield on the downwind (east) side of Downs Mountain has a substantially higher density than shallower snow at comparable elevations. The three other snow pits above 3800 m have a bulk density of 409 compared to 457 kg m⁻³ in the 4.0 m pit, a difference of about 12 %. Considering that one of these three other high-elevation pits is also fairly deep (2.1 m/438 kg m⁻³, near Hunters Hump in the Green River watershed), the drift density is even more pronounced when comparing it only to relatively shallow high-elevation snow (e.g., 387 kg m⁻³ for a 0.8 m pit near the upper edge of the Continental Glacier). Conversely, snow density increases rapidly at lower elevations, and even shallow snow is denser than the deep snowfield at altitudes only 200–250 m lower (e.g., 1.1 m/471 kg m⁻³ at 3630 m elevation and 0.8 m/490 kg m⁻³ at 3580 m elevation). Thus, deep wind drifts exhibit consistently higher densities compared to nearby shallower snow, but the absolute magnitude of drift density also depends on elevation and slope/aspect, and hence the snow metamorphosis legacy of the local microclimate. Similar patterns of high-elevation snow density heterogeneity are observed in other years (Fig. 2). However, the absolute magnitude of bulk density depends on the specific year, and high elevation (3800 m) drifts as deep as 3.2 m can have bulk densities as low as 398 kg m⁻³, or 59 kg m⁻³ lower than observed in 2025, at roughly the same time of year due to interannual variability in the processes that control densification (such as the onset of freeze-thaw cycles).

Despite the onset of ablation at most places in the WRR by the time of our 2025 surveys in late May and early June, some high-elevation snow remained colder than the freezing point, potentially contributing to density variability. Only two of our 36 snow pits show temperatures more than one degree below zero, which is the precision of our analog thermometer. These two cold locations are the Downs Mountain summit drift (mean -2.4 °C, range -0.5 to -5.0 °C) and the Downs Mountain deep drifted snowfield (mean -2.0 °C, range -0.5 to -4.0 °C). Interestingly, the second-highest snow pit (3980 m elevation) shows nearly isothermal temperatures (all temperatures between -1.0 and -0.5 °C), demonstrating that elevation (and air temperature, by proxy) is not the only control on snowpack temperature. Notably, this high-elevation snow pit is half as deep as the Downs Mountain summit drift (0.8 vs. 1.6 m). The coldest temperatures in the Downs Mountain summit drift are near the ground surface, and the snowpack warms to zero near the surface, which suggests that this difference in temperature may be the result of seasonal and/or diurnal insulation effects from the deeper snowpack. Likewise, the deep snowfield reaches a temperature of zero near the surface, though it also warms to zero near the ground, with the coldest snow at a height of 1.4–2.4 m (near the middle of the profile).

Snow temperature had an inconclusive relationship with density at the time of our 2025 surveys. There is a weak negative correlation between density and temperature in the deep snowfield (r=-0.37, p= 0.02), unlike the shallower summit drift, which has a strong positive density-temperature correlation (r= 0.73, p= 0.001). In the case of the shallower non-isothermal pit, we observe depth hoar (loose facets) near the ground, but the temperature profile is reversed compared to what we would expect for depth hoar formation, which is typically associated with colder surface temperatures and warmer snow near the ground. This temperature reversal could be indicative of the transition between the accumulation and ablation season, as the snowpack ripens and warms from the top down. In contrast, the deep snowfield does not exhibit depth hoar and shows a significant trend toward increased density with depth, which is rare in the WRR at the time of our spring snow surveys (Fig. 3). It is possible that the colder temperatures near the ground (Fig. 6) could be related to permafrost, which can be associated with patterns of snowpack redistribution in alpine environments (Luetschg et al., 2004; Kenner et al., 2017). All of the snow pits in the Downs Mountain vicinity have at least some ice layers, so melt-freeze processes and liquid water percolation probably also contribute to density variations, even though parts of the vertical profile remain below freezing. In 2024, the snowpack ripening elevation (above which temperatures fall below 0 °C) closely matched the fastest rate of change implied by the elevation-dependent term of the snow density regression (Boardman et al., 2025), and at the time of our 2026 surveys, the snowpack was mostly 0 °C at all elevations. The variable seasonal progression of elevation-dependent ripening is reflected in changes to the elevation-dependent terms of the density models (Table 1).

We now shift to analyzing statistical models of snow density heterogeneity, which are used to assess systematic variability and propagate point measurements to the watershed scale. We do not attempt to benchmark the “best” empirical models from the prior literature, and there are too many models for an exhaustive comparison in this study. Instead, we focus on a few widely used models with a range of complexity and predictor variables to conduct a broad sensitivity analysis.

A nonlinear regression model (Eq. 1) can satisfactorily explain density differences between most snow pits in our 2025 dataset. We exclude the two pits from the edges of the nivation hollow since they are not representative of most areas of the snowpack with similar depth. Across the other 34 snow pits, the fitted model has a root-mean-square-error (RMSE) of 26 kg m⁻³, which is 6 % error relative to the mean of 447 kg m⁻³. The coefficient of determination for density variability is R2= 0.76 across all 34 of these pits, which increases to R2= 0.94 considering only the densities of the eight pits > 2 m deep.

All of the fitted model parameters in Eq. (1) describe relationships that match physical expectations (Fig. 7). The model indicates higher densities at lower elevations, in deeper snow, and in open areas (conversely, density is lower in forested areas). Elevation is a proxy for air temperature, and lower elevations are associated with more melt-freeze cycles and an earlier start to snowpack metamorphosis, leading to higher densities. Snow depth is a proxy for wind drifting and avalanching, which can increase density by fracturing and sintering grains. We expect that forest canopy cover may be associated with lower density snow due to a combination of factors including wind shelter, shade, and nighttime longwave radiation, which could reduce the number of melt-freeze cycles. The model also identifies slightly lower densities on north-facing and east-facing slopes, consistent with fewer melt-freeze cycles caused by overall shading and reduced late-afternoon sun exposure when temperatures are warmest. Since Bayesian sampling (Sect. 2.2.1) inherently quantifies and propagates parameter uncertainty, we can robustly estimate confidence intervals for these effect sizes. The fitted model is represented by a multivariate posterior distribution (Fig. S3), and the spread of this distribution rigorously quantifies the uncertainty of the model with respect to the interaction of all five variables (Eq. 1). The linear depth-density relationship has an uncertainty of 9 % (CV of posterior parameter distribution for θ2 in Eq. (1), with a 95 % confidence interval of 18 to 26 kg m⁻³ increase in density per meter increase in depth (bulk density variability between locations, not vertical density trends). The parameter distributions for elevation and canopy cover are less straightforward due to nonlinearity, but 100 % of the parameter samples are negative for θ3 and θ6, indicating very high confidence that forests and high elevations produce lower densities. The linear relationships with north and east slope are more weakly constrained due to the limited slope of our observations (max 43°), but the posterior parameter distribution still provides 96 % confidence that north-facing slopes have lower densities and 90 % confidence that east-facing slopes have lower densities. Similar snow density relationships are detected by the 2024 and 2026 regression models (Table 1). Although the directionality of each effect is consistent across years, in some years snow density is more or less sensitive to different predictors, presumably reflecting interannual variability in the specific processes driving densification (e.g., warmer years may be less sensitive to elevation if the whole snowpack has ripened by the late-May timeframe).

No single variable dominates the prediction of snow density (Fig. 7). Instead, observed snow densities are the result of interactions between multiple factors, e.g., high-elevation deep snow may have the same density as low-elevation shallow snow. This finding is strikingly apparent in Fig. 7: although the directionality and magnitude of the response to each individual predictor is well constrained by the ensemble of Bayesian parameter samples, the observed snow densities vary widely across any single predictor. Together, however, the seven-variable model explains most of the observed spatial variability (r= 0.86), as shown by the scatterplot of modeled and measured snow density in the WRR 2025 panel of Fig. 8. The simpler model from 2024 (no slope/aspect dependence), which was fitted to data from 2023–2024 and published prior to the collection of the 2025 data (Boardman et al., 2025), also achieves a reasonably strong correlation (r= 0.76) with the 2025 data. Similarly, the 2026 WRR model (Table 1) correlates with the 2025 data (r= 0.82), but the 2026 model was supplemented with 11 high-elevation snow pits from 2023–2025 (due to weather constraints during the 2026 surveys), so the 2025–2026 comparison is not strictly independent. Nevertheless, the similarity of 2025–2026 models, and especially the completely independent 2024–2025 models, bolsters our finding that patterns of density heterogeneity repeat across years.

All five of the tested previous empirical snow density models systematically underestimate the heterogeneity of the WRR snowpack. Each of the panels of Fig. 8 illustrates a different approach to empirical density modeling reported in the literature (Table 1) and evaluated at our snow pit locations. All of these models include at a minimum a regression on snow depth, which itself has a reasonably strong correlation with density across our snow pits (r= 0.59, p< 0.001), though this relationship is scattered (cf. depth-sensitivity panel of Fig. 7). Since the models proposed by Tabler (2003), Jonas et al. (2009), and Sturm et al. (2010) exclusively use snow depth to infer spatial variability in snow density using various linear and non-linear equations, these models offer little to no additional predictive skill relative to the raw depth-density correlation. By accounting for additional meteorological effects, we would expect the Bormann et al. (2013) multiple linear regression models to account for additional variability. However, since site-specific meteorological data are not available (unlike at the stations where the models were developed), the interpolated gridded meteorological data offer minimal additional predictive skill for snow density. Thus, the Bormann et al. (2013) models, as adapted for this study, produce similar correlations as the other models (r= 0.57 for the alpine model and r= 0.59 for the combined global site model). The Bormann et al. (2013) global (all-site combo) model has lower bias than the alpine model in the WRR, so we use the global model to evaluate watershed metrics discussed subsequently. The Hill et al. (2019) model also uses gridded climate data (Table 1), but the regression is insufficiently sensitive to snow depth in the WRR, yielding the lowest correlation across all of the tested models (r= 0.31) and largest bias (-20 %). The underestimation of density by the Hill et al. (2019) model may partly result from the relatively low December–February estimated precipitation (mean 0.182 m across snow pit locations) compared to the relatively deep measured snow (mean 1.46 m across snow pit locations), since much of the Rocky Mountain snowpack accumulates in spring storms occurring after the December–February “winter precipitation” period (Serreze et al., 2001; Trujillo and Molotch, 2014). This low winter precipitation estimate (from gridMET) is roughly congruent with the map presented by Hill et al. (2019) in their Fig. 5B, and the resulting large density bias highlights the risk of relying on gridded climate data aggregated over an arbitrary time period (defined by specific “winter” months) as opposed to the more direct depth-density approach taken by most other models, which can implicitly account for the full accumulation season.

Surprisingly, none of the prior literature density models considered here quantitatively parameterize the effect of forest cover, though forest cover is sometimes considered qualitatively (Sturm et al., 1995) when choosing the appropriate climate class for the Sturm et al. (2010) model. Except for the Tabler (2003) model, which has similar bias between forested and open locations, the other models all have significantly more negative biases (greater underestimation of snow density) in non-forested locations, with a statistically significant difference in the bias at the p< 0.01 level for all other prior models except Hill et al. (2019), which has a significant bias difference at p< 0.02. The explicit consideration of forest cover in Eq. (1) of this study removes this systematic bias, i.e., the difference in forested and non-forested bias for our WRR 2025 and WRR 2024 models is not statistically significant (p= 0.44 and 0.45 respectively). The observed density contrast between low-density forest snow and high-density snow in open drifted areas is especially apparent in the Fig. 9 maps. Although the Tabler (2003) indirectly captures this effect because of the confounding effect of snow depth (i.e., forest snow is generally shallower), this model predicts the same low density for shallow snow everywhere, whereas our measurements indicate a systematically higher density in non-forested areas (Figs. 2 and 7).

Neglecting the high density of alpine wind drifts can lead to substantial underestimation of the total SWE volume in mountain watersheds. Noting again that most of the prior literature snow density models (Fig. 8) are not specifically calibrated to the WRR, applying these models to the 2025 lidar snow depth map can nevertheless inform expectations for the potential sensitivity of SWE to different assumptions about density variability. Indeed, some of these models (e.g., Hill et al., 2019) are explicitly intended to convert depth to SWE at large scales in support of water resource projects. In the three watersheds where our snow pits are located (Torrey Creek, Dinwoody Creek, and upper Green River watersheds; Fig. S1), the watershed-average snowpack storage on 1–2 June 2025, is 0.18, 0.21, and 0.26 m SWE, respectively, estimated using the lidar depth and the 2025 WRR Bayesian regression model (Eq. 1). Relative to this baseline, the Tabler (2003) model predicts 9 %–13 % less total SWE in these three watersheds, the Jonas et al. (2009) model predicts 3 %–5 % less SWE, the Sturm et al. (2010) model predicts 5 %–8 % less SWE, the Bormann et al. (2013) models predict 16 %–18 % less SWE (all-site model) or 22 %–24 % less SWE (alpine model), and the Hill et al. (2019) model predicts 18 %–20 % less watershed-total SWE.

Field-observed patterns of snow density heterogeneity repeat across years, although the sensitivity to different predictors (elevation, depth, etc.) varies year-to-year (Fig. 10) according to the WRR 2024–2026 regression models (Table 1). Considering all 3 m grid cells with snow depth > 0.2 m, the median density varies from 425–448 kg m⁻³ across 3 years (during the same late-May to early-June timeframe), which is a fractional variation of 5 %. The highest-density 99th percentile similarly varies by 6 % (521–553 kg m⁻³), but the lowest-density 1st percentile only varies by 3 % (355–364 kg m⁻³). In addition to changes in the snow density relationships across the terrain, these interannual variations are responsive to differences in the spatial extent of drifts and low-elevation forested snow each year. In light of the tremendous natural variability of mountain snowpacks and the wide range of measured densities, it is noteworthy that such similar patterns are observed year to year. Compared to the large range of discrepancies from the prior literature models (Figs. 8–9), a single year of targeted local field surveys can thus have a substantial impact on reducing lidar-based SWE quantification uncertainty. Nevertheless, snow density still exhibits notable spatial patterns of change between years at the watershed scale (Fig. 10). Compared to 2025, bulk snow density was 32 % and 45 % less sensitive to depth in 2024 and 2026, respectively (θ2 in Table 1). However, snow density was about 75 % and 28 % more sensitive to elevation in 2024 and 2026, respectively (comparing the predicted effect of elevation at 4000 vs. 3000 m elevation using median parameter values from Table 1). Interestingly, the impact of forest cover on snow density is the most consistent across years, with 19 % greater sensitivity in 2024 and < 1 % lower sensitivity in 2026 (comparing the predicted effect of 100 % forest cover, Table 1). Despite these differing sensitivities between years, the overall landscape-scale pattern of snow density remains more consistent due to the relative rarity of the most extreme environments. In other words, most of the landscape has similar snow density across years, while the deepest drifts and highest elevations are more sensitive to interannual variability.

Metrics related to deep snow storage are particularly sensitive to bulk density assumptions in the absence of local field data. Previous analysis of the WRR snowpack suggests that the amount of snow deeper than some minimum threshold, e.g., SWE > 2 m, is related to inter-watershed differences in late-summer streamflow and glacier mass loss (Boardman et al., 2025). Our 2025 SWE map based on ASO lidar and the WRR-specific density model (Eq. 1) indicates that 34 %, 26 %, and 10 % of the total snowpack is stored in areas with SWE > 2 m in the Torrey, Dinwoody, and Green River watersheds, respectively. Relative to this baseline, all of the tested prior literature density models underestimate the watershed-total amount of snow storage in places with > 2 m SWE by 14 %–57 % in Torrey Creek, 14 %–67 % in Dinwoody Creek, and 27 %–78 % in the upper Green River. Even deeper snow (SWE > 4 m) is relevant for constraining certain snow science topics such as snow transport (Boardman, 2025b). At the time of our surveys, we estimate that areas with SWE > 4 m account for about 6 %, 2 %, and 0.5 % of the total snowpack in the Torrey, Dinwoody, and Green River watersheds, respectively. Analogously, the prior literature snow density models underestimate the amount of snowpack storage in areas with SWE > 4 m by 41 %–90 % in Torrey Creek, 42 %–88 % in Dinwoody Creek, and 34 %–81 % in the upper Green River. However, it is important to note that these percentages are calculated only considering the grid cells with SWE exceeding a discrete threshold in each respective map, so this underestimation is compounded by the shrinking number of grid cells that qualify for inclusion with SWE > 2 m or SWE > 4 m when deep snow density is underestimated (i.e., the SWE is not underestimated by 90 % in any single grid cell, but the total volume in areas with SWE strictly exceeding 4 m is underestimated by up to 90 %).

Despite the underestimation of deep snow storage by the prior literature density models, all of the lidar-based SWE maps still exhibit much higher spatial heterogeneity compared to the other near-real-time gridded SWE datasets (Fig. 11, Table 2). It is no surprise that the lidar-based SWE maps capture more heterogeneity at the native 3 m resolution, since the other gridded datasets are two orders of magnitude coarser (500+ m) and thus incapable of capturing heterogeneity associated with nivation hollows, cornices, and similar fine-scale snowpack features. However, even after averaging all maps to the same 500 m resolution, the lidar-based SWE map exhibits considerable spatial heterogeneity that is not represented in the SNODAS, U. of Colorado, or U. of Arizona datasets, probably because mountain-scale snow-atmosphere coupling processes (Mott et al., 2018) are missing from the underlying gridded climate datasets and poorly represented by the small range of variability in the SNOTEL network. Within the combined three-watershed area, the lidar-based SWE map shows an area of 16.8 km² with > 1 m of average SWE at 500 m resolution, while the U. of Colorado dataset only shows an analogous area of 0.8 km² and the other two datasets have zero area with > 1 m SWE. The standard deviation of 500 m average SWE is 56 % higher, 75 % higher, and 33 % higher in the lidar-based map compared to the SNODAS, U. of Colorado, and U. of Arizona SWE maps, respectively. Remarkably, the SNODAS mean SWE is only 2 % lower than the lidar-based mean SWE, but the single deepest grid cell (500 m average) in SNODAS is 61 % lower, and SNODAS has 36 % more snow-free area, indicating a severe underestimation of spatial heterogeneity within the snow-covered area. Even when further averaging the maps to 1000 m resolution (Fig. S6), the lidar-based SWE map has 52 % higher standard deviation compared to SNODAS. Aggregated to the same 1000 m resolution, the lidar-based map shows an area of 8 km² with > 1 m SWE compared to zero area with > 1 m SWE in SNODAS despite the near-identical mean SWE in both datasets.

Snow density is an important control on watershed-scale SWE heterogeneity, but all of the lidar-based datasets exhibit more spatial heterogeneity (regardless of which density model is used) in comparison to any of the non-lidar gridded SWE datasets (Table 2). The non-lidar SWE datasets underestimate the watershed-average mean SWE by 2 % to 17 % compared to an underestimation of 12 % on average across all five prior literature density models. At 500 m resolution, the non-lidar gridded SWE datasets underestimate the deepest 90th, 99th, and 99.9th SWE percentiles by a mean of 25 %, 48 %, and 52 %, respectively, while the prior literature density models underestimate the same percentiles by a mean of 11 %, 12 %, and 17 % at 3 m resolution. When considered as a percentage uncertainty, deep snow metrics are relatively more sensitive to snow depth (lidar vs. non-lidar) than snow density (snow pits vs. literature models), and the depth becomes relatively more important compared to density at more extreme SWE percentiles. However, when considering the absolute sensitivity of the SWE distribution, there is a mean difference of 0.09 m SWE between the snow pit and literature-based density assumptions for the 90th SWE percentile compared to a 0.26 or 0.65 m difference at the 99th and 99.9th percentiles, respectively. Thus, the density constraint has the largest absolute impact on estimated SWE in the deepest parts of the snowpack, as density errors scale multiplicatively with depth.

In our WRR dataset, the highest bulk snow densities are associated with deep wind drifts at relatively low elevations. Although avalanche debris is the densest snow within the high-elevation Downs Mountain vicinity, even avalanche debris is less dense than the lower-elevation nivation hollow drift, both when considering the bulk density of each snow pit (483 vs. 585 kg m⁻³) or the highest density of any individual measurement (638 vs. 688 kg m⁻³). Unlike glacial firn, which densifies at greater depths (e.g., Stevens et al., 2024), the 5.9 m snow drift density profile exhibits no significant depth-density trend. Since snow near the surface of the drift exceeds the vertically integrated bulk density, it seems that overburden pressure is a negligible driver of densification in the nivation hollow drift. This is in contrast to common process-based models used with lidar data, such as iSnobal, wherein overburden pressure is stated as the assumed driver of densification with depth (Hedrick et al., 2018). Thus, snow lidar surveys relying on modeled density fields may mischaracterize the relationship between depth and density in heavily wind-affected regions such as the WRR. Our finding of constant high density with depth is congruent with expectations for dense packing of snow grains in wind drifts (e.g., Jellinek, 1959; Anderson and Benson, 1963), though ours appears to be the first literature account of a vertical density profile (at 10 cm resolution) from a 5–6 m drift in an alpine mountain environment. Thus, we are uniquely positioned to test hypotheses about drift density in a mountain environment that is important for western water resources (Bales et al., 2006; Li et al., 2017), as opposed to the Arctic tundra and ice sheet contexts of many older snow drift studies (e.g., Anderson and Benson, 1963; Benson and Sturm, 1993; Craven and Allison, 1998; Sturm et al., 2001; Libois et al., 2014; Parr et al., 2020).

Since our deepest drift profile is very dense even in sections without visible ice layers, and since adjacent snow with a similar microclimate is much less dense, we infer that melt-freeze processes are also unlikely to be the dominant driver of the density contrast between deep and shallow snow. Instead, wind-packing is the obvious explanation for the uniformly high bulk density, which is supported by the distinctive drift morphology (Fig. 4) and wind-topography interactions (Fig. 5). Kotlyakov (1966) found that freshly deposited Antarctic snow can exceed a density of 400 kg m⁻³ when deposited during winds of roughly 15–30 m s⁻¹ (visualized in Fig. III-4 of Mellor, 1964), far exceeding the typical density of new snow (Tabler, 2003). Consequentially, the correlation between snow depth and density (deeper drifts are generally denser) is not necessarily a causal relationship in windy locations, i.e., depth is not the cause of density variability. Rather, snow depth may function as a proxy for the intensity of wind-driven accumulation processes such as fragmentation during saltation (Comola et al., 2017), which causes wind-packing snow densification (Sommer et al., 2017). In other words, the process of drifting increases both the depth and density of certain snowpack locations. Although this finding might be a trivial consequence of spherical geometry from the perspective of grain-scale snow science (Jellinek, 1959; Anderson and Benson, 1963), Tabler (2003) emphasizes “the pressure of overlying snow” when introducing the relationship between drift depth and density, which would logically lead to greater densities at deeper depths (since there is no overlying snow at the surface), in contrast to our observations. This discrepancy in the prior snow drift literature indicates that our first-in-class complete density profiles from deep alpine drifts contribute additional evidence in support of wind-packing as the dominant driver of the depth-density correlation in drifted locations. Moreover, grain-scale wind-packing effects on snow density are rarely considered in operational near-real-time SWE quantification (i.e., not considered in the iSnobal density model used by Hedrick et al., 2018). In another example, a recent intermediate-complexity layer-wise snow density model (H2SWE) also does not consider wind-packing, instead parameterizing exclusively the effects of metamorphism and deformation (Winkler et al., 2021), and this class of model is predominantly applied in wind-sheltered areas (Magnusson et al., 2025). Obviously, the effect of wind-packing on snow density not captured by in-situ stations (i.e., SNOTEL) that are located predominantly in sheltered forest locations (Molotch and Bales, 2006).

Despite the strong depth-density relationship identified in our study, deep snow can be much less dense in wind-sheltered environments. Complete vertical density profiling of a snow pit at the Central Sierra Snow Laboratory (CSSL, California, USA) by the authors on 16 March 2023, yielded a bulk density of 377 kg m⁻³ despite a snow depth of 4.6 m (Fig. S7). This CSSL snow pit was located in a forest clearing below timberline, and the deep snow was the result of high precipitation in water year 2023, not the result of drifting. Compared to our deep snow drift profiles from the WRR (526, 568, and 585 kg m⁻³ bulk densities for drifts > 5 m deep), the CSSL profile provides further support for the importance of wind-packing, since deep snow in a wind-sheltered region exhibits a similarly low density to the shallow forest snow observed in our study. The depth-density relationship in our study is a consequence of the fact that snow transport (wind drifting and avalanches) drives depth variability in alpine regions.

Wind-packing densification also provides evidence for salient topographic controls on spatial snowpack patterns and the potential repetition of these patterns across years (Pflug and Lundquist, 2020). Our computational fluid dynamics modeling of nivation hollow wind speeds in Fig. 5 demonstrates how concave terrain features could plausibly create self-limiting snow drifts, potentially explaining the spatial density variations observed in these locations. The difference in wind speed between the ground and snow surface (Fig. 5C) indicates that as the drift grows, the nivation hollow becomes less effective at reducing near-surface wind speeds, limiting further growth of the drift and producing the nearly identical snow surface profile observed in different years (Fig. 4). This self-limiting behavior could explain why the lower edge of the drift matches the density stratification near the ground in the deep pit: early wind-packed snow covers both areas similarly, but the middle of the nivation hollow continues accumulating snow after the lower edge reaches its equilibrium profile. This behavior could also explain the lower density at the upper edge of the drift, since the top of the drift would only fill with snow after the rest of the drift volume has already accumulated, leaving less time for metamorphosis and densification at the upper edge. Similar drift-trapping behavior in topographic concavities is observed in other alpine environments (e.g., Niwot Ridge; Berg, 1986), but our study provides a unique opportunity to visualize this behavior with multiple years of lidar data and computationally verify the underlying wind behavior (Fig. 5). A self-limiting drift-trapping process could also explain why perennial ice patches in similar nivation hollows exhibit very thin annual accumulation layers (Meulendyk et al., 2012; Chellman et al., 2021; Alt et al., 2024). Regardless of the total annual snowfall, a wind-limited equilibrium drift profile could limit net accumulation to match the slow rate of annual compaction, producing the observed thin ice patch accumulation layers. In contrast to the self-limiting nivation hollow location, which exhibits near-identical snow accumulation each year (Fig. 5D–F), the higher elevation drifts in the Downs Mountain vicinity vary widely between years (Fig. 6C–D). The repeat application of lidar surveys across years can thus be used to distinguish filling and non-filling drift zones (Benson and Sturm, 1993; Parr et al., 2020).

The processes controlling snow drift density are relevant for SWE quantification even at watershed scales. Nivation hollows with drifts as large as the one in Figs. 4–5 are infrequent across the WRR, but snow drifts in general are common in alpine regions. Looking beyond watershed-scale SWE quantification, the nivation hollow drift is also interesting from the perspective of periglacial geomorphology and alpine ecology (Thorn, 1976; Wigmore and Molotch, 2024), since meltwater from this drift (which persists throughout most of the summer) contributes to accelerated weathering and sustains a downslope wetland (Fig. S5). In addition to our exceptionally deep drift profiles (Fig. S7), our numerous other complete snow pits from the WRR alpine environment constrain drift density across gradients of depth and topography. Neglecting the higher density of deep drifts can lead to systematic underestimation of watershed-scale snowpack storage by as much as 24 % (Sect. 3.5). This bias is exacerbated when considering metrics of snowpack heterogeneity, as watershed-total deep snow storage metrics are underestimated by 14 %–78 % (SWE > 2 m) and 34 %–90 % (SWE > 4 m) since the area meeting a respective threshold also decreases when SWE is underestimated in any particular grid cell (e.g., a grid cell underestimated at 3.9 m SWE no longer counts for the SWE > 4 m metric). Although these precise depth thresholds are merely illustrative, deep snowpack metrics in general are important for accurate understanding of snow accumulation processes (Boardman, 2025b) and snowmelt processes shaping glacier resilience and streamflow seasonality (Boardman et al., 2025), so underestimating the SWE heterogeneity in a watershed could severely hinder these derivative analyses. The lack of any snow pits deeper than 2 m in many prior lidar-based SWE mapping studies (e.g., Broxton et al., 2019b; Meehan et al., 2024) limits the utility of these prior datasets for snowpack quantification in alpine mountain regions, where drifts frequently exceed several meters in depth (Fig. 11).

Existing empirical snow density models systematically underestimate the spatial heterogeneity of snow density in the WRR, affecting snowpack analyses at scales ranging from individual snow pits to kilometer-scale watershed metrics (Figs. 8–10). We identify three primary sources of missing heterogeneity: (1) underestimation of snow density in drifted areas affected by wind-packing, (2) failure to quantify systematic differences in snow density between forested and open alpine areas, and (3) missing topographic and microclimatic controls on snow density.

Although each of the prior literature density models (Table 1) considers snow depth as a predictor variable for density, the models tend to underestimate the magnitude of this relationship in the WRR (Fig. 8). Across all 36 snow pits, the literature models have mean biases between -3 % and -23 %, which is a reasonable level of skill given that these models are based on mean annual densification rates measured in various locations globally, and the literature models are not necessarily intended for precise snow density prediction at a particular location in a particular year. Nevertheless, it is noteworthy that all five of the literature models systematically underestimate the density of the deep nivation hollow drift (Fig. 4) by -14 % to -35 %, even the models with minimal bias across all snow pits (Fig. 8). The underlying measurements upon which these models are based do not capture the full range of natural snowpack variability. For example, despite the vast amount of data (11 147 depth-density pairs) used by Jonas et al. (2009), there are no pairs with snow depth > 3.5 m (Fig. 3 of that study). Moreover, deep snow measurements are relatively infrequent in these large model-fitting datasets, leading to models that do not fully capture this heterogeneity (even within the maximum measured range) due to regression dilution and mean centralization. Despite the much smaller size of our dataset (36 snow pits), we structured our survey design to strategically include extreme endmembers, including two complete pits in high- and low-elevation deep drifts of 4–6 m (Figs. 4 and 6). Across the three-watershed region considered here (Fig. 11), 21 % of all non-zero lidar snow depth measurements have depth > 2 m, constituting an estimated 56 % of the total SWE volume, while 22 % of our snow pits are at least 2 m deep, suggesting that deep snow density may still remain under-represented. Prior studies have sometimes only achieved weak or inconclusive empirical models of snow density (e.g., R2= 0.29 to 0.39 for modeled vs. measured density, Broxton et al., 2019b), which could be partially attributable to sampling a small range of snowpack variability (e.g., snow pit depth range of 0.15 to 1.18 m in that study). Our wide range of sampled snow densities and depths (Fig. 2) manifests in an empirical model describing snow density variations across a relatively large range (R2= 0.76 for 34 snow pits or 0.94 for the eight pits > 2 m, Figs. 7–8).

Snow is systematically denser in open areas compared to forested areas (Figs. 2 and 7). This fact, which is trivial to anyone who spends much time walking in the mountains, is nevertheless rarely tested statistically. The forest-open density contrast has been recognized by other studies (Bonner et al., 2022; Meehan et al., 2024), but this distinction remains unaccounted in most large-domain empirical density models (Table 1). Surprisingly, it appears that gridded canopy cover datasets (e.g., RCMAP used here) have not been previously considered as a predictor for empirical snow density models despite the widespread availability of gridded vegetation datasets. Meehan et al. (2024) did consider gridded vegetation metrics derived from lidar (height and distance to vegetation) in their small-domain density extrapolation, but our use of a continental-scale forest cover dataset is more germane to water supply studies across large river basins. For example, gridded canopy cover would naturally fit into the Hill et al. (2019) modeling framework, but it was not considered in their initial list of potential variables (Eq. 6 of that study) despite the consideration of other gridded terrain datasets (e.g., elevation). Perhaps one reason for the historic oversight of gridded vegetation data in prior snow density regressions could be attributed to the overrepresentation of in-situ stations in forested regions, obscuring the forested-open density contrast in the convenience-sampling datasets upon which prior empirical models are often based (e.g., Bormann et al., 2013; Hill et al., 2019). Additionally, SNOTEL-type monitoring stations are typically located within forest gaps, blurring the distinction between “open” and “forested.”

An important caveat on our inter-model comparisons is that many of the tested models are being pushed beyond their intended scope. For example, we evaluate the Sturm et al. (2010) model using the regression for the alpine climate class based on the map from Sturm et al. (1995), but Sturm et al. (1995) note that mountain snowpacks are a special case that may not fit into any single other climate class because of the high degree of spatial heterogeneity. This caveat from Sturm et al. (1995) seems to present a roadblock to SWE estimation across large mountain domains, since the meteorological data required to choose the appropriate climate classes are not available at high resolutions over entire mountain ranges. However, our nonlinear regression with only five independent variables (Eq. 1) can explain most of the observed variability (Fig. 8), suggesting that the problem of mountain snow heterogeneity is far from unsolvable. Since these snow pit locations are strategically situated to capture likely endmembers of variability, and since similar patterns repeat across our WRR survey datasets from different years (Fig. 8), it appears that our regression approach can successfully explain the majority of density heterogeneity across a mountain region using a parsimonious and physically interpretable set of five variables that are available in any lidar-surveyed region (depth, elevation, forest cover, and north/east slope). Further evaluation of our regression model structure in other mountain ranges with different climates and vegetation could inform to what degree these relationships are generalizable beyond the WRR.

Much of the missing heterogeneity in prior snow density models (Sect. 4.2) appears to stem from the non-representativeness of in-situ monitoring sites where snow density data are routinely collected, since several of the empirical models are fitted to similar in-situ datasets. In the WRR specifically, all 13 of the currently active (and one other inactive) SNOTEL stations are located below treeline, with a maximum elevation of 3070 m, which is lower than 97 % of the SWE volume at the time of our 2025 surveys (Fig. 12). As a result, all but three of the WRR stations were snow-free at the time of our 2025 field surveys. These three stations show inconsistent day-to-day variability indicative of snow bridging and other quality control issues (Goodison et al., 1981), which is exacerbated by the shallow low-elevation snowpack at this time in the season (depth range 0.05–0.51 m). Further, colocation issues between the snow pillow (which measures SWE) and the snow depth sensor (which measures snow depth) could contribute to inconsistent density calculations, especially when the low-elevation snowpack is shallow and patchy (Goodison et al., 1981). More broadly, all of the 88 active SNOTEL stations in the State of Wyoming are located in forested settings, with adjacent trees visible in each of the station photos. Figure S3 to Boardman et al. (2025) further illustrates how SNOTEL stations do not capture the range of density variability observed in the WRR. Empirical density models based on SNOTEL thus lack the data necessary to constrain snow density variations above treeline in the WRR (Fig. 9). This missing heterogeneity highlights the statistical problem with the convenience sampling used by Bormann et al. (2013), Hill et al. (2019), and others, whereby the automatically collected and easily accessible SNOTEL dataset is used for regression modeling without consideration of whether the constituent stations actually constrain the relevant axes of variability (e.g., forested vs. open, high vs. low elevation, deep vs. shallow snow).

Our representativeness score introduced in Sect. 2.2.1 provides robust statistical evidence that our snow pits in deep alpine drifts are more representative of the watershed snowpack compared to the SNOTEL locations (Fig. 13). The two most-representative snow pits have snow depths around 2.6 m with densities around 490 kg m⁻³, more than twice as deep as the deepest ablation-season snow depth (after peak SWE) recorded at any SNOTEL in Wyoming in 2025 (1.16 m). As discussed previous (Sect. 4.1), the absolute deepest drifts are relatively rare across the landscape, but nevertheless the two deepest snow pits (4.0 and 5.9 m) have the third- and fourth-highest representativeness scores. From Fig. 13, it is obvious why extrapolating from the SNOTEL depth-density relationship (or bias-correcting physical models based on SNOTEL data) could lead to unreliable results in the WRR, because the SNOTEL data overlap the least-representative snow pit locations, with anomalously shallow depths and widely scattered densities. Databases built from manual field measurements also reflect the common practice of ignoring heterogeneity caused by snow drifts. For example, Hill et al. (2019) suggest that participants in the Community Snow Observations program should “avoid measurements in areas of significant wind scour or deposition” in order to “ensure their recorded depth measurements are as representative as possible,” but wind drifts can actually be the most representative types of snow in mountain environments (Fig. 13). Similar representativeness problems occur widely when comparing small-scale measurements (i.e., in-situ or lidar-derived) to kilometer-scale grids (Meromy et al., 2013). Although the low elevation bias of stations in the SNOTEL network (strictly below treeline) generally causes an underestimation of the snowpack in alpine terrain (Molotch and Bales, 2006), sampling exclusively flat locations can also lead to snowpack overestimation in some cases due to under-sampling of wind-scoured areas (Grünewald and Lehning, 2015).

Boardman (2025) previously argued that kilometer-scale snow transport and redistribution is an underappreciated driver of mountain snowpack patterning in areas with alpine plateaus, and our analysis here indicates that common spatial snowpack datasets are indeed missing a considerable degree of kilometer-scale heterogeneity in the WRR (Fig. 11). This missing heterogeneity is ultimately explained, at least in part, by extrapolation from non-representative SNOTEL measurements. All three widely available near-real-time gridded SWE datasets analyzed here (SNODAS, U. Colorado SWE, U. Arizona SWE) are based at least in part on the SNOTEL stations discussed previously (Figs. 11–12). The SNODAS product uses a physically based snow model that is “nudged” to better match SNOTEL data, satellite-derived snow-covered area, and potentially other observations when available, though the precise methodology for data selection and assimilation in a particular region or year remains opaque and subject to agency personnel decisions (Barrett, 2003). The U. Colorado SWE product (sometimes called SWE-fusion) uses generalized linear regressions to extrapolate spatial SWE from SNOTEL data using physiographic variables, historical SWE patterns (where available), and satellite-derived snow-covered area (Schneider and Molotch, 2016; Yang et al., 2022). The U. Arizona SWE product (sometimes called SWANN) uses artificial neural networks to extrapolate spatial SWE from SNOTEL data and other in-situ data using gridded climatology and physiographic variables (Broxton et al., 2016, 2019a, 2024). All of these snowpack datasets are available in near-real-time (i.e., within a few days of the nominal date) and are widely used in research across the western USA related to water supply forecasting and other watershed-scale snowpack analyses. Also of note, each of these SWE products is dependent (at least in part) on the SNOTEL network, which is not representative of alpine mountains (Figs. 12–13).

Despite the limitations of gridded SWE datasets extrapolated from sparse measurements, these products are commonly used to represent heterogeneity without testing their representativeness. For example, Raleigh et al. (2025) used SNODAS and U. Arizona SWE as supplemental datasets to explicitly investigate spatial SWE heterogeneity across watersheds, but in wind-drifted alpine environments, these datasets completely fail to capture the observed patterns at 500 m scale, rendering it futile to search for “hot spots” within datasets that are neither spatially accurate nor statistically representative (Fig. 11). Other studies have likewise used, or perhaps misused, these extrapolated gridded datasets to represent grid-scale spatial heterogeneity across watersheds: SNODAS was used to optimize monitoring station locations (Keum et al., 2018), SNODAS was used for calibration of spatially distributed models (Tiwari et al., 2024), and U. Arizona SWE was used to investigate spatial interactions between the snowpack and forest fires (Abolafia-Rosenzweig et al., 2022). It is unsurprising, yet critically important, that snow drifts and other deep accumulation zones are entirely missing from large-domain gridded datasets, because extrapolation from non-representative station data logically leads to a dramatic underestimation of spatial heterogeneity at watershed scales (Figs. 11–13 and Table 2). Researchers considering SNODAS, U. Colorado SWE, or U. Arizona SWE as interchangeable proxies for “spatial SWE datasets” need to understand that these datasets are extrapolated from sparse measurements, which are not necessarily representative of the mountain regions where most snow is stored (Figs. 12–13). Thus, these extrapolated spatial datasets may be useful for estimating a lumped regional snowpack index, but they are unsuitable for studies addressing heterogeneity in the WRR (and likely also in similar alpine regions across the western USA). Our study also underscores the importance of independent validation data for testing common snow data products, which can otherwise become effectively unfalsifiable due to assimilation of all available data (Hedrick et al., 2015). We do not intend to diminish the potential value of extrapolated SWE datasets as a first-order estimate of the lumped watershed-scale snowpack storage, but using these datasets for spatially explicit applications is misguided in alpine environments that lack representative in-situ data, like the WRR (Table 2).

Gridded SWE products that do not explicitly consider snow transport processes are unlikely to capture the spatial variability of SWE in windy landscapes (Lv et al., 2019), which poses a challenge for water resource assessments in the mountainous western USA (Li et al., 2017). For example, the red arrows in Fig. 11 indicate a particularly notable glacial cirque (43.370°, -109.671°) that collects snow from a roughly 3 km² upwind contributing “snowshed” (Boardman, 2025b). At 500 m resolution, this cirque location is represented by four grid cells with mean SWE of 1.1 to 1.7 m based on our 2025 lidar and snow pit density model. Within the same four 500 m grid cells, the non-lidar gridded datasets show mean SWE of 0.34–0.39 m (SNODAS), 0.25–0.46 m (U. Colorado), and 0.37–0.53 m (U. Arizona). Note that the U. Colorado dataset masks glacier locations, so those values are interpolated from neighboring cells (raw gridded SWE products are shown in Fig. S6). Comparing the total volume in these four grid cells, the non-lidar SWE datasets underestimate the deep snow storage in this cirque basin by about 65 –73 % despite much closer agreement with the mean SWE across the whole landscape (e.g., as low as 2 % difference for SNODAS, Table 2). Analogously, the five literature density models (Table 1) underestimate the snow storage in the same four grid cells by 9 %–27 %. Accurate measurement of snow depth (from lidar) greatly reduces the uncertainty of deep snow quantification regardless of the density assumption (Fig. 11), but underestimating the density of deep drifts and avalanche debris (Figs. 4 and 6) still causes a substantial underestimation of heterogeneity when applying the prior literature snow density models. Thus, accurate spatial SWE quantification requires accounting for both the extreme depth and extreme density of snow in cirque basins, nivation hollows, and other enhanced accumulation zones affected by both wind transport and avalanches (Fig. 6). Ultimately, these deep drift zones and glacial cirques create strong controls on hydrograph shape and upland water supply during the growing season (Boardman et al., 2025), so the enhanced heterogeneity of windy alpine landscapes cannot be ignored in sub-seasonal snow water supply forecasting and other water resource applications that are sensitive to runoff timing or peak flows (Pfohl et al., 2026).

In a scientific era emphasizing “big data” and remote sensing (e.g., Dozier, 2011), tedious manual field measurements remain essential. Deep drifts, avalanche runout zones, and steep slopes are globally underrepresented in snow monitoring networks (e.g., Bormann et al., 2013) for obvious practical reasons. Unfortunately, these extreme alpine settings are also typically difficult and/or dangerous to sample manually. To obtain the data in Fig. 6, the lead author assumed personal responsibility (outside the scope of any institutional affiliation) for entering extreme terrain using technical mountaineering gear. These actions were taken after informed acceptance of the associated risks by an individual with extensive personal mountaineering experience in the study region, and obtaining similar measurements could be considered reckless in other circumstances. Moreover, the large amount of manual labor required to sample a very deep drift (Fig. 4) is also prohibitively time-consuming in most circumstances (back-of-the-envelope calculations show that we moved on the order of 10 000 kg of snow just to sample the single deepest 5.9 m snow pit). Our approach also necessitates coordinating the timing of fieldwork with flight planning logistics, which can be challenging in mountain environments with variable weather and multi-day backcountry approach hikes. Thus, comprehensive field surveys of alpine snow density may not always be practical for real-time water supply forecasting applications.

However, even sparse opportunistic measurements from extreme alpine environments can help constrain patterns of snow density that repeat across years (Figs. 8 and 10). These heterogeneous snow density zones are missing from the in-situ station network (Figs. 12–13) and can advance our understanding of the physical processes underlying variations in snowpack stratigraphy and densification (Figs. 3 and 5). A direct application of our regression model (Eq. 1 and Table 1) to other regions is unlikely to yield the correct mean density due to site-to-site and year-to-year variability, but we anticipate that our model might be useful to indicate the potential degree of spatial heterogeneity in other alpine regions associated with deep drifts and forest versus open areas. We suggest that a closer collaboration between snow scientists and the technical mountaineering community would be useful to further constrain snow heterogeneity in rarely sampled extreme alpine settings globally, but we emphasize that such measurements must be conducted responsibly, perhaps through partnership with professional mountain guides or similar services.