ICESat-2 was launched in 2018 in a precessing orbit with an altitude of ∼500 km and inclination of 92°, which allows for measurements up to 88° N latitude with a 91 d orbital repeat cycle (Markus et al., 2017). ICESat-2 carries the Advanced Topographic Laser Altimeter System (ATLAS), which is a single wavelength (532 nm), high repetition rate (10 kHz) lidar system with photon counting detectors (Markus et al., 2017; Neumann et al., 2019). Each ATLAS laser pulse is split into 3 simultaneous beam pairs (one strong and one weak beam per pair) by a diffractive optical element. The 3 beam pairs are separated by about 3 km across track. Atmospheric backscatter is obtained by ATLAS using only the three strong beams, spanning from the surface to an altitude of 14 km, with an along-track resolution of approximately 280 m and a vertical resolution of 30 m. Each 280 m ICESat-2 atmospheric profile represents the aggregate of 400 individual ATLAS laser shots (Palm et al., 2021). In this study we use ICESat-2 strong beam 1 observations from version 6 of the ATLAS/ICESat-2 Level 3A (ATL09) calibrated backscatter profile product (Palm et al., 2023).

The algorithm used to detect blowing snow in ATLAS backscatter profiles is adapted from the CALIOP approach (Palm et al., 2011) and further detailed in Palm et al. (2021, 2022). When a surface return is identified and the 10 m wind speed from NASA's GEOS-5 FP-IT analysis exceeds 4 ms-1, the algorithm compares the near-surface atmospheric backscatter to the expected molecular (Rayleigh) signal. If the observed signal exceeds a fixed multiple of the molecular scattering, the algorithm steps upward through each vertical bin until the backscatter drops below an adaptive threshold (typically ∼2×10-5 m-1sr-1). To be flagged as blowing snow, the detected feature must touch the ground and be shallower than 500 m. Retrievals deeper than 500 m are classified as diamond dust, which can stretch for a kilometer or more vertically and frequently reaches the ground (Intrieri and Shupe, 2004). Further, we use the version of the blowing snow algorithm described in Robinson et al. (2025) which includes modifications to help alleviate several challenges unique to the Arctic. These modifications serve to (1) minimize the misidentification of low clouds as blowing snow and (2) correct for the attenuation due to transmissive clouds.

Once blowing snow is retrieved, its properties (geometric and optical depths) are logged. Optical depth (OD) is estimated as the sum of the backscatter within the blowing snow retrieval multiplied by the product of the bin depth (30 m) and the extinction to backscatter (lidar) ratio. A lidar ratio of 25 sr is used, which is a typical value for ice crystals in cirrus clouds (Chen et al., 2002; Josset et al., 2012). To infer blowing snow particle number density, transport flux, and sublimation flux from the observed ICESsat-2 backscatter we follow the same approach as described in Palm et al. (2017) and Robinson et al. (2025), which relies on meteorological fields (10 m wind speed, 2 m temperature, and 2 m relative humidity over ice) from the NASA GEOS-5 FP-IT analysis (run at 0.5° latitude × 0.625° longitude; Lucchesi, 2015) as well as assumptions about blowing snow particle size. As in Robinson et al. (2025) we use the formulation r(z)=5.05×10-5z-0.085 to estimate the particle radius (r, m) as a function of altitude (z, m). This fit was constrained by observations of blowing snow particle sizes during the 2019–2020 Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) campaign.

To improve signal-to-noise in sunlit conditions, we apply along-track averaging to the ICESat-2 observations when the solar elevation angle exceeds -7°, a threshold beyond which background solar photons begin to significantly degrade sensitivity. Under these conditions, which affect late February through April (Fig. S1 in the Supplement), increased solar background can reduce the detectability of low-backscatter features such as blowing snow. To mitigate this, we average the native 25 Hz (280 m) profiles to 1 Hz (∼7 km) resolution, effectively reducing solar background noise and enhancing the reliability of blowing snow retrievals. While this approach lowers spatial resolution, it reduces false positive detections and provides a more robust estimate of blowing snow properties under marginal lighting conditions without introducing significant biases in seasonal statistics.

SnowModel-LG is a physics-based snow-on-sea ice model forced by atmospheric inputs of air temperature, RH, winds, and precipitation from the NASA Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA-2; Gelaro et al., 2017) as well as sea ice inputs of concentration and parcel motion (Tschudi et al., 2019, 2020). At each 3 h timestep, SnowModel-LG performs mass-budget calculations (e.g., Eq. 1) where SWE depth evolution is accounted for by snow gains, losses, and sea ice dynamics (Liston et al., 2020).

The MicroMet module (Liston and Elder, 2006) is used to time average (1-hourly to 3-hourly) and distribute the MERRA-2 fields (0.5° latitude × 0.625° longitude) to the sea ice parcels. As part of this procedure, the MERRA-2 water equivalent precipitation is bias corrected (as described in Sect. 2.5 and Table 1 of Liston et al., 2020) and partitioned into snowfall and rainfall based on environmental conditions (Dai, 2008).

Blowing snow in SnowModel-LG is accounted for by SnowTran-3D (Liston et al., 2007, 2018; Liston and Sturm, 1998). The snow threshold friction velocity, u∗t, is calculated as a function of snow density, ρs, which is related to snow strength and hardness. Snow density evolution includes the history of temperature, precipitation, and wind-transport. When the friction velocity exceeds the threshold value, snow begins to be lifted off the surface, first into the saltation layer (several cm thick) and then into the turbulent suspension layer (several m thick). The vertical mass concentration in the blowing snow profile is estimated following Liston and Sturm (1998) and is combined with the environmental conditions to calculate transport and sublimation fluxes. Mass transport is related to the windspeed and vertical mass concentration. SnowModel-LG's blowing snow sublimation is calculated as a function of several factors, including the vertical mass concentration, temperature-dependent humidity gradients between the snow particles and the atmosphere, conductive and advective energy- and moisture-transfer mechanisms, particle size, and solar radiation. The SnowModel-LG blowing snow transport and sublimation fluxes represent column integrated values in units of kgm-1s-1 and cmSWEd-1, respectively. SnowModel-LG variables are output as 3-hourly values on an EASE grid with a resolution of 25 km.

We also include estimates of the bulk blowing snow sublimation rate (Qbs in Eq. 1) using the approach described by Déry and Yau (1999, 2001) and subsequently Yang and Yau (2007). Throughout the analysis we refer to this approach as DY2001. We chose to include it because it is computationally efficient and has been widely applied in studies of blowing snow aerosol production over sea ice (e.g., Gong et al., 2023; Frey et al., 2020; Huang et al., 2020; Huang and Jaeglé, 2017; Yang et al., 2008, 2019). Sublimation depends on several factors including surface windspeed, temperature, and humidity deficit.

Following Yang et al. (2008), sublimation is scaled by snow age A′ which accounts for the reduced ease of wind lofting as snow ages. For a full description of the sublimation calculation used here, we refer the reader to Sect. 2.1.1 of Yang et al. (2008). In our calculations, we adopt a representative mean snow age of 3 d over Arctic sea ice (Huang and Jaeglé, 2017).

A key factor controlling blowing snow occurrence in DY2001 is the threshold windspeed, which follows Li and Pomeroy (1997a). The threshold windspeed (Ut=6.975+0.0033[T2m+27.27]2) is estimated from the 2 m surface air temperature (T2m) and has a minimum value of ∼7 ms-1 at an air temperature of -27 °C. At both higher and lower temperatures, the threshold wind speed will be larger (maximizing at ∼10 ms-1 for temperatures near 0 °C). We estimate the DY2001 threshold windspeed and blowing snow sublimation using the same meteorology (10 m windspeed, 2 m temperature, and 2 m RH_ice) used to derive the ICESat-2 sublimation.

We aggregate the ICESat-2 observations to a National Snow and Ice Data Center (NSIDC) Equal-Area Scalable Earth (EASE) grid (Brodzik and Knowles, 2002) with a horizontal resolution of 100 km. This resolution balances spatial detail with observational coverage, ensuring sufficient ICESat-2 sampling within each grid cell while minimizing noise that would arise at finer resolutions due to the narrow swath of the lidar. Temporal resolution is determined by the duration of the binning period, allowing flexibility to examine daily, seasonal, or multi-year patterns.

Within each 100 km grid cell, the ICESat-2 blowing snow occurrence for a specified time window is computed as the number of profiles with a blowing snow detection divided by the total number of valid profiles. A valid profile is defined as one where the surface return is clearly detected, which excludes profiles with optically thick cloud cover (optical depth >3), where surface detection is unreliable or is not achieved. For blowing snow properties such as geometric and optical depths, only blowing snow retrievals are gridded.

For comparison with model estimates, we extract values from the SnowModel-LG fields (25 km resolution) by sampling the nearest-neighbor grid point to each valid ICESat-2 profile location. These sampled values are then binned to the same 100 km EASE grid alongside the ICESat-2 data. We apply the same procedure to the DY2001 estimates: values are first computed at the location of each valid ICESat-2 profile, and the resulting fields are aggregated onto the 100 km grid for direct comparison with both ICESat-2 observations and SnowModel-LG outputs.

Figure 1 highlights a blowing snow storm which occurred over the Central Arctic on 10 December 2022. During an orbit which transited from the Canadian Arctic Archipelago towards Svalbard, ICESat-2 retrieved blowing snow for roughly 1200 km along track, with depths up to 250 m and observed attenuated backscatter exceeding 1.50×10-4 m-1sr-1 (Fig. 1a). In this region, MERRA-2 windspeeds ranged from 7.5 to 15 ms-1 (blue line, Fig. 1b) and SnowModel-LG predicted intense blowing snow, with mass fluxes peaking at 4 Mgm-1d-1 (green line, Fig. 1b). The strongest ICESat-2 observed and SnowModel-LG predicted blowing snow occurred coincident with the strongest winds (middle of Fig. 1a and b). While ICESat-2 did retrieve blowing snow to the west of this maximum (left side, Fig. 1a) coincident with windspeeds >8 ms-1, SnowModel-LG predicted only minimal blowing snow mass transport.

At the location of intense blowing snow, MERRA-2 air temperatures ranged from -25 to -15 °C (black line, Fig. 1c) and the air was subsaturated with respect to ice (85 %–95 %; red line, Fig. 1c). This combination of meteorological factors resulted in substantial blowing snow sublimation inferred from ICESat-2 and predicted by SnowModel-LG (Fig. 1d). ICESat-2 sublimation maximized at 0.07 cmSWEd-1 coincident with the strongest winds and driest conditions (magenta line, Fig. 1d). In the same region SnowModel-LG sublimation reached 0.05 cmSWEd-1 (green line, Fig. 1d).

Winds in excess of 8 ms-1 covered much of the Central Arctic and coincided with tight sea-level pressure (SLP) gradients stretching from the Beaufort to Lincoln Sea (Fig. 2a). SnowModel-LG predicted blowing snow mass transport >0.20 Mgm-1d-1 over an area of 750 000 km2 (Fig. 2b), which is slightly larger in size than the state of Texas. Given a total Central Arctic area of roughly 3.2×106 km2, this storm impacted about a quarter of the basin.

To examine the spatial distribution of ICESat-2 profiles, we first gridded the ICESat-2 orbits to the 100 km grid (Sect. 2.4) and then assigned each grid cell to one of four categories: blowing snow, mixed, clear air, or cloud attenuated. If more than 70 % of all profiles were attenuated due to clouds, the grid cell was labeled as cloud attenuated. We assigned the other three categories based on the occurrence of blowing snow: blowing snow if more than 50 % of profiles were blowing snow, mixed if 15 %–50 % of profiles were blowing snow, and clear air if less than 15 % of profiles were blowing snow. ICESat-2 grid cells in the western Central Arctic were consistently classified as blowing snow (magenta colors, Fig. 2c), coinciding with the strongest winds and the highest SnowModel-LG predicted transport. The total area of ICESat-2 grid cells labeled as blowing snow was 740 000 km2, closely matching the SnowModel-LG predictions and confirming that the blowing snow was synoptic in scale, covering much of the Central Arctic.

Figure 3 shows the mean multi-year blowing snow occurrence and properties derived from the ICESat-2 observations for November through April 2018–2023. To generate the average maps, we grid each cold season independently (following Sect. 2.4) and then average the five cold seasons together. We found a significant fraction of the central Arctic experiences blowing snow frequencies >25 %, with maxima of near 35 % in the Fram Strait region (Fig. 3a). This is consistent with several previous studies which showed these regions have consistent influence (>15 % of the time) from storms entering the Arctic (e.g., Clancy et al., 2022; Valkonen et al., 2021). This is also evident in the spatial distribution of MERRA-2 windspeeds (Fig. 3d), where the region of high blowing snow occurrence frequency is collocated with average windspeeds >6.5 ms-1.

The delineation between first- and multi-year sea ice (Tschudi et al., 2020) lies north of the Fram Strait and visibly bifurcates the region of elevated ICESat-2 derived blowing snow occurrence (Fig. 3a). In a recent study focused on Svalbard, Li et al. (2025) found that observed increases in tropospheric bromine were correlated with air mass contact over sea ice under strong winds, and that boundary layer air masses reaching Svalbard spent more time over multi-year sea ice compared to first-year sea ice. The spatial pattern shown in Fig. 3a is consistent with these findings, where storms entering the Arctic drive strong north-northwesterly winds that preferentially advect air masses across the multi-year ice zone towards Svalbard.

The ICESat-2 occurrence frequency does not include shallow (<30 m thick) blowing snow layers, since these cannot be reliably detected at the vertical resolution of the atmospheric backscatter profiles. In addition, ICESat-2 cannot sample conditions where optically thick clouds prevent the surface from being detected. Regions of the Kara, Barents, and Greenland Seas are particularly susceptible to this under sampling, where the ICESat-2 cloud attenuated occurrence (% of all profiles where the surface cannot be detected) can exceed 50 % across much of the cold season (Fig. S2 in the Supplement).

The multi-year cold season ICESat-2 retrievals show blowing snow layers averaging ∼100 m in depth, ranging from ∼50 m up to 160 m (Fig. 3b). Our previous analysis of ICESat-2 observations near the 2019–2020 MOSAiC campaign demonstrated that low level turbulence often mixes blowing snow to the top of the surface inversion (Robinson et al., 2025), suggesting that blowing snow layer depth may serve as a useful indicator of Arctic inversion depth. Blowing snow optical depths average 0.12 across the Arctic, with maxima near 0.20 in the Fram Strait and southern Baffin Bay (Fig. 3c). These regions also experience thicker blowing snow layers on average. Figure 3 further shows that regions of deeper, optically thicker blowing snow are co-located with areas of high occurrence frequency and stronger winds.

Figure 4 shows that the ICESat-2 pan-Arctic blowing snow occurrence frequencies are consistent from year-to-year at 18 %–20 %. The spatial pattern of occurrence also remains fairly consistent, with the Central Arctic and Fram Strait displaying the highest frequencies and only moderate shifts in location. Despite this, the Central Arctic can display substantial year-to-year variability. For example, the highest (2019–2020) and lowest (2020–2021) pan-Arctic frequencies were observed in consecutive cold seasons.

The contrast between these two cold seasons appears closely aligned with large scale climate and atmospheric circulation patterns, particularly the Beaufort High and the Arctic Oscillation (AO). In early 2020, a record positive AO phase (+3.5, top row Fig. S3 in the Supplement) coincided with a collapse of the Beaufort High, enhanced cyclone activity (Ballinger et al., 2021; Rinke et al., 2021), and widespread blowing snow. From January to March 2020, MERRA-2 sea-level pressure (SLP) and windspeed featured an elongated region of consistently low pressure (<1000 hPa) extending from Iceland into the ice-covered Kara and Barents Seas (Fig. 5a). Over these regions and the Central Arctic, mean windspeeds reached 7–9 ms-1 (Fig. 5b). During this period, ICESat-2 observed several intense blowing snow episodes covering more than 25 % of sea ice area (blowing snow >1×106 km2; Fig. S4 in the Supplement), with mean pan-Arctic blowing snow frequencies of 21.8 %, reaching up to 50 % in the Central Arctic (Fig. 5c).

In contrast, the 2020–2021 season was marked by a strong negative AO (-2.4, top row Fig. S3) and a persistent Beaufort High (mean MERRA-2 SLP >1020 hPa across most of the Arctic basin, Fig. 5d), conditions known to suppress storm activity (Kenigson and Timmermans, 2021; Serreze and Barrett, 2011). Consistent with this pattern, MERRA-2 windspeeds were on average ∼2 ms-1 lower relative to January–March 2020 (Fig. 5e). From December 2020 to February 2021 ICESat-2 detected substantially less blowing snow (47 % lower relative to January–March 2020), with frequencies in the Central Arctic maximizing at only ∼25 % (Fig. 5f). Across all months, we find a moderately strong correlation between AO phase and ICESat-2 blowing snow occurrence (r=0.62; Fig. S3c). Composites highlight this relationship: positive AO months (N=18; Fig. 5g) exhibit 20 % more blowing snow than negative AO months (N=12, Fig. 5h), with particularly large differences (up to a factor of two) in the Fram Strait and Central Arctic.

In the following section we focus on the Central Arctic region during January–March, the region most well-sampled by ICESat-2 and months least affected by optically thick clouds (Fig. S2). To examine relationships between meteorological factors and blowing snow, we use daily 100 km grid-cell averages. Although this lowers the total number of samples compared to a profile-based approach, averaging helps to reduce noise.

Figure 6 compares the blowing snow occurrence as a function of windspeed and temperature. For comparison to ICESat-2 and SnowModel-LG, the blowing snow occurrence from Li and Pomeroy (1997b) is also shown (see their Eq. 7). The blowing snow occurrence from Li and Pomeroy (1997b) is based on a statistical analysis of observations for 16 stations on the prairies of western Canada and is a function of windspeed, temperature, and snow age (assumed in our analysis to be 72 h). It is also in contrast to DY2001, where the threshold windspeed essentially acts as an on-off switch for blowing snow. ICESat-2 retrievals indicate a 10 %–40 % blowing snow occurrence below the DY2001 threshold of ∼7 ms-1 (black dashed line, Fig. 6a), with a much stronger dependence on windspeed than on temperature (Fig. 6a). For example, at 8 ms-1, the ICESat-2 occurrence is 50 %–60 % across all temperatures, while at -25 °C it rises from 10 %–15 % at 4 ms-1 to >80 % at 15 ms-1. SnowModel-LG predictions (defined as blowing snow transport >0.20 Mgm-1d-1) display frequencies ∼10 % larger than ICESat-2 on average but capture similar features (Fig. 6b). The occurrence of blowing snow predicted from Li and Pomeroy (1997b) displays a narrower transition region, increasing sharply from <20 % to >60 % over the 8–10 ms-1 range (Fig. 6c).

The one-dimensional distributions (Fig. 6d–f) further emphasize the dominant control of windspeed, with all three datasets showing increasing occurrence with stronger winds. ICESat-2 and SnowModel-LG show a weak temperature dependence, with slightly lower occurrence at higher temperatures, especially for stronger winds, consistent with enhanced snow cohesion and bonding resistance (Fig. 6d and e). The Li and Pomeroy (1997b) formulation shows a stronger temperature sensitivity, ranging from 75 % at T<-30 to 20 % at T>-5 °C for a 10 ms-1 windspeed (Fig. 6e). The temperature dependence is likely stronger because of our assumption of a fixed snow age of 72 h. Snow age also influences bonding and cohesion, with older snow being more resistant to erosion. Because SnowModel-LG and ICESat-2 sample a range of snow ages, their apparent temperature dependence is likely weaker.

ICESat-2 blowing snow properties also show a strong dependence on windspeed (Fig. 7a). Median blowing snow layer height increases from 30 m at windspeeds of ∼4 ms-1 to more than 150 m at windspeeds >14 ms-1. Optical depth exhibits a similar relationship, rising from 0.02 to 0.26 over the same windspeed range. The spread in both height and optical depth (shading, Fig. 7a) also widens with increasing windspeed, which we attribute to increased noise from fewer observations in the highest windspeed bins (Fig. 7b).

The increase in blowing snow optical depth reflects a combination of increased blowing snow height and stronger backscatter signal (Fig. 7c). Across nearly 700 000 ICESat-2 retrievals in February 2022, near-surface blowing snow extinction increased by 40 % from 1.5×103 Mm-1 at 4 ms-1 to 2.1×103 Mm-1 at 14 ms-1. The enhancement is even larger aloft (a factor of 2–3). Together, these results indicate that stronger winds loft more blowing snow higher into the atmosphere, consistent with previous studies (Palm et al., 2011, 2018b; Robinson et al., 2025).

Figure 8a shows the spatial distribution of blowing snow transport flux inferred from ICESat-2. The flux is calculated by combining the ICESat-2 derived mass concentrations with the vertical profile of windspeed, integrated over the depth of the blowing snow layer. The pan-Arctic mean transport flux observed by ICESat-2 is 74 Mgm-1, with maxima >160 Mgm-1 in the Central Arctic, co-located with regions of frequent and intense blowing snow (Fig. 3). SnowModel-LG produces a similar spatial distribution but yields transport fluxes that are 2–3 times lower (Fig. 8b). This discrepancy likely arises because SnowModel-LG confines blowing snow to the lowest several meters of the atmosphere, where winds are weaker. In contrast, ICESat-2 detects blowing snow layers extending several hundred meters above the surface (Figs. 3b and 7a), where stronger winds enhance snow transport. To support this interpretation, we examined the pan-Arctic blowing snow burdens (mass per square meter; Fig. S5 in the Supplement) and found that they agree to within about 20 % between ICESat-2 (0.17 gm-2) and SnowModel-LG (0.14 gm-2). In the Central Arctic regions of enhanced transport, both datasets have mean blowing snow burdens of up to 0.40 gm-2.

Although the spatial pattern of transport broadly agrees, our seasonal values are smaller than those reported by Yang et al. (2010). Their simulations for December 2006–February 2007 suggested transport fluxes up to 800 Mgm-1 in the Central Arctic and >1000 Mgm-1 along Greenland's east coast. These higher values could reflect methodological differences: their model did not explicitly account for variable snowpack conditions, which could lead to an overestimate in blowing snow occurrence and transport, and was run at finer spatial (18 km) and temporal (5 s) resolutions, which could capture small-scale wind gradients and localized enhancements in snow redistribution. Despite these differences, both our results and those of Yang et al. (2010) indicate that blowing snow transport plays a relatively minor role in the basin-scale snow budget. For example, the divergence of ICESat-2 transport (Fig. 8c) is limited to a few tenths of mm SWE, with localized maxima near 1 mm SWE in regions of frequent blowing snow. We further examined the divergence separately for each cold season (Fig. S6 in the Supplement), finding that basin-wide averages remain <10-3 mm SWE. The divergence exhibits interannual variability which is largely tied to prevailing meteorological conditions and blowing snow occurrence. For example, we find substantially greater ICESat-2 inferred divergence during 2019–2020 compared to 2020–2021, consistent with the AO-phase dependence of blowing snow discussed above (Fig. 5).

Figure 9 shows the mean total annual blowing snow sublimation and snowfall for the 2018–2023 cold seasons. Pan-Arctic blowing snow sublimation totals from ICESat-2 (1.63 cm SWE) are in close agreement with SnowModel-LG (1.66 cm SWE) and within 30 % of DY2001 (2.07 cm SWE). All three estimates are broadly consistent with previous modeling studies (Chung et al., 2011; Liston et al., 2020; Yang et al., 2010). In the Central Arctic near Svalbard, ICESat-2 indicates the highest values of sublimation (3–4 cm SWE). A secondary maximum (>3 cm SWE) occurs in the Barents Sea, where blowing snow is retrieved half as often. This reflects the sensitivity of sublimation to temperature and humidity, because the marginal seas are generally warmer than the Central Arctic (Fig. S7 in the Supplement). Thus, the reduced occurrence of blowing snow is offset by higher temperatures and lower humidity, which enhance sublimation.

The average blowing snow sublimation derived from ICESat-2 over first-year sea ice is 1.47 cm SWE (SnowModel-LG: 1.66 cm SWE; DY2001: 2.02 cm SWE), compared to 2.06 cm SWE over multi-year ice (SnowModel-LG: 1.81 cm SWE; DY2001: 2.21 cm SWE). While multi-year ice constitutes only 25 % of our study area, it accounts for 30 %–35 % of the seasonal blowing snow sublimation. This disproportionate contribution is consistent with the spatial pattern of blowing snow shown above (Fig. 3), where the region of elevated winds and occurrence is split by the transition between first- and multi-year ice.

We compare blowing snow sublimation to total MERRA-2 snowfall over the cold season (12.41 cm SWE, Fig. 9d). On average, we find that blowing snow removes 13.6 % (ICESat-2), 14.1 % (SnowModel-LG), and 16.9 % (DY2001) of snowfall. The regional impact, however, varies strongly (Fig. 9e–g). In the Kara and Barents Seas, where snowfall is highest, sublimation removes only 5 %–10 % of snowfall. In the Central Arctic losses increase to 18 %–24 %, while in regions with more moderate snowfall, such as the Beaufort Sea, sublimation losses can exceed 30 % (e.g., 2–3 cm SWE of sublimation compared to 8–10 cm SWE of snowfall).

The fraction of snowfall removed by blowing snow sublimation inferred from ICESat-2 reaches 30 % in the Beaufort Sea north of the Canadian Arctic Archipelago (Fig. 9e). SnowModel-LG and DY2001 show a similar enhanced offset, though their maxima are shifted southwestward along the coast of Alaska (Fig. 9f and g). The 2018–2023 period was marked by several strong Beaufort High episodes, such as the 2020–2021 event highlighted in Fig. 5, which are typically associated with calm, dry conditions. Under such conditions, ICESat-2 retrievals may occasionally overestimate blowing snow. False positives could arise when low-level ice crystals (ice clouds or diamond dust) mix with blowing snow, leading the entire ICESat-2 backscatter signal to misattributed to blowing snow. This effect was most pronounced during winter 2021–2022, when exceptionally warm (T>-20 °C) and dry (RHice<90 %) conditions prevailed north of the Canadian Arctic Archipelago (Fig. S8 in the Supplement).

The fraction of snowfall removed by blowing snow sublimation, as inferred from ICESat-2, also varies by ice type. On average, values over multi-year ice (15 %–22 %) are 1.6 times larger than over first-year ice (9 %–14 %). The enhanced offsets reflects both stronger sublimation and overall lower snowfall over multi-year ice (11.43 cm SWE) compared to first-year ice (12.64 cm SWE).

Along Greenland's east coast, DY2001 predicts much higher sublimation fluxes (4–5 cm SWE, >70 % of snowfall) than either ICESat-2 and SnowModel-LG (2–3 cm SWE, 20 %–30 % of snowfall). This discrepancy likely reflects DY2001's simple threshold-based parameterization, which tends to overpredict blowing snow at the typical windspeeds in this region (6–8 ms-1, Fig. 3). Warmer and drier conditions in this region (Fig. S7) further amplify the sublimation predicted by DY2001.

Daily pan-Arctic time series (Fig. 10) show that blowing snow sublimation is nearly continuous throughout the cold season, punctuated by sharp peaks during major storm events. The most intense episodes (>0.04 cmSWEd-1 averaged over sea ice) occur only a few times per season and correspond to widespread blowing snow detected by ICESat-2 (Fig. S4). These storms contribute disproportionately to the seasonal total, with individual events removing up to 60 % of daily snowfall (Fig. S9 in the Supplement). Between storms, sublimation persists at lower but steady rates (0.01–0.02 cmSWEd-1) and these background losses accumulate to a substantial share (35 %–40 %) of the seasonal total.

The ICESat-2 inferred sublimation ranges from 1.4 to 2.4 cm SWE across the five cold seasons (Fig. 10, bottom row), corresponding to a 11 %–20 % offset of seasonal snowfall. Both snowfall and blowing snow sublimation vary by 1–2 cm SWE year to year, but the two do not always covary. For example, the 2021–2022 cold season had the lowest snowfall (11.9 cm SWE) yet the highest ICESat-2 sublimation (2.4 cm SWE, 20 % offset). Conversely, 2018–2019 featured higher snowfall (12.9 cm SWE) but relatively low sublimation (1.4 cm SWE, 11 % offset). These interannual differences highlight that sublimation depends not only on storm frequency and strength (which also drive snowfall) but also on atmospheric conditions which regulate blowing snow occurrence and sublimation efficiency. SnowModel-LG and DY2001 generally agree with the ICESat-2 sublimation, though DY2001 tends to predict slightly higher values.

Blowing snow sublimation exceeds surface sublimation by a factor of 4–5, underscoring the dominant role of blowing snow in sublimation-driven snow loss during much of the cold season. The cumulative surface sublimation timeseries (Qss in Eq. 1) predicted by SnowModel-LG is shown in Fig. 10 (red lines, bottom row). Seasonal total surface sublimation averages only 0.3–0.5 cm SWE, with nearly all of it occurring from late February through April, when solar radiation increases, near-surface air warms, and RH_ice decreases. These values are lower than the 1–2 cm SWE reported by Déry and Yau (2002), likely because their annual means included the warmer spring and summer months. Consistent with this, SnowModel-LG calculates an Arctic-wide annual mean surface sublimation of ∼1 cm SWE.

ICESat-2 likely underestimates blowing snow sublimation because it cannot observe blowing snow beneath optically thick clouds. These conditions are most frequent during winter storms, when strong winds can drive intense sublimation. To assess this sampling bias, we examine the 2018–2023 SnowModel-LG and DY2001 predictions under all conditions (i.e., regardless of whether ICESat-2 detected the surface). The all-conditions maps (Fig. S10 in the Supplement) show patterns similar to Fig. 9 but with magnitudes 16 %–25 % larger. Pan-Arctic blowing snow sublimation totals increase to 2.1 cm SWE for SnowModel-LG and 2.4 cm SWE for DY2001. Comparing these values to the seasonal snowfall from Fig. 9 (12.4 cm SWE) yields offsets of 17 % for SnowModel-LG and 19 % for DY2001. This comparison suggests that ICESat-2 captures the spatial pattern and temporal variability of blowing snow sublimation well but underestimates the total by roughly 20 % due to this sampling bias.