We operated four ground-penetrating radar (GPR) systems to cover the five field sites, including two surface-coupled 1.0 GHz center-frequency PulseEkko Pro single transceiver/receiver systems, an air-coupled 1.0 GHz center-frequency PulseEkko Pro single transceiver with dual polarization receivers system, and a surface-coupled 1.6 GHz GSSI single transceiver/receiver system. The PulseEkko Pro systems have a 6 dB bandwidth of 500–1500 MHz, whereas the GSSI system has a 6 dB bandwidth of 800–2400 MHz. Each GPR used a GNSS system to provide locations for collected traces: the air-coupled PulseEkko Pro used a Geode dGPS system (±0.5 m accuracy), the two sled-coupled PulseEkko Pro GPRs used Emlid RS2 rovers with Emlid RS2 bases located nearby for post-kinematic processing (±0.5 m accuracy), and the sled-coupled GSSI GPR used a mapping-grade GPS system (±3 m accuracy). In Fig. 1, each transect is marked by the associated GPR system.

For the three surface-coupled GPR surveys, the GPR was pulled across the surface of the snowpack in a sled, whereas the air-coupled GPR was carried by two people above the snow surface. GPR systems were operated in the common-offset configuration and we took care not to disturb the snowpack immediately below the GPR transect. Transect lengths ranged from 3–12 m. Methods for deriving GPR snow depths are provided in Sect. 3.2. Following GPR data collection, the snow along the full length of the transect was excavated and snowpack thicknesses, which excluded any void spaces or tree branches or logs in the snowpack, were measured at 0.25–1 m intervals. These snowpack thickness observations are unique, as typical in situ snow depths are measured with little-to-no knowledge of void spaces within the snowpack. Typical in situ snow depths are measured by using the probe to identify the ground, which has a large range of potential conditions (e.g., soft and mossy, hard and frozen, brushy) that can be challenging to interpret based on the ground-feel alone. The excavated snow depths are thus more representative of the true snowpack thickness, as demonstrated recently by . Pictures and/or video recordings were acquired for each transect and detailed notes were taken on vegetation and ground conditions. An example photomosaic of an excavated boreal forest transect is provided in Fig. 2. Each survey was completed within 15 m of a snow pit, wherein measurements of snow depth, snow density, SWE, and snow temperature were collected. In total, 17 surveys were collected at the boreal forest field sites and 27 surveys were collected at the tundra field sites on the North Slope.

At each field site, snow-on airborne lidar surveys were collected during the campaign and snow-off airborne lidar surveys were collected during the following summer. Snow-on and snow-off lidar surveys were used to generate 0.5 m snow depth and canopy height models . We note that lidar-derived snow depths based on acquisitions that occurred before or after ground surveys have increased uncertainty due to either a changed snow surface (i.e., accumulation, compaction, redistribution) between the lidar flights and ground surveys or snow disturbance caused by the ground observations that preceded lidar flights (n=20/44). Further processing details for the lidar snow depths are provided in . A list of lidar and ground survey dates is provided in Table 1 and snow accumulation measurements for every day of the campaign are provided in Table S1 in the Supplement.

For GPR systems, the transceiver emits a signal that transmits through the snowpack and reflects off boundaries of contrasting dielectric permittivities (e.g., vegetation, snow stratigraphy, the snow-ground interface). The receiver then records the amplitude and two-way travel times (twtt) of these reflections. Radargrams were processed to a trace spacing of ∼0.10 m. We provide an example of a boreal forest radargram in Fig. 2. Exported twtt of the snow-ground interface, typically identified as the first reflection at depth with the highest amplitude that is spatially coherent, were manually selected for all single-polarization GPR datasets e.g.,. For the dual-polarization GPR dataset, twtt was obtained by utilizing the coherent reflection between the two receivers further details provided in. The twtt can be converted to snow depth with an estimate of the snowpack radar velocity vs 1vs=cϵs, where c is the speed of electromagnetic energy in a vacuum and ϵs is the dielectric permittivity of the snowpack. Snowpack conditions were dry during our surveys, thus, the relative permittivity can be approximated from the bulk snow density (ρs) through an empirical relation 2ϵs=1+0.845ρs1000. Snow density was sampled in snow pits adjacent to the surveyed transects using 1000 cm³ wedge samplers at 10 cm intervals along two columns in the snow pit. If density measurements differed by more than 10 % for a single layer, a third measurement was acquired. For layers with three measurements, we excluded the measurement with the largest difference. We then calculated bulk snow densities as the column average. Finally, snow depth (Ds) is calculated as 3Ds=vstwtt2.

GPR surveys have several sources of uncertainty. GNSS positioning was estimated as ±0.5–3 m horizontal accuracy at the boreal forest sites and ±0.5 m at the Arctic tundra sites, which could complicate the comparison between GPR and lidar snow depths. GPR-derived snow depths may be reduced by up to 0.07 m due to removal and/or compaction of snow as the sled travels across the surface. Finally, site-specific ground conditions may add uncertainty to the GPR snow depths. A few examples include: (1) snow depths may be overestimated because of the presence of void spaces within the snowpack caused by vegetation, (2) snow depths may be underestimated when a dense layer of vegetation obscures the true bottom of the snowpack (e.g., a buried bent-over tree), and (3) snow depths may be overestimated if the radar signal penetrates into frozen soil without a strong reflection at the snow-ground interface.

We extracted the lidar snow depths (0.5 m × 0.5 m resolution) along each surveyed transect based on the GNSS coordinates recorded by the GPR. We aligned the excavated depth locations with the GPR depth locations based on field measurements, resulting in a high-quality co-registration between the GPR and the excavated depths. However, the co-registration between the lidar data and the transect surveys are subject to the GNSS uncertainty listed above. GPR, lidar, and excavated depths were then compared as individual profiles to evaluate any systematic differences. We calculated the mean and standard deviation of snow depth for the GPR, lidar, and in-situ measurements from each transect to estimate the overall root mean squared error (RMSE), bias, and Pearson's correlation coefficient (r) between measurement techniques in both the tundra and boreal forest sites Although both GPR and lidar snow depths likely contain errors and biases, we primarily emphasize bias in the results and discussion because the two are difficult to untangle due to the complex conditions in which these data were collected. Finally, lidar-derived vegetation heights were extracted along each transect in the boreal forest to determine whether accuracy was related to lidar-derived vegetation height.

At boreal forest sites, GPR mean snow depths overestimated excavated mean snow depths for 12 of 17 transects, whereas lidar mean snow depths underestimated excavated mean snow depths for 14 of 17 transects (Table S2, Fig. 3). Large differences (>0.20 m) between GPR and excavated depths occurred where sub-snow shrubs or tree branches caused void spaces in the snowpack (e.g., Fig. 3g, i), but close agreement was observed for the transect in the agricultural field where only crop stubble was present below the snowpack (Fig. 3b) and for transects below black spruce canopy (e.g., Fig. 3n–p). The average residual between GPR and excavated depths was +0.06 m (9 % of mean excavated depth) and the average residual between lidar and excavated depths was -0.12 m (16 % of overall mean excavated depth) when field surveys occurred after lidar acquisitions (Table S2). For these surveys, the lidar profile either mostly mimicked the excavated depths and agreed within ±0.07 m (e.g., Fig. 3j, p), or the lidar profile failed to capture both the spatial patterns and exhibited reduced accuracy when compared to the excavated depths (e.g., Fig. 3h, l). The lidar snow depth bias worsened to -0.16 m (24 % of mean excavated depth; Table S2) when surveys performed before the lidar survey date were included in the comparison. We visually inspected the boreal forest lidar snow depth rasters to verify the presence of all transects excavated before the 11 March lidar survey and we found that excavation activities clearly influenced five of these surveys. These transects are shown in Fig. 3b, d, e, f, and n, illustrating how sample timing can severely bias direct comparisons between field and remote sensing-based depth estimates.

Of the three boreal forest field sites, CPCRW yielded the lowest mean GPR residual (+0.03 m), whereas BCEF yielded the highest mean residual (+0.11 m). Three of the five CPCRW transects were performed below black spruce canopy, where mosses are the dominant substrate and shrub canopy is sparse, whereas shrubs and tussocks dominate the ground below BCEF transects. FLCF surveys (mean GPR residual =+0.06 m) were performed primarily below mixed deciduous/coniferous canopy and contained various shrubs, saplings, and tree branches within the snowpack. Compared to GPR, mean lidar residuals exhibited the opposite canopy-driven trend; BCEF surveys yielded the lowest mean residuals (-0.10 m) and FLCF/CPCRW yielded the highest mean residuals (-0.19 and -0.15 m). However, we note that three of four BCEF surveys were performed after lidar collection, whereas eight of 13 surveys at FLCF/CPCRW were performed before lidar collection (Table S2). Overall, we calculated an r of 0.55 and RMSE of 0.14 m for GPR snow depths and an r of 0.51 and RMSE of 0.22 m for lidar snow depths at the boreal forest field sites (Fig. 4a–b). The best Pearson's correlation coefficient, but highest RMSE was observed for the lidar vs. GPR comparison (r=0.66, RMSE = 0.26 m; Fig. 4c).

Vegetation structure was less complex at the Arctic tundra sites than at the boreal forest sites. At ACP, surveys had grassy ground cover (11 surveys with grass heights of 0.05–0.20 m) or ground ice (two surveys; Fig. 5i, m), whereas tussocks (four surveys with tussock heights of 0.08–0.25 m) and shrubs (four surveys with shrub heights <0.30 m) were the predominant ground cover at UKT. Based on the excavated depths, both mean GPR residuals (ACP =+0.03±0.04 m; UKT =+0.01±0.10 m; Tables S3–S4, Figs. 5–6) and the overall performance metrics (r=0.87; RMSE = 0.08; Fig. 4d) improved at the Arctic tundra sites relative to the boreal forest sites.

Compared to the boreal forest sites, the lidar snow depths from the Arctic tundra sites yielded a comparable Pearson's correlation coefficient, but an improved RMSE (r=0.48; RMSE = 0.16 m; Fig. 4e). Despite having more complex ground conditions than ACP, UKT yielded the lowest mean lidar residual (ACP =+0.19±0.05 m; UKT =-0.02±0.12 m; Tables S3–S4, Figs. 5–6). All ACP lidar profiles exhibited poor magnitude agreement with the excavated depths. Despite this, several of these profiles exhibited spatial patterns that were similar to the excavated depth profiles (Fig. 5c, f, i, m). Notably, all ACP surveys were performed after the lidar flight, whereas 13 of 14 UKT surveys preceded the lidar flight (Table 1). We visually inspected the transects in the UKT lidar snow depth raster and observed that only five transects were influenced by snow excavation (Fig. 6h–j, l–m).

We found that GPR captured snow depth variability in both boreal forest and Arctic tundra environments over length scales >0.5 m, but we observed larger uncertainty over shorter length scales (<0.5 m), likely due to the large sensor footprint ∼1.5 m radius for a snow depth of 0.6 m; e.g.,. GPR snow depths exhibited a modest positive bias (+0.06 m) relative to excavated depths in the boreal forest and a smaller positive bias in the Arctic tundra (overall mean residuals =+1–3 cm; Tables S2, S3). While it is possible that a portion of the observed GPR bias could be explained by differences between the GPR and excavated depth measurement footprints, we largely attribute the positive biases to discontinuous vegetation-induced void spaces at the base and within the snowpack e.g., Fig. 2;. Vegetation was dense and complex in the boreal forest, resulting in cluttered radargrams that were difficult to interpret, whereas Arctic tundra sites had less complex vegetation and better-defined ground reflectors.

Lidar at ACP was expected to have the highest agreement with excavated depths because the lidar surveys were conducted before the snowpack was disturbed. Instead, ACP lidar exhibited a +0.19 m average bias and failed to reproduce fine-scale snow depth patterns (e.g., Fig. 5a–b), despite having relatively simpler ground conditions at the time of the surveys. We attribute this to a strip alignment issue in the ACP bare earth elevation model collected in August 2022, which results in linear striping in the snow depths aligned with the lidar swaths. All ground surveys happened to be conducted within a single swath that exhibited higher snow depths (Fig. 7), which provides the likely explanation for the observed ACP bias in the snow depths.

Although UKT lidar exhibited lower mean bias, it also did not reproduce fine-scale snow depth patterns (e.g., Fig. 6e–f). Both ACP and UKT received <0.06 m of snow depth accumulation between ground-based and airborne surveys (Table S1). Based on field observations, blowing snow and wind redistribution e.g., likely caused the poor spatial pattern agreement between excavated depth and lidar depth profiles. Seasonal thaw, which has been observed to drive subsidence up to 0.05 m in this region , may also contribute, but is challenging to quantify with available observations.

Lidar at the boreal forest also exhibited higher uncertainties, yielding an RMSE that was nearly half of the average measured snow depth. However, demonstrated that the RMSE could be reduced to 0.12 m at CPCRW by removing anomalous snow depths and areas with anomalous snow depth patterns. Uncertainties tended to be higher for taller vegetation heights (>5 m; Fig. 4b), although we note a large range of accuracy for lower vegetation heights (<5 m). This weaker agreement could be due to one or more of four primary factors: (1) snow disturbance due to ground surveys caused errors within the lidar retrieval method, (2) differences in canopy structure between summer and winter acquisitions (e.g., bent-over trees from snow burden), (3) signal penetration through the dense canopy may be problematic for the method e.g.,, or (4) GNSS location uncertainty between the lidar and excavated depths. Despite the higher uncertainty for the lidar snow depths, the lidar coverage is spatially continuous and complete over the field sites, a clear advantage over the GPR measurements.

Although we identified a clear bias for GPR in the boreal forest, it appears that this bias exhibited some dependence upon the land cover, but we were unable to fully evaluate this point. Further work, potentially through a depth probe-GPR evaluation, is needed to identify systematic biases within the boreal forest for GPR operation.

ACP lidar snow depths appear to be affected by swath alignment artifacts, which users should be aware of before using the dataset. If the identified bias is removed from the lidar swath by simple subtraction, the ACP lidar RMSE improves from 0.20 to 0.05 m. However, the remaining lidar swaths may be challenging to correct given the lack of independent in situ observations in these swaths. At UKT, lidar snow depths exhibited low bias and relatively high accuracy, whereas lidar snow depths at the boreal forest sites exhibited a substantial positive bias. evaluated ICESat-2 lidar snow depths at UKT and FLCF using the airborne lidar snow depths and found better ICESat-2 performance at UKT (r2=0.84–0.92) than at FLCF (r2=0.14–0.58), further highlighting the boreal forest complexities. In the forests, the lidar snow depth bias is directly related to potential void spaces caused by shrub canopy interception and a lack of photon penetration below canopy and underbrush. In the Arctic tundra, lidar snow depths may be influenced by seasonal thaw cycles. Given the magnitude of snow disturbance associated with in situ surveys, we expect poor agreement between snow pit/excavated measurements and lidar measurements when the field survey preceded the lidar flight (e.g., Fig. 3e, n).