Here, we provide an overview of the key UAVSAR processing steps. For additional and more detailed information, we direct readers to Appendix A1–A2. During the 2020 and 2021 airborne campaigns, UAVSAR deployed a fully polarized L-band (1.26 GHz center frequency, 0.24 m wavelength), 80 MHz bandwidth, left-looking InSAR. The instrument was flown at an altitude of ∼12500 m and acquired data along a ∼40 km stretch with a 16 km swath width (area =∼640 km2; Fig. 1a; ). In 2020, overpasses were performed with a temporal baseline of 7 d for the first three acquisitions (12, 19, and 26 February) and 15 d for the final acquisition (12 March). In 2021, overpasses had varying temporal baselines (typically 5 to 8 d); due to other aircraft commitments, one acquisition had a longer baseline (20 d for 3–23 February). Poor coherence prevented phase unwrapping at the field sites for one InSAR pair (10–16 March 2021). The UAVSAR team at the NASA Jet Propulsion Laboratory processed the UAVSAR data and generated geocoded amplitude, interferogram, unwrapped interferogram, and coherence products at ∼5 m spatial resolution. We accessed the products from the Alaska Satellite Facility (ASF; ) and converted the products to GeoTIFFs using uavsar_pytools . InSAR measures phase deformation within a single ±π rad phase cycle, which equates to about ±108 mm SWE. Interferograms where ΔSWE exceeded a full phase cycle for some pixels require unwrapping for the accurate estimation of ΔSWE. Therefore, we focused on the unwrapped interferogram and coherence products and outline key workflow steps for calculating ΔSWE, rather than total SWE, in Fig. 2. Although we included all four polarizations, we present the horizontal-transmit and horizontal-receive polarization (HH) for all intervals except the 3–23 February 2021 interval, which used the vertical-transmit and horizontal-receive polarization (VH) due to incomplete phase unwrapping in the HH data product. Detailed radar SWE retrieval methodology is outlined in Appendix A2.

We tested for atmospheric delays following methods developed by . We identified snow-free pixels in the unwrapped interferograms using the normalized difference snow index NDSI; calculated from Sentinel-2 imagery Fig. S1; and regressed snow-free unwrapped phase pixels against the corresponding signal path lengths. Importantly, this method assumes that snow-free pixels are not undergoing any physical changes that would lead to a phase change. We tested whether an atmospheric correction was needed using the three criteria outlined in Appendix A2.2. Importantly, no unwrapped interferograms met all three criteria (Table S1). Therefore, we conclude that stratified atmospheric artifacts are either limited for all interferometric pairs or were more complicated than what our linear model identified. See Appendix A2.2 for a more detailed description of the atmospheric correction.

For these flights, UAVSAR had average look angles of 26–70° from the near to far range. We calculated incidence angles in uavsar_pytools Eq. A4; from the Copernicus 30 m DEM (rescaled to the UAVSAR grid) and the UAVSAR-provided look vector. The Copernicus DEM was chosen because it is the primary DEM used within the processing flow of ASF HyP3 and will be the basis for NISAR interferometric products. We evaluated incidence angles derived from the Copernicus DEM and the ΔSWE retrieval uncertainty caused by these incidence angles in Appendix A2.5.

UAVSAR acquisitions were collected during the winter over relatively short temporal baselines (<21 d). Therefore, we consider changes at the snowpack surface to be the primary driver of phase deformation in the unwrapped interferograms, but we provide a discussion of other potential sources of phase deformation in Appendix A2.1. Changes at the snow surface may include new snow accumulation, sublimation, redistribution, or melt. For both study periods, we conclude that the snowpack is dry, based on results presented in Sect. 4.1. Thus, for ΔSWE retrievals, we consider only the density of snow that accumulated between UAVSAR acquisitions. Surface densities were estimated by averaging density measurements of the snow that accumulated on the interval boards between UAVSAR acquisitions (Sect. 3.2). For instances where snow accumulation had occurred but had been removed from the interval board by, for example, wind redistribution, we used an average of the uppermost 10 cm of the snow pit-measured densities. For each interferometric pair, we converted surface densities to relative permittivity (Eq. A5). Relative permittivities, unwrapped phase, and incidence angles were then used to calculate snow depth changes (Eq. A6), which were subsequently converted to ΔSWE using the surface snow density (Eq. A7). Because the InSAR phase is relative , we estimated the absolute phase as the median difference between a 20 % set of randomly selected GPR ΔSWE retrievals (Sect. 3.2) and coincident UAVSAR ΔSWE retrievals for each interval. The median differences were then subtracted from the UAVSAR ΔSWE retrievals for each interval, and the 20 % of the GPR observations used in this step were removed from subsequent analyses. Finally, we supplemented our analysis by evaluating an InSAR ΔSWE retrieval method that approximates ΔSWE from the InSAR phase change and the incidence angle and is thus independent of snow density and relative permittivity measurements. The methods and results of this analysis are reviewed in Appendix A2.4.

UAVSAR coherence values from corresponding TLS and GPR pixels were used to evaluate coherence as a measure of noise for ΔSWE retrievals. Coincident GPR and UAVSAR ΔSWE retrievals were binned by coherence, and the root mean square error (RMSE) of the UAVSAR ΔSWE retrievals was calculated for each bin. The effect of temporal baseline upon coherence and UAVSAR ΔSWE retrieval accuracy was then evaluated by calculating the median coherence and RMSE for UAVSAR ΔSWE retrievals across all temporal baselines used in this analysis.

We obtained daily observations of snow depth, SWE, and air temperature from the Joe Wright SNOTEL station (ID 551) and daily observations of SWE from an additional three SNOTEL stations within the UAVSAR swath for the 2020 and 2021 seasons (Fig. 1a; Table S2). Daily snow depths were obtained from three automated stations (two with sonic sensors and one with a snow stake paired with a time-lapse camera) within 4.5 km of the Joe Wright SNOTEL station (Table S2). We converted the snow depths to SWE by calculating density from Joe Wright SNOTEL station measurements of SWE and snow depth. SWE estimates were then smoothed with a 5 d moving median filter to reduce the effects of new snow settlement.

We expanded our UAVSAR analysis beyond our relatively small field sites (∼0.2 km2 total area) to include measurements from the four SNOTEL stations and three automated stations within the swath (Table S2). We calculated the median UAVSAR ΔSWE within a 3×3 pixel grid (∼ 15 m × ∼15 m) around each station, added the ΔSWE retrievals for each interval, and matched the ΔSWE time series to the station time series by adding the station's SWE at the start of the UAVSAR flights to the UAVSAR ΔSWE time series. Because of spatially extensive missing data within the 10–16 March 2021 interferometric pair, we adjusted the UAVSAR ΔSWE time series at each station with the ΔSWE measured by the station. Median coherence was calculated within each 3×3 grid for the SNOTEL stations to evaluate the effects of coherence on the ΔSWE retrieval time series. Lastly, station-measured SWE was compared with cumulative InSAR SWE for the final dates of the 2020 and 2021 UAVSAR acquisitions.

UAVSAR flights coincided with 117 mm of SWE accumulation (18 % of peak SWE; Fig. 3a) during the 2020 campaign (4 weeks) and 282 mm of accumulation (48 % of peak SWE; Fig. 3b) during the 2021 campaign (9 weeks). SWE at the in situ interval boards increased on average by 34±12 mm and 31±29 mm per flight interval during the 2020 and 2021 campaigns, indicating that ΔSWE at the field sites was likely within a full phase cycle (±108 mm; Appendix A2.3) for most UAVSAR acquisitions. New snow density, used for UAVSAR ΔSWE calculations, ranged between 106 and 145 kg m-3 across all study dates in 2020 (Fig. 3c) and over a larger range, 118–219 kg m-3, in 2021 (Fig. 3d). Bulk density, used for GPR and TLS SWE calculations, increased minimally between most flights (mean =+20 kg m-3; Fig. 3c–d), with a notable exception being the 12–19 February 2020 pair (mean =+72 kg m-3).

Surface melting can lead to significant decorrelation of the radar signal and cause increased uncertainty in the ΔSWE retrievals. There were three notable warm periods during the campaigns (7–9 March 2020, 2–10 March 2021, and 21–22 March 2021), but median snow pit temperatures during our survey dates remained <-1.1 °C (Fig. 3e–f). We did observe near-surface melt–freeze crusts in the snow pits during certain surveys, but our observations suggest that liquid water content was absent or minimal during UAVSAR flight times (Table 1) at our study sites throughout the campaigns.

GPR SWE retrievals from the 2020 MR field site showed that median SWE increased by 127 mm between 12 February and 11 March (Fig. 4a), with the largest median ΔSWE occurring during the 12–19 February interval (+99 mm). The 2021 MR (Fig. 4b) and CP (Fig. 4c) field sites showed similar dynamic ranges, with GPR SWE retrievals increasing by 249 mm at the MR site and 233 mm at the CP site between 15 January and 22 March. For both sites, the largest median ΔSWE occurred during the 2–24 February interval (MR =+97 mm, CP =+110 mm). GPR SWE retrievals and SWE converted from depth probe measurements are highly correlated, with an overall Pearson's correlation coefficient (r) of 0.97 and an overall RMSE of 35 mm (Fig. S2).

UAVSAR ΔSWE retrievals along the GPR transect at the 2020 MR field site saw a mean cumulative increase of 40 mm for the three intervals (Fig. 5c–e; Table S3). The largest median ΔSWE occurred during the 12–19 February interval (median =+97 mm), with modest SWE increases observed for both the 19–26 February (median =+16 mm) and 26 February to 12 March (median =+8 mm) intervals. The largest ΔSWE retrieval range was observed for the 12–19 February interval (minimum =+60 mm, maximum =+149 mm). The expanded 2.7 km × 3.6 km region around the MR site reveals a somewhat different pattern than ΔSWE retrievals along the transect, with less accumulation for 12–19 February (+67 mm) and negligible SWE changes for 19–26 February (0 mm) and 26 February to 12 March (+1 mm; Fig. 5c–e).

UAVSAR ΔSWE retrievals along the GPR transects at the 2021 MR field site saw a median cumulative increase of 104 mm for six of the seven 2021 intervals (no data for 3–23 February 2021), whereas the median cumulative increase for the expanded 2.7 km × 3.6 km region was +143 mm (Fig. 5f–l). At the CP site, the median cumulative SWE across the seven surveys was 203 mm along the GPR transect and 171 mm from the 2.2 km × 3 km expanded region around the CP field site (Fig. 6c–i). The largest median ΔSWE for the expanded regions occurred during the 3–23 February interval (MR median =+103 mm, CP median =+107 mm). Minimum UAVSAR ΔSWE retrieval medians from the expanded regions were observed from 27 January to 3 February at CP (median ΔSWE =-2 mm) and 3–10 March at MR (median ΔSWE =-6 mm).

UAVSAR ΔSWE retrievals appear to capture detailed spatial distributions of ΔSWE across all dates at each field site. In particular, larger SWE accumulation is observed in the open meadows and avalanche paths in the MR study area than in the surrounding forests (mean difference =66 %, range of mean differences =-2 to +29 mm; Fig. 5). These patterns are particularly noticeable at the MR site for the 12–19 February 2020 interval (Fig. 5c), which recorded a median ΔSWE increase of +98 mm in open meadows and avalanche paths, whereas ΔSWE in the surrounding forests increased by a median of +69 mm. A similar spatial pattern exists at the CP site, as the burned area consistently recorded a larger ΔSWE than adjacent unburned forests. This is best observed in the 20–27 January 2021 interval (Fig. 6d). During this interval, we calculated an average of 31 mm ΔSWE in the burned area and 15 mm ΔSWE in the unburned forests. Median coherence across the time series is somewhat higher for unforested areas in both the MR and CP field sites (+0.05; Figs. 5b, 6b). This subtle difference is further illustrated within the CP field site, where median coherence of the 7 d baseline InSAR pairs increased from 0.56 in 2020 pre-burned forests to 0.60 in 2021 post-burned areas (p=<0.0001).

UAVSAR ΔSWE retrievals have a relatively low pixelwise correlation with GPR ΔSWE retrievals for any single InSAR pair (r=-0.24 to 0.20; Table S4). However, compiling the measurements across all surveys increases the ΔSWE dynamic range and correlation substantially (r=0.79; Fig. 7a). Here, we present a time series that includes only InSAR pairs from the HH polarization for all dates except the 3–23 February 2021 pair, which is represented by the VH polarization. For this time series, we observe RMSEs from 16–34 mm (Table S4) for single InSAR pairs, with an overall RMSE =22 mm (Fig. 7a). Although pixelwise comparisons between UAVSAR and GPR ΔSWE retrievals exhibit scatter, the box plot distributions for ΔSWE at co-located GPR–UAVSAR pixels are nearly identical, yielding absolute median differences between median GPR ΔSWE and median UAVSAR ΔSWE of 0–4 mm (Fig. 7b, c; Table S2). Although we primarily present analyses based on the HH polarization, we find that UAVSAR ΔSWE retrievals have nearly equivalent RMSE values across all four polarizations (RMSE =19–22 mm; Table S4).

We explored the possibility of coherence as an error metric for UAVSAR ΔSWE retrievals and found that RMSE exhibited a narrow range (21–25 mm) for coherence bins between 0.1 and 0.7 (Fig. 8a). However, the RMSE at very low coherence (0.0–0.1) is double the RMSE at very high coherence (0.9–1.0). Average coherence was highest for ∼ weekly baselines, but average coherence for the 15 d baseline (0.51) was within the range of average coherence for the 5-to-8 d temporal baselines (Fig. 8b). Of note, the 20 d baseline had average coherence >0.40 (Fig. 7d) but yielded the highest RMSE (33 mm; Fig. 8b).

TLS ΔSWE retrievals had median values of +9 mm for the MR field site during the 26 February to 12 March 2020 interval and +55 and +39 mm at the MR and CP field sites during the 10–24 February 2021 surveys (Fig. 9a, d, g). TLS SWE retrievals have a high correlation with SWE converted from depth probes, with an r of 0.83 and an RMSE of 66 mm (n=189; Fig. S3). For each set of TLS acquisitions, UAVSAR ΔSWE retrievals had median values of +6, +60, and +45 mm, respectively (Fig. 9b, e, h). Spatial patterns were similar between the two methods of ΔSWE retrievals. Large portions of data are missing in Fig. 9e due to coherence-related phase unwrapping errors. RMSEs were comparable between the 2020 survey (MR =20 mm) and the 2021 surveys (MR =15 mm, CP =20 mm). UAVSAR ΔSWE retrievals have an overall RMSE of 19 mm and an r of 0.72 when compared with TLSs. Coherence was used to color points on the UAVSAR–TLS comparison plots (Fig. 9c, f, i) and shows that scatter is approximately equal throughout the range of observed coherences.

UAVSAR SWE retrievals overestimated SWE accumulation for the 12–19 February 2020 InSAR pair by an average of 163 % at the automated stations but underestimated SWE accumulation by an average of 88 % between 19 February and 12 March (Fig. 10a–e). The 2021 cumulative UAVSAR SWE retrievals record net increases at all seven sites (+109 to +219 mm), which is similar to the net increases recorded by the stations (+101 to +242 mm; Fig. 10a–g). Median coherence for the 2020 season was lowest at the Lake Irene SNOTEL station (median coherence =0.30) and highest at the Phantom Valley SNOTEL station (median coherence =0.63), whereas median coherence for the 2021 season was lowest at the Montgomery Snow Stake (median coherence =0.49) and highest at the Lake Irene SNOTEL station (median coherence =0.60). The lowest median coherence for all sites was observed for the 26 February to 12 March 2020 interval (median coherence =0.31), an interval that yielded negative SWE retrievals for three of the five operating stations (-17 to -3 mm). At the end of the UAVSAR campaigns, cumulative UAVSAR SWE retrievals from the seven stations (n=12) yielded an RMSE =42 mm and an r=0.92 (Fig. 10h).

From our evaluation with GPR and TLSs, we established the RMSE for L-band InSAR ΔSWE retrievals as 19–22 mm for single InSAR pairs (Figs. 7, 9). For cumulative InSAR SWE, we estimated an RMSE of 42 mm at seven automated stations (Fig. 10). Previous studies have established that UAVSAR ΔSWE retrievals resemble the spatial patterns of lidar-derived ΔSWE retrievals, but differences between the two datasets were not systematic . evaluated UAVSAR ΔSWE retrievals over a 4 km2 relatively flat and non-forested region of Grand Mesa, Colorado, using airborne lidar and found very low error for the technique (RMSE = 9 mm). UAVSAR ΔSWE retrievals have been evaluated using GPR and automated station measurements in Valles Caldera, New Mexico , and from in situ and SNOTEL measurements of ΔSWE in the mountains of Idaho . Both studies identified and corrected significant atmospheric artifacts and contained at least one InSAR pair that was collected when liquid water content was present in the snowpack but estimated study-wide errors of similar magnitude found by our study (RMSE =15–40 mm; Table S4).

UAVSAR ΔSWE retrievals had higher RMSE in 2020 than in 2021 (Table S4), and the agreement between the InSAR time series and the automated stations was poorer in 2020 than in 2021 (Fig. 10). One potential explanation for the lower agreement in 2020 was the significant deviation (>10 m) from the cross-track and vertical baselines of the aircraft during the 2020 flights . UAVSAR ΔSWE spatial patterns are similar to those of TLS ΔSWE (Fig. 9), and the comparison of UAVSAR and GPR ΔSWE site-wide distributions reveals nearly identical medians (absolute median difference =2 mm; Fig. 7b–c). We found that low coherence did not substantially increase the RMSE of UAVSAR ΔSWE retrievals as the RMSE was less than 35 mm for >10 d temporal baselines (Fig. 8). However, lower coherence for InSAR pairs with >10 d temporal baselines exhibited issues with phase unwrapping. Collectively, these findings suggest a high degree of accuracy and reliability for InSAR ΔSWE retrievals, particularly in relatively simpler environments (i.e., dry snow conditions, non-forested areas, slopes <20°) and when atmospheric delays are limited.

The NISAR satellite mission holds promise for global repeat 12 d ΔSWE retrievals, providing the opportunity to evaluate the L-band InSAR technique in a range of environments and to better assess its uncertainties. In our evaluation, we used two ground-based methods that many snow community researchers have access to and showed that both methods are capable of assessing InSAR ΔSWE retrieval accuracy. Both methods can be used to derive spatially continuous SWE measurements over large areas and are therefore advantageous over standard in situ SWE measurement methods . Below we outline advantages, considerations, and challenges of GPR and TLSs for InSAR evaluation.

Few methods match the sophistication of InSAR for change detection. Of the two techniques we employed in our evaluation, lidar is the most applicable for change detection , but its methodology for ΔSWE retrievals is not straightforward. There are two conceptual paths for ΔSWE retrievals from lidar: (1) subtraction of two repeat snow-on lidar elevation surveys or (2) subtraction of two bulk SWE datasets derived from lidar. The first option is complicated by snow compaction, while the second option requires accurate bulk snow densities and a snow-off bare-earth digital terrain model, which may be difficult to acquire in densely vegetated areas. We chose the second option because bulk density variability is less of a concern for the relatively small areas surveyed by the TLSs . We found the best agreement between UAVSAR and TLS ΔSWE retrievals for surveys that were aligned on the same date, as differential SWE accumulation/redistribution increased uncertainty (Fig. 9). Note that if the TLS platform is set up on top of the snowpack, accurate TLS ΔSWE retrievals may be hindered by small shifts in the TLS platform as it settles in the snow .

Repeat GPR transects also have several challenges. Our survey methodology involved marking our transects and post-processing the onboard GPS sensor (±0.25 m accuracy), but it is likely that our tracks were offset by ±1–2 m from the transect for some surveys. Further, as SWE increases throughout a season, the twt to the ground reflector increases, effectively increasing both the GPR horizontal footprint and the potential for clutter in the radargram . Surface-coupled GPR has the potential to both compact the snow below the sled and remove snow from the surface e.g.,, which may further increase the uncertainty of GPR ΔSWE retrievals, particularly in low-density snow on the surface of the snowpack. These complications may explain the low Pearson’s correlation coefficients observed between UAVSAR and GPR ΔSWE retrievals for single InSAR pairs (r=-0.24 to +0.2; Table S4), as well as the low GPR–UAVSAR ΔSWE retrieval relation (r2<0.1) described by . However, as our analysis shows, repeat GPR transects are effective at evaluating the InSAR technique if there are enough data collected across a range of SWE accumulation magnitudes (Fig. 7).

A major difference between UAVSAR and planned NISAR interferograms is the spatial resolution (∼5 vs. 80 m), which may complicate future NISAR ΔSWE retrieval ground-based evaluations. GPR surveyed along transects scaled well to the resolution of UAVSAR, but a different survey design (i.e., spiral or grid) may be required to provide sufficient coverage of the NISAR pixels. Thus, GPR may have increased uncertainty in its scalability due to a difficulty of repeating complicated survey designs. Lidar is scalable to coarser resolutions e.g., 50 m;, and TLS and drone-mounted lidar e.g., may be valuable tools for evaluating InSAR ΔSWE retrievals at small field sites. However, at larger scales, comprehensive airborne lidar surveys will be required to fully evaluate NISAR ΔSWE retrievals.

L-band InSAR has been seen as a promising technique for high-resolution snow monitoring for over a decade , yet insufficient opportunities existed for robust evaluations. In the last few years, airborne InSAR campaigns over seasonal snowpacks have created opportunities for a more thorough evaluation of this technique. Our study and others show that this technique can have a high accuracy, but there are several areas of uncertainty that need to be considered, including forested environments, wet snowpacks, complex topography that results in steep incidence angles, spatially varying atmospheric delays, and the integration of InSAR ΔSWE retrievals with other remote sensing methods and models.

Recent UAVSAR studies , including this study, have largely focused on ΔSWE retrievals in open environments. We found that ΔSWE retrievals were 66 % higher on average in the open areas around the MR field site than below forest cover. Forest canopy interception and sublimation may play a role in this signal, because this process is known to drive a 20 %–30 % reduction of total snowfall at the nearby Fraser Experimental Forest . On the other hand, a contrast between lower snow surface densities in the forest compared with the potentially higher densities we measured in the open could explain a similar magnitude of the signal. Unfortunately, we are unable to validate the forest ΔSWE retrievals as only 20 % of GPR observations in 2020 and 10 % of GPR observations in 2021 were collected below spruce–fir canopy (15 %–70 % canopy cover). Forests interfere with the radar signal, reducing coherence and potentially biasing retrievals, particularly for longer temporal baselines . However, coherence only improved by +0.05 from forests to open areas at our field sites, and even the removal of canopy due to the Cameron Peak wildfire only increased coherence by +0.04. Thus, because of its canopy penetrative capabilities, the L-band InSAR ΔSWE retrieval technique may be the first satellite-based technique viable for SWE monitoring in forests.

At our site, UAVSAR flights occurred during the accumulation season when the snowpack was likely dry (Fig. 3e–f). However, SWE monitoring is needed for snowpacks that accumulate at or near 0 °C and for the melt season, making ΔSWE retrieval evaluation prioritized in wet snowpacks. Liquid water in the snowpack raises both the real and imaginary components of relative permittivity, which decreases the snowpack radar velocity and increases absorption of the radar signal, causing decreased signal penetration . Even if the backscattering interface is unchanged, reduced radar velocity causes ΔSWE retrieval overestimation if the liquid water content is not considered . evaluated ΔSWE retrievals with the Landsat fractional snow-covered area product and found reasonable snowpack ablation over a 14 d period in Valles Caldera, New Mexico, but found reduced ΔSWE retrieval accuracy in wet snowpacks. Wet snow detection techniques have been developed and implemented at the C-band e.g.,, and similar techniques should be evaluated at L-band frequencies e.g.,.