1. Introduction
Ice surface velocity is a key measurable parameter used in discerning the dynamics and discharge of glaciers [
1]. Therefore, error propagation from uncertainty in these measurements can have severe consequences. The measurement of glacier velocity has evolved with the development of technology leading to increased data, precision, and reduced uncertainty. Rudimentary methods involved tracking the movement of flagged stakes with a compass or theodolite using triangulation from an immobile observation point [
2]. In areas with damaged ice or other observable features, photogrammetric techniques can similarly derive velocity by displacement. Feeding a computer timelapse photography data increases the temporal and spatial resolution of these velocity fields, as well as reducing error [
3]. Modern GNSS is the gold standard for field measurements, capable of recording both vertical and horizontal displacement to the centimeter scale [
4] or finer. Yet field data collection often suffers from logistical difficulties and severe weather conditions, reducing spatiotemporal coverage. By far the largest and most wide-spanning ice surface velocity data come from cross-correlating displacement between image-pairs produced from sensors on polar orbiting satellites [
5,
6,
7,
8,
9,
10,
11,
12].
The launch of Landsat-1 in 1972 achieved continuous worldwide coverage in the visual range with a temporal resolution of 18 days (improved to 16 days in 1982 with the launch of Landsat-4). While velocity derivations from feature tracking on large bodies of ice began to appear in the 1980s [
13,
14,
15], finer scale measurements from improved spatial resolution were not available until the 1990s [
5,
16]. The capabilities of Landsat for tracking surface ice displacement improved after 2000 with the addition of the 15 m resolution panchromatic band onboard Landsat-7 and the development of pan-sharpening techniques. However, two common conditions in glaciated regions impede visual detection: cloud cover, for example, due to orographic precipitation for alpine glaciers and prolonged darkness, such as austral winter in Antarctica.
Unlike the passive visual spectrum sensors onboard the Landsat missions, synthetic aperture radar (SAR) actively detects microwave reflection signals that pierce through cloud cover and darkness [
17,
18,
19], making it a useful tool in alpine and polar glacier velocity determination. Spaceborne SAR sensors onboard the European Space Agency’s (ESA) ERS-1/2 missions launched in the 1990s created a tandem constellation of satellites that achieved sub-30 m spatial resolution with individual 35 day repeat cycles [
20]. Both temporal and spatial resolution improved in the 2000s with the deployment of TerraSAR-X, granting a sub-16 m pixel SAR scanning mode and an 11-day repeat cycle [
19]. In 2014 and 2015, the ESA launched the Sentinel-1a and Sentinel-1b (S1a/S1b) satellites, respectively, providing 10 m resolution coverage on a tandem repeat cycle of 6 days, further improving accuracy and precision [
18]. Individual researchers can obtain SAR-based velocity fields using the ESA’s SeNtinel Analysis Platform (SNAP).
For hosted dataset products, the methodologies applied to tracking ice velocity with visual spectrum and interferometric SAR (InSAR) are cross-correlation offset tracking, speckle tracking, and autonomous Repeat Image Feature Tracking (auto-RIFT) [
6,
9,
10,
11,
12,
20,
21,
22]. The result is contemporaneously growing datasets over various spatial dimensions, mostly contained within the National Aeronautics and Space Agency’s Making Earth Science Data Records for Use in Research Environments project (NASA MEaSUREs) [
11,
23,
24,
25,
26]. However, these MEaSUREs data are calibrated by filtering against prior satellite-based velocity datasets [
10], while the products outside of MEaSUREs are calibrated and filtered against Gardner’s MEaSUREs Inter-mission Time Series of Land Ice Velocity and Elevation (ITS_LIVE) [
12].
With the focus of the last decades on ice sheets, GNSS validation of these satellite datasets is largely restricted to the poles [
27,
28] where there is good agreement. However, SAR backscatter is sensitive to complex interactions in surface roughness, grain size, and liquid water content in snow [
17,
29]. Fewer validation efforts are undertaken on subpolar, temperate alpine glaciers where higher temperatures and rainstorms lead to SAR reflectance changes [
17]. In a changing climate [
30,
31,
32] more melt and liquid precipitation events are expected in glaciated regions—as noted with recorded rainfall on the summit of the Greenland Ice Sheet in 2021 [
33]. Therefore, validating satellite measurements on temperate alpine glacier regions is essential in establishing baseline error for future work. Indeed, considering the observed warming of the high latitudes and polar amplification [
34], it is imperative to understand the accuracy of satellite-measured ice velocity in temperate glaciers.
Study Area
The Juneau Icefield (JIF), centered on 58.625°N and 134.25°W, southeastern Alaska, USA, is a 4000 km
2 network of alpine glaciers concentrated around the Taku and Llewellyn glaciers in a subpolar region and the site of the second longest continuous annual record of glaciological observations (
Figure 1). The JIF equilibrium line altitude is 1425 m, but the icefield is subject to a highly variable transient snow line (800–1425 m) over the summer, typical of temperate alpine glacier systems. Since 1948, the researchers and students of the Juneau Icefield Research Program (JIRP) have collected repeated mass balance, velocity, snow density, and weather data from consistent locations and profiles [
4]. The scope of this work is to use GNSS measurements to validate the accuracy of optical and SAR-derived velocity products within a temperate alpine glacier setting since the launch of the S1 platforms. Satellite-derived velocities are determined to be accurate if their difference from spatiotemporally coincident GNSS-derived velocities is less than 0.05 m/day.
JIRP has conducted annual measurements and monitoring of surface velocities of the glaciers of the JIF across a network of transverse and longitudinal profiles [
36]. The survey profiles used here include Profiles 4, 8, 9a, and 13 (
Figure 2). Profile 4 is the longest running survey on the JIF and is located on the main trunk of the Taku Glacier at a mean elevation of 1121 m. Profile 8 is located on the upper Matthes Glacier approximately 4.5 km south of its divide with the Llewellyn Glacier at a mean elevation of 1802 m [
35]. Profile 9a is located on upper Matthes Glacier and extends westerly over a small divide onto the Vaughan-Lewis Glacier at a mean elevation of 1769 m. Profile 13 is located on the primary north–south trending branch of the Llewellyn Glacier at a mean elevation of 1410 m on the Canadian side of the JIF. Profiles 4, 8, and 9a are located in the accumulation zone of their respective glaciers, while Profile 13 is located well below the equilibrium line.
3. Results
One of the key components of this study was the direct temporal correlation between glacier deformation captured by the field surveys and satellite platforms. In 2016, velocities along Profiles 4 and 9a were measured in the field and velocity image-pairs satisfying the <30 day time separation were acquired from four datasets: ITS_LIVE v1, ITS_LIVE v2 (S2), SNAP, and RETREAT, leaving ITS_LIVE v2 (S1) as the only dataset that did not meet the search query (
Figure 4a). Of the four qualifying image-pair datasets, only two contained velocity measurements at the stake locations along both Profiles 4 and 9a: ITS_LIVE v2 (S2) and SNAP. RETREAT only had data coverage over Profile 4. No velocity estimations were available for comparison at the stake locations for ITS_LIVE v1. ITS_LIVE v2 (S2) provided velocity measurements at 28 of 30 and 9 of 16 stake locations for Profiles 4 (P4) and 9a (P9a), respectively, while SNAP covered 24 of 30 (P4) and 12 of 16 (P9a), and RETREAT covered 13 of 31 (P4).
In 2017, velocities along Profiles 4 and 8 (P8) were measured in the field and qualifying image-pair velocities were acquired from two datasets: ITS_LIVE v2 (S1) and SNAP, with no suitable data from RETREAT, ITS_LIVE v2 (S2), and ITS_LIVE v1 (
Figure 4b). Of the two qualifying datasets, only SNAP provided measurements at any of the stake locations, of which it covered 25 of 30 and 9 of 12 for P4 and P8, respectively. Image-pair data availability for 2018 provided the most robust assessment in the study. In 2018, velocities along P4, P8, and Profile 13 (P13) were measured in the field, and all five velocity image-pair datasets provided suitable data (
Figure 4c). Of the five qualifying datasets, only two provided data at stake locations across all three profiles (ITS_LIVE v2 (S2) and SNAP), one contained data on two profiles (RETREAT at P4 and P13), one provided data at a single profile (ITS_LIVE v1), and ITS_LIVE v2 (S1) did not cover any of the three profiles’ stake locations. ITS_LIVE v2 (S2) provided velocity measurements at 16 of 31, 5 of 12, and 8 of 8 stake locations at P4, P8, and P13, respectively. SNAP provided measurements at 23 of 31, 6 of 12, and 8 of 8 stake locations at P4, P8, and P13. RETREAT recorded measurements at 4 of 31 stake locations on P4, and 5 of 8 at P13. ITS_LIVE v1 covered all of the stake locations at P13, although there was a 5-to-9-day lag between the second Landsat flyover date (29 July) and the start (3 August) and end (7 August) of the P13 survey. A complete description of the velocities measured from GNSS and image-pairs is provided in
Figure 5.
The spatial distribution and variability of flow velocities measured using GNSS reflected the expected pattern of alpine glaciers with the fastest flow rates observed in the channel center and a progressive slowing toward the edges. The distribution and amplitudes of flow velocities derived by the image-pairs were widely varying and only partly agreeable with the GNSS results. The differences between the GNSS-measured and satellite-derived velocities (Δv) at each stake are shown in
Figure 6. The Δv for the eight velocity measurements made by ITS_LIVE v1 at Profile 13 varied from 0.01 to 1.31 m/day. with only one achieving an accurate result (<0.05 m/day) and the remaining seven having errors >0.20 m/day. Velocities derived by ITS_LIVE v2 (S2) had the largest range of Δv, varying between 0.005 and 1.15 m/day. Accurate velocity results (<0.05 m/day error) were primarily achieved at Profile 4, although they were greatly exceeded by the number of erroneous results. Of the 97 total velocity measurement attempts made by ITS_LIVE v2 (S2), 17 were accurate to within 0.05 m/day, 49 had errors greater than that threshold, with the remaining 31 retrieving no data. The spatial distribution of Δv for ITS_LIVE v2 (S2) was chaotic with highly inaccurate results (>0.20 m/day) occurring at both the slow-flowing edges of the glacial margins and the quicker central axis. The Δv of velocities derived by SNAP ranged between 0.001 and 0.62 m/day. Of the 140 total velocity measurement attempts made by SNAP, 27 were accurate to within 0.05 m/day, 79 had errors greater than that same threshold, with the remaining 34 retrieving no velocity data. The spatial distribution of Δv for SNAP showed a pattern of accurate velocities occurring along the outer edges of the glacier with error increasing toward the central flowline, particularly at Profiles 4 and 8. Of the 22 total measurements from RETREAT, 8 contained errors <0.05 m/day, 7 of which were located on Profile 4 toward the outer edge of Taku Glacier. The spatial distribution of Δv for RETREAT appeared similar to SNAP in that error appeared to increase toward the central flowline until the dataset ultimately failed to provide any velocity information.
K-means cluster analyses were performed on the optical image-derived (ITS_LIVE v1 and v2 (S2)) and SAR-derived (SAR/RETREAT) Δv and GNSS-measured data to assess the apparent relationship between actual flow velocity and the accuracy of satellite-derived techniques (
Figure 7). The
k-means clustering algorithm used the squared Euclidean distance metric (each centroid was the mean of the points in that cluster) and an optimal number of clusters (
k = 3). The clustering results for the optical image-derived data showed that there was not a significant relationship between the GNSS-measured velocities and the accuracy achieved (
Figure 7a). Errors approaching 0.60 m/day occurred for both slow (<0.10 m/day) and fast (>0.50 m/day) moving parts of the glacier. This agreed with the spatial distribution of Δv (
Figure 6), which showed a random pattern across the profiles. The clustering results for the SAR-derived data revealed a relationship between the accuracy achieved by the SAR technique (Δv) and the GNSS-measured velocity (
Figure 7b). Both the data (scatter plot) and centroids (or cluster means) showed a positive linear trend, with the three centroid positions located (X, Y) at 0.08, 0.10 (red cluster), 0.41, 0.16 (yellow cluster), and 0.51, 0.50 (blue cluster). These results agreed well with the spatial distribution of Δv, which suggested that the most accurate results were found along the slower-moving edges of the glacier.
Linear correlation coefficient statistics and root mean square errors (RMSEs) were calculated between the contemporaneous satellite-derived and GNSS-measured velocity datasets to assess and compare accuracies. The Pearson correlation coefficient (R) provided a measure of two random variables’ (A and B) measure of linear dependence for
N scalar observations and was calculated by:
where
μA and
σA are the mean and standard deviation of
A, respectively, and
μB and
σB are the same for
B.
p-values for each Pearson coefficient were calculated to determine significance, whereby a standard of
p < 0.05 was used. RMSE measured the deviation of the satellite-derived velocities from the GNSS-measured and is calculated by:
where
Ai is the GNSS-measured velocity and
Fi is the satellite-derived velocity.
A total of fifteen comparisons were made between the satellite-derived and GNSS-measured flow velocities spanning the four survey profile lines (
Figure 8;
Table 3). There were no correlation results for ITS_LIVE v2 (S1) owing to the paucity of coverage at every stake location in the image-pair velocity data. There was one correlation result for ITS_LIVE v1 at Profile 13 in 2018, which appeared to have some correlation (R = 0.47), but the ITS_LIVE v1 velocities were nearly an order of magnitude faster than the GNSS velocities. This significant deviation from the GNSS velocity was reflected by an RMSE of 0.8 m/day. ITS_LIVE v2 (S2) provided 5 of the 15 correlation analyses with results from all four profile lines. The correlation results at Profiles 8, 9a, and 13 were poor with R values of 0.002, 0.114, and 0.424 at Profiles 9a, 13, and 8, respectively, and corresponding RMSE values of 0.28, 0.3, and 0.88 m/day. The ITS_LIVE v2 (S2) velocities were also an order of magnitude faster at P8 compared to the GNSS velocities, which again was reflected by an RMSE of 0.88 m/d (the largest observed). The remaining two correlation results for ITS_LIVE v2 (S2) were acquired at P4 in 2016 and 2018. Both image-pair velocity datasets showed significant positive correlation to the GNSS-measured data, particularly the 2016 data, which had an R value of 0.737 and an associated
p-value of 5.02 × 10
−9 and an RMSE of 0.14 m/day. SNAP accounted for 7 of the 15 correlation analyses spanning all four profile lines. The correlation results at Profiles 4, 8, and 9a were very poor with R (RMSE) values of 0.04 (0.33), 0.139 (0.33), and 0.295 (0.3) at P4; 0.247 (0.26) and 0.257 (0.08) at P8; and 0.064 (0.18) at P9a. The SNAP velocity data also displayed an inverse relationship with the GNSS data in all three Profile 4 correlations. SNAP velocities at P13 achieved a strong R value of 0.89 (
p = 0.0004) and RMSE of 0.05 m/day. RETREAT provided the final two comparisons, with the 2018 survey at Profile 4 not being included owing to only 4 velocities of 31 being retrieved. The correlation at P4 in 2016 was strong with an R value of 0.79 and RMSE of 0.09. The P13 survey in 2018 showed similar results with an R value of 0.90 and RMSE of 0.12.
4. Discussion
A total of 35 velocity correlation analyses were possible given the 7 GNSS surveys and 5 image-pair velocity datasets. Only 15 of this total were (even partly) realized, mainly because of the lack of data availability in the ITS_LIVE v1 and ITS_LIVE v2 (S1) datasets (1 of 6) and the RETREAT datasets. The ITS_LIVE v1 velocity data showed significant gaps that were likely too expansive to be corrected by the autoRIFT interpolation procedure [
11]. Initially, these data gaps were interpreted to have been caused by cloud contamination as ITS_LIVE v1 was derived from multispectral Landsat imagery. The Quality Assessment bands, QA_PIXEL and CLOUD_QA, for the Landsat images used in the autoRIFT algorithm (
Table 2) were acquired to determine if this was the cause. The Pixel Quality Assessment Band (QA_PIXEL) contained the results of the C Function of Mask (CFMask) multipass algorithm that determines the presence of clouds, shadows, and snow/ice pixels in an image as well as confidence flags for each [
44], although it should be noted that the algorithm performance is degraded over bright targets such as snow/ice. The Surface Reflectance Cloud Quality Assessment Band (CLOUD_QA) determines the presence of clouds, cloud shadows, and proximity to clouds in image. The CLOUD_QA band indicated the presence of snow/water pixels across the P4 surface area in both 2016 images (
Figure 9a,b) with the QA_PIXEL band also revealing high confidence snow/ice in the same area (
Figure 9c,d). The QA bands for the 2018 image-pair, which, like 2016, contained significant data gaps, showed the same high confidence snow/ice pixel classification. We concluded that the significant gaps in the ITS_LIVE v1 velocity data were not a product of cloud contamination based on the ubiquitous presence of snow/ice pixels and complete omission of any clouds in the original Landsat imagery. It is more likely that the lack of data in the ITS_LIVE v1 datasets can be attributed to low coherence between the image-pairs and subsequent removal of those displacements by the Normalized Displacement Coherence filter. It should be noted that the term “coherence” is used to describe the signal-to-noise ratio of the displacement field generated by both optical and SAR image-pairs.
The accuracy of velocity results and the total region of interest (ROI) covered was influenced by many factors, such as the Ground Sample Distance (GSD) of the source imagery, sub-pixel oversampling ratio, time separation between the image-pair flyovers, surface texture, and elevation changes/DEM error [
22]. The GSD of the source imagery was directly related to the minimum displacement that could be detected using normalized cross-correlation with detection of small-scale changes improving with higher resolution imagery. autoRIFT used a fast Gaussian pyramid upsampling algorithm with an oversampling ratio of 64 to quantify sub-pixel displacements as small as 1/64 pixel [
22]. This amounted to minimum observable displacement values of 0.23, 0.15, and 0.06/0.23 m for Landsat (15 m/pixel), Sentinel-2 (10 m/pixel), and Sentinel-1 (4 and 15 m/pixel in range and azimuth), respectively. All the satellite-derived measurements made in this study were well above these minimum thresholds with total displacements ranging between 0.48 and 8.25 m across all profiles.
Temporal decorrelation is typical in radar image-pairs that have large temporal baselines approaching 36 days or greater. The accuracy of velocity measurements derived from radar pairs with temporal baselines of 6, 12, 18, and 24 days (corresponding to Sentinel-1 repeat intervals) have been shown to be relatively uniform [
22]. The Sentinel-1 image-pairs used here had temporal baselines of 12 or 24 days and as such, were likely not subject to temporal decorrelation. Optical imagery (Landsat and Sentinel-2) exhibited the opposite temporal correlation trend in that velocity results became markedly improved as the temporal baseline lengthened (a few months to a year). Given that a strict temporal baseline was set in this study to <30 days to ensure contemporaneous overlap with the field surveys, it is possible that the Landsat and Sentinel-2 velocity results could be markedly improved if a longer baseline were allowed. Lei et al. (2021) found that a temporal baseline of 368 days provided the best results in optical image-pairs but using this baseline would have made making direct comparisons to our 7–12 day field surveys impractical. Velocity results from cross-correlation are insensitive to DEM errors, and sensitivity analyses have shown that a DEM change or error amounting to tens of meters results in pixel mis-registration on the order of 0.001 pixels [
22]. As it applies here, co-registration error due to elevation changes between successive flyovers can result in displacement errors of 0.01, 0.015, and 0.004 m for Sentinel-2, Landsat, and Sentinel-1 imagery, respectively (determined by multiplying the pixel mis-registration error and spatial resolution of the imagery). Error offsets due to DEM/elevation changes well below (<1%) the measured displacements observed in this study were not likely to be a significant source of error.
Determining the accuracy of velocities derived by ITS_LIVE v2 (S1) was not possible owing to no data availability at the stake locations at any of the profiles across the three-year study period. This was a surprising result, given that the newer velocity product uses Sentinel-1 SAR imagery, which is not obscured by clouds or limited by solar illumination. The lack of data was likely due to widespread low-coherence displacement results achieved by the autoRIFT algorithm, which were then removed from the final velocity data product [
11,
22]. The lack of high coherence likely indicated that there was a significant shift in the backscatter amplitude of the glacier surface between the sequential Sentinel-1 flyover times (12 days apart). The distribution of amplitude values along Taku Glacier stretching from its border with the Matthes Glacier to the northeast and to the main trunk line in the southeast (
Figure 1) was used to determine variability in the Ground Range Detected SAR data. Analysis of >800,000 pixels spanning Taku Glacier in the S1 GRD data revealed a significant increase in the mean GRD amplitude between successive flyovers (
Figure 10a,b). The average detected amplitude for the 3 August 2018 SAR GRD image was 150.36, a 9.1% increase over the 22 July 2018 GRD image, which had an average of 137.73. This rapid shift in VV amplitude likely caused extensive temporal decorrelation and significantly limited the extent of velocities derived in the image-pair.
The backscattering coefficient of glacier surfaces, particularly in the accumulation zone, is widely variable depending on liquid water content. Owing to its high dielectric constant, liquid water significantly reduces the SAR backscatter properties of snow [
45,
46,
47]. Meteorological data were acquired from the Juneau Airport weather station (58.354°, −134.55606°) located approximately 37 km from Profile 4. It was likely that the study area received less precipitation than what was recorded at the JNU station based on known patterns of precipitation and the rain shadow effect of the Coast Range, but was suitable for detecting wet and dry periods on the JIF. Daily measurements of precipitation in July 2018 (corresponding with
Figure 10) revealed that 71 mm of rain fell in the twelve days leading up to the first Sentinel-1 flyover on 22 July (
Figure 11), saturating the glacier surface and reducing its mean backscatter amplitude. The twelve days following the first flyover were notably dry, with only 3.05 mm of rainfall occurring on 2 and 3 August. Much of the excess liquid water content on the glacier surface was likely lost to seepage and evaporation in the warm summer temperatures (average high: 17.2 °C), leading to the observed 9% increase in the mean backscatter and subsequent loss of coherence in SAR-derived image-pairs.
The correlation results presented here reflect the dynamic variability of SAR backscatter characteristics of the snow-covered surfaces previously described. Although no ITS_LIVE v2 (S1) velocity data were available to assess because of the filtering of low-coherence areas by autoRIFT, the same Sentinel-1 SAR image-pairs were used to derive velocities at all seven profiles using the offset tracking module within SNAP and were accounted for in the RETREAT dataset by offset tracking in GAMMA. Comparison to contemporaneous GNSS-measured velocities revealed widely varying performance, with no observed correlation at Profiles 4, 8, and 9a for SNAP (R < 0.3 and RMSE > 0.18) and anomalously positive correlation at Profile 13 (SNAP R = 0.89; RETREAT: R = 0.95). These results can be directly attributable to the glacier surface conditions that were present at each profile. Profiles 8 and 9a were in the accumulation zones of the Matthes Glacier, while Profile 4 on the Taku Glacier was in the middle of the highly variable transient snow line. The glacial surface in these zones is ubiquitously snow-covered and characterized by ablation hollows, or sun-cups, which serve as the primary features for tracking. Rainfall, snowmelt, and other drivers of liquid water content variability likely caused shifts in backscatter coefficients between successive flyovers as was observed for Taku Glacier in 2018 (
Figure 10). Variability in the backscatter amplitude likely reduced the success of patch intensity cross-correlation, and a SNAP correlation threshold value of 0.25 was not large enough to filter out the erroneous results, which stood in contrast to the sparse coverage from heavily filtered correlations provided by RETREAT. Profile 13 was in the ablation zone of the Llewellyn Glacier and was characterized by bare ice, crevasses, and abundant supraglacial debris on its surface. These glacial features served as persistent, backscatter-stable targets for offset tracking, resulting in the significant positive correlation with the GNSS-measured velocities.
In addition to the observed accuracy dichotomy between the accumulation and ablation zones, analyzing the spatial distribution of the satellite-derived velocity errors (Δv) for SNAP and RETREAT revealed that cross-correlation performance within the accumulation zone was sensitive to velocity magnitudes and hence position on the glacier. Both the spatial distribution of low Δv (
Figure 6) and
k-means clustering (
Figure 7) revealed that velocities derived by SNAP and RETREAT using Sentinel-1A/B imagery were accurate (Δv < 0.05) at locations where actual flow velocity was slow (i.e., along the glacial margins). The error of results obtained by SNAP increased nearly linearly with increasing flow velocity, especially when flow velocity > 0.45 m/day. The measured correlation at Profile 4 for RETREAT in 2016 (R = 0.79 and RMSE = 0.09) was much improved over SNAP, even though both used Sentinel-1 image-pairs. This was likely due to the substantial filtering by RETREAT, which followed a 3-step process developed by Lüttig et al. (2017) that detects smooth segments, removes statistical outliers, and removes velocities with directional anomalies [
48]. This created an apparent performance improvement based on the correlation analyses, but the two SAR-derived datasets were comparable after accounting for the inclusion of SNAP’s velocity estimates in the central flowline regions. Proximity to static features (bedrock) along the glacier margins was likely not a contributing factor to the derived slow velocities (and improved accuracy) given that similar amplitude (and accurate) velocities were derived near the central axis of Profile 8 at a distance greater than the patch size from the nearest nunatak. The spatial variability of Δv for the optical image-derived velocities was indiscriminate and did not show any apparent sensitivity to actual flow velocity or position on the glacier.
5. Conclusions
Glacial flow velocity is a commonly measured parameter used for discerning the changing dynamics and discharge of glaciers. Given the logistical difficulties, and often low priority, of measuring velocity in situ, cross-correlation of spaceborne radar and multispectral image-pairs has become the preferred technique for deriving displacement. Recent studies focused on assessing the accuracy of satellite-derived velocities have mainly focused on ice sheets, given the pronounced interest in those regions, and have shown strong agreement between the field- and satellite-derived velocities. Fewer validation efforts have been made in temperate alpine glacier regions, where higher temperatures and precipitation rates can cause SAR backscatter and multispectral reflectance variability of the glacier surface on very short timescales (hours to days) and uncertainty in cross-correlation of image-pairs.
This study provided a robust assessment of satellite-derived velocities based on direct comparison with spatially coincident and contemporaneous GNSS-measured velocity data on the Juneau Icefield (Alaska) from 2016 through 2018. Results presented here indicate that there are significant gaps in the displacement results derived from optical imagery (Landsat/Sentinel-2) that were not a result of cloud contamination but rather removal due to low coherence. Of the few optical image-derived velocities that were acquired at the stake positions, nearly all were deemed inaccurate by comparison to the GNSS-measured data. These results could be attributed to the very short temporal baseline (<16 days) used in this study, as the accuracy of velocities derived from optical image-pairs greatly improve with larger baselines. Therefore, using cross-correlation of optical image-pairs to derive flow velocity at temperate glaciers is perhaps best suited for longer frequency measurements of velocity (annual) rather than short-term changes. Investigation of ITS_LIVE v1 and ITS_LIVE v2 (S2) velocity data with low frequency baselines revealed substantial improvement in spatial coverage (very few data gaps) but differentiation of velocity values across the glaciers was nearly entirely lost, with the margins along Taku Glacier showing the same annual velocities as the central flow axis.
Comparing the cross-correlation results of SAR data (Sentinel-1) using multiple techniques (autoRIFT/SNAP/GAMMA) with the GNSS-measured velocity data revealed a two-tiered control on the performance of the technique. At a first order, there was a strong dichotomy in the accuracy achieved between the study areas located in the accumulation zone (Profiles 8 and 9a) and the area contained by the transient snowline (Profile 4) versus the study area located in the ablation zone (Profile 13). SAR-derived velocities in the accumulation zone/transient snowline area showed poor correlation to the GNSS-measured data, likely because of significant shifts in the backscatter amplitude of the homogenous, snow-covered glacier surface between the successive satellite flyovers that produced the image-pair. Contrastingly, SAR-derived velocities (using SNAP/GAMMA) in the ablation zone were highly accurate when compared to the GNSS data (R = 0.89/0.90 and RMSE = 0.05/0.12 m/day). This is likely attributed to the surface features of the ablation zone providing stable backscatter targets such as crevasses and supraglacial debris that encourage high coherence in the cross-correlation procedure. A second-order control was observed on the performance of SAR-derived velocities in the accumulation zone. SAR velocity results were anomalously accurate where GNSS velocities were below ~0.10 m/day, most commonly along the glacier margins. Above that velocity threshold, accuracies deteriorated linearly with increasing GNSS-measured velocities. Based on the observations made in this study, cross-correlation of SAR image-pairs is likely best suited for derivation of velocities in the ablation zone and across slow-flowing glaciers in a temperate alpine environment and should be subject to scrutiny for fast-flowing temperate alpine glaciers or those with active hydrologic surface systems.