Snow water equivalent retrieval over Idaho – Part 1: Using Sentinel-1 repeat-pass interferometry

Abstract

Snow water equivalent (SWE) is identified as the key element of the snowpack that impacts rivers' streamflow and water cycle. Both active and passive microwave remote sensing methods have been used to retrieve SWE, but there does not currently exist a SWE product that provides useful estimates in mountainous terrain. Active sensors provide higher-resolution observations, but the suitable radar frequencies and temporal repeat intervals have not been available until recently. Interferometric synthetic aperture radar (InSAR) has been shown to have the potential to estimate SWE change. In this study, we apply this technique to a long time series of 6 d temporal repeat Sentinel-1 C-band data from the 2020–2021 winter. The retrievals show statistically significant correlations both temporally and spatially with independent in situ measurements of SWE. The SWE change measurements vary between $-$5.3 and 9.4 cm over the entire time series and all the in situ stations. The Pearson correlation and RMSE between retrieved SWE change observations and in situ stations measurements are 0.8 and 0.93 cm, respectively. The total retrieved SWE in the entire 2020–2021 time series shows an SWE error of less than 2 cm for the nine in situ stations in the scene. Additionally, the retrieved SWE using Sentinel-1 data is well correlated with lidar snow depth data, with correlation of more than 0.47. Low temporal coherence is identified as the main reason for degrading the performance of SWE retrieval using InSAR data. We also show that the performance of the phase unwrapping algorithm degrades in regions with low temporal coherence. A higher frequency such as L-band improves the temporal coherence and SWE ambiguity. SWE retrieval using C-band Sentinel-1 data is shown to be successful, but faster revisit is required to avoid low temporal coherence. Global SWE retrieval using radar interferometry will have a great opportunity with the upcoming L-band 12 d repeat-pass NASA-ISRO Synthetic Aperture Radar (NISAR) data and the future 6 d repeat-pass Radar Observing System for Europe in L-band (ROSE-L) data.

1 Introduction

The seasonal snowpack provides water resources to billions of people worldwide . Snow is the primary source of water for river channel discharge in middle-to-high-latitude areas. Therefore, snow mass and snow cover has a great impact on global and regional water cycles. Large-scale mapping of snow water equivalent (SWE) with high resolution is critical for many scientific and economics fields. SWE is defined as the depth of water which would be obtained if all ice contained in the snowpack were melted. NASA SnowEx is a multi-year effort to improve SWE and snow surface energy balance measurements and estimates. SWE has been identified as the key variable for terrestrial snow by the SnowEx campaign and NASA's decadal survey.

Estimating SWE on a global scale with enough accuracy and resolution is still a challenge. Passive spaceborne sensors based on the microwave emission of the snowpack have a coarse spatial resolution on the 10 km scale. The technique saturates for SWE deeper than 150 mm, which makes their application in the mountains challenging. Nevertheless, passive microwave sensors represent the current state of the art of SWE retrieval methods. These sensors are applied operationally to generate daily estimates of SWE globally ; however, many products such as GlobSnow mask out mountainous areas, due to the saturation limit and resolution.

Airborne lidar has been successful in estimating snow depth . However, clouds and limited regional coverage are limiting factors for this method. This technique also needs a snow-density model to estimate SWE from the lidar snow depth, and there currently is not a path to space for global snow depth mapping at the temporal resolution required.

Active microwave sensors provide high-resolution and global coverage. There have been many efforts in the last 2 decades trying to estimate SWE or snow depth using active sensors mounted on a tower , airborne , or spaceborne systems . Backscattered power from active sensors is used to estimate SWE . A dual-band (X and Ku) SAR mission has been the focus of the European Space Agency (ESA) and Canadian Space Agency (CSA) for SWE spaceborne measurements . However, accurate a priori characterization of snow micro-structural parameters is of primary importance in the accuracy of SWE retrieval algorithms using backscattered power . The most common a priori characterization used for SWE retrieval algorithms using backscattered power is grain radius. This has been done using passive data; however, the methods are limited by passive retrieval errors and also mismatch between active and passive resolutions. The ratio of cross-polarized to co-polarized Sentinel-1 backscattered power has been used to estimate snow depth over mountainous regions with deep snow . Using Sentinel-1 backscattered power ratio is a unique approach showing the success of snow depth retrieval using the spaceborne radar time series data. However, the retrieval mostly works for deep snow in mountainous regions. The radiative transfer physics at C-band for this method are still poorly understood. The co-polar phase difference (CPD) between VV and HH polarization of X-band SAR acquisitions is used for estimating the depth of fresh snow .

Lightweight and portable frequency-modulated continuous-wave (FMCW) radar systems have been used to map snowpack properties (such as depth, SWE, and stratigraphy) rapidly over large distances and at high resolution . The system was normally deployed nadir looking and was a real aperture radar system. The resolution of FMCW system for SWE application is in centimeter scale. In order to achieve such high resolution, the bandwidth should be in gigahertz scale. Due to limitation on frequency bandwidth allocation of a spaceborne active sensor , FMCW systems cannot be used in spaceborne missions for global coverage due to their wide bandwidth.

The phase change of specularly reflected signals in signals of opportunity (SoOp) is shown to be strongly dependent on SWE changes for dry snow . The theory behind using SoOp for SWE retrieval is similar to repeat-pass interferometry that is explained in Sect. . The advantage of this method is that the stratigraphy of the snow has little impact on the SWE retrieval , similar to SWE retrieval explained in Sect. . Using the long wavelength signal at P-band in SoOp is very helpful for addressing the loss of temporal coherence and phase unwrapping challenges of this method. However, the phase sensitivity to SWE changes decreases at lower frequencies. There have been very limited data showing the success of this method at P-band. Achieving high resolution for SoOp data is another challenge .

As explained in detail in next section, the phase difference between two SAR observations is proportional to changes in SWE variation ($\mathrm{\Delta }\mathrm{SWE}$). We evaluated the performance of SWE retrieval using interferometry over Idaho. In Part 1 of this study (the current paper), we used Sentinel-1 interferometric time series data over Idaho. In Part 2 , we use Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) interferometric time series data over Idaho to evaluate the performance of this method. We explain SWE estimation using repeat-pass interferometry in Sect. . The details about different data sets used in this study are discussed in Sect. . Section  describes how we processed Sentinel-1 data and convert them to SWE. The retrieved SWE is compared with in situ and lidar data in Sect. . This work shows the success of SWE retrieval using long-time-series spaceborne InSAR data in winter 2021.

2 Using differential interferometry to estimate SWE

Differential SAR interferometry measurements have been used to detect small surface elevation changes over large areas with a vertical accuracy of a few millimeters . The measured phase difference is proportional and sensitive to changes in SWE variation ($\mathrm{\Delta }\mathrm{SWE}$) during the snow season . The main advantage of this method is its simplicity and a reduction in necessary a priori information.

The snow volume scattering affects the interferometric phase for very deep snow in Greenland at relatively high frequencies such as C-band . However, for the terrestrial snow, the effect of volume scattering of dry snow on the interferometric phase is very small compared to scattering from the ground at high frequencies. The snow refractive index delays the echo received from the ground. The signal delay caused by refraction can be measured with differential radar interferometry as 

1$$\mathrm{\Delta }\mathit{\varphi }=-\mathrm{2}{\mathit{\kappa }}_i(\cos \mathit{\theta }-\sqrt{\mathit{ϵ}-{\sin }^{\mathrm{2}}\mathit{\theta }})\mathrm{\Delta }d,$$ where $\mathrm{\Delta }\mathit{\varphi },{\mathit{\kappa }}_i,\mathrm{\Delta }d,\mathit{\theta }$, and $\mathit{ϵ}$ are the interferometric phase between two observation dates, incidence wavenumber, snow depth change, incidence angle, and permittivity of the snow, respectively. The change in the interferometric phase is used to calculate $\mathrm{\Delta }\mathrm{SWE}$ . Similar to the dual-polarization, dual-frequency retrieval algorithm , this technique relies on the dryness of snow in order to penetrate all the way to the ground so that the scattering from the snow layers and snow volume is minimized compared to the snow–ground return .

Using Envisat interferometric data to estimate SWE was not very successful mainly due to the large temporal baseline and, hence, low temporal coherence . A modified version of SWE estimation using InSAR is also introduced . The backscattering from the roughness in the ground and snow layers is combined with the interferometric phase to improve the accuracy . The sensitivity of the dry-snow refraction-induced InSAR phase to topographic variations is used to bypass the unwrapping problem . Airborne data collected over the Austrian Alps in 2021 showed good agreement between retrieved SWE using InSAR and mean in situ SWE. Root mean square differences of 4.0 mm for a small snowstorm of 14 mm snow depth at C-band and 11.2 mm for a big snowstorm of 66 mm at L-band were observed . The correlation of 0.76 was observed between the retrieved SWE change using L-band UAVSAR differential interferometry between 1 and 13 February 2020 and the collected lidar snow depth change between 1 and 12 February 2020 over the open regions of Grand Mesa in dry-snow conditions . SWE retrieval using Sentinel-1 interferometric data showed a mean accuracy of 6 mm over Finland for just two passes .

All these studies have proven the potential of this method but were limited in time or space for data collection or validation. In this study, we show the performance of SWE retrieval using a long time series of Sentinel-1 interferometric data in winter 2021. This study shows that SWE estimation using repeat-pass interferometry works well by validating the retrieved value with a large number of in situ stations and two regional lidar snow depth maps.

With the recent SnowEx 2020 campaign using UAVSAR L-band differential interferometry data, Sentinel-1 C-band differential interferometry, and future NASA-ISRO SAR (NISAR) L-band data, there will be more advances in the limitations and capabilities of this method.

2.1 Temporal coherence

The received radar signals at two different times will be correlated with each other if the set of scatterers in the resolution cell remain the same. However, the movement of the scatterers such as leaves and branches or sea ice particles decreases the temporal coherence . The loss of coherence between the observations is one of the main limitations for SWE retrieval using differential interferometry. Methods such as using two frequencies or shorter revisit time are used to overcome these problems . Melting and wind are the main reasons for low temporal coherence in snow . A medium mean temporal coherence of 0.41 is observed at L-band between two winter seasons in shrub-lands with 10.2 cm average snow depth . Temporal coherence decreases with increasing frequency . A median temporal coherence of about 0.5 is observed at 10.2 and 16.8 GHz even after 60 d . However, the spaceborne TerraSAR-X temporal coherence over snow at 9.65 GHz is reduced significantly in 11 d . This is probably due to random phase drifts over time that cannot be estimated and corrected in a spaceborne system compared to a ground radar. Vegetation cover decreases the temporal coherence significantly at high frequencies . A tower-based fully polarimetric InSAR studied the effect of air temperature, precipitation intensity, and wind on the temporal decorrelation at L-, S-, C-, and X-bands . The temperature was shown to be the most critical variable affecting the temporal coherence among other variables. Temperature above $\mathrm{0}$ ${}^{\circ }$C reduced the temporal coherence drastically . On the other hand, snow cover has a thermal insulation effect on the ground and underlaying layers . The insulation increases with the snow depth. Therefore, during the snow season we assume the ground remains frozen even when snow becomes wet. Hence, temporal decoherence from the ground is negligible. SWE accumulation retrieval was successful for short temporal baselines and low frequencies in non-vegetated areas. However, the error increased for high frequencies and long temporal baselines. The SWE profile retrieval using C-band data performs well using 12 h and 1 d repeat-pass data. The retrieval is poor using the 12 d repeat-pass data at C-band . The 6 d repeat-pass C-band data showed good performance for small SWE changes but poor performance for large SWE changes between the interferometric pairs due to phase ambiguity caused by large SWE change . The low temporal coherence and low penetration depth at frequencies higher than 10 GHz make L- and C-band desirable frequencies for differential interferometry.

2.2

Relationship between $\mathrm{\Delta }\mathrm{SWE}$ and $\mathrm{\Delta }\mathit{\varphi }$

With some approximation to Eq. (), showed a linear relationship between the interferometric phase and SWE change. The approximation is limited to a smaller range of incidence angle than Sentinel-1 incidence angle. However, approximation applies to a wide range of snow density up to solid ice density. Due to the wide range of Sentinel-1 incidence angle in a frame, we tried to make a more accurate approximation for a wider range of incidence angles and snow densities limited to terrestrial snow. The snow permittivity in Eq. () is dependent on snow density, $\mathit{\rho }$ (g cm${}^{-\mathrm{3}}$), and relatively independent of signal wavelength. Following , we use Mätzler's model for calculating $\mathit{ϵ}$ in Eq. () ($\mathit{ϵ}\left(\mathit{\rho }\right)=\mathrm{1}+\mathrm{1.5995}\mathit{\rho }+\mathrm{1.861}{\mathit{\rho }}^{\mathrm{3}}$ for $\mathit{\rho }<\mathrm{0.4}$ g cm${}^{-\mathrm{3}}$; and $\mathit{ϵ}\left(\mathit{\rho }\right)=\left(\right(\mathrm{1}-\frac{\mathit{\rho }}{\mathrm{0.917}})+\mathrm{1.4759}\frac{\mathit{\rho }}{\mathrm{0.917}}{)}^{\mathrm{3}}$ for $\mathit{\rho }\mathit{>=}\mathrm{0.4}$ g cm${}^{-\mathrm{3}}$). We can rewrite Eq. () as 2$$\mathrm{\Delta }\mathit{\varphi }=-\mathrm{2}{\mathit{\kappa }}_{i}C(\mathit{\theta },\mathit{\rho }){\displaystyle \frac{{\mathit{\rho }}_{\mathrm{water}}}{\mathit{\rho }}}\mathrm{\Delta }\mathrm{SWE},$$ where $C(\mathit{\theta },\mathit{\rho })=\cos \mathit{\theta }-\sqrt{\mathit{ϵ}\left(\mathit{\rho }\right)-{\sin }^{\mathrm{2}}\mathit{\theta }}$. Note that $\mathit{ϵ}\left(\mathit{\rho }\right)$ and consequently $C(\mathit{\theta },\mathit{\rho })$ are unitless, ${\mathit{\rho }}_{\mathrm{water}}=\mathrm{1}$ g cm${}^{-\mathrm{3}}$, $\mathrm{\Delta }\mathrm{SWE}$ is in meters (m), and ${\mathit{\kappa }}_i$ is in per meter (m${}^{-\mathrm{1}}$).

Figure 1

(a) $C(\mathit{\theta },\mathit{\rho })$ versus snow density for $\mathit{\theta }=\mathrm{0}$ and $\mathit{\theta }=\mathrm{70}$ shown by solid blue and red lines, respectively. The blue and red dashed lines show the linear fit to $C(\mathit{\theta },\mathit{\rho })$ with zero intercept $\widehat{C}(\mathit{\theta },\mathit{\rho })$. (b) The line slope in panel (a) versus incidence angle. The dashed line shows the fitted polynomial, $\widehat{A}$. (c) $\mathrm{\Delta }\mathrm{SWE}$ error percentage ($\frac{\mathrm{\Delta }\mathrm{SWE}-\widehat{\mathrm{\Delta }\mathrm{SWE}}}{\mathrm{\Delta }\mathrm{SWE}}\times \mathrm{100}$), assuming $\widehat{C}=\widehat{A}\mathit{\rho }$ versus incidence angle for snow density equal to 0.15, 0.3, and 0.45 g cm${}^{-\mathrm{3}}$, shown by blue, red, and yellow lines, respectively.

[Figure omitted. See PDF]

The blue and red lines in Fig. a show $C(\mathit{\theta },\mathit{\rho })$ versus snow density for incidence angles equal to 0 and 70${}^{\circ }$, respectively. As seen in this figure, there is approximately a linear relationship between $C$ and snow density. We fit a line to $C$ for different incidence angles as $\widehat{C}(\mathit{\theta },\mathit{\rho })=A\left(\mathit{\theta }\right)\times \mathit{\rho }$ for $\mathrm{0.15}\le \mathit{\rho }\le \mathrm{0.5}$ g cm${}^{-\mathrm{3}}$. The blue and red dashed lines show $\widehat{C}(\mathit{\theta },\mathit{\rho })$ at incidence angles equal to 0 and 70, respectively. As seen in Fig. a, the fitted line with zero intercept is a good approximation. The zero intercept approximation is essential to retrieve $\mathrm{\Delta }\mathrm{SWE}$ independent of snow density. The incidence angle mostly lies between 0 and 80 for Sentinel-1 data. The terrestrial snow density lies mostly between 0.15 and 0.5 g cm${}^{-\mathrm{3}}$. Therefore, we limit ourselves to incidence angle between 0 and 80 and snow density between 0.15 and 0.45 g cm${}^{-\mathrm{3}}$ in fitting a line to $C$. Solid blue line in Fig. b shows $A\left(\mathit{\theta }\right)$ versus incidence angle. By fitting a polynomial to A, we can write it as 3$$\widehat{A}=-\mathrm{0.6784}{\mathit{\theta }}^{\mathrm{2}}+\mathrm{0.2899}\mathit{\theta }-\mathrm{0.8473}.$$ The dashed blue line in Fig. b shows the fitted curve, $\widehat{A}\left(\mathit{\theta }\right)$. We can rewrite Eq. () as 4$$\mathrm{\Delta }\mathit{\varphi }=-\mathrm{2}{\mathit{\kappa }}_i(-\mathrm{0.6784}{\mathit{\theta }}^{\mathrm{2}}+\mathrm{0.2899}\mathit{\theta }-\mathrm{0.8473})\widehat{\mathrm{\Delta }\mathrm{SWE}}.$$

Figure c shows the $\frac{\mathrm{\Delta }\mathrm{SWE}-\widehat{\mathrm{\Delta }\mathrm{SWE}}}{\mathrm{\Delta }\mathrm{SWE}}\times \mathrm{100}$ versus incidence angle for different snow densities. As seen in this figure, the error in $\mathrm{\Delta }\mathrm{SWE}$ calculation using the approximation in Eq. () is less than 10 % for incidence angles less than 70. We use Eq. () for estimating $\mathrm{\Delta }\mathrm{SWE}$ using the interferometric phase, $\mathrm{\Delta }\mathit{\varphi }$, for the rest of this study. Using one equation for the entire Sentinel-1 frame makes the interferometric phase conversion to $\mathrm{\Delta }\mathrm{SWE}$ very convenient. However, we need to keep in mind that the approximation for lower-density snow has more than 10 % error for an incidence angle larger than 70.

3 Data sets

3.1 Sentinel-1

The Sentinel-1 radar operates at C-band at a central frequency of 5.405 GHz. It has four exclusive imaging modes with different resolutions (down to 5 m) and swath width up to 400 km. Sentinel-1 has dual-polarization capability and rapid product delivery. The Sentinel-1 constellation includes Sentinel-1A and Sentinel-1B. These two satellites are in the same orbit, with a $\mathrm{180}{}^{\circ }$ orbital phasing difference. The revisit time for each of the satellites is 12 d. However, revisit time can get to 6 d if both satellites make observations. The data are free and available through the Alaska SAR Facility (ASF) or the Copernicus Data Hub distribution service. We used the Interferometric Wide (IW) swath mode data with 5 and 20 m single-look resolution in the range and azimuth direction, respectively. The IW swath width is about 250 km. We used ASF on-demand processing to generate the interferometric phase and coherence at vv and vh (transmit-received polarization) polarization. The Alaska Satellite Facility's Hybrid Pluggable Processing Pipeline (HyP3) is a service for processing synthetic aperture radar (SAR) imagery . The workflow includes interferometric phase correction for ground topography and geolocation. The ASF HyP3 uses a minimum cost flow (MCF) algorithm for phase unwrapping. The unwrapped phase and interferometric coherence were used in this study. The resolution of the HyP3 phase and coherence is 80 m $\times$ 80 m. Sentinel-1 collects data every 12 d globally but has the capability to collect the data every 6 d over targeted areas, mainly over Europe and selected areas such as SnowEx sites. In order to validate our SWE retrieval using Sentinel-1 data, we use lidar data from the SnowEx campaign and SNOwpack TELemetry Network (SNOTEL) data as discussed in Sect. . We also use the average of SNOTEL data as a reference point for SWE retrieval, as seen in Sect. .

Figure 2

© Google Earth View. (a) Google Earth View of Sentinel-1 path 71, frame 444, in Idaho. (b) Zoomed to the Sentinel-1 path 71, frame 444, shown by the big green rectangle. Red boxes show the location of lidar data acquisition. The green diamonds show SNOTEL stations with an $\mathrm{\Delta }\mathrm{SWE}$ error of less than 2 cm in the entire time series. The red diamonds show SNOTEL stations with an $\mathrm{\Delta }\mathrm{SWE}$ error of more than 2 cm in at least one observation in the time series. Yellow squares are SNOTEL stations 1 and 11 used are a reference point. Blue diamonds show the location of stations with temporal coherence less than 0.35 or temperature more than $\mathrm{0}$ ${}^{\circ }$C in the entire time series.

[Figure omitted. See PDF]

The NASA SnowEx 2021 time series is the continuation of the multi-year effort to improve SWE measurements and estimates. The data acquisition for different sensors and in situ collections spread over different US sites in winter 2020. These sites span a range of snow climates and conditions, elevations, aspects, and vegetation. Flight paths were designed to include sites with ongoing snow research projects, existing ground-based remote sensing infrastructure (e.g., radar and lidar), snow-off and planned snow-on aerial lidar, and scheduled ground snow measurement. The 2021 time series data set covers fewer regional sites and more frequent temporal sampling compared to the 2020 campaign. The SnowEx campaign coordinated with the Sentinel-1 team to observe some of the SnowEx sites with 6 d revisit during the winter, which included the Idaho SnowEx sites.

Figure a shows one of these sites that was observed every 6 d with Sentinel-1 over Idaho. The green frame shows the geographic location of path 71, frame 444, of Sentinel-1 data. Figure b is zoomed to the Sentinel-1 frame in panel (a).

3.2 SNOTEL

SNOwpack TELemetry Network (SNOTEL) sites are located in remote, high-elevation mountainous regions in the western US. They automatically measure different snowpack characteristics and climate conditions. We used the United States Department of Agriculture (USDA) website to access hourly SNOTEL data (https://wcc.sc.egov.usda.gov/nwcc/inventory, last access: August 2023) over the region of interest shown in Fig. b. As the Sentinel-1 frame in Idaho is collected at around 06:00 local time, we downloaded the SWE, snow depth, and near-surface air temperature at 06:00 for each of the SNOTEL stations. Small red, green, and blue diamonds in Fig. b show the SNOTEL locations in the Sentinel-1 frame. Figure a shows the time series SWE of these SNOTEL sites starting from 1 December 2020 at 06:00. Different colors show different SNOTEL stations. The elevation of these stations varies between 975 and 2902 m. Therefore, the large spread of SWE between different stations in Fig. a is expected. The dashed vertical lines are the start date of each 6 d repeat Sentinel-1 data. As seen in this figure, there is a 6 d repeat data acquisition gap in Sentinel-1 data on 5 February 2021. Figure b shows the mean $\pm$ standard deviation (SD) of SNOTEL $\mathrm{\Delta }\mathrm{SWE}$ between the start date Sentinel-1 data in Fig. a and 6 d later. We used the SWE data from these in situ stations for (a) SWE retrieval validation by comparing retrieved $\mathrm{\Delta }\mathrm{SWE}$ with SNOTEL $\mathrm{\Delta }\mathrm{SWE}$ (as seen in Sect. ) and (b) the InSAR reference point by subtracting the average of two SNOTEL $\mathrm{\Delta }\mathrm{SWE}$ from the retrieved $\mathrm{\Delta }\mathrm{SWE}$ (as explained in Sect. ).

Figure 3

(a) The daily SWE (cm) of in situ stations shown in Fig. b from 1 December 2020 to 30 March 2021. The dashed vertical lines show the start date of Sentinel-1 observations. (b) The mean $\pm$ SD of in situ $\mathrm{\Delta }\mathrm{SWE}$ for Sentinel-1 observation dates shown in panel (a). Note that the $\mathrm{\Delta }\mathrm{SWE}$ is marked on the first day of each observation.

[Figure omitted. See PDF]

3.3 QSI lidar

Airborne lidar provides high-resolution snow depth maps. These data are reliable sources of validation data and a particularly powerful constraint for InSAR retrieval of SWE. We used the lidar data for validating the retrieved SWE results. The “SnowEx20-21 QSI Lidar DEM 0.5m” data set is part of the SnowEx 2020 and SnowEx 2021 campaigns . The data include digital elevation models, snow depth, and vegetation height with 0.5 m spatial resolution. Data were acquired over multiple areas in Colorado, Idaho, and Utah during February 2020, March 2021, and September 2021. The two red boxes in Fig. b show the location of lidar data acquisition. The big purple box is over Banner Summit and the small red box is over Mores Creek in Idaho. Figures a and a show the QSI snow depth over Banner Summit and Mores Creek, respectively. We used these data in Sect.  to compare with retrieved SWE using Sentinel-1 data.

4 SWE retrieval using the Sentinel-1 interferometric phase

As mentioned in Sect. , Sentinel-1 data were collected every 6 d over the region shown in Fig. b during 2020 and 2021, following coordination between the SnowEx campaign and the Sentinel-1 team. We used 6 d repeat Sentinel-1 time series data between 1 December 2020 and 30 March 2021. We selected this period to (a) capture most of the seasonal snowstorm and (b) avoid wet snow as much as possible. The main sources of error in the science and applications using Sentinel-1 repeat-pass interferometry are (1) tropospheric noise, (2) temporal decorrelation, and (3) phase ambiguity. We removed tropospheric noise from the unwrapped phase as explained in Sect. . The unwrapped phase is converted to $\mathrm{\Delta }\mathrm{SWE}$ using Eq. (). Temporal decorrelation is relatively high at C-band. The 6 d repeat time improves the temporal coherence significantly over snow compared to the normal 12 d Sentinel-1 repeat time. In this study, any pixel with temporal coherence more than 0.35 is considered reliable. Temporal coherence of 0.35 is arbitrary, but based on experience working with InSAR data, it is a reasonable threshold number. However, for the results in Sect. , we used all the time series data, including the data with low coherence, to calculate total SWE. The reason is that in order to compare the total SWE on a date close to the lidar acquisition date, we need the whole $\mathrm{\Delta }\mathrm{SWE}$ time series up to that date.

Figure 4

Retrieved $\mathrm{\Delta }\mathrm{SWE}$ using Sentinel-1 path 71, frame 444, interferometric phase data between (a) 1 and 7 December 2020, (b) 13 and 19 December 2020, and (c) 24 and 30 January 2021. Sentinel-1 path 71, frame 444, coherence between (d) 1 and 7 December 2020 (observation 1), (e) 13 and 19 December 2020 (observation 3), and (f) 24 and 30 January 2021 (observation 10). The small diamonds are in situ locations. The averages of in situ $\mathrm{\Delta }\mathrm{SWE}$ for panels (a), (b), and (c) are 0.01, 2.72, and 4.33 cm, respectively.

[Figure omitted. See PDF]

Phase ambiguity is still one of the big sources of error in some of our data as discussed in Sect. . The radar signal propagating through the ionosphere is delayed. The delay is a function of frequency of the signal, Earth's magnetic field, and total electron content (TEC) and affects the accuracy of the $\mathrm{\Delta }\mathrm{SWE}$ retrieval. The ionospheric error at C-band is much smaller than other sources of error, and we consider it negligible in this study.

The temperature is also an important factor. Equation () is valid for dry snow , and we use near-surface air temperature above $\mathrm{0}$ ${}^{\circ }$C as a metric that indicates wet snow in the snow season. Any retrieved SWE with SNOTEL near-surface air temperature more than $\mathrm{0}$ ${}^{\circ }$C is unreliable in our study. Similar to coherence filtering, for the results in Sect. , we used all the time series data, including the data with temperature more than $\mathrm{0}$ ${}^{\circ }$C. Similar to temporal coherence, the reason is that in order to compare the total SWE with lidar snow depth on the lidar acquisition date, we need the entire $\mathrm{\Delta }\mathrm{SWE}$ time series up to that date.

Another important factor in interferometric phase images is the reference point to calibrate the unwrapped phase or consequently $\mathrm{\Delta }\mathrm{SWE}$. In geophysics applications using InSAR, the reference point is a stable target with no displacement or known displacement in the time interval between acquisition of the two images. For $\mathrm{\Delta }\mathrm{SWE}$ estimation using InSAR, the reference point is chosen either by corner reflectors (cleaned of snow) with stable zero phase or using the average of in situ $\mathrm{\Delta }\mathrm{SWE}$ or using a snow-free region . As seen in Fig. b, there are a large number of in situ stations in this frame. In this study, we used the average of two in situ $\mathrm{\Delta }\mathrm{SWE}$ values to calibrate the retrieved $\mathrm{\Delta }\mathrm{SWE}$ images. The two selected in situ stations have reliable measurements (coherence more than 0.35 and temperature less than $\mathrm{0}$ ${}^{\circ }$C) for the entire time series. For the rest of this study we used in situ stations 1 and 11 $\mathrm{\Delta }\mathrm{SWE}$ values to calibrate the retrieved $\mathrm{\Delta }\mathrm{SWE}$. Stations 1 and 11 are shown by yellow squares in Fig. b.

Figure a, b, and c show retrieved $\mathrm{\Delta }\mathrm{SWE}$ between 1 and 7 December 2020, 13 and 19 December 2020, and 24 and 30 January 2021, respectively. The small diamonds show the location of in situ stations in this Sentinel-1 frame. The averages of in situ $\mathrm{\Delta }\mathrm{SWE}$ for Fig. a, b, and c are 0.01, 2.72, and 4.33 cm, respectively. The retrieved $\mathrm{\Delta }\mathrm{SWE}$ images in the top row of Fig.  show no SWE change in panel (a) and snowstorms in panels (b) and (c), which match the in situ measurements.

The bottom row of Fig.  shows the coherence of the images in the top row of Fig. . Interferometric decorrelation has different sources, such as temporal decorrelation, volume decorrelation, signal to noise ratio decorrelation, and geometric decorrelation, among others. The volume decorrelation is negligible due to the relatively small Sentinel-1 perpendicular baseline. Temporal decorrelation is the dominant source of decorrelation. For the rest of this study, we assume the observed interferometric decorrelation is approximately the temporal coherence. As shown in Fig. e and f, snowstorms reduce the coherence significantly, whereas no SWE change shows a very small decorrelation, as expected.

4.1 Tropospheric noise removal

A radio wave's differential phase delay variation through the troposphere is one of the largest error sources in interferometric synthetic aperture radar (InSAR) measurements, and water vapor variability in the troposphere is known to be the dominant factor. The differential delay present in a given interferogram may reach tens of centimeters. Various ways of mitigating tropospheric effects are routinely employed. Here, we used a global atmospheric weather model to predict the radar phase delay due to variations in atmospheric pressure and water vapor content between passes. Specifically, we used the European Center for Medium-Range Weather Forecasts (ECMWF) ERA5 model of atmospheric variables, which provides hourly estimates on a 30 km global grid based on assimilation of surface and satellite meteorological data. We used the Python-based Atmospheric Phase Screen (PyAPS) software to interpolate this grid and convert those variables into a radar phase delay. PyAPS is integrated into, and leveraged by, the Miami InSAR Time-series software in Python (MintPy) . We used MintPy to crop the atmospheric delays to match the spatial extent of the interferograms and projected the delays into radar line of sight (LOS). It should be noted that while the ERA weather models often provide a reliable method for representing atmospheric phenomena at $>\mathrm{30}$ km wavelengths (grid spacing), they are less accurate at finer spatial scales, where atmospheric conditions can vary as a function of topography. Model interpolation between grid nodes as a function of elevation were performed; however, some over-smoothing of atmospheric variations might still occur. More work is necessary to better determine the overall effectiveness of atmospheric phase removal, including whether tropospheric delay is completely mitigated or overcorrected and on what spatial scales.

Figure 5

Sentinel-1 path 71, frame 444, (a) line-of-sight (LOS) displacement (m) with atmospheric noise, (b) atmospheric noise (m), and (c) line-of-sight displacement (m) without atmospheric noise, between 13 and 19 March 2021.

[Figure omitted. See PDF]

Figure shows an example of how significant tropospheric noise can be in an InSAR image. Figure a shows the line-of-sight displacement with no atmospheric correction over our area of interest in Fig. b between 13 and 19 March2021. Figure b shows the atmospheric noise estimation using PyAPS. Figure c shows LOS displacement after tropospheric noise removal by subtracting panel (b) from panel (a). Comparing Fig. a and c, we can see that the atmospheric noise can affect the estimated $\mathrm{\Delta }\mathrm{SWE}$ by 5–10 cm (LOS displacement error converted to $\mathrm{\Delta }\mathrm{SWE}$) in the upper left of the images.

5 Results and discussions

In this section we compare retrieved SWE using the Sentinel-1 interferometric phase with in situ stations and lidar data.

5.1 Comparing retrieved SWE using Sentinel-1 and SNOTEL SWE

5.1.1

Comparing retrieved $\mathrm{\Delta }\mathrm{SWE}$ using Sentinel-1 and SNOTEL $\mathrm{\Delta }\mathrm{SWE}$

We used all the retrieved $\mathrm{\Delta }\mathrm{SWE}$ (using the Sentinel-1 data from 1 December 2020 to 30 March 2021) for in situ stations shown in Fig. b and compared them with corresponding SNOTEL $\mathrm{\Delta }\mathrm{SWE}$. As mentioned in Sect. , any retrieved value with temporal coherence less than 0.35 and temperature higher than $\mathrm{0}$ ${}^{\circ }$C is discarded. Note that the data shown in Fig.  are the SWE change between two consecutive Sentinel-1 data that are 6 d apart. We showed the $\mathrm{\Delta }\mathrm{SWE}$ for all stations and all consecutive observations between 11 December 2020 and 30 March 2021. As mentioned in Sect. , the resolution of the Sentinel-1 InSAR data from HyP3 is 80 m $\times$ 80 m. We used a $\mathrm{10}\times \mathrm{10}$ multi-look window of retrieved SWE and temporal coherence around the SNOTEL locations to reduce the speckle noise. Therefore, we compared the SNOTEL SWE with the 800 m $\times$ 800 m retrieved SWE around the SNOTEL site. The heterogeneity of the environment such as vegetation cover, vegetation fraction, land type, and SWE distribution in the 800 m $\times$ 800 m around the SNOTEL station affects our accuracy. We will analyze the effect of the heterogeneity of the environment on the SWE retrieval for SNOTEL stations in the future work of this study.

Figure 6

(a) Retrieved $\mathrm{\Delta }\mathrm{SWE}$ using the Sentinel-1 interferometric phase versus in situ $\mathrm{\Delta }\mathrm{SWE}$ for all the stations with temporal coherence more than 0.35 for the entire Sentinel-1 time series from December 2020 to March 2021. (b) Correlation (left axis) and absolute error (right axis) between retrieved $\mathrm{\Delta }\mathrm{SWE}$ using the Sentinel-1 interferometric phase and in situ $\mathrm{\Delta }\mathrm{SWE}$ for each in situ station. (c) Correlation (left axis) and absolute error (right axis) between retrieved $\mathrm{\Delta }\mathrm{SWE}$ using the Sentinel-1 interferometric phase and in situ $\mathrm{\Delta }\mathrm{SWE}$ for each interferogram. Note that the labels on the $x$ axis show the first date of each interferometric observation.

[Figure omitted. See PDF]

Figure a compares all the retrieved $\mathrm{\Delta }\mathrm{SWE}$ time series using Sentinel-1 data over all in situ stations with SNOTEL $\mathrm{\Delta }\mathrm{SWE}$. As seen in this figure, the retrieved and in situ $\mathrm{\Delta }\mathrm{SWE}$ are highly correlated (0.8), with an RMSE of 0.93 cm.

Figure b shows the correlation and RMSE between the entire time series of retrieved and in situ $\mathrm{\Delta }\mathrm{SWE}$ for each station, with blue and red circles respectively. As seen in this figure, the correlation is good (more than 0.6 for all stations except three). The RMSE is less than 2 cm for all stations and less than 1 cm for most stations. Note that station 4 has just one observation with temporal coherence more than 0.35. That observation is the first observation with zero SWE change. Therefore, there are not enough points to calculate ${\mathit{\rho }}_i$. Hence, the RMSE and correlation are zero.

Figure 7

Sentinel-1 wrapped phase path 71, frame 444, between (a) 19 and 25 December 2020 (observation 4) and (b) 25 and 31 December 2020 (observation 5).

[Figure omitted. See PDF]

Figure c shows the correlation and RMSE between the in situ stations and retrieved $\mathrm{\Delta }\mathrm{SWE}$ for each Sentinel-1 acquisition first date, with blue and red circles respectively. Note that the labels on the $x$ axis show the first date of each interferometric observation. The RMSE is again less than 2 cm for all dates and less than 1 cm for many dates. As seen in this figure, the correlation is more than 0.4 for some dates and poor (less than 0.4) for some others. Among the observation dates with a correlation of less than 0.35 (observation 1, 2, 4, 7, 9, 15, 16, and 17), observations 1, 2, 7, 9, 15, and 16 (first date of 1 December, 7 December, 6 January, 18 January, 7 March, and 13 March) have very small snow accumulation (the average $\mathrm{\Delta }\mathrm{SWE}$ is less than 0.5 cm, with $\mathrm{\Delta }\mathrm{SWE}$ close to zero for most stations). Therefore, the phase is not sensitive enough to SWE change, hence the low correlation. For observation 4 and 17 (first date of 19 December and 19 March), we observed that the low coherence degrades the phase unwrapping performance for these InSAR images. Figure a and b show the wrapped phase for observations 4 and 5, respectively. Note that the correlation between in situ and retrieved $\mathrm{\Delta }\mathrm{SWE}$ in Fig. c is 0.1 for observation 4 and 0.7 for observation 5. The average in situ $\mathrm{\Delta }\mathrm{SWE}$ between 19 and 25 December 2020 (observation 4) is 1.6 cm and that between 25 and 31 December 2020 (observation 5) is 2.3 cm. However, the interferometric fringes in Fig. a are very noisy compared to Fig. b. We observe that 4 out of 6 d between 19 and 25 December 2020 (observation 4) are relatively warm, including day 19 December 2020. All 31 stations have temperatures between $-\mathrm{7}$ and 6 ${}^{\circ }$C at 06:00 in those 4 d. The warm days cause a lot of melting and refreezing in those 4 d. Hence, we expect to have small temporal coherence and consequently noisier fringes. On the other hand, the temperature is relatively warm only on 26 December 2020. The rest of the 5 d between 25 and 31 December 2020 (observation 5) are mostly colder than $-\mathrm{7}$ ${}^{\circ }$C for all 31 stations, thus resulting in higher temporal coherence and less noisier fringes. We believe the noisy fringes degrade the performance of the unwrapping algorithm significantly. Therefore, the retrieved $\mathrm{\Delta }\mathrm{SWE}$ is more accurate for observation 5 compared to observation 4. One of the main future works of this study is to improve the phase unwrapping over images with low coherence.

5.1.2 Comparing retrieved total SWE using Sentinel-1 and SNOTEL total SWE

In this section, we used time-series-retrieved $\mathrm{\Delta }\mathrm{SWE}$ to calculate total SWE at each date compared to the start date of our time series (1 December 2020) by

5$$\mathrm{SWE}\left(t_{i+\mathrm{1}}\right)=\sum _{t_j=t_{\mathrm{1}}}^{t_i}\mathrm{\Delta }\mathrm{SWE}(t_j,t_{j+\mathrm{1}}),$$ where $t_{\mathrm{1}}$ is 1 December 2020. For instance, SWE at 25 December 2020 compared to 1 December 2020 is the summation of all four retrieved $\mathrm{\Delta }\mathrm{SWE}$ ($\mathrm{\Delta }{\mathrm{SWE}}_{\mathrm{1}-\mathrm{7}\mathrm{Dec}\mathrm{2020}}+\mathrm{\Delta }{\mathrm{SWE}}_{\mathrm{7}-\mathrm{13}\mathrm{Dec}\mathrm{2020}}+\mathrm{\Delta }{\mathrm{SWE}}_{\mathrm{13}-\mathrm{19}\mathrm{Dec}\mathrm{2020}}+\mathrm{\Delta }{\mathrm{SWE}}_{\mathrm{19}-\mathrm{25}\mathrm{Dec}\mathrm{2020}}$). Note that the $\mathrm{SWE}\left(t_{i+\mathrm{1}}\right)$ is measured compared to $\mathrm{SWE}\left(t_{\mathrm{1}}\right)$. For simplicity, we assume the SWE at time $t_{\mathrm{1}}$ is equal to zero.

Figure 8

Time series of total in situ and retrieved SWE using the Sentinel-1 interferometric phase shown by blue and red lines, respectively, for stations 12 (shown in panel a), 30 (b), and 20 (c).

[Figure omitted. See PDF]

Figure a, b, and c show the time series of total SWE for in situ stations 12, 30, and 20, respectively. Note that we used the average of stations 1 and 11 $\mathrm{\Delta }\mathrm{SWE}$ as a reference point in this study. The red and blue lines show the retrieved and in situ total SWE at each Sentinel-1 date acquisition compared to 1 December 2020. However, as mentioned in Sects.  and , we only used $\mathrm{\Delta }\mathrm{SWE}$ values with temporal coherence more than 0.35 and temperature less than $\mathrm{0}$ ${}^{\circ }$C. We had 18 observations for the entire time series. Discarding some observation due to low temporal coherence or high temperature changes the time series length. As seen in Fig. , we keep all 18 observations for station 20 but only 15 observations for station 12.

As seen in this figure, the time series of total retrieved SWE aligns closely with in situ values for stations 12 and 30. The error is less than 2 cm in the entire time series. However, the retrieved SWE for station 20 diverges from in situ values even though it follows the same pattern. The error in total SWE estimation is about 10 cm at the end of the time series. We think the main reason for divergence is the phase unwrapping error and phase ambiguity. As discussed in Sect. , the noisy fringes degrade the performance of the unwrapping algorithm. A similar problem is observed in tower-based studies. The retrieval diverges from the in situ values by phase ambiguity values over large snowstorms at C-band (Fig. 13c in ). However, even in these cases, the trends of SWE remain the same between retrieved and in situ values. We will investigate the reason behind the divergence of retrieved SWE from in situ SWE of these stations in the future work of this study.

Figure 9

(a) $\mathrm{\Delta }\mathrm{SWE}$ ambiguity versus incidence angle using Leinss's approximation (blue line) and Oveisgharan's approximation (red line) and (b) coherence for data acquired between 11 and 17 February 2021 (observation 11). Green diamonds show the location of stations with a total SWE error of less than 2 cm. Red diamonds show the location of stations with a total SWE error of more than 2 cm.

[Figure omitted. See PDF]

Figure a shows the Sentinel-1 $\mathrm{\Delta }\mathrm{SWE}$ ambiguity versus incidence angle. The red line shows the $\mathrm{\Delta }\mathrm{SWE}$ ambiguity using Eq. () ($\mathrm{\Delta }\mathit{\varphi }=\mathrm{2}\mathit{\pi }$). The blue line shows the $\mathrm{\Delta }\mathrm{SWE}$ ambiguity using the approximation ($\mathrm{\Delta }\mathit{\varphi }={\mathit{\kappa }}_i(\mathrm{1.59}+{\mathit{\theta }}^{\mathrm{2.5}})\mathrm{\Delta }\mathrm{SWE}$). As seen in this figure, $\mathrm{\Delta }\mathrm{SWE}$ ambiguity is between 1.5 and 3.5 cm depending on the incidence angle. The relatively small $\mathrm{\Delta }\mathrm{SWE}$ ambiguity of Sentinel-1 makes the unwrapping challenging for snowstorms. Figure b shows the temporal coherence between 11 and 17 February 2021. We can see very low coherence in the snowstorm regions, which degrades the unwrapping process. As mentioned before, one of the main future projects of this study is to work on improving the unwrapping phase.

For each station plot in Fig. , we also report the average RMSE error ($<\mathrm{\Delta }{\text{SWE Err}}_{\mathrm{station}\mathrm{\#}}>$) and correlation (${\mathit{\rho }}_{\mathrm{station}\mathrm{\#}}$) between retrieved and in situ $\mathrm{\Delta }\mathrm{SWE}$, as also plotted in Fig. b. We also report the average of temporal coherence for all the interferograms over that station ($<{\mathit{\rho }}_{\mathrm{temp}}>$) to show how reliable the measurements at that station are. For all three stations, the RMSE error for $\mathrm{\Delta }\mathrm{SWE}$ is less than 1.1 cm, the correlation between in situ and retrieved $\mathrm{\Delta }\mathrm{SWE}$ is greater than 0.8, and temporal coherence is greater than 0.5. The SNOTEL sites are shown by small diamonds in Fig. b. The green small diamonds have a total SWE error of less than 2 cm in the entire time series, similar to stations 12 and 30. The red diamonds have a total SWE error of more than 2 cm, similar to station 20. However, the retrieved SWE has a similar pattern to in situ SWE. Therefore, we think they have a phase unwrapping problem similar to station 20. These stations are also shown in Fig. a. As seen in this figure, the red diamonds are mostly located in regions with noisy fringes, which makes the unwrapping challenging. Among all 31 stations in the Sentinel-1 frame, 6 of them have temporal coherence less than 0.35 or temperature more than $\mathrm{0}$ ${}^{\circ }$C in their entire time series. These stations are shown by blue diamonds in Fig. b. Two stations are used for calibration of the phase. Hence, these two stations cannot be used for comparisons. So, there were 23 stations with more than two reliable observation dates in their time series. Among the 23 stations, 9 have an SWE error of less than 2 cm (green diamonds) and 14 of them have SWE error larger than 2 cm (red diamonds).

5.2 Comparing retrieved SWE using Sentinel-1 and lidar SWE

As mentioned in Sect. , the QSI lidar data were collected during the SnowEx campaign. There are two lidar data sets collected over the Sentinel-1 path 71, frame 444, in winter 2021. The locations are shown with red rectangles in Fig. b.

Figure 10

(a) QSI lidar snow depth over Banner Summit, ID, on 15 March 2021 (b). Retrieved total SWE using Sentinel-1 interferometric data from 1 December 2020 to 19 March 2021 over Banner Summit, ID. (c) Two-dimensional histogram of data in panel (b) versus data in panel (a).

[Figure omitted. See PDF]

Figures a and a show the lidar snow depth on 15 March 2021 over Banner Summit and Mores Creek, respectively. As shown in Fig. b, Banner Summit covers SNOTEL 2 and Mores Creek covers SNOTEL 21. These two SNOTEL stations are shown by diamonds in Figs. a and a. The terrain DEM is measured by a lidar sensor during September 2021. The DEM is used to measure the snow depth using the lidar data collected on 15 March 2021. The big purple rectangle in Fig. b corresponds to Banner Summit, and the small red rectangle corresponds to Mores Creek. We calculated the total SWE compared to 1 December 2020 on the closest day to lidar date acquisition. We used all the retrieved $\mathrm{\Delta }\mathrm{SWE}$ from 1 December 2020 to 19 March 2021 and calculated the total SWE on 19 March 2021 using Eq. (). Figures b and b show the retrieved SWE on 19 March 2021 over Banner Summit and Mores Creek, respectively. Panels (a) and (b) in Figs.  and have very similar patterns. The 2D histograms of these two images are shown in Figs. c and c, where the $x$ and $y$ axes show the lidar snow depth and Sentinel-1-retrieved SWE, respectively. The colors in panel (c) show the ${\mathrm{10}}^{\text{number of cells}}$ with lidar snow depth $x$ and InSAR SWE $y$. The correlation between these two data sets is 0.47 for Banner Summit and 0.59 for Mores Creek. Note that the lidar data show the snow depth, whereas Sentinel-1-retrieved data show the total SWE accumulated during the Sentinel-1 overpasses analyzed. On the other hand, lidar has a much higher resolution. The relatively good correlation (0.47 and 0.59) between the two independent measurements with different resolutions is a very good indication of the success of this method in estimating SWE.

Figure 11

(a) QSI lidar snow depth over Mores Creek, ID, on 15 March 2021. (b) Retrieved total SWE using Sentinel-1 interferometric data from 1 December 2020 to 19 March 2021 over Mores Creek, ID. (c) Two-dimensional histogram of data in panel (b) versus data in panel (a).

[Figure omitted. See PDF]

Figure 12

(a) Mean of Sentinel-1 temporal coherence between 1 December 2020 and 19 March 2021 over Banner Summit, ID. (b) (Left axis) Correlation between lidar snow depth and retrieved total SWE on 19 March 2021 using Sentinel-1 over Banner Summit for all points with mean temporal coherence greater than ${\mathit{\rho }}_{\text{temp, threshold}}$ versus ${\mathit{\rho }}_{\text{temp, threshold}}$. (Right axis) Number of points in Banner Summit with mean temporal coherence greater than ${\mathit{\rho }}_{\text{temp, threshold}}$ versus ${\mathit{\rho }}_{\text{temp, threshold}}$. (c) Mean of Sentinel-1 temporal coherence between 1 December 2020 and 19 March 2021 over Mores Creek, ID. (d) (Left axis) Correlation between lidar snow depth and retrieved total SWE using Sentinel-1 over Mores Creek for all points with mean temporal coherence greater than ${\mathit{\rho }}_{\text{temp, threshold}}$ versus ${\mathit{\rho }}_{\text{temp, threshold}}$. (Right axis) Number of points in Mores Creek with mean temporal coherence greater than ${\mathit{\rho }}_{\text{temp, threshold}}$ versus ${\mathit{\rho }}_{\text{temp, threshold}}$.

[Figure omitted. See PDF]

Figure a and c show the mean of all Sentinel-1 temporal coherence data between 1 December 2020 and 19 March 2021 over Banner Summit and Mores Creek, respectively. As seen in these figures, the temporal coherence varies between 0.2 and 0.9. As mentioned earlier in this section, the correlation between lidar snow depth data on 15 March 2021 and retrieved total SWE using Sentinel-1 data on 19 March 2021 is 0.47 for Banner Summit and 0.59 for Mores Creek. However, some of the points may have low temporal coherence and may not be viable for retrieval as discussed in Sect. . The left axis in Fig. b and d show the correlation between lidar snow depth data on 15 March 2021 and retrieved total SWE using Sentinel-1 data on 19 March 2021 for points with mean temporal coherence above ${\mathit{\rho }}_{\text{temp, threshold}}$. Figure b shows the correlation versus ${\mathit{\rho }}_{\text{temp, threshold}}$ over Banner Summit, and Fig. d shows the correlation over Mores Creek. Note that the correlation is 0.47 for Banner Summit and 0.59 for Mores Creek with no filter (${\mathit{\rho }}_{\text{temp, threshold}}=\mathrm{0.1}$) as reported in Figs. c and c, respectively. The right axis in Fig. b and d shows the number of points in the image in panels (a) and (c) with mean temporal coherence more than ${\mathit{\rho }}_{\text{temp, threshold}}$. There are 5 times more points in Fig. a compared to Fig. c. Therefore, we can better perform a statistical evaluation for panel (b) compared to panel (d). As seen in Fig. b, the correlation between lidar snow depth and retrieved total SWE increases by filtering out points with low temporal coherence, as expected. We need to investigate more to explain the reason for correlation decrease in the $\mathrm{0.74}<{\mathit{\rho }}_{\text{temp, threshold}}<\mathrm{0.8}$ interval.

The correlation between lidar snow depth and retrieved total SWE in Fig. d) is relatively constant with increasing ${\mathit{\rho }}_{\text{temp, threshold}}$. However, as we increase ${\mathit{\rho }}_{\text{temp, threshold}}$ to more than 0.46, the correlation gradually decreases to 0.4 at ${\mathit{\rho }}_{\text{temp, threshold}}=\mathrm{0.65}$ and remains relatively constant up to ${\mathit{\rho }}_{\text{temp, threshold}}=\mathrm{0.72}$. The number of points in the image with temporal coherence more than 0.72 is less than 20. Therefore, the correlation is not statistically very meaningful. Mores Creek has a lower elevation (6100 m at station 21) compared to Banner Summit (7040 m at station 2). Mores Creek is also warmer (mean temperature of $-\mathrm{4.3}$ ${}^{\circ }$C for the entire time series at station 21) than Banner Summit (mean temperature of $-\mathrm{8.8}$ ${}^{\circ }$C for the entire time series at station 3). We expect to have higher correlation with filtering low temporal coherence points as seen in Fig. b. We think the reason we do not see such a behavior in Fig. d is that the warmer temperature, melting, and refreezing degrade the retrieval performance even for highly correlated regions. More investigation is needed to better explain the constant or decreasing correlation with increasing ${\mathit{\rho }}_{\text{temp, threshold}}$ in Fig. d.

Figure 13

(a) (Left axis) Correlation between lidar snow depth and retrieved total SWE using Sentinel-1 on a specific observation date over Banner Summit versus observation number. (Right axis) $\mathrm{\Delta }\mathrm{SWE}$ (cm) for any specific observation date at station 2 in Banner Summit. (b) (Left axis) Correlation between lidar snow depth and retrieved total SWE using Sentinel-1 on a specific observation date over Mores Creek versus observation number. (Right axis) $\mathrm{\Delta }\mathrm{SWE}$ (cm) for any specific observation date at station 27 in Mores Creek.

[Figure omitted. See PDF]

Left axis in Fig. a and b shows the correlation between lidar snow depth on 15 March 2021 and retrieved total SWE for each observation between 1 December 2020 and 19 March 2021 over Banner Summit and Mores Creek, respectively. Observation 16 shows the correlation reported in Figs. c and c. The right axis in Fig. a and b shows $\mathrm{\Delta }\mathrm{SWE}$ (cm) for each observation between 1 December 2020 and 19 March 2021 at station 2 in Banner Summit and station 21 in Mores Creek, respectively. As seen in both figures, the correlation gradually increases after observation 7 or 8, as expected. On the other hand, the correlation is smaller for observation 16 compared to 15. Observation 16 shows the total SWE on 19 March 2021, and lidar data show the snow depth on 15 March 2021. There is about 1 cm $\mathrm{\Delta }\mathrm{SWE}$ for observation 16 that is not fully captured by lidar. Therefore, comparing total SWE for observation 15 with lidar snow depth is more appropriate. The correlation is high for observation 4 at the beginning of the snow season for both Banner Summit and Mores Creek. Observation 4 is after the first snowstorm of the season. It shows that the spatial variability of snow at the end of the snow season is captured by the first or second snowstorm. Although $\mathrm{\Delta }\mathrm{SWE}$ for station 21 is zero for the first observation, the correlation between lidar snow depth on 19 March 2021 and total SWE on the first observation is relatively high, as seen in Fig. b. As shown in Fig. a, station 2 is in the relatively low snow depth region. We believe there has been a snowstorm in the high-altitude region of Mores Creek. The high correlation is simply the correlation between lidar data on 19 March 2021 and the first snowstorm in Mores Creek.

6 Conclusions

In this study, we used Sentinel-1 time series to retrieve $\mathrm{\Delta }\mathrm{SWE}$ and consequently total SWE. We chose a frame in Idaho that covers several SnowEx 2020–2021 sites and 31 SNOTEL in situ stations. Lidar data are available for validating our results. Sentinel-1 data were collected every 6 d over this SnowEx site instead of the regular 12 d, which helps a lot with temporal coherence over snowstorms. This provides a unique dense time series of spaceborne data for studying the performance of SWE retrieval using InSAR.

We showed that retrieved $\mathrm{\Delta }\mathrm{SWE}$ between two consecutive Sentinel-1 observations is highly correlated (0.8) with in situ values, with an RMSE of 0.93 cm. For the reference point of the interferometric phase, we used two in situ stations with temporal coherence more than 0.35 and temperature less than $\mathrm{0}$ ${}^{\circ }$C for the entire time series. We subtracted the difference between the average of in situ and retrieved $\mathrm{\Delta }\mathrm{SWE}$ of these two stations from retrieved values to calibrate the retrieved $\mathrm{\Delta }\mathrm{SWE}$. The $\mathrm{\Delta }\mathrm{SWE}$ RMSE error is less than 2 cm for all stations and less than 1 cm for most stations. The correlation between retrieved and in situ $\mathrm{\Delta }\mathrm{SWE}$ is more than 0.6 for most stations. Interferograms with a small average of in situ $\mathrm{\Delta }\mathrm{SWE}$ show low correlation between retrieved and in situ $\mathrm{\Delta }\mathrm{SWE}$. We demonstrated that low temporal coherence not only degrades the SWE retrieval performance, but also the unwrapping algorithm performance. We showed that big melting events between two Sentinel-1 acquisitions make the interferometric fringes noisy and the unwrapping algorithm challenging. The retrieved total SWE has an RMSE error of less than 2 cm compared with in situ values in the entire time series for 9 stations and an error of more than 2 cm for 14 stations.

The highlight of the results of this study is the similarity between two independent measurements retrieved SWE using Sentinel-1 data and lidar snow depth data. We used Sentinel-1 data between 1 December 2020 and 19 March 2021 to retrieve $\mathrm{\Delta }\mathrm{SWE}$ time series. By adding the entire time series of $\mathrm{\Delta }\mathrm{SWE}$, we calculated the total SWE on 19 March 2021. Total retrieved SWE values using Sentinel-1 interferometric data and lidar snow depth images over two regions in Idaho show similar patterns and are correlated by more than 0.47. We showed that the correlation is higher for regions with higher temporal coherence in Banner Summit.

Considering all these validations, we show for the first time that SWE retrieval using time series of InSAR spaceborne data is a very promising candidate for the future SWE mission.

We also show that the main constraints for this method are temporal coherence, phase unwrapping, and phase ambiguity. We show that snowstorms reduce the temporal coherence significantly. Low temporal coherence reduces the accuracy of the interferometric phase and unwrapping algorithm. This study all shows that melting due to warm temperature reduces the temporal coherence and the performance of the unwrapping algorithm. Small SWE ambiguity at C-band (1.5 to 3.5 cm) makes the phase unwrapping more challenging. We think using an in situ station as the reference point helps reduce the phase ambiguity error, at least locally, compared to other methods for referencing the interferometric images. If the temporal coherence is large enough for the entire image to reduce the phase unwrapping error, using the in situ SWE as the reference point reduces the phase ambiguity error in a larger region. Using a snow-free point or snow-free corner reflector as the reference point cannot address the phase ambiguity in regions with deep snow. Going from C-band to lower frequencies such as L-band improves both the temporal coherence and SWE ambiguity. With the L-band NISAR launch coming next winter, the new data set would be a great opportunity for global SWE retrieval.

Data availability

The Sentinel-1 data are publicly available through the Alaska SAR Facility (ASF) or the Copernicus Data Hub distribution service. The SNOTEL data are free and available at the United States Department of Agriculture (USDA) website (https://wcc.sc.egov.usda.gov/nwcc/inventory, ). Finally, we collect the lidar data from the National Snow and Ice Data Center (10.5067/VBUN16K365DG, Adebisi et al., 2022).

Author contributions

SO conceptualized the overall study. SO processed the data using HyP3. RZ applied atmospheric correction to the data. SO analyzed the data. ZH and HPM provided fruitful discussions and feedback and lidar data.

Competing interests

The contact author has declared that none of the authors has any competing interests.

Disclaimer

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors.

Acknowledgements

The research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration (grant no. 80NM0018D0004). The authors would like to thank Silvan Leinss and Jorge Jorge Ruiz for their very useful reviews that improved Sects. 2.2, 4, and 5.2 significantly.

Financial support

This research has been supported by the Jet Propulsion Laboratory (R&TD 2022/2023).

Review statement

This paper was edited by Nora Helbig and reviewed by Silvan Leinss and Jorge Jorge Ruiz.