As described in Sect. , interferometric phase measurements are used to estimate changes in snow water equivalent (SWE). Following the formulation in Oveisgharan et al., the SWE change between two Sentinel-1 acquisitions is expressed as : 1Δϕ=-2κi(-0.3385θ2+0.0486θ-0.8647)ΔSWE^ where Δϕ,κi,ΔSWE, and θ denote the interferometric phase between two acquisition dates, the incidence wavenumber, SWE change, and the local incidence angle, respectively. The local incidence angle is computed by accounting for both the satellite viewing geometry and terrain topography (slope and aspect), thereby representing the angle between the radar line-of-sight and the local surface normal. Temporal coherence is generally low at C-band; however, the 6 d Sentinel-1 repeat cycle significantly improves coherence compared to the nominal 12 d repeat, particularly over snow-covered regions. Consequently, SWE retrievals were primarily performed for areas with available 6 d repeat data, as discussed in Sect. . Equation () is valid for dry-snow conditions . Near-surface air temperature exceeding 0 °C was therefore used as an indicator of wet snow, where this assumption does not hold.

A critical step in ΔSWE estimation is selecting a suitable reference point for phase calibration. In InSAR-based geophysical studies, this point is typically a stable target with negligible or known displacement between acquisitions. Following , we used the average of one or two in situ ΔSWE measurements exhibiting a correlation above 0.35 and subzero temperatures throughout the time series as the reference point for phase calibration . Note that reference point selection does not affect comparisons between retrieved total SWE and LIDAR-derived snow depth, since it only introduces a constant phase offset across all dates, preserving spatial variability.

Atmospheric phase delays were estimated using the European Center for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis product, which provides hourly atmospheric variables on a 30 km global grid. The Python-based Atmospheric Phase Screen (PyAPS) tool was used to interpolate these data and convert them into radar phase delays. PyAPS is integrated into the Miami InSAR Time-Series software in Python , which was used to crop the delays to the interferogram extent and project them into the radar line-of-sight (LOS). Ionospheric errors at C-band are small relative to other uncertainty sources and were therefore considered negligible. Phase ambiguity for Sentinel-1 data ranges from approximately 1.5 to 3.5 cm, depending on incidence angle (Oveisgharan et al., 2024), making it one of the dominant error sources during large snowstorms associated with significant ΔSWE .

The coherence between two signals is computed as the magnitude of the normalized cross-correlation of two complex SAR images . We use HH polarization Sentinel-1 complex data to calculate coherence. The radar signals acquired on two different dates remain correlated if the scattering elements within a resolution cell remain unchanged. In practice, however, temporal variations in scatterers – such as freeze/thaw, soil permittivity change, the movement of leaves, branches, or snow particles, reduce this correlation, leading to temporal decorrelation . Loss of temporal coherence is one of the principal limitations for SWE retrieval using differential interferometry, as discussed in Sect. .

In snow-covered regions, melting and wind are the dominant drivers of temporal decorrelation . While soil moisture can influence temporal coherence, its impact is expected to be limited during the snow season, as the ground is typically frozen under dry-snow conditions. In situations where the ground is not frozen, the snowpack is likely to be wet, violating the assumptions required for InSAR-based SWE retrieval. Therefore, the effect of soil moisture on temporal coherence is considered negligible for the dry-snow conditions analyzed in this study.Temporal coherence also decreases with increasing radar frequency , making longer wavelengths such as L- and C-band more suitable for differential interferometric applications. Vegetation cover can further degrade temporal coherence, particularly at higher frequencies . This is because higher frequencies (shorter wavelengths) are more sensitive to small-scale variations in the scattering medium, such as snow microstructure and vegetation elements. These small scatterers introduce phase variations within a resolution cell that change between acquisitions, leading to increased decorrelation. Ruiz et al. conducted tower-based fully polarimetric InSAR experiments at L-, S-, C-, and X-bands to assess the effects of air temperature, precipitation, and wind on temporal decorrelation, identifying temperature as the most influential variable .

The standard deviation of the SWE error due to total coherence can be derived using: 2σΔSWE=-λ4π(-0.3385θ2+0.0486θ-0.8647)⋅12N1-ρ2ρ2 where σSWE is the SWE uncertainty due to coherence, ρ is total true coherence, λ is the radar signal wavelength, and N is the effective number of looks. This relationship is derived from the Cramér–Rao bound and is widely used in InSAR error analysis . We used effective looks of 43 for Sentinel-1 data to calculate the true correlation from measured correlation . The temporal coherence is part of the total coherence. We use Eq. () in Sect.  to calculate the uncertainty in retrieved SWE.

In this study, we extend the temporal coherence analysis using a larger spatial and temporal dataset to evaluate the effects of multiple environmental and surface parameters on temporal coherence, as discussed in Sect. .

Sentinel-1 is a C-band radar mission with four imaging modes, providing resolutions down to 5 m and swath widths up to 400 km. It offers dual polarization, a 12 d repeat cycle, and rapid data access, with products freely distributed through the Alaska Satellite Facility (ASF). Differential interferometric phase and coherence from VV and VH polarizations can be generated globally at 12 d intervals using ASF’s On-Demand Processing system. Interferometric processing is supported by the Hybrid Pluggable Processing Pipeline (HyP3), which applies topographic phase and geolocation corrections and employs the Minimum Cost Flow (MCF) algorithm for phase unwrapping. In this study, we employed the unwrapped phase and interferometric coherence derived from Sentinel-1 observations. The constellation consists of two satellites, each with a 12 d global repeat cycle; when operated in tandem, the effective revisit interval decreases to six days over selected regions, particularly across Europe.

The NASA SnowEx2020 and 2021 Time Series represents a continuation of the multi-year initiative to advance snow water equivalent (SWE) measurement and estimation. Data acquisition during winter 2020 and 2021 included multiple sensors (e.g., radar and LIDAR) and in situ collections. As part of this effort, the SnowEx campaign collaborated with the Sentinel-1 team to obtain six-day revisit observations over selected sites throughout the winter season. The blue frames in Fig. a show the geographic locations of Sentinel-1 data that cover the available LIDAR data in 2020 or 2021. The blue frames in Fig. b show the geographic locations of all Sentinel-1 data with 6 d revisit time during the winter of 2020 and 2021 that cover more than 5 SNOTEL stations. Our analysis indicated that snow water equivalent (SWE) retrieval using Sentinel-1 data performs better with 6:00 am acquisitions compared to 6:00 pm data. Consequently, the majority of our study focuses on results derived from descending Sentinel-1 acquisitions collected at 6:00 am local time; however, a performance comparison between ascending (6:00 pm) and descending (6:00 am) data over two LIDAR sites is provided in Sect. .

Airborne LIDAR provides high-resolution snow depth maps and serves as a reliable source of validation data, offering a particularly strong constraint for InSAR-based retrieval of SWE. For this study, we surveyed all publicly available LIDAR datasets and selected those collected during the snow accumulation season. The “SnowEx20-21 QSI LIDAR DEM 0.5m” dataset, part of the SnowEx 2020 and 2021 campaigns, includes digital elevation models, snow depth, and vegetation height at 5 m spatial resolution . To enable direct comparison with Sentinel-1-derived SWE, the LIDAR data were resampled to 80 m resolution. These data were acquired over sites in Colorado, Idaho, and Utah during February 2020, March 2021, and September 2021. In addition, the Airborne Snow Observatory (ASO) program has been operating since 2013, providing over a decade of high-resolution snow observations across the western United States. ASO products combine LIDAR, imaging spectrometry, and physically based modeling to generate snow depth and snow water equivalent estimates . The snow depth measurements have an accuracy of approximately 3 cm with a spatial resolution of 3 m. While early ASO campaigns focused on the melt season, current operations span a broader period from January through July, reflecting the importance of both accumulation and melt processes for water resource management. For this study, we used all publicly available ASO datasets collected between January and March. All ASO datasets included in Table  were processed and provided by the ASO program.

As mentioned LIDAR snow depth data is a very valuable dataset for validating our retrieved SWE. The red boxes in Fig. a show the location of LIDAR data acquisition during the dry snow regions.

Figure a–j show the zoomed Google Earth View of LIDAR scenes, Banner Summit, ID (BS), Mores Creek, ID (MC), Dry Creek, ID (DC), Reynolds Creek, ID (RC), Merced, CA (ME), Tuolumne, CA (TU), Fraser, CO (FR), Little Cottonwood Canyon, UT (LC), East River, CO (ER), Taylor River, CO (TR), respectively.

Table  summarizes all LIDAR scenes used in this study, consisting of data collected between January and March of 2020 and 2021. The first column lists the LIDAR site ID, followed by the data provider in the second column. The third column reports the date of LIDAR acquisition, while the fourth and fifth columns provide the revisit time and the corresponding Sentinel-1 path/frame information for the time series covering each LIDAR scene. The LIDAR data are subsequently used in Sects.  and to evaluate the performance of SWE retrievals derived from Sentinel-1 observations.

The SNOwpack TELemetry (SNOTEL) network consists of stations located in remote, high-elevation mountain regions across the western United States. These stations automatically record a range of snowpack and meteorological variables. For this study, hourly SNOTEL observations were accessed through the United States Department of Agriculture (USDA) website . Because we used descending Sentinel-1 acquisitions, collected at approximately 6 am local time, we extracted SNOTEL measurements of SWE, snow depth, and near-surface air temperature corresponding to that time. The locations of the SNOTEL stations are shown as purple diamonds in Fig. b. Notably, during winter 2021, only three Sentinel-1 frames with 6 d repeat cycles covered more than five SNOTEL sites. These in situ measurements were employed for (a) establishing InSAR reference points (Sect. ), (b) evaluating the performance of SWE retrievals (Sect. ), and (c) analyzing the influence of environmental parameters on SWE retrieval accuracy (Sect. ).

As summarized in Table  and illustrated in Figs.  and , a total of 10 LIDAR scenes with snow depth measurements were available between January and March of 2020 and 2021. Among these, the Banner Summit (BS) and Mores Creek (MC) sites contain data for both years. For each LIDAR scene, we used the corresponding Sentinel-1 time series (path and frame information provided in Table , column 5) to retrieve ΔSWE following the approach described in Sect. . Following , time-series-derived ΔSWE was then used to estimate total SWE at the date closest to the LIDAR acquisition by 3SWE(ti+1)=∑tj=t1tiΔSWE(tj,tj+1) where t1 is the first Sentinel-1 data collected after December first of the corresponding winter and ti+1 is the closest Sentinel-1 date to LIDAR acquisition date. For simplicity, we assume the SWE at time t1 is equal to zero. The assumption of zero SWE at t1 may introduce a bias in the reconstructed total SWE in regions where early-season accumulation is present. However, this assumption primarily affects the absolute magnitude of SWE rather than its spatial variability. For example, based on in situ measurements distributed across the Sentinel-1 frame (p71, f444), the spatial standard deviation of SWE is approximately 5 cm at the beginning of December and increases to about 23 cm by the end of March. This indicates that early-season SWE exhibits relatively low spatial variability compared to peak winter conditions. Given that the LIDAR scenes cover a much smaller spatial extent (∼ 16 km × 15 km) within the larger Sentinel-1 frame (∼240 km × 240 km), the spatial variability of SWE at the beginning of December is expected to be even smaller at the LIDAR scale. Therefore, neglecting the initial SWE mainly introduces an approximately constant offset in the reconstructed SWE and is not expected to significantly affect the spatial correlation with LIDAR snow depth. However, the reconstructed SWE values may be systematically biased low in absolute terms.

It is important to note that the Sentinel-1-based retrieval estimates SWE, whereas the LIDAR observations provide snow depth. As a result, the comparison between retrieved SWE and LIDAR snow depth is not strictly one-to-one. Although LIDAR snow depth could be converted to SWE using an assumed snow density, such a conversion would introduce additional uncertainty associated with density estimation. The correlation metric used here implicitly assumes a constant snow density across each scene. Despite this simplifying assumption, the results show reasonable agreement as seen in Sect. . This limitation should be considered when interpreting the comparisons. The effective density implied by the comparison between retrieved SWE and LIDAR snow depth is further evaluated in Sect. .

Another important consideration is the difference in spatial resolution and sampling geometry among the datasets. The Sentinel-1 products used in this study have an effective spatial resolution of approximately 80 m after multilooking, whereas the LIDAR data are available at 0.5 and 3 m resolution, and SNOTEL measurements represent point-scale observations. To improve consistency, the LIDAR data were spatially averaged to 80 m resolution prior to comparison. Nonetheless, residual differences in spatial representativeness remain. In addition, the observation geometries differ: Sentinel-1 is a side-looking system with look angles of approximately 29–46°, whereas LIDAR measures snow depth in the nadir direction. Consequently, the sampled snowpack, vegetation, and ground characteristics are not identical between the datasets. Spatial variability in snow properties, vegetation structure, and terrain within the 80 m comparison scale may further contribute to discrepancies. Terrain slope in complex topography can affect both LIDAR measurements and InSAR-derived SWE. For LIDAR, steep slopes may introduce sampling biases and noise in snow depth estimation, while for InSAR, slope influences the local incidence angle and the effective radar propagation path through the snowpack, thereby affecting the interferometric phase and the retrieved SWE. In addition, steep terrain can introduce radar-specific effects such as layover and shadow, which further degrade signal quality and coherence and can impact retrieval accuracy. Because the radar footprint is significantly larger than that of LIDAR, these effects can be more pronounced in the InSAR observations. A more rigorous treatment of slope effects, including explicit modeling of radar–snow interaction geometry, is beyond the scope of this study. However, the impact of slope is partially addressed by including terrain slope as one of the parameters in the retrieval performance analysis (Sect. ). Its influence is also evident in the comparison between ascending and descending acquisitions discussed in Sect. . These factors should be considered when interpreting validation results, as some observed differences or dependencies may arise from mismatches in measurement geometry and resolution rather than from retrieval performance alone.

To evaluate performance, the analysis was divided into two groups: LIDAR scenes covered by Sentinel-1 data with 6 d revisit intervals and those with 12 d revisit intervals. As shown in Table , eight scenes fall into the 6 d repeat group, while four scenes correspond to 12 d repeats.

As described in Sect. , Sentinel-1 data were acquired every 6 d over three frames containing more than five SNOTEL sites (Fig. b) during 2020 and 2021, coordinated between the SnowEx campaign and the Sentinel-1 team. For this analysis, we used 6 d repeat Sentinel-1 time series from 1 December 2020 to 30 March 2021.

The left, middle, and right columns of Fig.  correspond to frames (p71, f444), (p71, f450), and (p42, f461), respectively. Figure a1–c1 show SNOTEL SWE time series from 1 December 2020 to 31 March 2021 at 6:00 am for all stations within each frame, with different colors indicating individual SNOTEL sites. SWE increases relatively uniformly in (p71, f444), whereas sharp increases are observed in (p71, f450) and (p42, f461). Snowpack is also generally shallower in (p71, f450).

Figure a2–c2 present the average near-surface temperature across SNOTEL stations in each frame. Among them, (p42, f461) is consistently warmer than (p71, f444) and (p71, f450). Figure a3–c3 display the average Sentinel-1 temporal coherence time series, with black squares marking SNOTEL locations. Low-coherence regions align with snow-covered areas. Frame (p71, f450) exhibits higher coherence due to its shallower snowpack, while (p42, f461) shows lower coherence compared to (p71, f444).

Figure a4–c4 compare retrieved ΔSWE from Sentinel-1 time series with ΔSWE from SNOTEL stations over the study period. It is important to note the differences in spatial resolution and sampling geometry between the datasets. The Sentinel-1 products used in this study have an effective spatial resolution of approximately 80 m after multilooking, whereas SNOTEL measurements represent point-scale observations. To reduce the impact of spatial heterogeneity and improve comparability, we further averaged Sentinel-1-derived SWE change over a 10 × 10 pixel window centered on each SNOTEL site, corresponding to an effective spatial resolution of approximately 800 m. This spatial averaging helps mitigate mismatches between point measurements and distributed estimates. The correlation coefficients are 0.81, 0.55, and 0.54 for frames (p71, f444), (p71, f450), and (p42, f461), respectively. The corresponding RMSE values are 0.78, 1.32, and 1.17 cm. The reduced performance in (p71, f450) and (p42, f461) is attributed to limited SWE change in the former and warmer temperatures in the latter.

In this section, we assess how various parameters influence the performance of SWE retrieval from Sentinel-1 by comparing retrieved SWE against LIDAR snow depth. This analysis builds on the limited LIDAR validation presented in by incorporating a larger number of LIDAR scenes across diverse environments, enabling a more comprehensive assessment of retrieval performance and its controlling factors. Performance is quantified using the correlation between the two datasets. The last column of Table  reports the correlation for each LIDAR scene over the entire image. To further investigate, we partition the data according to specific parameters and recompute the correlations.

Figure a shows the relationship between correlation and mean Sentinel-1 temporal coherence (ρtemp) for sites with available 6 d repeat acquisitions. Different colors represent different LIDAR sites. As seen in the figure, the correlation generally increases with higher temporal coherence across most sites, with the exception of RC20. For example, in BS21 the overall correlation between retrieved SWE and LIDAR snow depth is 0.47 for the entire image, but ranges from ∼0.1 at low coherence to ∼0.78 at ρtemp∼0.75. This pattern is consistent across most scenes, underscoring the critical role of temporal coherence in SWE retrieval: higher coherence enables more reliable phase measurements and, consequently, more accurate retrievals. At the same time, the presence of exceptions indicates that temporal coherence alone does not fully explain retrieval performance, and that additional environmental and observational factors must also be considered. Therefore, temporal coherence is used here as a primary metric for evaluating retrieval quality, while its interaction with other controlling parameters is examined in Sect. . Although TU21 exhibits a low overall correlation (0.13), Fig. a (blue dashed line) shows that correlation still rises to ∼0.5 in regions with higher temporal coherence. This demonstrates that reliable retrieval is possible even in generally low-performing sites when coherence is sufficiently high. However, because TU21 has low overall coherence across the scene, we exclude it from the parameter impact analysis. The exception is RC20, where temporal coherence does not improve retrieval performance. As discussed previously, this is likely due to frequent melt events and overall low SWE, which degrade retrieval accuracy despite coherence levels.

Figure b illustrates the correlation between retrieved SWE time series and LIDAR snow depth at the end of the time series. In most sites, correlations are relatively low early in the snow season but increase steadily, reaching near-maximum values between mid-January and early February, and attaining their highest values at the end of the time series. The fact that maximum correlation occurs at the LIDAR acquisition date indicates that the relationship is not random; rather, the correlation strengthens as the retrieval approaches the validation date. This trend also suggests that the primary spatial pattern of snow accumulation is largely established by February.

Figure a–e illustrate the influence of snow depth, vegetation height, ground topography, slope, and aspect on the correlation between retrieved SWE and LIDAR snow depth. Different colors represent different LIDAR sites. Vegetation height is derived from the QSI LIDAR dataset, which provides co-registered measurements of vegetation structure and snow depth. Elevation is obtained from the Copernicus DEM (GLO-30), used within the ASF HyP3 processing workflow, with a spatial resolution of 30 m. Terrain slope and aspect are calculated from this DEM. All parameters are subsequently aggregated to match the 80 m spatial resolution of the multilooked Sentinel-1 data used in this analysis. Vegetation height data are only available for the QSI LIDAR products. As shown in Fig. a, total snow depth has little effect on retrieval performance. In contrast, Fig. b demonstrates that correlation decreases with increasing vegetation height, likely because vegetation obscures the snowpack and reduces retrieval accuracy. However, retrievals may remain reliable even in areas with tall vegetation, possibly due to canopy gaps or snow accumulation on top of the canopy. Correlation tends to increase with ground elevation within each site (Fig. c). However, this relationship is local rather than global; there is no specific elevation threshold beyond which performance improves uniformly. Instead, within a given site, higher elevations generally correspond to colder conditions, leading to drier snow and thus more reliable retrieval. As shown in Fig. d, correlation generally decreases with increasing ground slope, except in RC20 and ME21. Steeper slopes are expected to promote snow displacement, reducing temporal coherence and retrieval performance. However, RC20 and ME21 have relatively shallow snowpacks, which may limit downslope snow movement and explain the lack of slope dependence. Figure e shows that correlation is highest (except in RC20 and ME21) for aspect angles near 90°, corresponding to east-facing slopes. A possible explanation is that these slopes receive solar radiation earlier in the day, when temperatures are lower, thereby minimizing melt compared to west-facing slopes that receive afternoon sunlight under warmer conditions. In addition, as discussed in Sect. , observation geometry also contributes to this pattern. East-facing slopes tend to face toward the right-looking descending Sentinel-1 acquisition (6 am), resulting in higher backscatter, signal-to-noise ratio and improved temporal coherence. The combination of favorable viewing geometry and colder, drier snow conditions leads to improved SWE retrieval performance over east-facing slopes. In contrast, RC20 and ME21 exhibit consistently low correlations across all aspect angles, consistent with their shallow snowpack and warmer conditions.

In this section, we examine how different parameters affect the correlation between Sentinel-1-retrieved and SNOTEL ΔSWE using 6 d repeat acquisitions. We also assess the influence of these parameters on Sentinel-1 temporal coherence. In contrast to the site-specific validation in , this section leverages multi-site SNOTEL time series to investigate how environmental variability affects SWE retrieval performance and temporal coherence.

Figure a–c illustrate the effects of temporal coherence, temperature, and ΔSWE on the correlation between retrieved and SNOTEL ΔSWE, with different colors representing individual Sentinel-1 frames. For each frame in Fig. 1b, we use the full time series of parameters and calculate the correlation for each subcategory. For example, in Fig. b the correlation between retrieved and SNOTEL ΔSWE is about 0.9 for all SNOTEL observations within the temperature range -15 to -10 °C in (p71, f444).

Figure a shows that the correlation between retrieved and SNOTEL ΔSWE increases with temporal coherence for (p71, f444), but decreases for (p71, f450) and (p42, f461). This behavior differs from the more consistent positive relationship observed between temporal coherence and performance in the LIDAR-based analysis (Fig. a), where higher coherence generally leads to improved agreement. It is important to distinguish between the two validation approaches used in this study, as they assess different aspects of retrieval performance. The SNOTEL-based analysis evaluates SWE change (ΔSWE) between individual interferometric acquisitions, and therefore reflects the temporal performance of the retrieval at specific time intervals. In contrast, the LIDAR-based analysis compares total retrieved SWE with total LIDAR-derived snow depth at a given date, which effectively integrates SWE changes over time and represents an aggregate measure of performance. As a result, the two approaches differ in their sensitivity to errors. In the LIDAR comparison, total SWE is obtained by summing ΔSWE over time. Any error in the reference point used for phase calibration propagates as a constant offset in the reconstructed total SWE. Such an offset does not affect spatial correlation with LIDAR snow depth, which primarily reflects the similarity of spatial patterns. In contrast, the SNOTEL-based analysis compares ΔSWE across multiple dates, and is therefore more sensitive to reference point errors and temporal noise. For example, estimating a small ΔSWE for large ΔSWE can significantly affect correlation with SNOTEL measurements, whereas in the LIDAR comparison the same error would contribute only a constant bias with no on spatial correlation. More broadly, the LIDAR-based analysis evaluates “spatial” correlation of “total SWE”, while the SNOTEL-based analysis evaluates “spatio-temporal” correlation of “SWE change”. Consequently, the relationship between temporal coherence and retrieval performance differs between the two approaches. In the LIDAR comparison, higher average temporal coherence over the time series generally leads to improved performance, as it reflects more reliable phase measurements accumulated over time. In the SNOTEL analysis, however, the relationship is more complex. High temporal coherence often occurs during periods of small ΔSWE, when retrieval performance is inherently more sensitive to noise and may degrade despite favorable coherence conditions. Therefore, the LIDAR-based analysis captures the integrated, average behavior of the retrieval, while the SNOTEL-based analysis reveals the instantaneous performance and its dependence on temporal variability. Together, these complementary perspectives provide a more complete understanding of the factors controlling SWE retrieval performance. If a temporally consistent LIDAR time series were available, co-registered with the InSAR acquisitions, the comparison would become more analogous to the SNOTEL-based analysis, as both would involve evaluating SWE and snow depth changes between acquisition dates and their dependence on temporal coherence. In that case, the relationship between coherence and retrieval performance would be expected to follow similar behavior to that observed in the SNOTEL comparison.

As expected, higher temperatures reduce retrieval performance, as shown in Fig. b. Warmer conditions promote melt, which lowers the correlation between retrieved and SNOTEL ΔSWE, consistent with the behavior reported in . In contrast, the magnitude of SWE change does not significantly impact correlation, as shown in Fig. c.

We also evaluate the influence of temperature and SWE change on temporal coherence using SNOTEL data. Figures a and 13b show that temporal coherence decreases with increasing temperature and with larger ΔSWE. Higher temperatures make the snow wetter and increase snow particle motion, reducing coherence. Similarly, larger SWE changes are often accompanied by strong winds and substantial modifications to the scattering medium, which also degrade temporal coherence. This study moves beyond demonstrating the feasibility of InSAR-based SWE retrieval by providing a systematic evaluation of its performance across varying environmental conditions. The results identify the dominant controls on retrieval accuracy and define the conditions under which the method is reliable, which are critical steps toward operational SWE monitoring using current and future SAR missions.