The Copernicus Sentinel-1 C-band SAR mission consisted of two satellites for the analysis period of 2014 to 2020. Sentinel-1A was launched in April 2014, and Sentinel-1B followed in April 2016. In order to provide global coverage, the Sentinel-1 mission operates in a 12 d repeat cycle over most land surfaces.

The capability of Copernicus Sentinel-1 for FT products has been demonstrated by, e.g., and . The primary data source for the benchmarking is therefore Sentinel-1 C-band SAR data acquired in IW (interferometric wide swath) terrain observation with progressive scan (TOPS) mode at VV and VH polarization. Acquisition types and coverage vary considerably across the Arctic . The required data are unavailable for Greenland and the Canadian High Arctic . Comparably good coverage is available for Alaska, as well as for Scandinavia. These areas overlap with in situ measurement sites relevant to this study. All scenes starting from 30 July 2017 to 31 December 2020 have been processed for northern Alaska (covering all automatic weather stations and borehole locations; see Fig. ), and all scenes for August 2016 to August 2018 cover the Kaldoaivi area in northern Finland (Fig. ).

Three FT datasets based on relatively coarse-spatial-resolution PMW information are considered for benchmarking (Table ). The currently listed FT Fundamental Climate Data Record (FCDR) in the catalog of CEOS (Committee on Earth Observation Satellites) Working Group on Climate is the MEaSUREs (Making Earth System Data Records for Use in Research Environments) Global Record of Daily Landscape Freeze/Thaw Status, Version 3 , which is associated with the ECV (Essential Climate Variable) soil moisture. It is based on various passive microwave missions (SMMR, SSM/I, and AMSR-E/AMSR2 (referred to as AMSR in the following)) and goes back to 1979. It is provided with 25 km gridding, separation of AM and PM, and >80 % mean annual spatial classification accuracy. FT information is supplied as a binary (either frozen or unfrozen). Vegetation-related use of the product is suggested apart from soil moisture masking (https://climatemonitoring.info/ecvinventory/, last access: 16 September 2024). An updated version (version 5), which is available through the National Snow and Ice Data Center NSIDC;, was used for this study. Version 5 is also considered for directly deriving land surface temperature for the production of climate records within the European Space Agency CCI LST project Climate Change Initiative Land Surface Temperature,. The AMSR time series of the MEaSUREs dataset is available as a 6 km resolution record in addition to the 25 km record which includes time series from all sensors. The 6 km dataset covers the Northern Hemisphere and the Southern Hemisphere, currently spanning the period 2002–2021 . However, only the 25 km product (consistently with SSM/I) was used for this study. AM data were chosen where available since most available SAR data for the Alaskan benchmarking site (97 %) were acquired at AM times.

Further products exist from SMAP (2015 to present) and SMOS (2010 to present) and are available on an operational basis . Both are L-band PMW missions and therefore provide information on the upper centimeters of the soil. The SMAP mission specifically targets freeze–thaw, complementing soil moisture, and requirements have been defined for the mission . They have been phrased for the radar instrument (which is currently not in operation). The surface binary freeze–thaw state in the region north of 45° N latitude, which includes the boreal forest zone, was targeted with a classification accuracy of 80 % at 3 km spatial resolution and 2 d average intervals. A 2 d precision at the spatial scale of landscape variability (≈3 km) was originally targeted. The available SMAP PMW product considers four stages: frozen, thawed, transitional, and inverse transitional . The two transitional categories refer to days with differences between AM and PM acquisitions. The SMAP product shows >70 % accuracy for the 36 km grid version according to . Accuracies of 78 % and 90 % were determined for descending (AM) and ascending (PM) orbit observations, respectively, in relation to independent surface-air-temperature-based FT estimates from ∼5000 global weather stations . The SMOS mission FT product addresses three stages, namely thawed, partially frozen, and frozen, covering the Northern Hemisphere daily, 25 km;.

The different sensors represent different frequencies (see Table ), and different strategies were followed for the FT retrieval. The MEaSUREs algorithm is based on a seasonal-threshold approach. The threshold was derived annually on a grid-cell-wise basis using an empirical relationship between brightness temperature and daily surface air temperature records from global model reanalysis. The SMMR-SSM/I-SSMIS record was developed by merging the Scanning Multichannel Microwave Radiometer (SMMR), Special Sensor Microwave Imager (SSM/I), and Special Sensor Microwave Imager/Sounder (SSMIS) 37 GHz frequency (vertical polarization) brightness temperature records . The AMSR records represent 36 GHz. The global 25 km datasets are provided as a cylindrical, equal-area projection, which translates to 10km×62.5km at the latitude of the benchmarking sites (Fig. ).

SMAP FT state was determined using a seasonal-threshold approach . A normalized polarization ratio (using V and H brightness temperature; 1.41 GHz) is assessed. Winter and summer references are derived from frozen and thawed soil conditions. These are calculated for each year and averaged over the entire SMAP period according to the documentation of. These values are used for the derivation of a threshold and a seasonal scale factor. In cases of low SCV (single-channel vertical polarization) correlation with physical surface temperature, SCV retrieval is used, and, in some cases, an AMSR-E mitigation scheme is applied . The dataset is provided in Lambert azimuthal equal-area projection and is available at 36 and 9 km (enhanced level 3) gridding. Only the latter was used.

Thaw and freeze references are also used in the case of SMOS FT retrieval . For FT state estimation, the normalized polarization ratio (NPR, 1.400–1.427 GHz) values are scaled using the reference NPR values from frozen and thawed soil conditions. All potential observations from the frozen and thawed soil conditions are collected from the winter and summer periods when reanalysis air temperature data indicated <-3 and >+3 °C, respectively. From these collected potential values, the 50 most extreme ones are stored, and their median is used as a reference value. These are pixel-wise values. Then two thresholds are used to differentiate between three FT states from the scaled NPR. The dataset is provided in Lambert azimuthal equal-area projection, with 25 km gridding.

Several active microwave products have been published previously but are limited temporally and spatially and are thus not considered for benchmarking. These include an experimental product from MetOp ASCAT C-band scatterometer, 12.5 km gridding,, starting in 2007 and aimed at permafrost and flux applications . It is available at 12.5 km nominal resolution and considers three states: frozen, unfrozen, and melting. In addition, a C-band SAR product at 1 km was developed based on ENVISAT ASAR GM and was used to assess the ASCAT product .

Benchmarking of the PMW datasets has been carried out over 17 SMOS grid cells (and overlapping SMAP and MEaSUREs cells) for Alaska and 12 cells for northern Finland (Fig. ).

Temperature records from in situ observations include long-term measurement sites such as boreholes and automatic weather stations (AWSs), as well as short-term and spatially distributed observations using iButton data loggers. The latter network was established for validating surface conditions derived from C-band satellite data for details, see. These data have been used for inter-comparing Sentinel-1 with a coarser regional dataset based on MetOp ASCAT, covering the region of Kaldoaivi in northern Finland. This site is located within a palsa mire and represents a zone of sporadic permafrost. The temperature sensors have been distributed in the proximity of a permafrost borehole site. The majority of the iButtons are located outside of permafrost-affected landscape elements. The snow water equivalent (SWE) varies across the sites in winter, but at most sites, snow depth is sufficient for insulation, and temperatures remain close to 0 °C below the snowpack. The available record includes 24 sites, spanning a period of 2 years (August 2016 to August 2018). The iButton loggers were placed approximately 2–3 cm below the surface to avoid direct warming influence by the sun.

Records from boreholes and automatic weather stations (AWSs) across Alaska , including the North Slope and representing continuous to discontinuous permafrost, have been used for calibration (Fig. ). Data are available from 26 boreholes, with depths varying between 1–75 cm, and from 22 AWSs providing air temperature. A total of 3 years of data (2018 to 2020) have been used. ERA5 reanalysis air temperature data for the same time period were also used for the evaluation of the fused dataset to allow comparability with previous studies.

First, the benchmarking dataset was created using Sentinel-1. The implementation of surface state classification, building on intermediate products of the wet-snow retrieval, as proposed in , requires a threshold determination as a first step. For calibration, in situ data from Alaska limited to the upper few centimeters of the soil have been used. The target land cover type (as determined by the potential applications of an FT product) are soils. Water surfaces, bare areas, and glaciers therefore need to be masked with a dataset of similar spatial resolution . The circumpolar land cover units by fulfill this requirement and have therefore been used for masking.

Eventually, the frozen fraction is derived individually (spatial subsetting) from the masked data for each cell of the PMW FT products listed in Table . The resulting time series are used for the evaluation and the development of the fusion scheme. Fused records are assessed with the benchmark dataset in a similar way to the input records and the in situ records. The fused-record examples are combined with a dataset providing mid-winter thaw and refreeze of snow , where such events occur, to facilitate the discussion.

The original wet- and dry-snow detection utilizes repeat pass dual-polarization Copernicus Sentinel-1 SAR data (acquired in IW swath mode) from different tracks. The method includes the generation of reference images, applying a statistical pixel-wise analysis of tens of co-registered repeat pass Sentinel-1 images . To exploit the dual-polarization capabilities of Sentinel-1, the backscatter ratio of the snow-covered period and reference data are calculated separately for co- and cross-polarization and are then combined. The combination accounts for differences in the angular backscatter behavior of co- and cross-polarized backscatter data by applying the local incidence angle as a weighting factor. The result is a weighted ratio combination RC based on the backscatter ratios of VV (RVV) and VH (RVH). 1RC=WRVH+(1-W)RVV

The weight (W) varies with the local incidence angle (θ) by applying the following rule: 2IF (θ<θ1)→{W=1.0},IF (θ1≤θ≤θ2)→{W=k1+(θ2-θ)(θ2-θ1)},IF (θ>θ2)→{W=k}.

For the datasets in this study, we use k=0.5, θ1=20°, and θ2=45° following . The SAR data have been sampled to 100m×100m.

More than 1900 scenes have been preprocessed in total over both regions. In situ locations in Alaska were covered by 52 to 179 scenes each. A total of 115 scenes have been used over Finland for comparison with in situ records. The time period covered for Finland corresponds with the in situ data availability (August 2016 to August 2018, with acquisitions every 12 d). Data processed for Alaska span 3 years, from 2018 to 2020 (with, on average, an acquisition every 9 d using overlapping orbits). This results in differing sampling intervals between the regions and lower sampling compared to the global FT records.

The last step of the wet-snow retrieval scheme of is the segmentation of wet-snow areas versus snow-free and dry-snow areas based on the RC images . This step was adjusted for the FT retrieval. Specifically, a suitable threshold Thr needed to be defined for the segmentation of frozen versus unfrozen soil.

The threshold determination followed the approach originally suggested for VV backscatter time series by (scatterometer) and adapted by (SAR). The threshold was defined based on a logistic function fit between backscatter and temperature data and the threshold Thr determined from the inflection point. This was based on reanalysis data in . tested the calibration based on in situ air temperature over a limited number of sites and validated the results with in situ near-surface soil temperature data. Here, we also used in situ data for the air and soil but from a larger dataset and also for calibration. Air and soil temperature time series were compared to the RC value of each corresponding 100m×100m SAR pixel. The inflection point of the fitted function was derived for both types, and the corresponding backscatter value was used as a threshold to distinguish between frozen (≤Thr) and unfrozen (>Thr) conditions. Here, we applied it to the combined VV and VH ratio values (RC) as defined in . The logistic function was defined based on the processed RC data on the Alaska North Slope. The fitting procedure included the consideration of the maximum value of RC and the average for all values acquired under conditions below -5 °C. For the borehole records, the uppermost sensor below the ground was selected. Validation of the threshold approach was carried out based on the near-surface temperature observations available for northern Finland (Kaldoaivi). For this purpose, daily averages of 0 °C and lower were defined as frozen, and all values above 0 °C were defined as unfrozen.

A disadvantage of using backscatter intensity at C-VV is the temperature dependence under cold conditions . The potential influence was therefore also investigated using the available in situ records. In addition to RC, C-VV and C-VH backscatter intensity input values were compared with the temperature measurements for this purpose.

The available global products represent varying grids. Therefore, the Sentinel-1 backscatter ratio analyses needed to be carried out over sufficiently large areas in the proximity of the in situ locations (overlap of grids, Fig. ). The benchmarking extent was chosen considering both in situ data availability and the distance from the coast in order to limit potential quality issues in the PMW datasets. Both sites are labeled as high-quality areas in the MEaSUREs dataset. SMOS quality flags indicate, temporarily, up to 5 out of the 17 cells as being potentially affected by radio frequency interference (RFI) over northern Finland. No RFI issues are listed for the Alaskan sites. SMAP flags indicate retrieval with SCV (single-channel vertical polarization) due to a correlation of <5 % with temperature data and, in some cases, the use of AMSR or TB (brightness temperature) mitigation. The latter occurred only outside of the transition periods, in mid-summer and mid-winter (Fig.  in the Appendix). For Alaska, this includes mostly the end of December to the end of March and, in some cases, also July and August. For Finland, flagging occurred from July to September (where freeze-up starts in October).

Permanent snow- and ice-covered areas, water, and barren surfaces (bedrock) have been extracted from land cover for the masking of “none-soil” fractions in the SAR images. The none-soil fraction can be very large in permafrost environments. Figure  shows lakes and bare area as derived from a land cover map . Both land cover types were excluded from the frozen-fraction calculations.

The frozen and unfrozen fractions were derived for each of the overlapping footprints of the PMW products (grid cells as shown in Fig. ). The proportion of unmasked Sentinel-1 pixels classified as frozen has been determined for each acquisition. Basic statistics were derived for each global product separately (including for the presentation of results as violin plots).

Criteria for the preference of a certain dataset were eventually defined for a fusion scheme based on the benchmarking. The year is split into three phases for this purpose: DOY (day of year) 1–209, 210–300, and >300. In a first step, the datasets to be considered for a certain period were selected. In a second step, the conditions were defined (considering agreement between the datasets and/or the results of benchmarking). The combined dataset (referred to as the fusion dataset) was eventually also compared to the benchmark dataset derived from the Sentinel-1 frozen fraction. The fusion of the daily records was made on the basis of the SMOS grid (25 km). All comparisons of the fusion dataset were also made on the basis of the SMOS grid. The partially frozen flag was analyzed separately from the other flags (thawed and frozen).

Data from three borehole locations on the Alaska North Slope were used for the assessment of the unfrozen and frozen flags (temperature sensors at 5–7 cm depth) of the fusion product. In addition, ERA5 reanalysis data (air temperature) were investigated at the same sites (for locations, see Fig. ). The frozen state retrieval from the soil and ERA5 air temperature data was based on the following daily mean temperature thresholds: values <-1 °C are considered to be frozen, and values >1 °C are considered to be unfrozen, while values in between are classified as partial thaw.

Days with midwinter thaw and refreeze of the snowpack or wet snow were not part of the benchmarking but can potentially also be added using existing products. As an example, mid-winter thaw and refreeze, as available from the combination of MetOp ASCAT and SMOS , have been added, and a flag value has been assigned. Such data are applicable for November to February (day of year 300 to 60).

The influence of temperature on C-VV backscatter as described in and does not apply to VH (Fig. ). Backscatter remains at a similar level for all temperatures below 0 °C at this polarization combination. Differences between the years due to changes in snow structure, as can occur, in particular, due to rain-on-snow events , are also less pronounced at VH. The temperature influence is visible in the combined VV and VH ratio values (RC increasing with decreasing temperature), inherited from VV, but the magnitude is lower than in the VV input data.

The determined value of the inflection point from the logistic function fitting varies across sites (examples for comparison with soil temperature in Fig. ). The spreads of the ratio values in winter and in summer are large, on the order of 1 to 2 dB, but they are largely distinct.

Only minor differences are found between results using soil temperature and air temperature data (Fig. ). The mean values of the inflection point differ by only 0.12 dB. The range is higher for air temperature. All values are higher than the suggested threshold for wet snow -2 dB,. Most inflection point values even exceed it by more than 3 dB. A threshold of 1.34 dB based on the soil data (median) is suggested for regions with a soil layer that can contain water and/or ice (basic requirement for FT detection).

The applicability of the threshold could be confirmed using independent records by means of the in situ data from northern Finland (Kaldoaivi region; see Table ). In total, more than 1800 samples of daily average temperature from 24 sites have been available for the Kaldoaivi region. A total of 93.2 % and 91.2 % (average 91.8 %) of all samples within the 2-year time period were correctly classified for the unfrozen and frozen conditions, respectively.

Only grid cells of the PMW datasets that are completely covered by the Sentinel-1 scenes have been cross-compared (Alaska: Fig. ; Finland: Fig. ). The shown results are limited to two periods, namely from April to June (from completely frozen to completely unfrozen) and from August to November (from completely unfrozen to completely frozen), in order to identify potential issues for the different transition types. Violin plots have been chosen to visualize the data distribution (of frozen fraction) for thawed and frozen conditions as defined in the products. For SMOS, the fraction was also analyzed for the thawing flag, which represents partially unfrozen conditions. Time series examples are provided for two distinct borehole locations, namely Happy Valley and Sagwon (Fig. ). The frozen flags of each product were translated into values of 10 (unfrozen), 20 (partially frozen), and 30 (frozen) for comparability in these cases. Near-surface soil temperature and frozen fraction are shown together with the new flag values. In spring, the frozen fraction as detected by Sentinel-1 drops when the in situ temperature (1 cm depth) starts to rise above the late-winter level at Happy Valley. This drop does not coincide with the beginning of the thaw in the PMW cells. The start of the frozen-fraction increase does, however, agree with the drop below 0 °C in the 2020 example. This is similar for the second site (Sagwon, 8 cm depth, Fig. ). Both borehole records show the so-called zero-curtain effect at freeze-up. Temperatures can remain at almost 0 °C for weeks to months due to latent heat, which maintains the temperatures . The frozen fraction is already at 100 % weeks before the ground temperature eventually lowers.

The fraction comparison clearly shows a frozen-state detection issue for SMAP during the winter on the Alaska North Slope (Fig. ). A higher number of grid cells is defined as unfrozen than for the other datasets. This effect is more pronounced for April to June than for August to November. This leads to a premature thaw detection in spring. The timing of complete freeze-up is, however, generally captured when the flag switches from unfrozen to frozen, as can be demonstrated for the Happy Valley borehole location for 2020 (Fig. ). The unfrozen flag is assigned to the SMAP grid cell from the beginning of the depicted time period, starting on 1 April 2020. The frozen fraction as determined by Sentinel-1 drops below 100 % towards the end of April and reaches 0 % in June. The unfrozen-flag issue at Happy Valley returns in mid-November. This happens only occasionally at the Sagwon site (Fig. ). AMSR and SSMI flags switch from frozen to unfrozen by the end of May, when the Sentinel-1 frozen fraction is at almost 0 %. The partially frozen flag in SMOS only occurs here in autumn (coinciding with a frozen fraction of about 80 %); this can also be shown for a further borehole location for the autumn transition in 2019 (Sagwon, Fig. ) and agrees with the fraction comparison results from the North Slope (Fig. ).

SMOS captures the middle to the finalization of thaw on the North Slope well and is closer to the start of thaw for northern Finland (unfrozen flag includes a higher fraction of frozen conditions, Fig. ). Freeze-up is detected with a delay. The majority of days with the partially thawed flag show almost complete freeze-up (spatially) in Sentinel-1. The MEaSUREs AMSR and SSMI products largely agree with each other (time series examples in Fig. ). The spring switch occurs at the start to the middle of the thaw at the Happy Valley borehole location. The average frozen fraction with the unfrozen flag in spring is lower than for the other products.

The benchmarking results suggest the need for a fusion of all available products (Table ). This allows a start of the time series in 2015 only (SMAP launch). The start and end of thaw can be represented back to 2010 (SMOS). The MEaSUREs records back to 1979 allow good detection of the start to the middle of the thaw.

SMOS and MEaSUREs (SSMI or AMSR) are recommended for consideration in all three time periods (Table ). The performance of these products was shown to be similar for early winter; therefore, either of the products can be used. For late winter and during the thaw period, differentiation needs to be made in terms of whether the different products agree with each other or not. SMAP is recommended for consideration in the definition of the partially frozen flag during freeze-up, specifically when its flags disagree with the other products.

The PMW products were fused following the rules defined in Table . SMOS and MEaSUREs are recommended for consideration in all three time periods. The SMOS grid cell definition was used as a basis (see also Fig. ). The previously introduced scheme for flag values (10 – unfrozen, 20 – partially frozen, 30 – frozen) was applied.

Records were fused for the same grid cells as for the initial frozen-fraction comparison. The assessment procedure followed the same procedure in the first step. The frozen fraction from Sentinel-1 (SMOS grid) was compared by flag and separately for the thaw and freeze-up periods.

In addition, the partially frozen flag has been assessed over three selected borehole locations. The frozen fractions for days flagged in the SMOS product were compared to the fused dataset in order to evaluate the added value of MEaSUREs flags. The frozen and unfrozen flags of the fused product were evaluated considering soil temperature from boreholes at two sites. Results based on reanalysis air temperature data (ERA5) representing an average over a larger area are provided in addition.

The use of combined ratios of backscatter intensity using VV and VH polarization, as previously suggested for wet-snow detection, provides a means of reducing the incidence angle effects, but the impact of temperature variations on backscatter intensity at VV during frozen conditions remains (Fig. ). The combined ratio RC also shows this linkage. The potential influence of temperature on FT retrieval at C-VV has been previously discussed . Our results demonstrate that this is specific for VV but not present in VH (Fig. ). The retrieval based on VH alone would, however, require an appropriate approach regarding the influence of the incidence angle, such as the normalization of σ0 or the retrieval of γ0 . An implementation based on VH would also require a location-specific threshold determination (instead of a global threshold), which is not feasible for the purposes of this study and for potential regional applications of the FT detection approach.

The combined-ratio approach does, nevertheless, provide good-quality results. The overall agreement with in situ data of 92 % at the site in northern Finland is similar to the previously reported accuracy of 94 % using the location (pixel)-specific threshold determination approach . The validation was carried out over a different region (northern Finland) than the calibration (Alaska North Slope), confirming the transferability of the method. The performance is also similar to a CNN (convolutional neural network) approach using both polarization bands 88 %, with calibration and validation over the same region in NE Canada;.

We tested the combined-ratio approach over tundra only. It can be expected that the performance would be lower over forested regions as it relies on the use of VH. A method using VV only, such as that suggested by , might perform better for forested areas.

The SMOS–Sentinel-1 comparison results agree with findings of . Offsets in the FT timing compared to in situ measurements were described and attributed to the ability of the L band to represent larger soil depths. SMAP, which is also based on the L band, does, however, represent the end of the autumn freeze-up period well. The AMSR and SSMI frozen flags also indicate a later freeze-up in some cases (multiple switching from frozen or unfroze; see Fig. ), although a much smaller wavelength is used.

Differences in wavelength and/or frequency (see Table ) are expected to contribute to the deviations between the results in addition to differences in the FT classification method, although all of this depends on thresholding. The MEaSUREs scheme relies on the use of one polarization only, with, at the same time, a comparably small wavelength and high frequency. Both SMOS and SMAP FT products are based on the use of the NPR, which uses two polarizations but at a comparably large wavelength and small frequency. Whereas a single threshold is used to separate thawed from frozen conditions for SMAP, two thresholds are used to derive partially frozen conditions in the case of SMOS. The thawed fraction derived from Sentinel-1 is, however, rather low in most cases of the partially frozen condition (Table ). This applies specifically for the spring transition (Fig. ). Once SMOS has indicated a frozen soil condition, the mid-winter values are forced to remain frozen under cold air temperature conditions , which may contribute to the better agreement with the benchmarking dataset during this period than for SMAP. Both SMOS and SMAP can be affected by RFI. The SMOS quality flags indicate a potential impact for the Finland site but not for northern Alaska. Thus, the different results for Alaska, as shown in the example of Fig. , may originate from the threshold definitions and the consideration of a single channel (V) under certain conditions in the case of SMAP. The latter was developed for low latitudes according to the product documentation .

SSMI and AMSR records showed very similar performances (e.g., Table ). The AMSR time series, as part of the MEaSUREs dataset, is also available at 6 km nominal resolution. The use of this version might be of benefit, replacing both the 25 km AMSR and SSMI records. Further analyses would, however, be required for the quantification of the added value of the improved spatial resolution.

The comparison with the Sentinel-1 frozen fraction and in situ data with the seasonal freeze–thaw prototypes indicates an improvement through the fusion of the different products. However, a range of issues, which differ between the evaluation sites (continuous permafrost over Alaska versus sporadic permafrost for northern Finland), remain. Neither the actual end of thaw in spring nor the start of freeze-up in autumn can be captured. These issues cannot be solved based on the existing passive-microwave-based CDRs.

The frozen fraction derived from Sentinel-1 only applies to areas without open water and barren surfaces. Therefore, the deviation of the PMW products could also be analyzed with respect to the open-water fraction, as has previously been investigated for active microwave freeze–thaw , in order to provide more insight into product issues. An offset in freeze–thaw timing can be expected in relation to lake ice thaw and formation.

Previous analyses of C-band scatterometers indicated the applicability of active microwave data for FT retrieval in high-latitude regions e.g.,. Specifically, the fusion of SAR and scatterometers is promising for the characterization of the transition period actual determination of the start and end;. The new SAR-based method which has been developed in our study enables operational retrieval of surface state from sensors such as Sentinel-1 and theoretically allows us to meet the requirements for a multipurpose FT product if sufficient acquisitions (daily) are available. In addition to soil FT, wet-snow detection could also be addressed as it is based on the same preprocessing scheme. The availability of Sentinel-1 data is, however, constrained, and the polarization differs across the Arctic. The required data are unavailable for Greenland and the Canadian High Arctic. In the remaining regions, repeat intervals are not dense enough (daily required). However, demonstrated the use of historic (1-year) Sentinel-1 data for the calibration of a freeze–thaw fraction derived from MetOp ASCAT. FT monitoring considering fraction could therefore be implemented in regions with VV and/or VH availability. Retrieval with HH (horizontally sent and received) and/or HV (horizontally sent and vertically received) remains to be tested. An implementation based on MetOp ASCAT would allow the production of a CDR going back to 2012 (MetOp-A and MetOp-B availability).

More advanced FT algorithms using dielectric models to represent non-linear behavior in highly organic soils common in the Arctic may be able to detect changes in the amount of unfrozen liquid water remaining in soil under sub-zero temperatures rather than relying on simple binary FT thresholds e.g.,. Another advanced method employed by shows potential for directly retrieving continuous soil permittivity data over cold regions, as has currently been demonstrated over a test site.

The current benchmarking exercise was limited to AM data in the case of MEaSUREs and SMAP due to the availability of Sentinel-1 data at the selected sites during this period. Diurnal thaw and refreeze are, however, common e.g., during transition periods. The added value of PM data should be addressed in future studies considering other sites or further relevant SAR missions. The future L-band mission NISAR can be expected to be of high value for such benchmarking as it represents a similar frequency as that used for SMOS and SMAP, and a comparably high temporal sampling will be available. Our analysis is limited to non-forested areas due to the use of C-band SAR. L-band SAR provides the possibility to extend research into forests.

Current FT products aim for the identification of an average surface state condition within a footprint (e.g., , using SSMI; , using ASCAT; , using SMAP and , using SMOS), but specific applications, including permafrost, soil moisture, and greenhouse gas fluxes, require information on the start of thaw and/or freeze-up and the completion of thaw and/or freeze-up Tables –,.