We developed and validated the F/T detection algorithm based on reference soil (Tsoil) and air (Tair) temperature data (Fig. 2).

Soil temperature (Tsoil) was measured using low-cost sensors (i-Buttons) deployed within the IMA over 2 years from July 2018 to mid-July 2019 in ten sites; and from mid-July 2019 to August 2020 in seven of the ten previous sites. The i-Buttons measured the soil temperature every three hours and have been placed at three depth ranges in the soil: from 2 to 4 cm; from 10 to 20 cm and from 20 to 30 cm in the following sites (IPX): Hydric sedge fen (IP2), Sedge fen (IP7, IP8 and IP14), Productive sedge fen (IP6 and IP11), Subhygric communities (IP5 and IP10), Turfy mountain avens – Curly sedge (open <25 %) (IP15), and Unvegetated (open >90 %) (IP13) as detailed in McLennan et al. (2018). Only the i-Buttons located at the surface (first 4 cm of the soil) were used in the study, since C-band measurements are sensitive to changes near the surface and show limited soil penetration (Rowlandson et al., 2018). A fixed uncertainty of 0.5 °C was considered on the data since the observed biases are within the manufacturer's precision (Prince et al., 2019). Soil temperatures were averaged daily since daily soil temperature variability is of the order of the i-Buttons uncertainty. The surface was considered frozen when Tsoil≤0.5 °C and thawed when Tsoil>0.5 °C according to i-Button data, considering its accuracy and the clear zero curtain observed around 0 °C during the freezing in fall (Fig. 2). The zero curtain refers to the extended period during the freezing or thawing of the soil where the temperature stays around 0 °C due to the release of the residual latent heat or water in the soil (Domine et al., 2018a). Surface freezing (Dfr) and thawing (Dth) reference transition day of the year (DOY) were derived from the daily mean Tsoil. Dfr(soil) was defined as the first day of a series of 7 consecutive days when the surface was considered frozen for every i-Buttons, and Dth(soil), as the first day of a series of 7 consecutive days considered thawed.

Average daily air temperature (Tair) data for 2018 to 2019 were obtained from the Environment and Climate Change Canada meteorological station (ECCC, https://climat.meteo.gc.ca/historical_data/search_historic_data_f.html, last access: April 2026) located at the Cambridge Bay airport, about 15 km from the IMA. Global reference transition DOY (Dth(air) and Dfr(air)) were calculated from Tair as the first day of a series of 7 consecutive days classified as frozen (Tair≤0 °C) or thawed (Tair>0 °C). Since Tair is considered constant over the study site, reference transition seasons were derived as ±30 d around Dth(air) and Dfr(air) (grey area in Fig. 2).

In this study, we used the C-band SAR backscatter obtained from the sensors of two satellite missions, namely the Sentinel-1A-1B constellation (available since 2016), and RADARSAT-2 (made available for this study between August 2018 and December 2019). Both SAR operates at a center frequency of 5.405 GHz. This study focuses on the period spanning from August 2018 to December 2019, where the overlap between both data source allows for quasi-daily revisit time in the transition seasons of Fall 2018 (F-2018), Spring 2019 (S-2019) and Fall 2019 (F-2019). Using both sensors, the average revisit times are respectively 1 d for F-2018 and F-2019, and approximately 2 d for S-2019 (see Table 2 in Sect. 3.1). Both datasets were preprocessed independently with a standardized processing chain (Sect. 3.1) and resampled to a standard 50 m × 50 m grid to be combined into one multisensor time series. Note that all Sentinel-1A-1B and most RADARSAT-2 observations were taken around 13:00 UTC. Hence, only 14 % of SAR observations were taken around 00:00 UTC (all RADARSAT-2 observations).

We used 270 Level-1 GRD Extra Wide Swath (EW) images from Sentinel-1A-1B in dual polarization (HH, HV) with a grid spacing of 40 m × 40 m. Due to the combination of measurements from multiple orbits, the acquisition incidence angles ranged from 20 to 46°. Descending/ascending scenes were used together.

We used 200 ScanSAR wide SAR Georeferenced Fine product (SGF) RADARSAT-2 images in dual polarization (HH, HV) with a 50 m × 50 m grid spacing. We also used multiple orbits with an acquisition incidence angle ranging from 20 to 49°, including both descending and ascending scenes.

This study also presents a case study relying on a unique ecotype map created from unsupervised classification of 2011 WorldView-2 multispectral imagery (Ponomarenko et al., 2019) based on the Canadian arctic-subarctic Biogeoclimatic Ecosystem Classification (CASBEC) defined in McLennan et al. (2018). The CASBEC classification is based on terrestrial ecosystems (i.e., ecotypes) and includes vegetation characteristics, as well as the soil moisture regime (SMR). The ecotypes are mapped for Greiner Lake's watershed at a spatial resolution of 10 m and contain 19 map units regrouping 21 ecotypes. Resampling the 10 m × 10 m map to a 50 m × 50 m grid allowed to match the SAR imagery time series resolution. The analysis only included the pixels where 90 % of the original 10 m × 10 m pixels within a 50 m × 50 m pixel have the same ecotype value. Waterbodies and ecotype classes with pixel count smaller than 20 were also excluded from the case study analysis. Table 1 present the number of pixels left from the resampling for the remaining ecotype classes. More details on the ecotypes can be found in Ponomarenko et al. (2019) from which a summary of the key characteristics relevant to this study are presented in Table 1: the soil moisture regime (SMR) and the vegetation height. The soil moisture regime is defined by a scale ranging from 0 to 9, where 0 is very xeric and 9 is aquatic.

The optimized threshold, defined with the ten reference sites, was then used to create the surface F/T maps covering the IMA and its surroundings (IMA+; Fig. 1). The algorithm was applied on a cell-by-cell basis to retrieve the transition onset for F-2018, F-2019 and S-2019. We applied a hydrographic mask to remove any non-terrestrial pixel. Pixels with equal or higher frozen than thawed reference values (σfr≥σth) were also discarded. The temporal evolution of the backscatter in those cases did not allow for the detection of the dielectric discontinuity between soil state during the transition seasons, potentially linked to the presence of rocks of higher rugosity inside the pixels. Also, since Tair was considered constant across our study area, detected freezing and thawing dates falling outside the 60 d periods defined around Dth(air) and Dfr(air) were considered false detections, and were therefore discarded during the creation of the surface F/T maps. We chose to use the Total H power for the creation of the F/T maps to decrease the impact of the target orientation (e.g., vegetation distribution). Furthermore, by using this approach, we also increased the signal-to-noise ratio (SNR; Entekhabi et al., 2014), linked to the increase in signal quantity.

To analyze the impact of ecotypes on the surface F/T DOY, we compared the surface F/T maps derived from HH+HV to the resampled ecotypes map, and calculated the mean surface and standard deviation of surface F/T DOY per ecotype. A generalized least square model (GLS) was used to predict values of surface freezing and thawing DOY in function of the ecotype class. The gls function in R was used. This function fits a linear model using generalized least squares and considered when the ordinary least square regression assumptions are not met, i.e. errors are correlated and/or have unequal variances We created 3 models, one per transition season (freezing transition 2018 and 2019 and thawing transition 2019). A variance and spatial autocorrelation structure were defined in the models. F/T transition DOY derived from the models were used to establish if the differences of the timing in transition DOY between ecotype class were statistically significant.

The two reference values σfr and σth were calculated independently for each of the reference sites. Table 3 shows that for the combination of the reference sites, the performance of the algorithm is not impacted by the approach used to determine the reference values over the complete study period, with a difference of less than 0.5 % in detection accuracy. The same similarities were observed when calculating the accuracy per transition season for each reference value method. Those similarities between methods confirm that C-band signal can easily discern the two thermodynamic stages of the soil independently from the method chosen. Even though we observed close to no difference, the “average-5” method could be more sensitive to the presence of residual noise or outlier backscatter measurements left in the time series, just as the “average” method could be. Therefore, we used the “median” method for the following analysis.

Figure 5 shows the accuracy of the detection algorithm as a function of the threshold between 0 and 1 (at a 0.01 increment) on HH, HV and HH+HV datasets. The accuracy is calculated for the ten reference sites combined, with Tsoil (solid line) and Tair (dashed line) reference values over (a) the entire time series; (b) the freezing transition periods for 2018 and 2019 combined; and (c) the thawing transitions for 2019. The algorithm applied over the entire time series (Fig. 5a) suggests high accuracy (>95 %) for threshold values ranging between 0.40 and 0.65 for all polarizations and temperature datasets. Figure 5b and c show the difference in the retrieval accuracy along the increment of thresholds during the freezing (11 August to 10 October 2018) and thawing (11 May to 10 July 2019) periods respectively. Results show that a clear threshold value maximizing the accuracy exists between 0.50 and 0.65 for the classification of the freezing period (Fig. 5b) for all datasets. For the thawing transition, multiple threshold values across a larger range yield a high accuracy (Fig. 5c).

The signal classification during the thawing transition is largely insensitive to the threshold over a wide range of increments, where the accuracy remains mostly stable (0.2 to 0.6 for HH, 0.3 to 0.7 for HV and 0.2 to 0.7 for HH+HV polarization). A common threshold for both transition periods could then be considered for the surface F/T detection. Figure 5c also suggests that in the classification accuracy, the soil temperature performs better by almost 10 % when compared to air temperature over the range of thresholds presented. We therefore chose the optimized threshold from the best accuracy compared to Tsoil measurements for the combination of both transition seasons. A threshold of 0.56 was defined as maximizing the accuracy of transition seasons for HH polarization and of 0.53 for HV polarization (Table 4).

Since both polarizations show similar accuracy results for the F/T detection, the combination of the two polarizations also gives similar high results. The same threshold optimization was done on the Total H power time series and a threshold of 0.62 was found to optimize the accuracy of detection (Table 4).

The temporal evolution of Tsoil for the reference sites is quite different for F-2018 and F-2019 due to different meteorological conditions during those freezing transitions. Figure 8a and b show the temporal evolution of Tsoil at two locations (IP8 and IP13) along with Tair. We chose the site IP8 and IP13 to illustrate the difference between moist and dry ecotypes respectively. The site IP8 is composed of a Sedge fen ecotype, which is a hydric ecosystem characterized as very moist to wet (McLennan et al., 2018). The clear zero-curtain effect in higher moisture ecotype is highlighted during F-2018 in IP8 (Sedge fen), with the soil temperature remaining close to 0 °C over a longer period, before freezing completely (Domine et al., 2018a). This is different compared to the drier IP13 site located in the Unvegetated (open >99 %) ecotype. In addition to the difference due to soil moisture, we examined snow cover information from the IMS Daily Northern Hemisphere Snow and Ice Analysis (U.S. National Ice Center, 2008). This dataset provides information at 1 km to determine the snow onset (>40 % coverage of 1 km) for these two transition seasons. We found that snow arrived almost one month later in 2019 (9 October 2019) than in 2018 (11 September 2018). The zero-curtain period is increased when snow covers the ground and isolates the soil from the cold air temperatures (Yi et al., 2019). We observed this with the soil temperature of the Sedge fen ecotype during F-2018 (Fig. 8a) when compared to the same location during F-2019 (Fig. 8b). On the other hand, the absence of snow cover increases cooling of the soil from the air, leading to a more direct freeze for both ecotypes (Fig. 8b). This is illustrated in Fig. 8b, when both i-Buttons freeze concurrently with Tair falling below freezing in F-2019. A clearer distinction between states is reflected in the C-band signal with a faster and greater decrease of backscatter between thawed and frozen soil in the F-2019 time series for both the Sedge fen and the unvegetated ecotype as shown in Fig. 8d compared to Fig. 8c. Interestingly, despite the difference in the soil moisture regime between the two sites, the backscattering trends are similar, with slightly higher σ0 at IP13 in winter, probably related to the lower fraction of ice in the soil. These results show a possible impact of summer water balance and soil moisture at the end of the season on the initial backscattering coefficient at the beginning of the freezing season. Our results show that even if the initial soil moisture change the backscattering at the beginning of the freezing seasons, the contrast between frozen and thawed backscattering remain large enough to use the proposed simple thresholding approach. Another possible impact on freezing identification is the presence of vegetation (Cohen et al., 2024). However, vegetation high and biomass in our study site is low with limited impact on backscattering as shown from studies using ASCAT (Liu et al., 2023). Hence, if vegetation senescence in fall as an impact on backscattering, the effect is much lower than the soil FT processes as shown by the accuracy over 93 % obtained in this study.

Furthermore, Fig. 8a shows that the decrease in the backscatter signal agrees with the start of the zero-curtain effect as measured at 2 cm in the soil. This result highlights that even though Tsoil is not below zero, like for F-2018, the microwave signal decreases following the dielectric discontinuity of the soil surface when Tsoil is close to 0 °C, as observed in Rowlandson et al. (2018), thus making the C-Band signal sensitive to the beginning of the zero curtain. Since the signal at C-band frequency is sensitive to the soil surface, we hypothesis that, at the beginning of the zero curtain and around 0 °C, a thin frozen layer (i.e., a change in the water phase) appears at the very top of the soil, creating a dielectric discontinuity and thus resulting in a backscatter decrease. Also, we chose the i-Buttons location to be representative of a homogeneous patch of land cover. However, the SAR pixels cover a soil surface of 2500 m2. During F-2018, the start of the gradual backscatter decrease observed before Tsoil reached 0 °C could be explained by the averaging of the soil contribution inside the pixels. Indeed, the punctual measurements of Tsoil do not represent the full extent covered by the pixels especially when isolating snow is present on the ground, increasing the zero-curtain effect for higher ground moisture (Domine et al., 2018a).

As for the detection capacity of the algorithm, the faster decrease in the signal, showing a clearer distinction between surface states, resulted in higher accuracies for F-2019 classification (Fig. 9), meaning less uncertainty in the detection, than for F-2018 for every threshold increment. Nonetheless, the C-band signal was sensitive to the freezing onset of the surface with a decrease of 3–4 dB for the two fall seasons, with an accuracy higher than 93 % independently from the presence of snow compared to Tsoil reference values.

Dividing the signal into freezing and thawing transition seasons during the threshold optimization showed a clear difference between the signal behavior for those two seasons, linked to the different processes driving those transitions. Regardless, Fig. 5b and c showed that we could use a common threshold definition for both transition seasons since, for the thawing classification, the accuracy is nearly independent from the change in threshold. This can be explained by the behavior of the signal during the thawing season, for which we observe a clear difference in backscatter between frozen and thawed surface. Generally, it is expected that the surface thaw onset happening under wet snow would be difficult to monitor, because wet snow is mostly opaque to microwave (i.e., surface scattering) (Ulaby et al., 1986). The dielectric properties of snow are defined in two distinct phases: (1) dry snow, having a low dielectric constant related to the absence of liquid water within it; and (2) wet snow, having a high dielectric constant related to the presence of liquid water in the air-ice mixture (Langlois et al., 2007). On one hand, the impact of snow depth in the winter season when the snow is dry should be minimal on C-band backscatter. Studies showed that C-band was not able to capture snow depth variations lower than 1 m (Lievens et al., 2019; Hoppinen et al., 2024). Snow depth observations performed by our group between 2015 and 2019 showed that maximum snow depth maximum reach around 0.35 cm (±0.17 cm) (Meloche et al., 2022), which limit the possible impact of dry snow on backscattering signal. On the other hand, the C-band signal would then decrease with the presence of wet snow on the ground (Tsai et al., 2019), which decreases the penetration depth and increases the surface scattering of the air and snow interface as air temperatures rise above zero in the spring. In fact, in the SAR temporal series, the backscattering coefficient decreases when air temperature rises above 0 °C. However, that decrease is rather small because the backscatter of frozen tundra soil is already low. This signal intensity decrease is followed by a sharp increase of the order of 3–5 dB related to ground surface thawing. Hence, the increase in Tair during spring matches the start of snowmelt. However, our results show a strong agreement of this increase with soil temperature rising above 0 °C, suggesting that the sharp backscatter increase is an indication of wet snow turning into exposed thawed soil. Even though the C-band signal is affected by the presence of wet snow on the ground (Cohen et al., 2024), the fact that the wet snow signal and the thawed soil seem to happen almost at the same time, could indicate that the detection of soil surface thawing during the spring using C-band SAR observations is possible considering the strong agreements obtained with the soil surface temperature in the Arctic's tundra environment. Nevertheless, to fully understand the thawing signal in such environment, the use of a more comprehensive dataset that include snow depth, snow temperature and snow liquid water content for different conditions (years and area) would be necessary. Studies using ground-based radar could also help to better understand the processes leading to the thawing signal in spring for arctic environments (King et al., 2018; Rowlandson et al., 2018).