We used data from three automatic weather stations which are part of the network of the intercantonal measurement and information system (IMIS): SLF2 (1563 m a.s.l.), DAV5 (2315 m a.s.l.) and WFJ2 (2536 m a.s.l.), see Fig. . These stations provide automatic measurements every 30 min of: snow depth, air and surface temperature, wind speed and wind direction, relative humidity, reflected shortwave radiation, ground temperature, snow temperature at 25, 50 and 100 cm above the ground, and precipitation (unheated rain gauge) for SLF2 and WFJ2 . The variables indicated with “measured” in Table  are the ones we consider from the IMIS station measurements for this study.

To compare our data with relevant snowpack parameters, we used output data from the snow cover model SNOWPACK . This model was developed to support an operational avalanche warning service based on IMIS station data . SNOWPACK simulates the detailed stratigraphy of the snowpack based on meteorological input data. Specifically, the model uses local meteorological measurements, such as air temperature, snow depth or snow surface temperature, to predict the snow microstructure, density, temperature and liquid water content of the layers in the snowpack (parameters measured at IMIS stations mentioned above). The model produces a detailed description of snow properties including weak layer characterization , phase changes, water transport in snow using a simplified model or with full Richards Equations and water vapour transport in snow . SNOWPACK is used by several countries for their operational avalanche warning services , but also in fundamental and applied research studies. SNOWPACK is provided alongside its meteorological preprocessor MeteoIO and Graphical User Interface Inishell under an open source licence (LGPLv3). The whole dataset of reruns of the simulations performed for the operational avalanche warning service of Switzerland is available in .

In this study, we used the output data from the SNOWPACK version of December 2022 (git version: 98a23cd, ) in an operational setup. The SNOWPACK output includes both measured data at the station, providing its meteorological forcing, as well as modelled parameters and provides data every three hours from January 2018 to August 2021. Due to a sensor failure, no SNOWPACK simulations were available at the SLF2 station for the 2018–2019 season. Table  shows all data provided by the model, including abbreviations and the information on which are measured at the three reference sites and which ones are modelled.

We used SAR data from the European Space Agency's Sentinel-1 (S1) satellites between January 2018 and August 2021. The mission provides C-band data with a central frequency of 5.405 GHz in the two polarizations vertical-vertical (VV) and vertical-horizontal (VH). Over central Europe, S1 has long had an orbital repeat cycle of six days, though this is not always maintained due to the failure of S1B in December 2021 up to the launch of S1C .

We processed SAR data from the descending orbits 066 and 168 as well as the ascending orbits 015 and 117 (Fig. ) acquired in interferometric wide (IW) swath mode. The single-look-complex (SLC) data has a native pixel spacing of ∼ 2.3 m in slant range and ∼ 14.1 m in azimuth . Over Davos, the images were acquired at 05:34 (track 066), 05:27 (track 168), 17:15 (track 15), and 17:07 (track 117) UTC over the course of four days (in consecutive order; see Appendix Fig. ).

A digital terrain model (DTM) provided by swisstopo with a ground sampling distance 5×5 m was used in SAR data and considered to perform aspect dependency analysis in four aspect categories (N: 315–45°, E: 45–135°, S: 135–225°, W: 225–315°) .

Data from S1 is also the basis of the freely available Copernicus SAR Wet Snow (SWS) product. We downloaded the SWS products (version: 2025, ) for the period between January 2018 to July 2021 via the WEkEO platform (https://wekeo.copernicus.eu/, last access: 12 February 2026), the EU Copernicus reference service providing environmental data and virtual processing environments. The SWS dataset has a resolution of 60×60 m and is based on ground range detected (GRD) imagery, i.e. already focused SAR data, which was detected, multi-looked and projected onto an ellipsoidal model WGS84 . The data product offers classified information derived from S1 backscatter, specifically, wet snow (signal loss in comparison to stacked winter reference, class 110), dry snow, no snow or patchy snow (class 125). The product also includes pixel classes that are masked due to (i) unsuitable radar geometry, (ii) water, (iii) forest, (iv) urban area, (v) non-mountain area and no data available. We treated classes (i) to (v) as voids. We then mosaiced same day products (different footprints) to cover the entire site and when the footprints were overlapping for valid pixels (meaning class wet snow 110 or dry/no snow 125), used the latest assigned class on that day.

Our snow avalanche reference was the data from the Davos Avalanche Mapping Project (DAvalMap, detailed description in ). This inventory was created by the avalanche warning service SLF through systematic mapping of field observations within a perimeter covering roughly 180 km² (see Fig. ; ). The dataset includes information on the release date, avalanche type (wet or dry snow avalanche) and the elevation zone at release. Small avalanches (50 m for slab and glide-snow avalanches, 100 m for loose snow avalanches) were generally not recorded in the catalogue, thereby setting a minimum size of avalanches within the data set.

We defined avalanche conditions to be dominated by wet or dry snow avalanches, when the majority was within the corresponding category (priority to wet avalanches, when equal). In the season 2019–2020, the size restriction of the recorded avalanches was handled less strictly, also including smaller loose snow avalanches, which would have been excluded in the other years. The number of avalanches by type and season used in this work is summarised in Table  and the used information as a subset of the database can be found in the Supplement.

To create LRW composites from S1 data, we first performed a radiometric terrain correction (RTC; using the terrain-flattened γ0-convention, ) on the S1 acquisitions and then combined the data into LRW composites. These composites were calculated by applying a weighted function to the RTC images based on the local incidence angle from at least two different flight tracks (in our site we had a maximum of two descending and two ascending tracks). Weighting was applied according to the local spatial resolution in the corresponding flight track . We masked out areas that lied in radar shadow in all tracks, following the processing chain by and . In cases where information was unavailable in some tracks, i.e. lying in radar shadow, only the available orbits were considered as in. RTC values below -30 dB were classified as noise below the sensor sensitivity threshold and were therefore not considered in the LRW calculation .

Images with less than 75 % of the acquired scene containing valid pixels were discarded. This resulted in 444 LRW composites: 186 entailing all tracks, 29 three-track composites, and 6 two-track composites per polarisation. Two additional dates had only one track available therefore did not improve over RTC level. The maximum timespan covered by a composite was less than 84 h.

To generate wet snow maps we applied the common approach of comparing acquisition scenes to a no-snow or dry snow reference and identified a signal loss of 2 dB as an indication for the presence of wet snow . No standard approach for the selection of a reference image has yet been established, leaving space for potential differences in results . used a single image from summer, where the least amount of (melting) snow was assumed to be present in the scene. The revised version of the SWS product uses the mean of the winter months (December, January, February) as a dry snow reference. However, when working with large elevation ranges – as is the case in our study area – even images from December, January and February can contain wet snow and “contaminate” the image. Conversely, a summer image is also likely to contain wet snow at high elevations and more likely to be affected by changes in seasonality across the imagery (e.g. vegetation related).

Therefore, rather than relying on a single image, or a purely calendar-based winter-month approach, we calculated a median backscatter value per pixel over the entire available time series (indicated in black in Fig. ). With this approach, we account for the challenge of finding a suitable reference image, which becomes less accurate the longer the time span between the reference and acquisition becomes.

We then mapped wet snow on a pixel-by-pixel basis wherever the relative backscatter dropped below the 2 dB threshold at both polarisations. To increase robustness of the final product, we determined wet snow to be present only if it was identified in both polarisations.

To evaluate the factors influencing SAR backscatter time series, we performed correlation analyses between various measured and modelled parameters over time (see Fig. , Appendices and ). Since single pixel analyses are prone to radar noise (e.g. speckle) and uncertainties, we tested the correlation of values recorded on a 3×3 pixel window (indicated in Fig. C). We used the time closest to S1 acquisition on the corresponding day, where both IMIS and SNOWPACK data was available: namely 06:00 to match descending SAR image acquisitions and 18:00 to match the ascending acquisitions (UTC+1).

For the correlation analyses we calculated Pearson's correlation coefficient and associated RMSE values, along with Spearman's rank coefficient. Pearson’s was used to assess the potential strength and direction of linear relationships, while Spearman’s captured monotonic trends between variables. To minimise the correlation coefficient bias due to the effect of over-represented numerical values (e.g. zero is used by SNOWPACK when void), correlations were only calculated during the snow cover period (defined as snow height above zero within SNOWPACK). The significance level for both tests was set to p-values below 0.05.

We assessed the time difference between the onset of the main melting season and its detection in the S1 LRW data to better quantify potential uncertainties introduced by the temporal binning inherent to using the multi-track LRW data. We defined peak melt onset as the first date on which the 3 d median runoff exceeded the runoff threshold (set to 1 kg m⁻², from virtual lysimeter data in SNOWPACK). This date when the runoff exceeds the threshold, was then compared to the current or first following composite time on which wet snow was detected in the LRW time series. For example, in spring 2018 the 3 d median runoff first exceeded the threshold on 25 May; this date lies within the concurrent multi-track LRW composite (22–25 May 2018), in which wet snow was already detected (24 May 2018), corresponding to an offset of −1 d.

To quantitatively compare the S1 LRW wet snow product with the SWS dataset by taking all parts of the confusion matrix into account, we calculated Matthews Correlation Coefficient (MCC) between the two binary products over all years (similarly applied by e.g. ). Further, evident false positive wet snow detections at WFJ2 were separately quantified by calculating the frequency with which wet snow was detected, by either the multi-track LRW approach or the SWS product, when the measured snow depth at the station was zero.

After mapping wet snow across the study site, its relative occurrence to dry/no snow was quantified by assessing elevation-dependent melting over time across elevation bands. Elevation bands were defined in 100 m increments from 1500 to 3000 m a.s.l. Data was binned into six day increments, starting with the first LRW per season. To focus on areas that are potential avalanche release zones we excluded areas with slopes below 28° . For homogeneity in the intercomparison with the coarser resolved SWS dataset, the latter has been oversampled using nearest neighbour interpolation onto the 5×5 m grid.

For each elevation band, we computed the wet snow ratio as: 1Rwet(z)=Nwet(z)Ntotal(z), where Rwet(z) is the wet snow ratio at elevation band z, Nwet(z) is the number of wet snow pixels, and Ntotal(z) is the total number of non-void pixels in that elevation band.

By comparing the wet snow detection from S1 with the event-based snow avalanche catalogue (DAvalMap), we retrieved information on the wet snow avalanche preconditions. After discarding duplicates and data collected outside our study area from the catalogue, the avalanches were matched to the wet snow ratio (see Eq. 1) according to the elevation of the release zone and the date of release. Temporal matching to the LRW composites was performed to simulate an operational scenario, where avalanche occurrence is evaluated based on the most recent information on snowpack wetting. Avalanches were therefore assigned to the latest available LRW composite preceding their occurrence (illustrated in appendix Fig. ).

To assess the potential for this wet snow ratio to indicate the transition from dry to wet snow avalanche dominated conditions, we computed histograms and Gaussian distributions of recorded dry and wet snow avalanches in relation to the wet snow ratio. To enable a performance comparison, this analysis was performed for both the LRW-based dataset presented here and the downloaded SWS product. We then related the dominant avalanche types to the wet snow ratio and determined the thresholds at which conditions transition from dry-dominated to wet-dominated and finally to wet snow avalanche only. The wet snow avalanche only was defined as the threshold of wet snow ratio above which wet snow ratio there were never more dry snow avalanches recorded than wet snow avalanches.

Because the SWS product does not require temporal increments as the LRW product, we were able to match the avalanches from the catalogue to near-daily increments. Avalanches that occurred when no data was provided were carried over to the next previously available date (see Appendix ). We then extracted the thresholds from the distribution of wet snow ratios and the dominated avalanche type as before. This allowed us to compare and assess the feasibility between the two products for practice.

In general, the LRW backscatter signal fell below the wet snow threshold slightly after runoff commenced and measured snow height began to decrease (Fig. ). Wet snow detected by the S1 dual-polarisation approach preceded the rise of runoff above 1 kg m⁻² by no more than one LRW acquisition interval (6 d) (Fig. : -1, +1.6, +5.1, +5.6 d delay for 2017–2018, 2018–2019, 2019–2020, and 2020–2021, respectively). Every season, backscatter in both polarisations (VV & VH) fell 3 dB or more below the median in late spring/early summer, coincident with the onset of snow melt at high elevations and temperatures around the melting point (see Fig. E). Sporadic decreases of backscatter during periods where SNOWPACK data did not suggest the presence of wet snow were observed in both polarisations in all years. However, for all such spurious events the drop only occurred in one polarisation but not in the other. Overall, the combination of VV and VH polarisation (indicated by overlaid shadings in Fig. D and E) shortened the time-span of detected wet snow and resulted in somewhat delayed detections (especially in season 2020–2021), but minimised the potentially false detections during periods when snow melt likely did not occur.

Statistical comparisons between the RTC level backscatter time series and the station measurements at all stations revealed a negative correlation between S1 backscatter and modelled liquid water content as well as measured runoff (January 2018 to August 2021, Fig. ). For runoff, the correlations ranged from -0.13 (VV; SLF2) to -0.49 (VH; WFJ2) in Spearman's rank and from -0.28 to -0.26 for Pearson's correlation (averaged over the tracks per station per polarisation (Fig. , and Appendices and )). This and the slight correlation found with the liquid water content (0.15 in VH at SLF2 -0.49 in VV at WFJ2 for Spearman's and -0.01 in VV at SLF2 to -0.4 at WFJ2 for Pearson's) indicates the expected sensitivity of S1 to water within the snowpack .

Among the SNOWPACK variables, Spearman rank correlations (restricted to snow-covered conditions) revealed a strong positive relationship between runoff and liquid water content (0.74). Liquid water content showed a weak positive correlation with snow height (0.26), while no significant relationship was found between runoff and snow height (−0.03). This indicates that runoff is primarily controlled by the availability of liquid water within the snowpack rather than by snow depth itself. While the stations at lower elevations showed a much lower signal stability (Appendices and ), similar tendencies of correlations were found.

Comparing our detection of wet snow to the SWS product is not trivial due to the differences in spatial and temporal resolution and the lack of a spatially continuous reference data. However, the overall MCC value of 0.90 between SWS and LRW at WFJ2 station suggested a general agreement on the main melting phases (Fig. C for SWS and shadings in Fig. D and E for detected from LRW product), with yearly MCC values varying between 0.87 (2017–2018 and 2019–2020) and 0.94 (2018–2019)).

Overall, the SWS layer provides spatial coverage of 30 %–60 % of the study area (median: 48.6 %), whereas the multi-track LRW approach achieves 94 %–100 % coverage (median: 99.9 %). The reduced coverage of the SWS product is primarily attributable to masking strategies applied to minimise artifacts arising from radar geometry effects, from land cover types such as forest and water bodies (visual in Appendix ).

At the WFJ2 station, the SWS product resulted in 17 false positive wet snow detections across the time series, instances where wet snow was indicated in the absence of snow cover, whereas the dual-polarised multi-track LRW approach produced none.

Across the whole study region, the wet snow ratio shows some distinct seasonal patterns that are consistent across years. At the beginning of the winter season, the elevation of areas with moderate wet snow ratios decreases, indicating the decrease of the snowline and a gradual transition to wide-spread dry snow. Once the melt season starts, there is both an overall rise of wet snow ratio as well as a rise in elevation of areas with high wet snow ratios (Fig. ). Across the three analysed seasons, the 2018–2019 and 2020–2021 spring seasons reached higher wet snow ratios than 2019–2020, reaching 70 %–80 % in mid or late March, whereas 2019–2020 peaked at about 60 %. Both these two years saw warm phases in February with wetting occurring quite early compared to other years. In 2020–2021, elevated wet snow ratios (> 30 %) appeared earlier in the season (November–February) between 1500 and 2200 m a.s.l., followed by a second peak of up to 70 %–80 % in mid-May.

Figure A shows a histogram of avalanche types as a function of wet snow ratio. Figure B then displays the wet snow ratio against the proportion of wet snow avalanches among all avalanches. The transition between dry snow avalanche dominated conditions and wet snow-dominated conditions was found at a wet snow ratio of 0.17. The transition to exclusively wet snow avalanches occurred at a ratio of 0.35 when extracted from the LRW data (Fig. A–B).

When parsing the data by aspect ranges (Appendix ), we found the highest sensitivity (lowest wet snow ratio threshold for transition to wet snow-dominated or only-wet snow) in south facing slopes (0.07–0.21), similar in east and west facing slopes (East: 0.23–0.39 and West: 0.20–0.37) and an immediate transition from dry dominated to wet snow avalanches only for north facing slopes at 0.13. The narrower transitioning phase in south and north facing slopes than in east and west facing slopes, indicates higher sensitivity in those aspect directions.

For the SWS product, the general transition between the avalanche release conditions was found to be less distinct (Fig. C–D). Although dry snow avalanches were more prevalent at lower wet snow ratios (below approximately 0.5), periods characterised exclusively by dry snow avalanche activity could not be distinguished. The transition to a state dominated solely by wet snow avalanches occurred at a higher threshold, when compared to the multitrack LRW-based product, i.e. 0.81 (Fig. C–D).

We compared the produced wet snow time series to weather station data and modelled snow parameters extracted from SNOWPACK, using statistical analyses to assess the sensitivity of the radar data to changes in the snowpack. In the following, we first discuss the findings and limitations related to measured variables from the weather stations, then evaluate the relationships obtained from modelled SNOWPACK variables, and finally combine both perspectives to interpret the radar signal response.

Among the measured variables, snow height (HS) in combination with snow surface temperatures (TSS) provided an indication of the seasonal evolution of the snowpack and the onset of melt. Periods with non-zero snow height together with temperatures around the freezing point can be interpreted as conditions favourable for snow wetting. However, the in-situ stations directly measure physical quantities such as snow depth and snow surface temperature, whereas the SAR-derived wet snow product infers the presence of liquid water indirectly from changes in backscatter intensity: the two approaches therefore do not measure the same quantity, which explains the rather low correlation statistics between S1 backscatter and measured variables at WFJ2. While single-point measurements cannot fully capture the spatial variability of alpine terrain, the temporal agreement found between the observed snow depth, snow surface temperature, and the S1-detected wet snow periods nonetheless supports the feasibility of our approach.

However, the relationship between these variables and snow wetness remains indirect, explaining the rather low correlation statistics between S1 backscatter at WFJ2 and measured snow surface temperature. While single-point measurements cannot fully capture the spatial variability of alpine terrain, the temporal agreement found between the observed snow depth, snow surface temperature, and the S1-detected wet snow periods nonetheless supports the feasibility of our approach.

In contrast to the measured parameters, certain parameters extracted from SNOWPACK showed stronger relationships with the SAR backscatter signal. Across all tests, liquid water content and snow runoff exhibited the strongest correlations with the SAR derived snow wetness (Fig. and Appendices and ). This supports the assumption that decreases in SAR backscatter are linked to liquid water within the snowpack. Already at a liquid water content of 5 %, the water acts significantly as an absorbent and specular reflector to C-band frequencies, resulting in low backscatter values .

Combining the measured and modelled data allowed us to define the most likely onset of the melting season at the station and to subsequently determine the time difference between that and the wet snow detection. Interestingly, we observed a lag between the decrease in snow depth, the increase in modelled liquid water content, and the detection of wet snow from the LRWs. By contrast, the onset of runoff was matched without time lag (see Fig. ). This can be explained by the fact that runoff onset is itself delayed relative to the initial moistening and ripening phases, as the first meltwater is retained within the pore space of the snowpack before drainage occurs . The S1-derived wet snow signal therefore likely reflects a more advanced stage of the wetting process, by which time runoff onset is also imminent. Compared to , who identified all three melt phases (i.e. moistening, ripening, and runoff) from multi-temporal SAR backscatter, this suggests that our approach may not fully capture the early moistening phase. This is consistent with , who found liquid water content to be the dominant control on S1 backscatter at the onset of the melting season, with surface roughness becoming increasingly influential thereafter. Nevertheless, the detection of runoff onset may still provide valuable information that can support the early warning of wet snow avalanches.

In both the correlation analysis and the time series generation, data from the WFJ2 station outperformed the other two stations, but we believe that this is largely a consequence of local conditions and the effect of less pronounced snow conditions. The DAV5 station station (Appendix Fig. ) is on a north facing slope and is surrounded by avalanche protection structures on three sides. These structures very likely affect the radar backscatter. At SLF2, the snowmelt in spring 2020 is not captured at all. This station is situated at the lowest elevation of the three, so seasonal transitions are expected to be less pronounced and less intense melt events are more likely to be missed. The shallower snowpack and less pronounced winter conditions around SLF2 station is making this station-location most prone to the influence of surface roughness, which governs the S1 signal during patchy snow conditions . Also at SLF2, proximity to infrastructure (> 15 m) and temporary constructions may have caused signal interference. An example of this is likely the increased backscatter observed early in the year coinciding with the World Economic Forum, plausibly linked to temporary metal constructions for security around the nearby helicopter landing zone (Appendix Fig. , especially in 2018). The comparatively poor performance at the DAV5 and SLF2 stations highlights challenges faced by the product under specific local conditions. However, this is expected to have limited implications for the general applicability of the product to avalanche release assessment, as release zones are typically situated at higher, more exposed elevations, where conditions more closely resemble those at WFJ2 station, at which the method performed well.

The mapping of wet snow from S1 SAR data involves key methodological choices for which widely-used standards do not yet exist. This is particularly true for generating the reference image, defining the wet snow threshold, and the handling of the different polarisations. Each of these can have an influence on the final product.

No standard approach for generating a reference image is established. The first version of the SWS layer was based on a reference image calculated as the median from a stack of summer acquisitions assumed no snow;. Other studies as well as the current SWS dataset calculate the mean over dry or no snow images e.g., or use a single reference image featuring either dry or no snow conditions e.g.,. Increasing temporal baselines between the image of interest and the reference image, results in greater errors due to land use and land cover changes over time. In the application of wet snow avalanche forecasting and nowcasting, time efficiency and accuracy is of major importance. By choosing an “all season” median, we have wet snow contamination in the reference; however the results of our study indicate applicability of the method (see Fig. ) and follow the process applied in . The advantage of reference image generation in that way is a high level of automation, no arbitrary reference selection and a simplified update-ability.

Once relative changes are computed, a threshold below which wet snow is present needs to be set. This threshold is usually between 2 and 3 dB signal loss relative to a reference image or all-season median (this work) and is based on histogram analysis e.g.,. Adaptive and dynamic thresholds based on the underlying land cover class or other additional datasets have recently been suggested. However, we used the established and widely applied threshold of 2 dB for S1 data results indicated in Fig. . Although an adaptive backscatter analysis sounds promising, there were no widely available and regularly updated land cover products with suitable spatial resolution available in our case, and the most robust algorithm minimises requirements on the input data.

The selection of polarisation modes from S1 data has varied between wet snow observation studies. While e.g. , , , used VV polarised data, used VH polarised and , , , used a dual-polarised approach. In our results, the correlation values showed similar trends between the different polarisation modes (Fig. and Appendices and ). In contrast to we performed the analysis on separate polarisation states before combining the wet snow masks, this mitigated the issue of using a thresholding approach developed by on a dB scale on a ratio value as implemented by . The strongest divergence between the wet snow ratio extracted from VV and VH polarisation was observed in late summer/early autumn, whereby the VH polarisation indicated higher ratio values than VV. This might be related to a remaining influence of seasonal activity of vegetation (e.g. by occurrence of shrub forest), which is known to influence backscatter signal or by the ripening which found to have a depolarizing effect on the signal. The time series showed similar results for both polarisation modes. The applied combination of both polarisations resulted in fewer misclassifications and therefore higher robustness.

Due to their complex topography, high mountain environments pose several well-known challenges for monitoring with SAR data, and minimising these effects is crucial for operational applications.

While the multi-track LRW approach and commonly available S1-based snow products such as the SWS or Wet/Dry Snow layers showed strong overall seasonal agreement (MCC = 0.9), a direct quantitative comparison of both products against observed runoff carried too much uncertainty to be conclusive. The combined arbitrariness in defining melt season boundaries in the noisier SWS time series, the sensitivity of the chosen runoff thresholds, and the need to account for interruptions such as the runoff pause during the 2018–2019 season, rendered such a comparison insufficiently stable. While the overall agreement is strong, the multi-track LRW approach addresses more subtle differences in spatial detail and coverage.

Instead of the commonly used GRD data, the usage of SLC data in combination with a high resolution DEM and LRW compositing allows for a higher spatial resolution, resulting in more detailed representation of snow wetness patterns. Multi-track LRW allows the mitigation of typical SAR geometry effects such as layover and shadowing as well as the minimisation of noise in the dataset. By combining acquisitions from different viewing geometries, noise can be filtered and areas that would otherwise lack data can be partially recovered, increasing the spatial coverage of the resulting product. This is especially relevant in steep terrain, where single-track acquisitions often lead to systematic data gaps. This increased data availability also stabilises the derived wet snow ratios, as a larger number of valid pixels can be included in the calculations, reducing the influence of outliers. In addition to spatial improvements, the use of LRW composites enhances the temporal robustness of the time series. By integrating multiple acquisitions, the impact of missing individual scenes is reduced, allowing for more consistent monitoring of snow conditions over time. So while SWS offer a grid sampling of 60 m , which may not sufficiently resolve the variability of mountainous topography, our S1 LRW approach offers more spatial detail.

These improvements are particularly relevant for applications such as avalanche forecasting, where both spatial detail and temporal consistency are critical, especially for capturing the onset of snow wetting across different slope aspects . With the community increasingly moving towards the provision of large-scale LRW datasets such as already available from e.g. , making use of this data will become more efficient and operational applicability more feasible.

To assess the applicability of SAR-based wet snow detection for the retrieval of the avalanche release conditions, we compared the wet snow ratios derived from both the multi-track LRW product and the standard SWS to the DAvalMap wet snow avalanche catalogue. This approach provides an indication of how well remotely sensed wet snow conditions may provide indication on avalanche type occurrence.

The available DAvalMap catalogue contains field-based observations of avalanches and their type (wet, dry, unknown). Even though this dataset is not without limits, it provides a valuable opportunity to relate avalanche activity to SAR-derived snow wetting. In the following, we discuss a few aspects which must be considered when interpreting the wet snow ratio in relation to this dataset. The catalogue does not permit a full evaluation of the detection performance beyond the observed avalanche sample. In other words, a lack of record does not necessarily mean that wet snow avalanche conditions were not present. Also, the DAvalMap perimeter is smaller than the study area used for LRW production, meaning that avalanches outside of this region were not systematically recorded. As a result, elevation bins above approximately 2800 m a.s.l. contain little to no documented avalanche activity.

All avalanches included in the dataset were classified into wet, dry or unknown-type snow avalanches, where the classification pertained to the conditions in the starting zone. However, this classification is challenging from a distance, and some events may have been misclassified, particularly in cases when avalanches started as dry snow avalanches, but developed into wet snow avalanches during their descent . Lastly, the inventory is based on in-situ observations that depend on good visibility and are biased towards days with lower avalanche risk, as observations are less likely to be made during hazardous conditions . This can lead to uncertainties in the assigned release dates of up to several days (avalanches are reported when they have been observed instead of when they have been triggered . However, this issue is less pronounced for wet snow avalanches occurring on sunny days during warming. Given the applied six d temporal resolution of our product, the impact of these timing uncertainties is expected to be limited. However, such misclassifications are likely to have a stronger effect on the daily temporal matching to the SWS product. Even with these observational shortcomings, this dataset is one of the most comprehensive avalanche activity datasets .

The initial clusters of wet snow avalanches in spring 2019 and 2020 did not coincide with a pronounced increase in the wet snow ratio across those two seasons (see Fig. ). These events were the most susceptible to misclassification in the DAvalMap dataset, given their higher potential at the start of the melting season to initiate as dry releases before transitioning into wet snow avalanches. Furthermore, they are the most likely to occur during short-lived wetting phases that may go undetected given the temporal resolution of S1 imagery (shown in daily grid and unavailable images in Fig. ). Rainfall showed no significant dependency with wet snow avalanche occurrence over the time series (r<0.2; mean daily rainfall: 17 kg m⁻²; Appendix Fig. ), indicating that rain-on-snow events had limited influence on wet snow avalanche releases in this dataset, excluding this as potential factor. This underscores the need for improved temporal coverage of the SAR mission to enhance the timing accuracy of remotely sensed snowpack transitions.

All these effects impact the threshold extraction which marks the transition from dry to wet snow avalanche release conditions. Additionally, we found an aspect dependency on the sensitivity of the thresholds (Appendix Fig. ). This can be attributed for one to (a) the difference in melting, since the north-facing slopes are typically later in melting and then more intensive (visible in the elevation dependent melting diagrams, Appendix Fig. ), but also (b) the inherent flight geometry of the sensor, which covers the south facing slopes the best, and offers least coverage for north facing slopes.

Ideally, a clear threshold of the wet snow ratio would identify the transition between dry and wet snow avalanche release conditions, indicating that the ratio provides a robust measure to differentiate between these conditions. To mark the uncertainty range, we also provide a threshold for wet-dominated conditions, where wet snow avalanches prevail but dry snow avalanches can still occur. However, the use of the SWS product showed increased noise and lower reliability with a transition to wet-only conditions at a very late stage. This highlights the dependency of the threshold on the underlying radar data processing. Consequently, the thresholds derived in this study should be interpreted as a range that needs to be selected according to the specific product used in practical applications.

We encourage future research to make use of not only multi-platform but also multi-frequency data such as a combination of X- and L-band with e.g. the recently launched NISAR mission . This would allow the capture of different stages of the melting due to differences in sensitivity to water presence with varying wavelengths, thereby enabling better estimates of the liquid water content and not only the presence of liquid water.