This study uses snow depth data from two permanent low-cost terrestrial laser scanning stations (B1 and B2) at the test site Braemabuel, in the Dischma valley near Davos, Switzerland (see Fig. ). The observed slope represents the release area of the Wildi avalanche path, is northeast-facing, and ranges from approximately 2150 to 2250 m a.s.l., with inclinations between 30 and 45°. The terrain is characterized by heterogeneous subalpine vegetation, including grass, low shrubs, and a few small, scattered trees, as well as rocky patches.

At each of the laser scanning stations, we operate a Livox Avia lidar sensor. Jointly, these sensors cover the typical release area of the Wildi-avalanche. The lidar sensors measure at a wavelength of 905 nm which is optimal for sensing snow . We installed the sensors on 23 November 2023 and configured them to scan every hour. The specified maximum measurement range of the lidar sensors is 450 m, but in our region of interest we achieve around 150 m, due to the suboptimal measurement geometry. The data cover an area of about 14500m2, with an average resolution of 0.1 m. Accounting for measurement accuracy, registration errors and rasterization/interpolation errors the data has an accuracy of 0.1–0.3 m. We processed the data to a gridded raster data of 0.5 m resolution. A more detailed description of the sensors, setup, data processing, and data quality is given in .

Across the approaches for modelling snow depth distribution in our study, we use wind direction, wind speed, precipitation amount and snow density data. To ensure comparability, robustness and applicability of our approaches, we primarily rely on data from numerical weather prediction models, which offer greater consistency and coverage compared to station-based measurements. Specifically, we use the Kenda-CH1 analysis data of the ICOsahedral Non-hydrostatic (ICON) modelling framework as meteorological input for all models in the main analysis, including precipitation as well as wind speed and wind direction, available on a 1.0 km grid at 10 m above ground in Switzerland. We use the hourly values of the grid cell, whose centre is closest to our study site. This choice was made intentionally to ensure a consistent basis for comparing the model approaches and to evaluate their applicability in settings where local weather stations are not available. To our knowledge, there has not been a specific validation of the ICON-derived HNWs. Given the improved score for precipitation of ICON over COSMO , we expect that the bias in snowfall amounts is reduced compared to the -19 % found for COSMO .

In addition to the main ICON-based analysis, we use wind data from local weather stations to assess the sensitivity of model performance to the choice of wind input. Close to our test area there are three weather stations that measure, among other parameters, wind speed and wind direction: Our TLS measurement stations B1, B2 (see Fig. c) and the wind station HFY on top of the close-by peak (500 m horizontal, and 200 m vertical distance, see Fig. b). The data of the latter is accessible through the platform holfuy .

For the conversion between snow water equivalent and snow depth, we additionally require the density of new snow. The closest measurement station at similar altitude, where all information is available to derive snow density, is at Weissfluhjoch (WFJ), located about 5.5 km from our test site Braemabuel (see Fig. b). At WFJ researchers measure the height and water equivalent of new snow (HN and HNW) on a daily basis, from which we derive the density of new snow.

The terrain data used in this study are digital surface models (DSMs), derived from photogrammetric drone flights with a Wingtra One Gen II fixed-wing UAV using a Sony RX1R II 42 megapixel camera and Post Processing Kinematics (PPK) postitioning. We processed the photogrammetric data as described in , , and using the software Agisoft Metashape Professional, Version 1.6.5. We use the DSM acquired on 18 October 2022, as snow-off reference, and one from 19 December 2023, which is a DSM of the snow-covered terrain. The DSMs have a spatial resolution of 0.1 m, and we downsample them to 0.5 m to meet the resolution of the TLS data. The accuracy of the photogrammetrically derived DSMs is not explicitly assessed in this study. Based on previous studies using comparable photogrammetric workflows, we expect a vertical accuracy in the order of 0.1 m depending on illumination, surface texture, and image geometry .

The TLS data available to date cover the the winter seasons 2023/2024–2025/2026. In this study we focus on the winter 2023/2024, where we identified 15 potentially interesting events with a significant amount of new snow (defined as more than 0.1 m from start to end of the event). For the selection of TLS measurement epochs, we use the number of classified ground points per scan as proxy for snowfall in the monitored area, with the assumption that less ground points, mean heavier snowfall (some epochs are missing from station B2 due to sensor outages, and some from station B1 due to the scanner being snow covered). We were looking for rather short events with TLS measurements available directly before and after, so we can assume that processes other than snowfall and wind-driven redistribution, e.g., settlement or melting, have little influence on snow depth changes during the events. Among those, we selected the events without avalanche activity but with continuous availability of locally collected meteorological data (wind sensor not frozen).

With all these criteria applied, three events stand out. We indicate them as E1, E2, and E3, and analyse them, herein. They are E1: 24 November 2023, 14:00 UTC–26 November 2023, 08:00 UTC, E2: 21 December 2023, 09:00 UTC–23 December 2023, 03:00 UTC and E3: 22 February 2024, 22:00 UTC–23 February 2024, 16:00 UTC. They are distributed over the winter season and show a variety in average wind speed. With the latter we expect the presence of different processes per event (mainly preferential deposition and snow redistribution), for which we want to specifically evaluate the models. Figure  gives an overview of the selected events.

Due to the varying wind speeds and directions, we further divide the events into sub-events. For this, we rely on the availability of suitable lidar data (i.e. epochs which cover more than half of the test site). According to the available lidar data and best possible consideration of wind speeds and wind direction, we use the TLS acquisitions of 25 November 2023, 02:00 and 10:00 UTC during E1 and 21 December 2023, 19:00 UTC, 22 December 2023, 03:00 and 09:00 UTC during E2 (see Fig. ). We do not have suitable lidar scans to further sub-divide E3, therefore we use only the full event herein.

During the event in November 2023 (E1) the ICON model outputs an accumulated precipitation of 27.9 mm and average wind speeds of 5.7 m s⁻¹, mainly from north-west. At measurement station B1, which is located in the slope of interest, we recorded an average wind speed of 4.0 m s⁻¹ and gusts of maximum 11.2 m s⁻¹. Station B2, which is on the local ridge, and HFY, located at the local peak, (see Fig. ), measured higher values (Table ). The threshold for snow transport is 4–11 m s⁻¹ for dry snow, with an average of 7.5 m s⁻¹ for fresh snow . However, this threshold depends on many factors beyond wind speed, in particular on snow surface properties. Based on the locally measured wind speeds and evaluation of the differences of the TLS scans, we assume that during the event a mix of preferential deposition and snow redistribution lead to the variability of new snow depths in E1. With further dividing the event into sub-events, we assume the first hours where a mix of preferential deposition and redistribution, followed by a few hours of snow redistribution and then a period of preferential deposition in the second half of the event (see Table ). In the event in December 2023 (E2) we retrieve an accumulated precipitation of 36.6 mm from the ICON model and an average wind speed of 9.1 m s⁻¹, from west- and north-west directions. During the first hours of the event we recorded low wind speeds of 1–3 m s⁻¹ (B1), with rather warm temperatures (while below zero). After that, the wind speeds picked up and there were a few hours with high wind speeds, on average 9.3 m s⁻¹ and gusts up to 18.8 m s⁻¹ (referring to station B1). From this we assume that there was a rather homogeneous accumulation with preferential deposition of new snow in the beginning (although due to the warmer temperatures most likely including other processes like settlement as well), which was afterwards transported and redistributed, at the same time combined with more new snow. We classify this event as well as a mixture of mainly preferential deposition and snow redistribution, with sub-events of mixed processes, redistribution and preferential deposition (also see Table ). During the event in February 2024 (E3), the accumulated precipitation was 34.6 mm and the average wind speed reached 3.3 m s⁻¹ with great variation in wind directions, according to the ICON model. As the wind speeds recorded within the slope (on average 3.4 m s⁻¹) are below the reported threshold for snow transport, we assume that there was no snow redistribution, but mainly preferential deposition of the new snow.

The topographic position index TPI describes the relative elevation of a point, compared to the mean elevation of the surrounding terrain, in an application-specific radius. This allows the use of TPI to indicate slope positions and landforms, such as depressions, plains or ridges and hilltops. Depending on the selected radius, different scales of topography are described .

The TPI is calculated as follows: 1TPIjR=zj-z¯jR where j is the index of the cell Xj(xj,yj,zj) for which the TPI is evaluated, and z¯jR is the average elevation of all points within a circle of radius R around point Xj. For simplicity, we will subsequently omit the superscript R, herein. In our implementation we exclude a margin near the borders and represent missing values with NaN. TPI is only computed where the fraction of NaN values within the radius R is below 10 %, an empirically chosen threshold. For a DSM resolution of 0.5 m and R=10m, this yields an effective border exclusion of approximately 7 m (see Fig. ).

Positive values for TPI indicate ridges and hilltops, whereas negative values imply landforms such as gullies and depressions, suggesting a negative correlation with snow depths. Figure  (top left) shows the resulting TPI, using a radius of 10 m. In Sect.  we describe our derivation of change in snow depth ΔHS from TPI.

Snow depth distribution is dependent not only on the local terrain geometry, but also on the wind direction. Snow gets eroded on the windward side, and then deposits on the leeward side of a ridge. developed an indicator, whether a cell is wind exposed or wind sheltered, by evaluating the terrain elevation of a cell in relation to terrain elevations of all cells in a certain wind direction and up to a user-specified distance. Whether the elevation along the assessed wind direction is higher or lower than the elevation at the point of interest provides a basic approximation of whether it is exposed or sheltered from the wind in this direction, which we assume to be positively correlated to snow depths (see Sect.  for derivation of ΔHS from wind shelter index Sx).

The windshelter index Sx of a cell Xj(xj,yj,zj) describes the maximum upward slope of all cells Xd(xd,yd,zd), along an azimuth angle A, and a maximum search distance D with: 2SxjA,D=max⁡arctan⁡zd-zj(xd-xj)2+(yd-yj)2, When taking into account a number of azimuths nA, resulting from incrementing with a defined step size from A1 to A2, Sx can be extended to Sx‾ by taking the mean of all Sx per cell j with: 3Sx‾jA,D|A1A2=1nA∑A=A1A2SxjA,D, We take the approach a step further and instead of equally weighting all considered wind directions in the above Sx‾, we compute a wind rose and calculate a weighted mean Sx‾w, according to the relative occurrence of the observed wind directions within azimuth bins of a specified angle width. This ensures that dominant wind directions get more weight and less frequent wind directions are taken into account, but are not over emphasised. In the remainder of the paper we use Sx‾w for our analysis, but use the notation Sx, for simplicity. We calculate Sx in hourly time steps, scale each time step using the corresponding mean water equivalent of new snow HNW‾ from ICON, apply weighted means from 5° wind direction angular intervals, and then accumulate the results to the time intervals, as defined with the events. In Fig.  (top center) we show the Sx angles for an example epoch of 25 November 2023, 21:00–22:00 UTC, where the average wind direction was from 315°.

developed a statistical snowfall downscaling scheme based on parameters derived from a comprehensive 30 m new snow database, generated using 3D atmospheric model wind simulations over synthetic topographies spanning a broad range of slope characteristics, together with a snow transport model to compute preferential deposition. This dataset describes how snowfall is non-uniformly deposited on the ground in mountainous terrain due to interactions between near-surface wind, snow particles, and topography. Processes such as snow erosion, saltation, drifting snow sublimation, and snow melt or settlement were intentionally suppressed and are therefore not represented in the PD model. The downscaled snowfall (PD), or water equivalent of new snow j HNWjPD per cell is calculated by scaling HNW‾ by a local downscaling factor Xdsc,j: 4HNWjPD=HNW‾Xdsc,j Xdsc,j describes primarily the correlation of preferentially deposited snowfall with local vertical wind speed wj (with updraft on the windward side of mountains and downdraft on the lee side), in line with earlier studies reporting correlations between vertical wind speed and preferential deposition . further showed that preferential snowfall deposition scales with the local slope parameter μj, which is calculated for each cell j, from its partial derivates ∂xz and ∂yz (μj=((∂xz)2+(∂yz)2)/2). Xdsc,j is 5Xdsc,j(wj,μj)=erfc(awj(wj+|wj|))b⋅(1-cwj+dwj3)(1+eμjf), with the following constant values for a to f: a=0.4825, b=0.03418, c=0.592003, d=0.004452, e=0.24714 and f=2.24223. Xdsc,j requires local vertical wind speed w. When spatially fine-scale w is unavailable, as in our case, introduced a statistically downscaled vertical wind speed wdsc,j for use in Eq. () (the “aspect scheme”). wdsc,j is obtained by scaling the spatially averaged horizontal wind speed vh‾ with the local downscaling factor Ydsc,jaspect: 6wdsc,j=vh‾Ydsc,jaspect(Δψj,μj). Ydsc,jaspect is a function of μj and the local terrain aspect parameter Δψj, defined as the local terrain aspect angle ψj relative to the low-resolution wind direction wd. Ydsc,jaspect was derived from the 3D wind simulation dataset over synthetic topographies: 7Ydscaspect(Δψ,μ)=(a′-b′(Δψ)+c′erf(d′(Δψ))(e′+μf′), with constant parameters a′=-0.087122, b′=0.4788, c′=2.068, d′=0.6298, e′=-0.046577 and f′=0.72451. wdsc,j requires vh‾, which we do not have available. We use a subgrid parametrization for the spatial mean horizontal wind speed vhsgp, parametrizing the unresolved drag over topography by scaling low resolution horizontal wind speed vh with the subgrid sky view factor Fsky,sbg(L/ξ,μ) : 8vh‾=vhsgp=vhFsky,sbg(L/ξ,μ). Fsky,sbg(L/ξ,μ) is parameterized as in and relies on domain size L, domain mean slope parameter μ‾ and terrain correlation length ξ in the domain. For further information on the PD model we refer to .

We calculated the HNW distribution based on PD on the same snow-off DSM as TPI and Sx, and calculate each event in 1 h time steps. Due to the locations of the measurement stations B1 and B2 we omit the sky view factor Fsky,sbg(L/ξ,μ), when using these data as model input. Consistent with the PD model development database, grid cells with slopes exceeding 60° were excluded from the analysis and set to NaN. The final HNWs are the accumulation of the fine-scale precipitation per computation step. We convert HNW PD to HN PD using Eq. () adapted with HNWj,EPD. In Fig.  top right we show an example epoch of the terrain scaling parameter Xdsc,j (Eq. ), and scaled to HNW in the bottom row. For consistency with the TPI- and Sx-based results, we use the notation ΔHSW and ΔHS in the following, although the PD model is intended to describe only the accumulation of new snow and therefore corresponds more directly to HNW, i.e. positive ΔHSW.

The model approaches introduced above yield different output variables and units. While the PD model directly provides spatially distributed new snow amounts, the indices TPI and Sx are not directly comparable to snow depth change ΔHS and therefore require an additional scaling step.

We take a mean water equivalent of new snow per cell HNW‾ and determine a factor, depending on TPI or Sx, to increase or decrease the amount of HNW relative to the mean value. In our study we use the same value of HNW‾ for each cell, which is the value of the closest ICON cell. Based on , MeteoIO and we determine HNW of a cell with index j, depending on its Sx value, which we scale by the mean μ and standard deviation σ (over all grid cells), with: 9HNWjSx=HNW‾1-SxjμSx-σSxifSxj<0HNW‾1+αSxjμSx+σSxifSxj>0withα=∑j:Sxj<0|SxjμSx-σSx|∑j:Sxj>0SxjμSx+σSx In the above mentioned references each cell of Sx is scaled by the Sx domain minimum (if Sx <0) or maximum (if Sx >0). We found the minimum and maximum to be very sensitive to the extent of the evaluated area, and therefore apply a scaling by mean and standard deviation in this study.

We introduce the same principle for TPI, but we assume that areas with TPI>0 are depleted and TPI<0 are enriched (relative to HNW‾): 10HNWjTPI=HNW‾1-TPIjμTPI+σTPIifTPIj>0HNW‾1+αTPIjμTPI-σTPIifTPIj<0withα=∑j:TPIj>0TPIjμTPI+σTPI∑j:TPIj<0TPIjμTPI-σTPI We convert HNW to height of new snow HN, by using the temporal mean snow density ρ‾ for each event E = {E1, E2, E3} obtained from the WFJ observations (Fig. ), which we assume to be constant over the evaluated area. The computed HN per pixel j and event E, derived from Sx, is therefore calculated with: 11HNj,ESx=HNWj,ESx/ρE‾ The same equation holds for HNj,ETPI, but using TPI to calculate HNWj,ETPI and for HNj,EPD with HNWj,EPD.

Strictly speaking, the results of the scaling by mean and standard deviation are not HNW, since the values can become negative. We therefore refer to these model outputs as snow depth change ΔHS and, in water equivalent, ΔHSW. In the bottom row of Fig.  we show all scaled terrain parameters for an example epoch on 25 November 2023, 21:00–22:00 UTC, using a HNW‾ of 0.85 mm.

We compute our reference for change of snow depths ΔHS, by subtracting the DSM at the start of an event, from the the DSM at the end of the event, using the DSMs derived from TLS measurements. Figure  shows ΔHS for the three selected main events and their sub-events. In the first event (E1) we recorded a mean increase in snow depth of 0.53 m, with a standard deviation over the measured area of 0.27 m in a time of 42 h. The spatial distribution of new snow is varying, from up to 1.5 m snow accumulation in some areas to up to 0.15 m decrease in others. After the event of December 2023 (E2), we measured a mean increase in snow depth of 0.31 m, with a standard deviation of 0.23 m. The distribution pattern of snow depths is similar to E1. We find accumulations and depletions in similar areas, in both cases the prevailing wind direction during the event was north-west. The maximum accumulation of new snow is of a similar magnitude, but larger decreases occur. Several areas show a decrease in snow depth up to around -0.45 m, compared to before the event. In the last event of February 2024 (E3) the measured mean increase in snow depth is 0.37 m, with a standard deviation of 0.11 m. Due to the overall low wind speed during the event, we measured a rather homogeneous distribution of accumulated new snow without depletions.

This section first presents the parameter selection for TPI and Sx, followed by the resulting modelled snow depth distributions for the presented events.

In Fig.  we show the differences between the measured and the modelled ΔHS. Generally, the mean differences indicate an underestimation of modelled ΔHS, compared to ΔHS TLS, with exception of the TPI and PD model in E2. The error patterns are similar in the models for E1 and E2, underestimating areas with large, positive ΔHS, and overestimating areas with small, or negative ΔHS (see insets in Fig. ). This suggests that the main spatial structures are captured similar to the measured ΔHS, while the amplitude of the spatial variability is underestimated. In the case of E3, both ΔHS TPI and ΔHS Sx show large residual distributions, whereas ΔHS PD yields smaller and more homogeneous residuals, indicating a better agreement with the observed, spatially rather uniform snow depth distribution.

Figure  shows a comparison of the measured ΔHSs by TLS and modeled ΔHSs, using TPI, Sx and PD as scatter plots, with their Pearson and Spearman correlations coefficients (scatter plots for the sub-events are shown in Figs.  and ). The highest correlations per event are achieved by the snow depths derived from the terrain indices TPI and Sx. The spatial correlations in E3 are very low for all methods. We point out the narrow distributions of ΔHS Sx in E2 and ΔHS PD in E3. The ΔHS Sx distribution is influenced by the ratio of wind exposed to wind sheltered areas, while the ΔHS PD distribution shows minimal variation in snow depths due to the low wind speeds during E3 and partially opposing wind directions during the event, which can even out the distribution of ΔHS. Overall the correlations are moderate, with a generally higher value for the Spearman correlation, compared to Pearson. This suggests that there is no purely linear correlation.

In Table  we summarize all performance metrics of the models, compared to the measurements. The metrics for the sub-events are in the appendix (Tables  and ).

In this study we tested the performance of various approaches to estimate the distribution of new snow or snow depth changes in an avalanche release area at very high temporal and spatial resolution. The results show that the tested low-complexity approaches can reproduce parts of the observed spatial snow depth patterns, but none of them fully captures the variability of ΔHS across all events. This reflects both the simplified process representation of the models and the fact that real snowfall events are always a mix of several processes. The following discussion therefore focuses on the applicability and limitations of each approach.

The simplest, most general approach is the use of topographic position index TPI. However, it depends only on one parameter, the radius of evaluation. It is thus not sensitive to any wind variations in snowfall events. This is visible in Fig. , where the snow depth distribution pattern is the same for all events. There is only slight variation in scaling, according to the input of precipitation. This can be viewed as strength, as it makes the model more generic which could be an advantage for mixed type events. However, it is also a limitation, because the distribution of ΔHS for preferential deposition or redistribution events strongly depends on the wind speed and wind direction and the approach using only the TPI can not be adapted to individual meteorological conditions.

The wind shelter index Sx mainly depends on the wind direction, which makes it adaptable to different events, although the wind speeds are not considered. On this rather small evaluation area, the rescaling of TPI and Sx to ΔHS strongly depends on the scaling method and the evaluated domain extent. This becomes evident for Sx in the example of E2, where there are only little areas that are wind exposed, so the scaling ratio becomes very small (α in Eq.  or ), leading to very little variation of snow depths in the wind sheltered areas. In this study we use a scaling by mean and standard deviation, because when using the minimum and maximum (as suggested by e.g., ) we saw almost no variations of ΔHS in our ROI, due to some large (or small) values in the evaluated extent of DSM. Another approach for improvement could be to link the scaling factor to wind speed. For example, the scaling factor leads to too much variation in ΔHS for E3, but it seems reasonable for E2, as wind speed and therefore ΔHS variability are greater there.

The presented model for ΔHS based on PD takes all lacking parameters into account that are mentioned above, by relying on statistical downscaling of coarse meteorological parameters (precipitation, wind speed and direction). This allows an adaption to individual snowfall events, and the downscaling is designed to be broadly applicable, i.e., it is very little dependent on the resolution or extent of the model domain. However, the PD model has been developed for a spatial resolution of 30 m, and is the first time applied at the high resolution of 0.5 m. As the model name suggests, the PD model is limited on the representation of the process of preferential deposition. In reality the distribution of new snow is always a combination of multiple processes, so especially with higher wind speeds when also snow redistribution occurs, or during warm periods impacting the snowpack, the model cannot describe the resulting snow distribution (see mixed and redistribution events in Fig. ). For example, ΔHS PD overestimates the absolute ΔHSs in E2. Given the rather high wind speeds as well as the warm period at the beginning of this event, other processes than preferential deposition, have likely contributed to shape the snow distribution during E2. However, since the model downscales coarse precipitation and coarse wind, the overestimation of ΔHS could also be from inaccuracies in precipitation and wind input. The overall similar spatial bias pattern among the three models indicates that the distribution patterns agree, which suggests a possible improvement of the model performance by adjusting the absolute scaling to wind speed and precipitation. An explanation for the low spatial variability of ΔHS in E3, calculated with the PD model, is the change of wind directions during the period of the event. Opposite wind directions with similar wind speeds can even out the accumulated result.

All presented models strongly depend on the input of precipitation (HNW‾) and the scaling to ΔHS, using the density ρ of new snow. In this study we use ρs, derived from manual, daily measurements of HN and HNW at the measurement site WFJ. For comparison, we derived ρs for the same evaluation periods from the automated weather station at WFJ, which is placed a few meters next to the manual measurement location. The correlation values do not change, since ρ is only a linear scaling factor, but we found variations in the bias of up to 0.18 m across the models and events, in comparison to ρ derived from the manual measurements. An additional source of uncertainty is the ICON-derived HNW‾, which is taken from the hourly KENDA-CH1 analysis data. These data provide an observation-constrained and spatially consistent estimate of precipitation on a 1.0 km grid, but local snowfall amounts in steep alpine terrain can still vary substantially within a single grid cell. In our modelling framework, HNW‾ mainly acts as scaling input. Its uncertainty is therefore expected to affect primarily the absolute magnitude of modelled ΔHS. For this reason, we focus on the interpretation of the model results primarily on their ability to reproduce spatial snow distribution patterns.

The correlations between measured and modelled snow depth distributions in this study are broadly consistent with previous studies, although direct numerical comparison is limited because correlation metrics, spatial resolution, aggregation scale, and calibration strategies differ. For example, reported coefficients of determination (r2) of approximately r2=0.47–0.73 for measured snow depth changes and r2=0.32–0.53 for absolute snow depths, depending on search distance, terrain representation, and case study. Researchers obtained a higher agreement when measured snow-transport information was incorporated, such as reported correlation coefficients of R=0.78–0.86 for a modified terrain parameter combined with blowing-snow flux measurements. At coarser aggregation scales, found that locally calibrated statistical models explained 30 %–91 % of snow depth variability.

These limitations also indicate where future model development could build on the lidar dataset. A promising direction for improvement is the use of data-driven models that exploit the high temporal resolution of the lidar dataset. While machine-learning-based approaches for snow distribution have already been explored in previous studies, hourly lidar observations over several winter seasons provide a new opportunity to relate snow depth changes to meteorological and terrain parameters at event and sub-event scale. This makes it possible to test both simple regression-based approaches and more complex deep learning models for learning event-dependent snow distribution patterns from the observations themselves. Such approaches may help to better represent mixed situations, in which preferential deposition, redistribution, and additional snowpack processes act simultaneously. In addition, machine learning could also be used to improve specific components of the current workflow, for example by emulating local wind fields from terrain and coarse meteorological input, as in ongoing work based on the deep-learning model Devine . The dataset presented here therefore offers considerable potential not only for model validation, but also for the development of more flexible, data-driven snow depth models.

A promising application of the presented models could be the spatio-temporal completion of TLS measured snow depth maps. In Fig.  we show a first attempt, where we show the TLS-derived ΔHS, complemented with the ΔHSs modelled with TPI. Such applications may be particularly valuable when measurements are incomplete because of adverse weather conditions or terrain-related shadowing.

Our validation data is limited to a rather small area, due to the maximum measurement range of the TLS. Currently we use data from only one study site and evaluated three specific snowfall events. This has the consequence that we have only a certain variation of topography and meteorological conditions we can evaluate our models on. Furthermore, the measurements have local uncertainties in the range of 0.1–0.3 m , which has to be considered in the interpretation of error metrics.