The evaluation of snowpack models capable of accounting for snow management in ski resorts is a major step towards acceptance of such models in supporting the daily decision-making process of snow production managers. In the framework of the EU Horizon 2020 (H2020) project PROSNOW, a service to enable real-time optimization of grooming and snow-making in ski resorts was developed. We applied snow management strategies integrated in the snowpack simulations of AMUNDSEN, Crocus, and SNOWPACK–Alpine3D for nine PROSNOW ski resorts located in the European Alps. We assessed the performance of the snow simulations for five winter seasons (2015–2020) using both ground-based data (GNSS-measured snow depth) and spaceborne snow maps (Copernicus Sentinel-2). Particular attention has been devoted to characterizing the spatial performance of the simulated piste snow management at a resolution of 10 m. The simulated results showed a high overall accuracy of more than 80 % for snow-covered areas compared to the Sentinel-2 data. Moreover, the correlation to the ground observation data was high. Potential sources for local differences in the snow depth between the simulations and the measurements are mainly the impact of snow redistribution by skiers; compensation of uneven terrain when grooming; or spontaneous local adaptions of the snow management, which were not reflected in the simulations. Subdividing each individual ski resort into differently sized ski resort reference units (SRUs) based on topography showed a slight decrease in mean deviation. Although this work shows plausible and robust results on the ski slope scale by all three snowpack models, the accuracy of the results is mainly dependent on the detailed representation of the real-world snow management practices in the models. As snow management assessment and prediction systems get integrated into the workflow of resort managers, the formulation of snow management can be refined in the future.
The Alpine ski industry plays a central economic role in many mountain regions and is important for regional development. About 13.6 million people live in the European Alpine region, with around 60 to 80 million tourists visiting every year. The ski resorts generate a high turnover in the winter tourism destinations
Beyond the timescale of weather forecasts, which are generally reliable for a time frame of a few days, ski resort managers have to rely on various and scattered sources of information, hampering their ability to cope with highly variable meteorological conditions. In the framework of the funded European Union's Horizon 2020 (H2020) project PROSNOW
To increase the adaptation capacity of the skiing industry, there is a great need to combine weather and climate forecasting, snow modeling, and observations and to promote existing products to demonstrate their value for professional decision-making. In this context, in situ observations as well as optical and microwave remote sensing have proven to be mature technologies
Overview of the nine PROSNOW ski resorts, the period of available Sentinel-2 and measured GNSS snow depth data, and the used snow management configurations for the simulations based on the paper by
The objective of this paper is to evaluate the accuracy of the piste snow management module refined and implemented in the framework of the H2020 PROSNOW project embedded within several snowpack models to simulate snow management in general and for each individual PROSNOW ski resort in high spatial resolution. In a first step, the results of the snowpack simulations were compared both with remotely sensed satellite snow cover maps and with snow depth measurements spatially distributed along the ski slopes acquired with specific GNSS systems. The impact of elevation, slope, and aspect as well as temporal aspects within a season or amongst different years was evaluated. In the second step, the simulation domain was spatially discretized into defined ski resort reference units (SRUs)
Locations of the PROSNOW ski resorts.
Within the PROSNOW project we focused on the following nine ski resorts, which are also all part of this study: Seefeld (cross-country part) and Obergurgl in Austria; La Plagne and Les Saisies in France; Garmisch Classic in Germany; Colfosco, San Vigilio, and Livigno in Italy; and Arosa Lenzerheide in Switzerland. This selection of ski resorts represents a large diversity of geographical, climatical, and snow-making practices and equipment. Figure
Snowpack simulations are performed with AMUNDSEN for the Austrian and the Italian resorts (Colfosco, Obergurgl, San Vigilio, and Seefeld), with Crocus for the French resorts (La Plagne and Les Saisies) and with SNOWPACK–Alpine3D for the remaining resorts in Switzerland, Germany, and Italy (Arosa Lenzerheide, Garmisch Classic, and Livigno). We used for each ski resort different settings for the parameters concerning wet-bulb temperature, snow depth threshold, timing, and density of grooming. A detailed description of the functionality and parameters of the snow-making and grooming modules, which are used in all of the snowpack models, is shown in the study by
All three models require spatial input data for the snow management simulations consisting of a digital elevation model (DEM) covering the study sites, the locations of the ski slopes, and the locations and types of the snow guns (snow lances or snow fans – corresponding to different production rates for given ambient conditions as defined in Table 5 in configuration 2: no snow production; simulations based on a natural snow-only configuration, however with grooming activity configuration 7: snow production with a minimum required snow water equivalent (SWE) of 150 kg m configuration 11: snow production with a minimum required SWE of 150 kg m configuration 23: snow production with a minimum required SWE of 250 kg m configuration 31: snow production with a minimum required SWE of 250 kg m
Meteorological forcing data for the simulations are based on measurements from automatic weather stations close to or within the study sites and from the SAFRAN analysis for Crocus model runs
The models are equipped with a machine-made snow production and grooming module, which can be used for the operational applications. A set of core parameters can be used for very detailed simulations of snow management practices in single ski resorts. They take into account snow demand, the meteorological conditions including information on wet-bulb temperature and wind speed, and the ski resort infrastructure in terms of the amount of snow that can be produced in a given time step at a certain location within the resort. For the simulations it is assumed that for a given snow gun, all of the produced snow is distributed immediately and evenly over a predefined slope section. Additionally, the grooming module allows the distinct properties of groomed snow on ski slopes to be accounted for depending on the amount of snow present and a defined grooming schedule. It assumes that grooming has no effect on the distribution of snow, e.g., shifting of snow from one place to another, but rather only compacts it
Real-time simulations with a very fine spatial resolution (i.e., 10 m) require a very high computational demand, and such fine spatial resolutions are also often not necessary for the overall day-to-day resort management. Spatial clustering of slopes and slope sections is often sufficient for the snow managers working in an operational mode. Therefore, we additionally discretized each ski resort in ski resort reference units (SRUs). In a post-processing step, we aggregated the initial 10 m pixel size to larger areas. We defined different SRU sizes of the individual pistes by slicing them into the following elevation step ranges: 50, 100, 200, 300, and 400 m. Local snow management plays a major role in the SRU sizes as explained in more detail in Appendix
An example of a ski resort discretization into different SRU elevation bands: 50, 100, 200, 300, and 400 m. The figure shows the western part of Arosa Lenzerheide. The different colors represent the different SRU areas.
The model results for all ski resorts were compared with remote sensing images; for this study, Sentinel-2 (S2) data were used. The processing of the S2 snow-covered maps was done in three main stages: (i) calibration to top-of-the-atmosphere (ToA) reflectances; (ii) re-projection, resampling, and co-registration with the model grid with a final resolution of 10 m; and (iii) classification with a support vector machine (SVM) classifier trained with an active learning procedure
Since the S2 snow maps are compared to the snow simulations, particular attention was devoted to obtain accurate results. For this purpose, three main steps were performed to address the main problems related to snow classification from optical images, which are the detection of (i) particular cloud conditions; (ii) the mixed snow pixels, i.e., pixels in which classes other than the snow contribute to the observed spectral response; and (iii) the snow under the canopy of the forests.
First, we performed a visual analysis of all the S2 images for excluding the scenes presenting complex cloud conditions. In particular, semitransparent clouds, which are thin, high-altitude clouds composed of ice crystals, were detected from the S2 band acquired at 1.375
The SVM classifier was trained in a way that a pixel with at least 50 % snow coverage is classified as snow. This means that, for example, during the snow-making production at the beginning of the season, a pixel in which significant snow production is ongoing is classified as snow even though not all of the pixel area is covered with snow. Additionally, we identify the shadowed areas from where the multispectral sensor on board S2 is not able to record sufficient energy for distinguishing between snow and snow-free areas. This happens when the sun is low at the horizon approximately from mid-November to mid-February, and the terrain is extremely steep. In all the other shadow cases, the SVM classifier was trained to detect the snow presence. Hence, the output of the procedure was a classified map with four classes, i.e., snow, snow-free, shadows and clouds.
Since the detection of snow under forest canopy is a challenging research topic from both the remote sensing and modeling point of view, we conservatively masked out forested areas for all the ski resorts based on the land cover classification provided by OpenStreetMap (OSM) (resolution of 30 m)
A detailed overview of the number of available S2 scenes for each year and ski resort is presented in Table
More and more ski resorts are relying on spatially distributed snow depth measurements performed with modern Global Navigation Satellite System (GNSS) technology for an efficient management of their slopes. This technique relies on differential GNSS signals, comparing the snow-free (i.e., zero snow depth) reference signal with those obtained during the snow season to obtain snow depth. The sensors are installed on top of the groomers, and thereafter snow depth can be tracked as a positive side effect whilst grooming the pistes. This technology ensures a snow depth measurement accuracy down to the centimeter level and at a spatial resolution of 1 m, which also allows the tracking of snow redistribution with the groomers.
For our study, rasterized data were provided by the companies SNOWsat and Leica Geosystems AG and were resampled to a resolution of 10 m using the average value in order to be directly comparable with the model outputs. The GNSS snow depth data were available for all ski resorts except La Plagne. We considered for the analysis the measurements spanning from 1 December to 31 March with a daily temporal resolution when GNSS data were available. The data have been preprocessed to eliminate outliers and to check their consistency. Table
In this section, we describe how we evaluate the snowpack simulations carried out for the PROSNOW ski resorts. This includes (i) the evaluation of snow-covered area and (ii) the evaluation of simulated snow depth. In detail, the simulated snow depth was compared with the GNSS-derived measurements over a number of ski slopes, whereas the snow-covered area is evaluated by comparing the model snow-covered area with the S2 snow maps. The metrics used for assessing the agreement between the simulations and S2 snow maps are the confusion matrix and the snow persistence index defined below.
The evaluation analysis for both snow depth and snow-covered area was conducted by stratifying the data according to temporal and topographical constraints. Moreover, a differentiation was made between natural snow, i.e., snow outside the pistes, and managed snow, i.e., snow inside the pistes. In the following subsections more details on how the evaluation metrics were calculated are presented.
The latitude, longitude, and elevation of a ski resort have a big impact on the timing of snow accumulation and melt. Therefore, comparing snow patterns between regions is challenging despite the widespread application of remotely sensed methods for snow research. The snow persistence (SP) is a snow metric that can be used to map snow zones globally
Example of the confusion matrix used in the “Results” section. The analysis was split into three periods: beginning (B: October–November–December), middle (M: January–February), and end (E: March–April–May) of the season. TP: true positive; FP: false positive; FN: false negative; TN: true negative; OA: overall agreement.
Two SP indices were extracted considering both S2 snow maps and model simulation. They were calculated pixel-wise as the ratio between the number of snow-covered days derived by S2 or from the model, divided by the total number of S2 observations (snow or snow-free). The values of SP were always between 0, i.e., always snow-free dates, and 1, i.e., always snow-covered dates. If an S2 snow map pixel is classified as a cloud the corresponding snow model output is masked out, preserving the one-to-one correspondence between the two SP indices.
In addition to the SP index, for each ski resort a confusion matrix was computed to assess the quality of the S2 and snowpack simulations. We refer to the confusion matrix as modeled vs. observed variables. The confusion matrix has the form indicated in Table
We distinguish between natural snow and snow on the slopes. Furthermore, the analysis was split into three periods: beginning (B: October–November–December), middle (M: January–February), and end (E: March–April–May) of the season. A pixel can be either true (snow in S2 data and model, no snow in S2 data and model) or false (snow in S2 data and no snow in model, snow in model and no snow in S2 data). With the accumulation of all pixels assigned to be true, an overall agreement OA (%) was calculated for each period and catchment:
The metrics used for this assessment are the mean deviation (MD) and the root mean square deviation (RMSD) over time for each ski resort. Regarding the snow-covered area evaluation, the binarization of the simulated snow depth to the snow-covered map was done by imposing a threshold of 0.05 m; i.e., every value above this threshold was identified as snow, while in contrast all the values below 0.05 m were identified as snow-free. This threshold is in line with previous works
In this section, the simulated results for all ski resorts were compared with the S2-measured snow cover and with the GNSS-measured snow depth data on the ski slopes. Model runs were performed for five winter seasons (2015–2020) from 1 October until the end of May. The simulations were carried out for all ski resorts using the default snow management configurations accounting for both fan guns and lance guns as well as different temperature thresholds and base-layer production targets, given in Table
Confusion matrix for all ski resorts referring to snow on the pistes. The values in the parentheses are referring to natural snow. The metrics in the square brackets also refer only to snow at the beginning, middle, and end of the season.
The S2 algorithm produced accurate snow maps with an overall accuracy above 80 %, for both high-Alpine and lower-lying mountainous regions and different stages of the season like season start, mid-season, and end of season. A first approach to assess the skills of the models using S2 maps as a reference was a confusion matrix. Table
Snow persistence index difference between Sentinel-2 data and the model data (on the left) and SP indices for the simulation and Sentinel-2 (on the right) for each ski resort:
The snow coverage quality of the snowpack simulations using the snow management modules was further assessed using the S2 data and SP indices. The SP indices were calculated for the simulations and S2 data, and the relative differences are shown in Fig.
The overall accuracy was mainly impacted by the elevation and slope. An interesting analysis was represented by the trend of the overall accuracy over 100 m elevation classes,
A closer look at the SP index showed an elevation dependency, but a clear slope or aspect dependency is hard to detect. A better accuracy is obtained at high altitudes due to the fact that the ratio of snow cover area is near 1. Figure
Comparison for 100 m elevation bands,
The simulated default snow management configurations reproduced the actual conditions well at all ski resorts. Figure
Figure
Root mean square deviation (RMSD) (upper subplots: solid line, left axis) and mean deviation (MD) (upper subplots: dashed line, right axis) averaged over space between GNSS-measured snow depth (SD) (lower subplots: solid line, left axis) and simulated SD (lower subplots: dashed line, right axis) over time for the ski resorts. Within the period 2016–2020 we considered all valid GNSS-measured snow depth data which were available.
Overview of the root mean square deviation (RMSD; hollow symbols), mean deviation (MD; filled symbols), and standard deviation (
Discretizing the ski resorts in coarser clusters tends to mask the variability in the error in terms of RMSD due to averaging effects between the simulations and GNSS-measured snow depth between 8 % and 45 % independent of the season and cluster size (see Table
In addition, we tested the spatial variability within the pistes. For a visualization example presented in Fig.
This study presents a new high-resolution evaluation of snowpack simulations including snow management modules for mountain ski resorts to assess the quality of the simulations. The simulated results showed a high overall accuracy of more than 80 % compared to the Sentinel-2 data and a root mean square error in the GNSS-measured snow depth of below 0.6 m. The simulated results for all ski resorts are plausible and robust on the ski slope scale.
For every ski resort, a large number of S2 images classified with high accuracy were available to assess the quality of the simulations in terms of snow coverage. More than 62 S2 images at each ski resort and even more than 150 for two resorts were analyzed. The specific machine learning algorithm to derive information with low uncertainty about the presence or absence of snow from the Copernicus Sentinel-2 images allows the generation of snow maps across the Alps with relatively low manual effort. Additionally, the very detailed forest mask applied to the evaluation allowed us to simultaneously (i) avoid situations for which the information provided by S2 is insufficient to produce accurate results and (ii) be able to extract information about the snow cover for the pistes crossing dense forests, as is often the case for Garmisch. This allowed us to minimize the pixel loss due to canopy shading. The highest pixel losses with respect to the total piste area were in Arosa Lenzerheide (5.7 %) and Seefeld (5.5 %), followed by Obergurgl (2.4 %), Livigno (0.2 %), and Colfosco (0.2 %). For the other ski resorts it was zero.
Example of spatial variability within a piste in Garmisch Classic (22 January 2018). On the left, RMSD represented as pixel-wise difference between measured and modeled snow depth for a 10 m pixel resolution. The smaller figures represent the original modeled and measured snow depth. On the right, RMSD represented as averaged difference between measured and modeled snow depth for a 100 m band SRU resolution. The nature of the large measured snow depth by GNSS at around 1400 m elevation is due to a dip in the ground surface, which was filled with snow to level the piste. This led to a higher GNSS-measured snow depth compared to the simulations.
The underestimation of the overall accuracy at the beginning and the end of the seasons was due to the fact that sometimes the exact snowline in the S2 data was hard to detect. Some snow lines were obscured by shadows, and they often did not appear as continuous lines, which may slightly bias our regional estimates, especially in areas which span different land uses. Also white rock types or illuminated wet rocks can lead to brighter pixels in the S2 data and to an underestimation of the overall accuracy. Because the spectral signal of these pixels is similar to that of snow, the algorithm detects an ice–snow boundary and classifies these pixels as part of the snow lines. Since these patches were situated in ski resorts in rocky areas, these misclassified pixels introduced negative biases in the overall accuracy estimates, and they were filtered out manually.
By inter-comparing the model simulations and the S2 images we encountered some recurrent errors. As described in the previous section, on average the accuracy of the simulations is high, and the snow coverage simulated by the PROSNOW models is consistent with observations. However, wrong discretization and/or missing meteorological input and lack of snow managing or land use information were the main sources of errors. In particular, (1) ephemeral snow (i.e., snow that lasts a few days either at the beginning or at the end of the season) is difficult for the models to simulate correctly; (2) rain–snow transition (e.g., ensuing rapid snowmelt inside the catchment) is difficult to simulate accurately; (3) due to unknown snow-making strategies, which are then not incorporated into the PROSNOW models, snow-making at the beginning of the season and delayed and anticipated snow melting at the end of the seasons are not correctly modeled over managed slopes; and (4) the heterogeneous landscape at 10 m resolution plays a role in the snow accumulation and melting dynamics (e.g., towns, lakes, and roads are visible and change the snow distribution). This information is generally not addressed by the PROSNOW models. These are just confirmations of the expected limitations of the state-of-the-art snow cover models. However, for the first time a systematic and extensive evaluation at high model resolution was performed. The details of the analysis with all the different recurrent errors encountered for each ski resort are shown in Table
The GNSS data can only be used as ground observations with some restrictions. There are several problems which might affect the quality of the GNSS data: (i) the digital elevation model (DEM) profile of all ski resorts is prone to change every year due to earthwork and adaption of the slopes and is always considered correctly in the extraction of the snow depth, and (ii) the inclination of the groomer has a large impact on the GNSS-measured snow depth. For example if the calculated snow depth is 0.5 m, the effect of 30 % inclination would be 6 cm, which means that the calculated snow depth would be 12 % higher than in reality. Furthermore, the work of the groomers is not only to measure the local snow depth but also to fill sinks, compensate humps, or level out the snow production or redistribution by the skiers. This might lead to very small-scale variances in the GNSS-derived snow depth for pistes both with and without technical snow production, whereas the model shows less pronounced variability. Therefore, especially the small-scale deviations between the simulated and GNSS-measured snow depth are not well applicable to reflect the model accuracy. The comparison rather shows the degree to which the default snow management configuration was applicable for the individual ski resorts. As each ski resort spontaneously adapts its snow management production due to changing weather and snow height conditions, it is not possible to consider the spontaneous snow management adjustments during the season in the simulations. It does not make a big difference between pistes with and without technical snow production even though the models do not explicitly simulate the snow redistribution. However, the RMSD between the modeled snow depth and GNSS measurements over time for the ski resorts is in the uncertainty range of the simulated configuration.
Some recurrent errors are encountered by inter-comparing the model simulations and the GNSS-measured snow depth. The quality of the simulations is high, and they showed plausible and robust results on the ski slope scale. However, there are still sources of errors: (1) extreme snowfall situations, (2) snow redistribution by the groomers, (3) rapid ablation (e.g., south-, southwest-exposed pistes) due to high solar radiation, (4) levering of the pistes to reduce the accident risk of the skiers, (5) snow redistribution by the skiers, and (6) systematic errors due to the wrong snow-making strategy. These errors are already known, but it is too complex and not straightforward to consider this in the simulations in the current state. In detail, slightly different recurrent errors are encountered for each ski resort by inter-comparing the model simulations with GNSS-measured snow depth, which is shown in Table
Scale and data aggregation has important effects on the simulations and interpretation of snow depth data, which is also true for snow on pistes. Studies in different research fields clearly demonstrate that spatial variability and statistics are dependent on scale
Nevertheless, the identification of guiding principles for researchers to combine data and models at different spatial and temporal scales and to extrapolate information between scales still remains a challenge. By going from fine to coarse scales, aggregation and generalization set in. The rate of information loss is influenced by small-scale spatial snow production and grooming patterns. Heterogeneous snow production for instance leads to more information loss than aggregations at coarser scales. The small-scale spatial effects of moving snow by the groomers or skiers, for instance, disappears slowly with decreasing resolution, and those that are dispersed are lost rapidly. This leads to an under- or overestimation of the simulated snow height. Therefore, a methodology needs to be developed to find out how much the loss of information takes place. Multiscale analysis was necessary to show that variability for different aggregation types of the snow height is inherently different and that for each SRU this can be different, as shown in Fig.
However, we demonstrated a consistent evaluation procedure for all the PROSNOW ski resorts that can be useful for snowpack modelers and ski resort managers. An understanding of the nature of scaling effects is needed when spatial or temporal scale is an independent variable. In landscapes with homogeneous snow depths, where snow measurements can be summed directly, such scale problems may not occur. However, in snow-distributed landscapes like at ski resorts, snow measurements obtained at fine scales often cannot be summed directly to produce regional estimates. Therefore, reasonable measures are not always given by weighted averages because heterogeneity in snow production and distribution may influence scaling processes in nonlinear ways. In such cases, increasing the level of spatial heterogeneity may also increase the difficulty of extrapolating information across scales
The use of remote sensing and GNSS data allows an evaluation procedure to be defined for the snowpack models and helps to improve the resource management of the ski resorts. The GNSS data provide means to collect accurate snow depth points in the field for precise correction of the simulated snow depths. Further, using snow depth measurements over an entire season together with snowpack simulations is a powerful tool in the long term. It allows the estimation of the minimum snow depth required at various slope sections. This ensures that slopes are optimally prepared and groomed right up to the end of the season. The spatial remote sensing images are needed to improve the simulated snow-covered area, especially at the beginning or end of the season or for lower-situated ski resorts, where natural snow precipitation is low. Further, it allows the correction of the simulated ablation process at the end of the season and the snow produced at the beginning of the season. However, further studies to determine whether the models can simulate snow depth with sufficient accuracy to enable the resort mangers to maintain the optimum and minimum viable snow depth in a more efficient way are needed and will be attempted in the future
The combination of both techniques allows the evaluation and initialization of the simulations: imagery can be used for primary digitization of the snow cover where GNSS can be used for in situ observation of the snow height for the simulations. A detailed analysis of the differences between the two methods will allow us to make better decisions about when and how much snow is distributed by both groomers and skiers. Further, the effect of snow melting and snow redistribution by wind on the slopes can be extracted. This allows us to improve the models. However, this is not easy to extract and was not within the scope of this paper but should be considered for further studies.
The initiative for this study emerged within the H2020 PROSNOW project to evaluate the snow simulations over the nine PROSNOW ski resorts by comparing model outputs with local and remotely sensed measurements in terms of snow coverage, persistence, and snow depth. The three snowpack models AMUNDSEN, Crocus, and SNOWPACK–Alpine3D include all piste management modules and were evaluated using both ground-based data (GNSS-measured snow depth) and spaceborne Copernicus Sentinel-2-derived snow maps. We evaluated five winter seasons (2015–2020) from 1 October until the end of May and performed this evaluation in a stratified manner in order to assess the performance of the snow simulations under different conditions. Particular attention has been devoted to characterizing the spatial performance of the snowpack models with integrated snow management modules. Our presented results show high accuracy of the simulations, representing the “reality” well.
An inter-comparison of the three snowpack models applied to the same resort would be a logical next step from the model development perspective. Differences in the simulated results of the three models for a given ski resort would be mainly due to the different implementations of the snow management configurations into each model and due to the different snowpack energy and mass balance approaches. Such effects should have to be disentangled and discussed accordingly to have a reasonable comparison. However, this is not straightforward and is out of the scope of this paper but should be considered for the future. Nevertheless, this work showed that all three snowpack models applied for piste management reproduced plausible and robust results on the ski slope scale, and the overall accuracy of the results is mainly dependent on the degree to which the real-world snow management practices are integrated. Additionally, a detailed analysis to show the accuracy of the GNSS system installed on groomers to measure the snow depth is needed to validate the system. Moreover, integrating a snow redistribution model and an avalanche dynamics model into this system would help to point out where the biggest differences due to snow gliding or avalanches are between the Sentinel-2 data and the simulations. Further studies on the topographic complexity of the snow-free terrain and the rather smooth piste surface are needed to, for example, implement an index of surface smoothing compared to the bare ground. Future studies investigating how skiers redistribute snow under certain meteorological conditions in combination with topographic conditions (e.g., aspect, slope angle) would also help to overcome further potential errors.
In our approach, the spatial representations of ski areas and of the interpolated meteorological fields as well as the simulated snowpack information differ in their spatial representation: the geometry type used for the ski slope is a vector-based polygon, whereas the input and output of the snowpack models are based on a discrete approach, using regular grids of points. The challenge for PROSNOW was to define an intermediary spatial object. This should be consistent with representations balanced between the heterogeneity of meteorological conditions within ski slopes, the accuracy of snowpack models, and the computation resources and data volume required for a daily update. Also it should take into account the localization of the snowmaking facilities. The SRU, standing for ski resort reference unit, aims at fulfilling all of these requirements. It can be end-user-defined, including specific needs of the ski resort, but it can also be processed automatically by chaining several operations. We stored the vectorial geographic information system (GIS) and attribute data of all nine ski resorts in a PostgreSQL 10.7 database management system (DBMS), and the crossing operations between rasters and vectors were performed with Python 2.7 with the packages GDAL/OGR along with NumPy. For automatic processing of the SRUs we considered the following steps:
The association of each snow gun with a single ski slope is based on the spatial relation of the nearest neighbor to the ski slope. The slope area covered by snowmaking is calculated by determination of the upper- and lower-altitude bounds for each snow gun. Considering that the mean surface covered by a single snow gun is approximately Once the snow type attribute (“grooming only” or “with snowmaking”) is defined for every slope, the according areas were then divided into smaller parts based on the elevation resolution. The initial DEM topography was reclassified with respect to the targeted SRU resolution. A value in the according numeric series was assigned to a pixel whose value is in the range of the target value of approximately half the resolution (i.e., for 50 m resolution, the value 300 will be assigned for all DEM pixels whose value is more than or equal to 275 m and less than 325 m and between 150 and 450 m for 300 m resolution). Contiguous pixels with the same values were merged and polygonized. As small SRUs might occur by applying the abovementioned steps, e.g., at the beginning or ending of single pistes, we merged them in a post-processing step with other small adjacent piste fragments having the same snow type attribute. The final operation consisted of filling the missing snow type attributes by calculating the average area value for each polygon from the DEM and its derivatives (slopes and aspects) and matching the output from snow models (the snow type attributes of value, average altitudes (min and max altitude too), slopes, and aspects are also stored).
Overall agreement over time between simulation and S2 data for each ski resort for natural snow outside of the pistes. We considered images ranging from the winter season 2015/16 to the winter season 2019/20, except for Livigno, where simulations are not available for the first season 2015/16 due to gaps in the meteorological forcing data.
Overall accuracy trends over time between simulation and S2 data for each ski resort for machine-made snow on the pistes. We considered images ranging from the winter season 2015/16 to the winter season 2019/20, except for Livigno, where simulations are not available for the first season 2015/16 due to gaps in the meteorological forcing data.
Comparison for 100 m elevation bands,
Total number of available Sentinel-2 data for each ski resort.
Effect of discretization of the simulated results (10 m
Inter-comparing the model simulations with Sentinel-2 data and GNSS-measured snow depth for each ski resort.
Datasets related to this article can be obtained at
We used SNOWPACK version 20181109.1697, Alpine 3D version 20181116.472, SURFEX/Crocus version 8.1, and AMUNDSEN version 1.2.
PPE, FK, VP, CM, and ML developed the paper concept and methodology. PPE, FK, VP, CM, FH, CMC, HF, FM, and OH contributed to the data collection, and VP, CM, and HF processed these data with input from PPE and FK. The final figures were produced by VP, and snowpack simulations were performed by PPE, FK, FH, and CMC. US, SM, and ML supervised this work, and SM acquired the funding. The draft was written by PPE, FK, VP, CM, and ML, and all authors contributed to the refinement of the paper scope and revised the paper.
The authors declare that they have no conflict of interest.
PROSNOW is a project aiming at producing an operational climate service in order to transfer it as a commercial service.Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank our project partners and ski resorts for many constructive discussions and providing data to improve the manuscript.
This project has received funding from the European Union's Horizon 2020 research and innovation program under grant agreement no. 730203.
This paper was edited by Masashi Niwano and reviewed by Richard L. H. Essery, Paul A. B. Bartlett, and one anonymous referee.