Skitourenguru (https://www.skitourenguru.com/) is a popular online service that supports backcountry skiers in the selection and planning of suitable backcountry trips. It provides avalanche risk assessments for thousands of predefined backcountry ski routes across the Alpine region using an algorithm, which processes information from the current avalanche forecast and terrain characteristics . The website is freely accessible to all users and does not require a registered account. Users can search for ski routes based on criteria such as travel distance from home, elevation gain, route difficulty, or avalanche risk. Additionally, users who have registered for a free account can upload GPS tracks of their own tours . Both datasets used in this study were collected by Skitourenguru GmbH and are introduced in the subsequent sections.

Between 2013 and 2024, over 6800 GPS tracks were sampled from the platforms https://www.skitourenguru.com/, https://www.gipfelbuch.ch/ (last access: 26 February 2026) and https://camptocamp.com (last access: 26 February 2026) . The GPS data cover 9 winter seasons and 126 out of 128 Alpine warning regions, though many warning regions only contained a few tracks over the whole study period. On average, roughly 770 tracks were recorded in each season without a noticeable trend over time. Skitourenguru and Gipfelbuch are mainly used by German-speaking recreationists, while camp2camp is predominantely used by French- and Italian-speaking recreationists . By using all three we ensure coverage of German, French and Italian speaking regions of Switzerland. On these websites, users with a free account can post GPS tracks and condition reports of ski tours and other outdoor activities. These posts are visible to anyone visiting the websites. As the data was sampled manually by Skitourenguru GmbH, there are two data gaps in seasons 2021/22 and 2022/23 and there is no user information, such as a user ID, available. Although this dataset only represents a small fraction of real-world skiing activity, it reflects a user-base from different websites and language-regions.

This dataset has been previously used to study avalanche risk taken by backcountry skiers under different avalanche conditions . To preserve privacy, the coordinates of the GPS tracks were aggregated to the spatial granularity of warning regions and timestamps to 1 d (obfuscation). Figure b shows one example GPS track before obfuscation. After obfuscation, each track is represented by a single data point, holding information about the warning region, the mean elevation of the track and the date it was carried out.

On Skitourenguru, engagement data is collected by logging clicks on pre-defined ski routes (see Fig. b). This dataset contains over 8 million clicks on 2666 unique ski routes covering 122 of the 128 warning regions and a time period of 9 years between 2015 and 2024. Since 2017, clicks have been associated with a unique ID, which is retained as long as browser history or cookies are not cleared. We used these unique IDs to estimate the number of users (570 000) and the average distance between clicked routes per ID and day (15 km) to account for users clicking on multiple tours during trip planning. As the distance was smaller than the average warning region size, we assumed that the clicks typically fall within the same region and do not cause spatial distortion.

Every click can be related to exactly one geo-referenced route, from which terrain characteristics and the warning region can be inferred. Analogous to the GPS tracks, all clicks are aggregated to the spatial level of warning regions and to daily intervals. After the re-design of the website in 2020 and the related connection to other websites such as the website of the Swiss Alpine Club (SAC), the popularity of the website and the number of resulting clicks increased greatly. Due to this increase, data before and after 2020 are difficult to compare. Therefore, only data from the season 2020/2021 onwards is included for modelling and prediction, which results in ≈ 7.3 million clicks and represents 90 % of the initial dataset. However, all click data is used for the correlation analysis of GPS tracks and clicks to maximize temporal overlap between both data sets.

The variables used to predict skiing activity are linked to the four factors that contribute to avalanche release as introduced in Sect. , as well as by a literature research in the domain of outdoor recreation and specifically backcountry skiing. A list of all variables, a short description and the data source they were derived from, can be found in Table .

The selected variables can be divided into three temporally and spatially dynamic categories (weather, snow, temporality) and one spatially variable category (environment) (Table ), which reflect the different sides of the avalanche triangle. Weather and snowpack are directly represented by weather variables and snow variables. Terrain suitability is represented by environmental variables. Finally, patterns of human behaviour are captured through environmental and temporal variables, reflecting preferences related to accessibility, weekdays, holidays and seasonality.

There is rich literature on the importance of weather variables for outdoor activities . For instance, found that virtually all skiers access a weather forecast when planning a tour and that it can even deter them from ultimately going outside. Further, temporal variables relating to weekday, holiday and seasonality are often used for predicting behaviour in recreation and tourism and have shown to be an important driver for backcountry usage patterns . Snow conditions and the avalanche forecast are crucial for backcountry skiing and play an important role in the decision-making process. They can sometimes deter people from undertaking backcountry skiing trips, for instance when avalanche conditions are expected to be dangerous , while also enhancing activity due to the desire to ski an untracked slope of fresh snow, which is for many skiers the ultimate goal of a ski tour . Accessibility is a pre-requisite for recreation which is commonly used to predict recreational activity or recreation supply, and is a crucial factor for terrain-selection of backcountry skiers . Further, recreational activities can significantly disturb wildlife, the existence of protected zones therefore influences the regions where backcountry activities are undertaken .

The clicks and the GPS tracks have the same spatial (warning regions) and temporal (daily) resolution. Both datasets were enriched with the predictor variables aggregated to these resolutions.

Meteorological variables were derived from gridded datasets interpolated from SwissMetNet Stations . We used daily average temperature, the daily relative sunshine duration and the daily precipitation sum. Meteorological variables vary according to topographic elevation . Since backcountry skiing usually takes place at higher elevations within a region, mean values for precipitation and sunshine duration were calculated based on the grid points that lie in an elevation band within ±100 m of the mean track elevation for the track data, respectively the mean route elevation in a given region for the click data (Fig. b). To account for the snow fall line, we used the elevation belt around the minimum elevation, which is generally the warmest part of the tour, for the temperature calculation. Daily measurements of new snow and absolute snow height were available for 226 automated (IMIS) and 126 manual measuring stations (BEOB) . Most of the stations are concentrated in inner-Alpine regions, therefore some warning regions at the Alpine edge contain few or even no measuring stations. Further, some stations contain substantial measurement gaps. Due to the broad spatial resolution to which variables needed to be generalized, a spatial interpolation of the measurements would have been unnecessarily complex. Therefore, we opted to use the mean of the five nearest measuring stations for each warning region. If more than five stations lay within a region, those with the smallest elevation difference from the mean ski track elevation were selected. Further, we used the daily forecast avalanche danger communicated through the 5-level danger scale (1 = low, 2 = moderate, 3 = considerable, 4 = high, 5 = very high) as published by the WSL Institute for Snow and Avalanche Research SLF. For the remaining variables, we calculated road density by dividing total length by area, ski route density by dividing the number of ski routes per area, census density by dividing total number of inhabitants by area, accessibility by multiplying road density and census density (e.g., ) and we used the proportion of protected wildlife area per warning region. Season start was determined using the first day of the season on which an avalanche forecast was issued and we used day of the season as the number of days since 1 November, to allow comparison between seasons. Finally, for holidays we included all Swiss National holidays, as well as single days between public holidays and weekends (commonly referred to as “bridge days”), as well as the week between Christmas and New Year (see Appendix  for a complete list).

Different models can be used for prediction tasks, such as fully explainable, linear models (e.g., GLM/GAM: ), partially explainable machine learning models (e.g., random forests: ) and deep learning models (e.g., neural networks: ). Choosing the right model involves trade-offs: while more complex models like machine learning or deep learning models can better capture non-linear relationships, they are harder or even impossible to interpret. Simpler models on the other hand offer a high level of interpretability but have limited power with non-linear and potentially correlated data. Considering the characteristics of our training data, which is noisy, non-linear, inter-correlated and relatively small in size, we chose to use random forests.

Random forests have proven to be an efficient and effective tool to predict visits to outdoor recreation areas or map recreational ecosystem services . They have a number of advantages in that they are well suited to non-linear and correlated data and agnostic with respect to data types such as numerical and categorial data . Compared to deep learning architectures like neural networks, random forests are however relatively easy to interpret as the algorithm consists of a set of decision trees that make the prediction based on majority voting . Moreover, they provide an estimate of the variable importance as well as of how different values of a variable influence the outcome. In other words, random forests provide a level of interpretability that most other machine learning algorithms fail to provide . Additionally, they work well for relatively small and noisy data sets because they are not prone to overfitting due to the large number of trees that are grown .

We used the track data and the click data to train two separate random forests using the “randomForest” library in R . Because the track data was far less abundant than the click data – many regions only contained a few tracks over an entire season – we used it to train a binary classifier, with “presence” (when at least one track was recorded) and “absence” (when no track was recorded). The click data was used to train a regression model, where the response variable was the daily click count per warning region. Both models had identical spatial (warning regions) and temporal (1 d) resolution. For the remainder of this article we use the terminology “track model” for the binary classification model derived from the GPS track data and “click model” for the regression model derived from the click data.

While correlated variables do not impact the predictive power of a random forest, they can hinder the accurate estimation of variable importance as measured by variable permutation . Moreover, they may lead to increased computation time without contributing significant additional information. Therefore, a correlation analysis was carried out to exclude strongly correlated variables (r > 0.4). Additionally, variables with near zero importance values were excluded to speed up computation. Further, data points were excluded when they were recorded outside the winter season (June–October), or when no weather or snow data was available for the given day, since random forests do not accept NA values as input. Accordingly, 2.5 % of click data and 1 % of track data was filtered out.

Since both datasets included only presence data, we inferred absence by adding data points for days and regions without clicks or tracks, assuming absence of evidence implies evidence of absence – on the premise that no record signals fewer people in the field. For modelling, we assigned a click count of 0 or a track label “absence” to these generated points.

For the track data, the resulting absence points outnumbered the presence points by a ratio of 30 : 1. Class imbalance is a frequent problem when working with real life data and can be challenging for machine learning algorithms. When fed with imbalanced data, most algorithms fail to yield equally good performances in both the minority and the majority class since, depending on the performance measure chosen, the algorithm prioritizes accuracy of the bigger class . To address this, the two classes were artificially balanced by downsampling the absence class to train the track model. This in turn meant that we expected our model to overpredict presence, since presence counts were artificially inflated.

In typical machine learning applications, training and testing data are created by randomly partitioning the dataset. However, if temporally autocorrelated processes are present, a random split violates the assumption of independence between training and test sets . Since temporal autocorrelation was clearly present in our data, we used an entire season as the test set while training the model on data from all other seasons. This approach resulted in four training runs, cross-validated with four different winter seasons for the click model, and nine training runs, cross-validated with nine different winter seasons for the track model.

Hyperparameters were fine-tuned using a grid search to find the best possible parameter values for mtry (the number of variables randomly selected at each node of a tree) and sampsize (the number of data points sampled for each tree), which are the most common parameters used for tuning random forests (Figs. S6–S7). As the generalization error generally decreases with a higher number of trees and consequently more trees lead to a more stable prediction, we opted for a forest of 1000 trees for each model .

After training the models on click and track data, they were applied to unseen test data, repeating for each cross-validation run. Classification performance was assessed using sensitivity, specificity, balanced accuracy and the Hanssen–Kuipers Skill Score (KSS). Sensitivity and specificity were calculated according to . Balanced accuracy is the geometric mean of sensitivity and specificity and is frequently used when classes are imbalanced . To account for class imbalance, we additionally used KSS, a measure developed in meteorology and suitable for imbalanced prediction problems where the minority class is the focus . R2 and RMSE were used to assess performance of the regression model (e.g., ). Further, we calculated the prediction delta for both models, which we defined as the difference between predicted and observed tracks, respectively clicks to assess the spatial and temporal distribution of errors.

To assess how different variables impact the prediction, variable importance values were calculated using the built-in function for variable importance in the “randomForest” R library . Variable importance was calculated using a permutation-based method, measuring the average decrease in model accuracy and therefore predictive power, when a specific variable was excluded. To examine how each variable influenced activity, we calculated permutation-based partial dependency (PD) using the R package “pdp” . PD isolates a variable’s effect by holding all other variables constant and thereby assessing its impact on the probability for a given outcome of the response variable .

To demonstrate how the models are spatially influenced by altering one variable, we created idealized scenarios where all but one variable was held constant. For each scenario, a reference value was defined, and the variable of interest was systematically altered, while all other variables were fixed at reference values. The resulting differences in model predictions were visualized to highlight the spatial heterogeneity in variable influence. This approach allowed us to map the response of model predictions to changes in individual variables in a spatial context on an exemplary basis.

Residuals show that the track model consistently overpredicted activity across all regions (Fig. a). The underlying driver for this lies in the model training with artificially balanced presence and absence points. When verified with real-life and therefore unbalanced data, the model predicted more presence than was observed. The click model on the other hand both over- and underpredicted activity depending on the region (Fig. b). When comparing temporally aggregated predicted and observed counts per region, we find a nearly perfect correlation close to a 1 : 1 relationship for the click model (ρ = 0.98) and a strong correlation for the track model (ρ = 0.68). Scatterplots can be found in Fig. S8.

For the binary track model, errors were autocorrelated within regions and largely followed the distribution of the initial training data, with larger absolute errors in regions with more recorded tracks and smaller absolute errors in regions with very few recorded tracks. Contrastingly, the errors of the click model were neither autocorrelated, nor did they follow the underlying distribution of the training data. Generally, residuals approached zero in most regions, with some regions with larger absolute errors dispersed across the whole study area. Visually, the only slight spatial trend was that the click model underpredicted activity slightly more often in the northeastern and eastern part of the study area, which coincides with regions that received more clicks overall.

Figure  shows the predicted and observed track and click counts aggregated over the whole study area for one example season. The click model captured weekly and seasonal cycles, with higher predicted activities on the weekends and in the middle of the season, coinciding well with observations (actual clicks). The magnitudes of peaks were often underpredicted, while periods of lower activity were overpredicted. Overall, the predicted clicks reflected a smoothed version of the observed clicks. The track model on the other hand produced a very noisy prediction and systematically overpredicted activity. Nonetheless, track predictions correlated fairly well with click predictions (2020/2021: ρ = 0.69, 2023/24: ρ = 0.71, p < 0.001). Most predicted click peaks and some predicted track peaks visibly aligned in their temporal locations with the observed peaks. However, the predicted magnitude, especially for tracks, frequently did not match observations well. This was also reflected in the prediction delta (i.e., difference between model predictions and actual counts) (Fig. b and d), which was continuously positive for the track model while alternating between positive and negative for the click model. The track prediction delta was almost zero in the early stages of the season (November), which coincides with almost zero recorded tracks, hence the model performed best when there was no activity. This was in line with the spatial distribution of errors, as smallest errors were found in regions with few tracks. For the click model, periods of over- and underpredictions alternated over the season with highest absolute errors occurring in the middle of the season where highest click counts also occurred. The time series of click prediction errors suggests that there might be other factors at play that are not captured by the model, which leads to a temporal clustering of over- and underprediction. These factors may include weather variables that are not currently in the model (e.g., wind), or school holidays, which vary across cantons or even individual municipalities in Switzerland.