Marc Ruesch

Snow height measurements are still the backbone of any snow cover monitoring, whether based on modeling or remote sensing. These ground-based measurements are often realized using ultrasonic or laser technologies. In challenging environments, such as high alpine regions, the quality of sensor measurements deteriorates quickly, especially under extreme weather conditions or ephemeral snow conditions. Moreover, the sensors by their nature measure the height of an underlying object and are therefore prone to returning other information, such as the height of vegetation, in snow-free periods. Quality assessment and real-time classification of automated snow height measurements are therefore desirable for providing high-quality data for research and operational applications. To this end, we propose CleanSnow, a machine learning approach to automated classification of snow height measurements into a snow cover class and a class corresponding to everything else, which takes into account both the temporal context and the dependencies between snow height and other sensor measurements. We created a new dataset of manually annotated snow height measurements, which allowed us to train our models in a supervised manner and quantitatively evaluate our results. Through a series of experiments and ablation studies to evaluate feature importance and compare several different models, we validated our design choices and demonstrated the importance of using temporal information together with information from auxiliary sensors. CleanSnow achieves a high accuracy of almost 98 % and represents a new baseline for further research in the field. The presented approach to snow height classification finds its use in various tasks, ranging from snow modeling to climate science.

Ground-based snow depth measurements are often realized using ultrasonic or laser technologies, which by their nature measure the height of any underlying object, whether it is snow or vegetation in snow-free periods. We propose a machine learning approach to the automated classification of snow depth measurements into a snow cover class and a class corresponding to everything else, which takes into account both the temporal context and the dependencies between snow depth and other sensor measurements. Through a series of experiments we demonstrate that our approach simplifies the detection of seasonal snowmelt and corresponding onset of plant growth, which we used to assess climate-change related phenological shifts in otherwise rather poorly monitored high alpine regions.

In recent years, the integration of physical snowpack models coupled with machine-learning techniques has become more prevalent in public avalanche forecasting. When combined with spatial interpolation methods, these approaches enable fully data- and model-driven predictions of snowpack stability or avalanche danger at any given location. This prompts the question: Are such highly detailed spatial model predictions sufficiently accurate for operational use? To explore this, we assess the performance of interpolated, model-based predictions of snowpack stability and avalanche danger, comparing them to human-generated public avalanche forecasts during the 2023/2024 winter season in Switzerland. To do so, we compare human forecasts and model predictions for locations in avalanche terrain (considering coordinates, aspect, elevation) where skiers triggered avalanches (244 events) or which were skied but where no avalanche was triggered (non-events, 3177 data points from GPX tracks). While this data reflects human behavior to some extent, we consider the event ratio as a proxy for the probability of avalanche release due to human load. We observed that with increasing model-predicted danger level or decreasing model-predicted snowpack stability, the event ratio increased. Comparing model predictions with human-made forecasts showed that the predictive performance of two operationally used models was similar to the performance of human avalanche forecasts: both predicted a strong increase in the probability of human-triggered avalanches. In summary, our results indicate that models capture regional patterns of snowpack (in)stability or avalanche danger well, and that these model chains should therefore be systematically integrated in the forecasting process.

Machine learning (ML) models are revolutionizing avalanche forecasting, providing increasingly reliable predictions. At present, however, such model output can hardly be integrated into the traditional operational forecast workflow. To close this gap, the Swiss Avalanche Warning Service has developed the Snow Avalanche Forecast Editor (SAFE), allowing a flexible integration of both human forecasts and model predictions. In SAFE, each hazard source (e.g., an avalanche problem) is assessed and described as a separate layer. This allows an entirely independent characterization of the severity of avalanche problems including their spatial extent, in line with the European Avalanche Warning Services standards. Once the assessment has been completed, all existing layers are automatically combined and published as regional avalanche forecasts. Forecast accuracy can be increased by minimizing noise. Decision theory advises to calculate an expected value from multiple independent assessments. For this reason, each of the two or three forecasters on duty assesses the regional avalanche danger in the forecast domain independently of each other. Another estimate comes from the danger-level model. To combine these, forecasters and models must assess the same hazard source using the same target variable. SAFE merges these independent assessments to expected values of the danger level including sub-levels, and the particularly affected elevations and aspects. This includes clustering into a limited number of danger regions, characterized by similar expected avalanche conditions. Following the discussion between forecasters, clusters can be refined. Lastly, the avalanche danger of each cluster is manually described. SAFE is conceived as open-source platform and a first version has been operational within the Swiss avalanche warning since winter 2023/24.

The step towards interactively displaying the avalanche bulletins twice a day in four languages raised the need for specifically designed software tools to support the avalanche forecasters in their daily work of producing the avalanche bulletin and make them publicly accessible. Due to the various forms of visualization and location specific bulletins, the final products (printed reports, interactive web-pages, interactive mobile applications) cannot be produced manually. Data entry graphical user interfaces are required to provide the avalanche forecasters with a platform to define and describe the current situation and store the abstracted data into a database. Based on this abstracted data, other software tools will then produce the intermediate (maps, web-page fragments, text blocks) and final products. Finally, web technology and mobile applications present the generated avalanche bulletin in an interactive form to the end user. We present a technical insight into the specifically developed software required to generate and publish the avalanche bulletin, starting with the data entry and ending with the visualization of the final products to the end user.