Data and code working together is fundamental to machine learning (ML), but the context around datasets and interactions between datasets and code are in general captured only rudimentarily. Context such as how the dataset was prepared and created, what source data were used, what code was used in processing, how the dataset evolved, and where it has been used and reused can provide much insight, but this information is often poorly documented. That is unfortunate since it makes datasets into black-boxes with potentially hidden characteristics that have downstream consequences. We argue that making dataset preparation more accessible and dataset usage easier to record and document would have significant benefits for the ML community: it would allow for greater diversity in datasets by inviting modification to published sources, simplify use of alternative datasets and, in doing so, make results more transparent and robust, while allowing for all contributions to be adequately credited. We present a platform, Renku, designed to support and encourage such sustainable development and use of data, datasets, and code, and we demonstrate its benefits through a few illustrative projects which span the spectrum from dataset creation to dataset consumption and showcasing.

Wet-snow avalanches are triggered by the infiltration of liquid water which weakens the snowpack. Wet-snow avalanches are among the most destructive avalanches, yet their release mechanism is not sufficiently understood for a process-based prediction model. Therefore, we followed a data-driven approach and developed a random forest model, depending on slope aspect, to predict the local wet-snow avalanche activity at the locations of 124 automated weather stations distributed throughout the Swiss Alps. The input variables were the snow and weather data recorded by the stations over the past 20 years. The target variable was based on manual observations over the same 20-year period. To filter out erroneous reports, we defined the days with wet-snow avalanches in a stringent manner, selecting only the most extreme active or inactive days, which reduced the size of the dataset but increased the reliability of the target variable. The model was trained with weather variables and variables computed from simulated snow stratigraphy in 38$^\circ$ slopes facing the 4 cardinal directions. While model development and validation were done in nowcast mode, we also studied model performance in 24-hour forecast mode by using input variables computed from a numerical weather prediction (NWP) model. Overall, the performance was good in both nowcast and forecast mode (f1-score around 0.8). To assess model performance beyond the stringent definition of wet-snow avalanche days, we compared model predictions to wet-snow avalanche activity over the entire Swiss Alps, based on the raw data over 8 winters. We obtained a Spearman correlation coefficient of 0.71. Hence, our model represents a step toward the application of support tools in operational wet-snow avalanche forecasting.

Even today, the assessment of avalanche danger is by and large a subjective yet data-based decision-making process. Human experts analyse heterogeneous data volumes, diverse in scale, and conclude on the avalanche scenario based on their experience. Nowadays, modern machine learning methods and the rise in computing power in combination with physical snow cover modelling open up new possibilities for developing decision support tools for operational avalanche forecasting. Therefore, we developed a fully data-driven approach to assess the regional avalanche danger level, the key component in public avalanche forecasts, for dry-snow conditions in the Swiss Alps. Using a large data set of more than 20 years of meteorological data measured by a network of automated weather stations, which are located at the elevation of potential avalanche starting zones, and snow cover simulations driven with these input weather data, we trained two random forest (RF) classifiers. The first classifier (RF 1) was trained relying on the forecast danger levels published in the official Swiss avalanche bulletin. To reduce the uncertainty resulting from using the forecast danger level as target variable, we trained a second classifier (RF 2) that relies on a quality-controlled subset of danger level labels. We optimized the RF classifiers by selecting the best set of input features combining meteorological variables and features extracted from the simulated profiles. The accuracy of the models, i.e. the percentage of correct danger level predictions, ranged between 74 % and 76 % for RF 1 and between 72 % and 78 % for RF 2. We assessed the accuracy of forecasts with nowcast assessments of avalanche danger by well-trained observers. The performance of both models was similar to the agreement rate between forecast and nowcast assessments of the current experience-based Swiss avalanche forecasts (which is estimated to be 76 %). The models performed consistently well throughout the Swiss Alps, thus in different climatic regions, albeit with some regional differences. Our results suggest that the models may well have potential to become a valuable supplementary decision support tool for avalanche forecasters when assessing avalanche hazard.

<p>Avalanche hazard forecasting is essential to reduce the risk for people and infrastructure in mountain areas. Among the different types of avalanches, wet-snow avalanches are particularly challenging to predict due to the poor understanding of their release mechanism. We therefore trained a random forest model to predict wet-snow avalanche activity based on weather and snow measurements and downstream SNOWPACK simulations provided by automated weather stations. The model was trained on a database covering 20 years of avalanche observations (avalanche type, size, location, slope aspect) in the context of operational avalanche forecasting in Switzerland. The prediction performance (F1-score: harmonic mean between recall and precision) for wet-snow avalanche active days is around 76% (recall: 73%, precision: 80%), and is 99% for days with no activity.  The model not only well reproduced the onset, but also the end of wet-snow avalanche periods. Operational testing during winter 2021-2022 allow to evaluate differences in model performance between nowcast derived from meteorological measurements and forecast from numerical weather prediction models. Overall, the results are promising and are an important step forward a more reliable forecast of wet-snow avalanche activity.</p>

Avalanche forecasting implies predicting current and future snow instability in time and space. In Switzerland, avalanche bulletins are issued daily during the winter season to warn the public about the avalanche hazard, described by region with one of five danger levels. Assessing avalanche danger is by large a data-driven, yet experience-based decision-making process. It involves analysing a multitude of data diverse in scale – time and space, and concluding by expert judgment on the avalanche scenario. Numerous statistical models were developed in the past, but rarely applied due to limited usefulness in operational forecasting. Modern machine learning techniques open up new possibilities for developing support tools for operational avalanche forecasting. With this aim, we developed a data-driven approach based on the supervised Random Forest (RF) classifier to automatically predict the danger level for dry-snow avalanche conditions in the Swiss Alps. A large database of more than 20 years of meteorological data and modelled snow stratigraphy data obtained with the numerical snow cover model SNOWPACK were used to train the RF algorithm. We optimized the model and selected the best set of input features that combine meteorological variables and features extracted from the simulated profiles, resampled at the same daily resolution as the forecasts. Our target variable was the regional danger level forecast in the public bulletin. We evaluated the predictive performance of the RF model with an independent test set with data of two winter seasons (2018-2019 and 2019-2020). The test set accuracy was 72 %, which is slightly lower than the accuracy estimate of the public forecasts (about 76 %). Given this uncertainty in our target variable, we trained an optimized RF model on a subset containing so-called verified avalanche danger levels. The test set accuracy then increased to 80 %. During the winter season 2020-2021, both RF models were tested in operational setting and automatically predicted a ‘nowcast’ and a ‘forecast’ in real-time. In parallel, we also tested a deep recurrent neural network model, which used a 7-days time series with 3-hours time resolution as input and also predicted the avalanche danger level. We present a comparison of the performance of the three models. This is one of the first times that a data-driven approach is tested in real-time as a feasible tool for operational avalanche forecasting.