Forecasting avalanche danger: human-made forecasts vs. fully automated model-driven predictions

Abstract. 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 detailed spatial model predictions sufficiently accurate for use in operational avalanche forecasting? We evaluated the performance of three spatially-interpolated, model-driven forecasts of snowpack stability and avalanche danger by comparing them with human-generated public avalanche forecasts in Switzerland over two seasons as benchmark. Specifically, we compared the predictive performance of model predictions versus human forecasts using observed avalanche events (natural or human-triggered) and non-events. To do so, we calculated event ratios as proxies for the probability of avalanche release due to natural causes or due to human load, given either interpolated model output or the human-generated avalanche forecast. Our findings revealed that the event ratio increased strongly with rising predicted probability of avalanche occurrence, decreasing snowpack stability, or increasing avalanche danger. Notably, model predictions and human forecasts showed similar predictive performance. In summary, our results indicate that the investigated models captured regional patterns of snowpack stability or avalanche danger as effectively as human forecasts, though we did not investigate forecast quality for specific events. We conclude that these model chains are ready for systematic integration in the forecasting process. Further research is needed to explore how this can be effectively achieved and how to communicate model-generated forecasts to forecast users.