A framework for the quantitative assessment of new data streams in avalanche forecasting

Abstract

Data used by avalanche forecasters are typically collected using weather stations, manual field-based observations (e.g., avalanche events, snow profiles, stability tests, personal observations, public observations, etc.) and weather forecasts ("traditional observations"). Today, snow cover observations can be delivered via remote sensing (e.g., satellite data, UAV, TLS, time-lapse camera etc.). Forecasting operations can also use statistical forecasting, weather models, and physical modeling to support decisions. This paper presents a framework and methodology to quantify the impact these new, complex data streams have on the formulation of, and associated uncertainty of, avalanche forecasting. We use data from a case study in Norway. Avalanche forecasters in Norway assessed size (D), likelihood, avalanche problem, and hazard level for a highway corridor in Grasdalen, Stryn Norway. The control groups were given access to traditional observations. The experimental groups were given access to the same traditional data, but also near-real-time snow surface LiDAR data ("RS+"). In case study one the RS+ (n=10) consensus findings were a hazard level two steps lower than the control group (n=10). In case study two the traditional (n=10) and RS+ groups' (n=7) consensus findings assessed the northeastern avalanche path at the same hazard level. Assessing the southwestern slide path, the traditional group (n=10) and RS+ group (n=9) had the same consensus finding for hazard level. In 2 of 3 case studies, the RS+ groups had fewer selections for size, likelihood, and avalanche problem which indicates reduced uncertainty in their forecasts. Throughout the 2022-2023 winter season Norwegian Public Roads Administration avalanche forecasters performed a real-time experiment throughout the season - with and without additional RS+ data when forecasting. They agreed on hazard level in 6 of 10 forecasts. In the other 4 forecasts, RS+ forecasters assessed the hazard level higher than traditional data forecasts. When RS+ data reveals aspects of conditions that traditional observations did not detail, RS+ forecasters adjust their selections in the hazard matrix, resulting in greater clustering of their predictions, indicating reduced uncertainty. Due to uncertainty associated with avalanche forecasting, this framework for assessment should be used to track avalanche forecast efficacy and build a qualitative and quantitative historical record.