A nearest-neighbor model for regional avalanche forecasting

Abstract

AbstractThe application of nearest-neighbor algorithms to the task of regional avalanche forecasting in Switzerland is presented in this paper. The database used for the development of the model consists of snow and weather data from 60 manual weather stations and conventionally estimated avalanche-hazard levels. All these data are collected by the Swiss Federal Institute for Snow and Avalanche Research on a daily basis during winter. Data between 1987 and 1996 (10 winters) are used for our study. For the manual weather stations a nearest-neighbor model has been developed: NXD-VG calculates the 10 nearest-neighbor days by using a Euclidean weighted distance metric. A regional avalanche-hazard map is calculated by interpolating the results of NXD-VG between the stations. The avalanche forecasters can access the results of the model calculation directly because they are integrated into a program for bulletin construction. The model was validated using three complementary approaches. First, the database is cross-validated for all the winters available to estimate an unbiased prediction error of the models for two selected stations. Second, selected situations of the database are recalculated. Third, the model output is compared daily to the official forecasts published during winter 1999/2000.