Data-driven Avalanche Forecasting - Using automatic weather stations to build a data-driven decision support system for avalanche forecasting
MetadataShow full item record
In this paper, a decision support system for avalanche forecasting based on data from automatic weather stations is developed and tested. 17 years of avalanche and weather observations from Senja in Northern Norway are processed and analyzed to identify meteorological factors important for avalanche formation. Current snow depth and precipitation the preceding days of an avalanche release are found to be most important. Further, a simple model based purely on snow depth, a logistic regression model and a random forest model are fitted to training data and used to forecast the probability of an avalanche on test data. The results show that the logistic regression model and random forest model performs better than the simple snow depth model. Random forest is able to detect 12 out of 19 avalanches, obtaining a true skill score of 0.6. This is better than logistic regression that detects 9 out of 19, obtaining a true skill score of 0.43. The study shows that it is possible to develop a decision support system for avalanche forecasting using already existing infrastructure. However, the results also shows that the models have their limitations. Many avalanches are not detected, and hence, a system based on these models should only act as decision support system and not be relied on solely. At last, a prototype is developed and tested live. Live testing showed that reliability of the weather stations in use is important for operational usage.