Item: MACHINE LEARNING AND PATTERN RECOGNITION FOR PUBLIC AVALANCHE FORECASTING
-
-
Title: MACHINE LEARNING AND PATTERN RECOGNITION FOR PUBLIC AVALANCHE FORECASTING
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Andrew Schauer [ Chugach National Forest Avalanche Center, Girdwood, AK, USA ]
- John Sykes [ Chugach National Forest Avalanche Center, Girdwood, AK, USA ] [ Simon Fraser University Avalanche Research Program, Burnaby, BC, Canada ]
Date: 2026-09-28
Abstract: Avalanche forecasters in the U.S. and Canada use the five-level North American Public Avalanche Danger Scale (NAPADS) as a tool to communicate avalanche risk to backcountry recreationists. Forecasters consider the avalanche problem type, distribution, sensitivity, size, and likelihood to assign an appropriate danger rating for a given forecast period. While there is ongoing progress in standardizing this decision-making process, there remains a large degree of subjective judgment based on professional experience. This project develops two novel decision support tools for assigning daily danger ratings. The first incorporates weather station data, previous avalanche forecasts, and reassessed avalanche forecast data to identify established patterns in public avalanche forecasts and predict daily danger ratings. We assess the performance of Extreme Gradient Boosting (XGB) and Super-Organizing Maps (SOM) and compare the predicted danger ratings using the machine learning models to those issued by human forecasters. We find 73-79% agreement between the models and professional forecasters, with the XGB model outperforming the SOM. The second tool uses a national record of avalanche forecasts to identify relationships between avalanche size, likelihood, and assigned danger rating, fully implementing the workflow described in the Conceptual Model of Avalanche Hazard (CMAH), and clearly defining the relationship between avalanche size/distribution/sensitivity and predicted danger rating. The final products of this research are (1) a machine learning tool that uses weather and avalanche variables to predict the avalanche danger rating for an avalanche center in Southcentral Alaska, and (2) an online interactive tool to visually guide the user through the CMAH workflow, linking avalanche size and likelihood to the predicted danger rating using data from 24 avalanche centers across the U.S. Both of these forecasting tools incorporate visual representations of uncertainty in the predicted danger rating based on historical records. These tools aim to assist backcountry avalanche forecasters in predicting avalanche danger ratings and quantifying uncertainty in the assigned rating. We believe this work will leverage modern machine learning methods and archived forecast data to capture the institutional knowledge of avalanche centers, which can support the decisions of newer forecasters and help seasoned practitioners identify personal biases in their forecasts.
Object ID: ISSW2026_P2.39.pdf
DOI: https://doi.org/10.15788/1790098992
Language of Article: English
Presenter(s): Andrew Schauer
Keywords: Machine Learning, Decision Support, CMAH, NAPADS
Page Number(s): 1402 - 1407
-