Roof-Snow-Falling Hazard Estimation Using Machine Learning and SNOWPACK Outputs with Snow-Disaster Data

YukioroSignal was developed to mitigate house damage and accidents during snow removal, which is the most common cause of casualties in snow and ice disasters. It provides the spatial distribution of snow weight over a wide area to help people make decisions about roof snow removal. However, the current YukioroSignal cannot estimate the danger level of roof snow falling. SNOWPACK, a numerical snowpack model used in the YukioroSignal system, calculates the detailed layer structure of snow cover, and therefore has the potential to estimate this hazard. In this study, we attempted to develop a machine learning method for estimating roof-snow-falling hazard using the snow stratigraphy simulated by SNOWPACK and information on roof-snow-falling accidents from the snow and ice disaster database. Individual roof-snow-falling cases were classified into wet type and dry type based on the SNOWPACK simulation results. The hazard model was constructed using the past roof-snow-falling accident data as training data. As a result, the model showed a certain level of accuracy for both wet type and dry type during the fitting stage. In contrast, the evaluation conducted separately on training and test data showed some accuracy only for wet-type events and indicated the need for improvement for practical use. Based on the constructed machine learning model, a prototype of a roof-snow-falling hazard distribution estimation system was developed in conjunction with YukioroSignal and used to provide hazard distributions during the winter of fiscal year 2024.