{"title":"Machine learning-based identification and assessment of snow disaster risks using multi-source data: Insights from Fukui prefecture, Japan","authors":"Zhenyu Yang , Hideomi Gokon , Qing Yu","doi":"10.1016/j.pdisas.2025.100426","DOIUrl":null,"url":null,"abstract":"<div><div>Understanding the driving factors behind snowstorm risk and their nonlinear effects is critical for developing effective response strategies. This study, focusing on the 2018 Fukui snowstorm in Japan, integrates multi-source data, including mobile GPS data, Digital Elevation Model (DEM) data, road data, urban data, and traffic congestion data, to develop an interpretable model for quantifying high-risk areas and examining key nonlinear relationships and threshold effects influencing snowstorm impact occurrence, offering actionable insights for mitigation strategies. We employed four machine learning models—Decision Tree, Random Forest, Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost)—to capture complex nonlinear relationships among influencing factors and applied SHAP (SHapley Additive exPlanations) theory to interpret variable contributions. The results reveal that: (1) compared to Random Forest, Decision Tree, and MLP models, the XGBoost model demonstrates superior performance with a prediction accuracy of 0.8225; (2) factors such as elevation, slope, road density, and road width exhibit significant nonlinear impacts and threshold effects on snowstorm impact occurrence; (3) Urban areas with elevation below <span><math><mn>51.9</mn><mspace></mspace><mi>m</mi></math></span>, slopes exceeding <span><math><msup><mn>9.9</mn><mo>°</mo></msup></math></span>, a density of major roads (Road Type 1) less than <span><math><mn>443.75</mn><mspace></mspace><mi>m</mi><mo>/</mo><msup><mi>km</mi><mn>2</mn></msup></math></span>, a density of minor roads (Road Type 2) less than <span><math><mn>550.25</mn><mspace></mspace><mi>m</mi><mo>/</mo><msup><mi>km</mi><mn>2</mn></msup></math></span>, and where rural roads (Road Type 3) are nearly absent, along with population fluctuations ranging between <span><math><mfenced><mrow><mo>−</mo><mn>0.25,0</mn></mrow></mfenced></math></span>, are particularly vulnerable to snow disasters. In contrast, areas with flat terrain and high densities of rural roads are less likely to be affected; and (4) snow disaster resilience in mitigating traffic congestion can be improved by monitoring GPS data for early warnings and optimizing the sp. configuration of major and minor roads.</div></div>","PeriodicalId":52341,"journal":{"name":"Progress in Disaster Science","volume":"26 ","pages":"Article 100426"},"PeriodicalIF":2.6000,"publicationDate":"2025-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Progress in Disaster Science","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2590061725000237","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"ENVIRONMENTAL SCIENCES","Score":null,"Total":0}
引用次数: 0
Abstract
Understanding the driving factors behind snowstorm risk and their nonlinear effects is critical for developing effective response strategies. This study, focusing on the 2018 Fukui snowstorm in Japan, integrates multi-source data, including mobile GPS data, Digital Elevation Model (DEM) data, road data, urban data, and traffic congestion data, to develop an interpretable model for quantifying high-risk areas and examining key nonlinear relationships and threshold effects influencing snowstorm impact occurrence, offering actionable insights for mitigation strategies. We employed four machine learning models—Decision Tree, Random Forest, Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost)—to capture complex nonlinear relationships among influencing factors and applied SHAP (SHapley Additive exPlanations) theory to interpret variable contributions. The results reveal that: (1) compared to Random Forest, Decision Tree, and MLP models, the XGBoost model demonstrates superior performance with a prediction accuracy of 0.8225; (2) factors such as elevation, slope, road density, and road width exhibit significant nonlinear impacts and threshold effects on snowstorm impact occurrence; (3) Urban areas with elevation below , slopes exceeding , a density of major roads (Road Type 1) less than , a density of minor roads (Road Type 2) less than , and where rural roads (Road Type 3) are nearly absent, along with population fluctuations ranging between , are particularly vulnerable to snow disasters. In contrast, areas with flat terrain and high densities of rural roads are less likely to be affected; and (4) snow disaster resilience in mitigating traffic congestion can be improved by monitoring GPS data for early warnings and optimizing the sp. configuration of major and minor roads.
期刊介绍:
Progress in Disaster Science is a Gold Open Access journal focusing on integrating research and policy in disaster research, and publishes original research papers and invited viewpoint articles on disaster risk reduction; response; emergency management and recovery.
A key part of the Journal's Publication output will see key experts invited to assess and comment on the current trends in disaster research, as well as highlight key papers.