Projection and identification of vulnerable areas due to heavy snowfall using machine learning and K-means clustering with RCP scenarios

Abstract

Heavy snowfall is a natural disaster that causes extensive damage in South Korea. Therefore, predicting heavy snowfall occurrence, identifying vulnerable areas, and establishing response plans to reduce risk are crucial. In this study, to project heavy snowfall, meteorological and geographic data from the past 30 years were collected, and four machine learning algorithms were trained and compared: multiple linear regression, support vector regression, random forest regressor (RFR), and extreme gradient boosting. We observed that the RFR model (R2 = 0.64) demonstrated the most optimal performance in projecting snowfall compared to other models. Representative concentration pathway (RCP) scenario data was input into the RFR model to generate projection data up to 2100. Projection results of more than 48.2 cm based on heavy snowfall events in the past 20 years were observed 17 times in RCP2.6, 19 times in RCP4.5, 16 times in RCP6.0, and 17 times in RCP8.5. The annual GIS-based projected snowfall images for the RCP8.5 scenario were classified into five distinct groups using K-means clustering. These groups were then further divided based on the vulnerability of regions, including Gangwon-do, Jeollabuk-do, and northern Gyeonggi-do. Our study can aid decision-making on policies related to heavy snowfall disaster prevention standards, snow removal plans, budgeting, and the establishment of mid- to long-term climate change adaptation plans for government, public institutions and private organizations.

Keywords

• Snowfall projection
• Machine learning
• GIS
• RCP
• Image clustering