Abstract
This study predicts the average 6-hourly number of mountain accidents using data from 18 mountainous national parks in Korea, including Jirisan, Seoraksan, and Sobaeksan. Specifically, to achieve both fine-grained prediction and identify important variables, we divide mountain regions into grids, enabling risk prediction at both the mountain level and the specific grid level. Additionally, a Hurdle model is applied to address zeroinflated data, as mountain accidents often do not occur in many regions due to sparse population or generally safe areas. The Hurdle model is implemented via a generalized linear model, random forest, and gradient boosting decision trees (XGBoost, LightGBM, and CatBoost). An extensive exploratory data analysis is also conducted to enhance prediction accuracy and validate our analytic approach. Through a feature importance analysis, we find that climate-related variables are important for predicting the probability of an accident, while geological factors (slope and elevation) and temporal information are key contributors to modeling the count of accidents.

https://doi.org/10.5351/KJAS.2025.38.4.531