Development of a Machine Learning Model for Predicting Casualties from Heavy Rainfall Disasters

By combining damage data from heavy rainfall disasters with rainfall characteristics in South Korea from 1999 to 2019, this study developed and evaluated a machine learning (ML) model to predict human damage. The damage and rainfall data were obtained from the Disaster Yearbook of the Ministry of the Interior and Safety and Korea Meteorological Administration’s ASOS records, respectively. The ML models employed included Random Forest, Decision Tree, and XGBoost, with the period from 1999 to 2015 and 2016 to 2019 designated as the training and validation phases, respectively. A comparison of the performance of various models revealed that XGBoost attained the highest level of accuracy, with a value of 0.974, followed by Random Forest and Decision Tree, with 0.963 and 0.932, respectively. While all models exhibited remarkably high specificity for negative predictions, ranging from 94.9% to 99.3%, their predictive performance for positive cases was comparatively low, with recall values ranging from 0.34 to 0.37. This phenomenon was attributed to an imbalance in the occurrence of human damage between the training and validation periods. The significance of not only rainfall characteristics but also damage amounts to facilities in casualty prediction models was corroborated. Subsequent studies have revealed that enhancing prediction accuracy can be achieved through data imbalance correction and the implementation of time series-based models.