The optimal rainfall thresholds and probabilistic rainfall conditions for a landslide early warning system for Chuncheon, Republic of Korea

Abstract

AbstractThe purpose of this study is to establish the criteria for a landslide early warning system (LEWS). We accomplished this by deriving optimal thresholds for the cumulative event rainfall–duration (ED) and identifying the characteristics of the rainfall variables associated with a high probability of landslide occurrence via a Bayesian model. We have established these system criteria using rainfall and landslide data for Chuncheon, Republic of Korea. Heavy rainfall is the leading cause of landslides in Chuncheon; thus, it is crucial to determine the rainfall conditions that trigger landslides. Hourly rainfall data spanning 1999 to 2017 from seven gauging stations were utilized to establish the ED thresholds and the Bayesian model. We used three different calibration periods of rainfall events split by 12, 24, 48, and 96 non-rainfall hours to calibrate the ED thresholds. Finally, the optimal threshold was determined by comparing the results of the contingency table and the skill scores that maximize the probability of detection (POD) score and minimize the probability of false detection (POFD) score. In the LEWS, by considering the first level as “normal,” we developed subsequent step-by-step warning levels based on the Bayesian model as well as the ED thresholds. We propose the second level, “watch,” when the rainfall condition is above the ED thresholds. We then adopt the third level, “warning,” and the fourth level, “severe warning,” based on the probability of landslide occurrence determined via a Bayesian model that considers several factors including the rainfall conditions of landslide vs. non-landslide and various rainfall variables such as hourly maximum rainfall and 3-day antecedent rainfall conditions. The proposed alert level predicted a total of 98.2% of the landslide occurrences at the levels of “severe warning” and “warning” as a result of the model fitness verification. The false alarm rate is 0% for the severe warning level and 47.4% for the warning level. We propose using the optimal ED thresholds to forecast when landslides are likely to occur in the local region. Additionally, we propose the ranges of rainfall variables that represent a high landslide probability based on the Bayesian model to set the landslide warning standard that fits the local area’s characteristics.