Assessment of temporal probability for rainfall-induced landslides based on nonstationary extreme value analysis

Climate change may alter the frequency and intensity of rainfall, and thereby significantly affect the frequency and magnitude of shallow landslides. To accurately evaluate the temporal probability of landslide initiation, it is therefore important to consider the effect of climate variation. Although various approaches have been proposed to estimate the temporal probability of landslides to date, most of them were based on the stationary assumption, i.e., that the statistical properties of the historical rainfall data are time-invariant. However, if historical rainfall data show nonstationary characteristics such as a trend or an abrupt change, the stationary assumption is no longer valid and induces a miscalculation. In this study, we propose a new approach that can estimate the temporal probability of future rainfall-induced landslide occurrence while incorporating the nonstationary characteristics of the rainfall data. In assessing such data, a nonstationary generalized extreme value distribution was used to evaluate the temporal probability. Then, by combining the derived nonstationary temporal probability with landslide susceptibility results obtained from the random forest model, probabilities of landslide occurrence were calculated for future periods, from 1 to 50 years, and compared with the results based on a stationary model. The results showed that the stationary model underestimated the landslide probability compared with the nonstationary approach. This is because an increasing trend in local rainfall, taken from a gauge in the study area, was not considered in the stationary analysis. Thus, climate change has ongoing consequences for landslide occurrence. To reflect the impacts of climate change, a nonstationary approach capable of coping with climate variation should therefore be considered in any landslide hazard analysis.