Temporal clustering of precipitation for detection of potential landslides

Abstract

Landslides are complex phenomena that cause important impacts in vulnerable areas, including the destruction of infrastructure, environmental damage, and loss of life. The occurrence of landslide events is often triggered by rainfall episodes, single and intense ones or multiple ones occurring in sequence, i.e., clustered in time. Landslide prediction is typically obtained via process-based or empirical thresholds. Here, we develop a new approach that uses information on the temporal clustering of rainfall to detect landslide events and compare it with the use of classical empirical rainfall thresholds. In addition, we evaluate the performance of the two approaches combined together as a case study in the region of Lisbon in Portugal. We consider a dataset that categorizes landslides into shallow and deep events and a review of empirical rainfall thresholds that makes a good benchmark for testing our novel method. We show that the new approach based on temporal clustering overall has a good power of detecting landslide events but has a skill comparable with the classic rainfall threshold method. While there is no clear outperformance of one method, the novel clustering-based method has a higher sensitivity despite a lower precision than the threshold-based method. For all approaches, the potential detection is better for deep landslides than for shallow ones. The results of this study could help to improve the prediction of rainfall-triggered landslides.