How to enrich training data for machine learning-based landslide hazard prediction with spatio-temporal precipitation information?

Landslide hazard maps indicate the spatio-temporal likelihood of landslide occurrence based on the prevailing spatial precipitation patterns. In order to generate precipitation-informed hazard maps by means of Machine Learning, we need to gather precipitation information for both landslide and non-landslide sites. For landslide locations, precipitation data associated with the initiation of the events at these sites is used. For the non-landslide locations, precipitation data is typically sampled randomly in both space and time from a precipitation database. To ensure that suitable precipitation values are sampled even when there are fewer non-landslide locations where representative random sampling may not be feasible, we test three approaches to assign precipitation values in a less random manner. As a case study, we examine the shallow landslide hazard associated with the intense rainfall event in Switzerland in August 2005. We evaluate the effectiveness of the three approaches by analysing the variation in spatial hazard distribution and the overall size of hazardous areas across the daily hazard maps produced under different precipitation conditions as well as the temporal rate of the observed changes. We conclude that probabilistically determined precipitation values offer a promising alternative to enrich non-landslide locations with precipitation-information in landslide hazard prediction.