Abstract
Rapid assessment of spatial distribution of earthquake-induced landslides could provide valuable information in the emergency response phase. Previous studies proposed global analyses with the aim of predicting earthquake-induced landslide distributions in near real-time. However, in all those studies, mapping units are constituted by pixels, which do not reflect homogeneously distributed physical property for a given terrain unit and whose size do not match the resolution of existing thematic data at global scale. Moreover, none of the existing analyses considers sampling balance between different inventories or categorizing the inventories to construct a training set with higher statistical representativeness. We develop an improved global statistical method to address these drawbacks. We use slope units, which are terrain partitions attributed to similar hydrological and geomorphological conditions and to processes that shape natural landscapes. A set of 25 earthquake-induced landslide-events are selected and categorized based on the similarity between causal factors to determine the most relevant training set to make a prediction for a given landslide-event. As a result, we develop a specific model for each category. We sample an equal number of landslide points from each inventory to overcome the dominance of some inventories with large landslide population. We use seven independent thematic variables for both categorizing the inventories and modeling, based on logistic regression. The results show that categorizing landslide-events introduces a remarkable improvement in the modeling performance of many events. The categorization of existing inventories can be applied within any statistical, global approach to earthquake-induced landslide events. The proposed categorization approach and the classification performance can be further improved with the acquisition of new inventory maps.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have