Challenges in Sampling Extreme Events: A Case Study of Probabilistic Earthquake-Induced Liquefaction Hazard EvaluationThomas Oommen Empirical liquefaction models are the standard approach for predicting the occurrence of soil liquefaction. These models are typically based on in situ index tests, such as the Standard Penetration Test (SPT), Cone Penetration Test (CPT), and Shear Wave Velocity (Vs30) performed based on earthquake reconnaissance. In earthquake reconnaissance, the liquefaction instances are identified based on the spatial extent of some surficial expression such as sandboils or ground cracking etc. However, identifying the non-occurrence of liquefaction is more difficult because the absence of a surficial expression does not guarantee that liquefaction did not occur. Moreover, in reconnaissance it is common for the hazard event (i.e. instances of liquefaction) to be sampled much more frequently than the non-hazard event. This results in databases with class imbalance that have a large number of events from one class. In addition to class imbalance, the class ratio (i.e. ratio of hazard and non-hazard event) in the sample is often different from the population and this difference is referred to as sampling bias. Studies have shown that class imbalance and sampling bias critically affect the predictive capability of empirical models. In this study, using a rich dataset associated with the 17 January 1995 Hyogo-ken Nanbu earthquake (M = 6.9), I explore the impact of class imbalance and sampling bias for the predictability of liquefaction hazard and the spatial extent of observed liquefaction. The Hyogo-ken Nanbu earthquake is one of the best studied earthquakes both due to the extensive damage that resulted from the event and the broad data collection efforts that preceded and followed the event. Extensive geotechnical and damage data were collected after the event but have remained generally unavailable to the wider research community. The data include geotechnical, geographic, geological, and earthquake damage data. The geotechnical data include the Jibankun database with over 7000 borings with stratigraphic descriptions, related in-situ and laboratory tests, and velocity profiles. The geographic and geologic data include a digital elevation map, shoreline and water bodies, and a regional geologic map. The earthquake damage data include building damage, infrastructure damage, and liquefaction surface effects. The observed liquefaction surface effects from the earthquake indicate that a class imbalance exists in the existing liquefaction datasets that are used for model development. Moreover, the results indicate that the spatial extent of liquefaction should be considered in developing empirical liquefaction models. |
