Evaluation of Multivariate Adaptive Regression Splines for Prediction of Kappa Factor in Western Turkiye

Abstract

The recent seismic activity on the west coast of Turkiye, including the Aegean Sea region, indicates that a closer focus is necessary on this region. Located in an active tectonic regime of north-south extension with multiple basins on soft soil deposits, the region has a high seismic hazard. Recently, as a combination of basin effects and building vulnerability, the October 30, 2020, Samos event (Mw = 7.0) caused localized significant damage and collapse in Izmir city center despite the 70 km distance from the earthquake source. In spite of this activity, studies on site characterization and site response modeling, including local velocity models and kappa estimates, are still limited in this region. Kappa values exhibit regional characteristics, which necessitates local kappa estimates from past earthquake data for use in region-specific applications. To make the prediction, we used three-component strong ground motion records from accelerometer stations with known VS30 values in western Turkiye that are a part of the Disaster and Emergency Management Presidency's Turkish National Strong Ground Motion Observation Network. Multiple linear regression (MLR) and multivariate adaptive regression splines (MARS) algorithms have been implemented to build the prediction model. Three factors, such as distance, magnitude, and site class, are included in the kappa evaluation process. The performance of the models in kappa evaluation is calculated based on wellknown accuracy measures. The MARS model showed better performance compared to MLR over the selected sites concerning all performance measures. This finding may challenge the most commonly assumed linear models of kappa in the literature.