Investigations Into Non-Stationary Correlation Models for Site Terms of Ground-Motion Models due to Basins

Abstract

Spatial correlation models of ground motions are typically isotropic ad stationary, meaning the correlation of observations at different stations depends only on the absolute inter-station distance. This is a simplifying assumption. In this work, spatial correlation models of ground-motion model site terms are developed. Non-stationary effects due to basins are introduced into the models, first as a separation due to basin index, and then also dependent on a predictor variable, which n this case is the distance to the basin boundary. Different models are compared via the generalization error estimated by cross-validation. Unsuprisingly, spatially correlated models perform better than a non-spatial base models.
Models where the spatial correlation is separated inside and outside of a basin see a small improvement over a stationary models. Non-stationary models that depend on the distance to the basin boundary do not generally perform better. In this case, results depend slightly on the cross-validation scheme and the functional dependence on the predictor variable. The models oulined in this work provide a conceptual and computational framework to incorporate non-stationarity into spatial correlation models for ground motions.