Assessment of the uncertainty in spatial-correlation models for earthquake ground motion due to station layout and derivation method

The evaluation of the aggregate risks to spatially distributed infrastructures and portfolios of buildings requires quantification of the estimated shaking over a region. To characterize the spatial dependency of ground motion intensity measures (e.g. peak ground acceleration), a common geostatistical tool is the semivariogram. Over the past decades, different fitting approaches have been proposed in the geostatistics literature to fit semivariograms and thus characterize the correlation structure. A theoretically optimal approach has not yet been identified, as it depends on the number of observations and configuration layout. In this article, we investigate estimation methods based on the likelihood function, which, in contrast to classical least-squares methods, straightforwardly define the correlation without needing further steps, such as computing the experimental semivariogram. Our outcomes suggest that maximum-likelihood based approaches may outperform least-squares methods. Indeed, the former provides correlation estimates, that do not depend on the bin size, unlike ordinary and weighted least-squares regressions. In addition, maximum-likelihood methods lead to lower percentage errors and dispersion, independently of both the number of stations and their layout as well as of the underlying spatial correlation structure. Finally, we propose some guidelines to account for spatial correlation uncertainty within seismic hazard and risk assessments. The consideration of such dispersion in regional assessments could lead to more realistic estimations of both the ground motion and corresponding losses.

Automated summary

This study examines the application of geostatistical methods based on the likelihood function in estimating the spatial correlation of ground motion intensity measures. Maximum-likelihood methods may provide correlation estimates that are not dependent on the bin size and could outperform least-squares methods.