Multi-source geographically weighted regression for regionalized ground-motion models

This work proposes a novel approach to the calibration of regionalized regression models, with particular reference to ground-motion models (GMMs), which are key for probabilistic seismic hazard analysis and earthquake engineering applications. A novel methodology, named multi-source geographically-weighted regression (MS-GWR), is developed, allowing one to (i) estimate regionalized regression models depending on multiple sources of non-stationarity (such as site- and event-dependent non-stationarities in GMMs), and (ii) make inference on the significance and stationarity of the regression coefficients. Unlike previous approaches to the problem, the proposed framework is non-parametric – in the sense of the distribution of the errors – the inference being based on a permutation scheme. MS-GWR is here used to calibrate a new regionalized ground-motion model for predicting peak ground acceleration in Italy, based on a large scale database of waveforms and metadata made available by the Italian Institute for Geophysics and Vulcanology (INGV).