Supervised and unsupervised machine learning techniques to map seismic site amplification

Switzerland experienced earthquakes mainly in the Basel area and within its Alpine region, with the Canton Valais standing out as one of the most active zones. The Rhone Valley, crossing the entire canton, is characterized by sediment deposits with a thickness reaching up to 800 meters. The valley’s topography and the significant contrast in seismic wave velocities between sediments and the surrounding rock, make it susceptible to 2D/3D effects, leading to significant site amplification phenomena. To develop local amplification models that integrate geological and geophysical data, specific areas – of relevance from the risk point of view - in the Rhone Valley were selected. One area is Sion, where geophysical data were acquired during the earthquake risk model for Switzerland project (ERM-CH). The dataset encompasses 313 single station noise measurements and seismic records from 10 seismic stations. The single station measurements were employed to compute horizontal to vertical spectral ratios (HVSR), while earthquake recordings were utilized to derive empirical spectral modelling amplification functions (ESM). Our approach involves the application of the canonical correlation (CC) statistical method, which explores the correlation between two sets of variables by identifying linear combinations that exhibit maximum correlation. Specifically, we conducted CC analysis between the sets of HVSR and ESM using as calibration dataset of 172 free-field and urban free-field stations run by the Swiss Seismological Service over the entire Swiss territory. Using canonical correlation, we developed a method to predict the ESM for a specific site based on its HVSR information. Additionally, we employed a correlation analysis based on the Pearson cross-correlation coefficient as an alternative method. This approach was utilized to group the Sion HVSR, with the seismic station HVSR, for which ESM is available, serving as the centroid. This grouping resulted in the assignment of each of the 313 HVSRs to one of the 10 amplification functions. Consequently, we extrapolated amplification values to various locations and employed kriging for interpolation to generate amplification maps at specific frequencies. The utilization of different amplification models at defined frequencies allows for the assessment and definition of epistemic uncertainties in our findings.