GNGTS 2017 - 36° Convegno Nazionale

632 GNGTS 2017 S essione 3.2 Application of a method to determine S and P wave velocities from surface waves data analysis in presence of sharp lateral variations F. Khosro Anjom 1 , A. Arabi 1 , L.V. Socco 1 , C. Comina 2 1 DIATI (Dipartimentoo di Ingegneria dell’Ambiente, del territorio e delle infrastrutture) Politecnico di Torino, Italy 2 Dipartimento di Scienze della Terra, Università di Torino, Italy Introduction. Seismic surface waves are becoming a leading method to estimate S-wave velocity (VS) through dispersion curve (DC) extraction and inversion (Socco et al. , 2010). Socco et al. (2017) showed that a relationship between wavelength and depth (W/D) of surface waves can be determined and used to directly transform the DC into the time-average VS model, avoiding the inversion process. Surface waves are usually considered unsuitable for estimating P-wave velocity (VP), or Poisson�� ������ ��� �� ��� ��� ����������� �� ���������� ����� �� ����� ’s ratio, due to the low sensitivity of dispersion curve to these parameters (Nazarian,1984). Socco and Comina (2017) found that this W/D relationship is sensitive to Poisson�� ������ � ������ ��� ��������� ��������� �� ��������� ���� ��� ����� ’s ratio. A method was therefore developed to determine also the time- average VP by using the W/D relationship. These methods are robust in terms of accuracy and efficient in terms of computational cost. They can be applied to a set of dispersion curves over a seismic line to provide a pseudo 2D time-average VS and VP model along the line. The methods require a reference DC and its associated VS profile to get the W/D relationship. However, using a unique reference curve for the whole line becomes problematic when sharp lateral variations or zones with different velocities are present. In this work, a hierarchical clustering algorithm is developed to select ensembles of DCs within a uniform zone that can be interpreted using the same reference DC. Then, reference W/D relationships are estimated in each uniform zone and combined with DCs to determine time-average VS and VP models over the whole area. In the following, the clustering method is first briefly explained, then, the application of the method to field data from a test site (CNR in Turin) is shown and discussed. Methodology. Ahierarchical clusteringmethodhasbeenused togroup theDCs. Inhierarchical clustering, there is no need to set a predefined number of clusters, this is a great advantage since we want a method able to identify uniform zones without knowing a priori the expected lateral variability of the area. The result of hierarchical clustering is a dendrogram which shows the nested grouping of observations and similarity levels at which grouping change (Maimon and Rokach, 2000). Hierarchical methods can be divided into two subgroups (Maimon and Rokach, 2000): i) agglomerative - a bottom up approach in which every observation is considered a cluster and similar clusters are identified and merged together to make bigger clusters; ii) divisive - a top down approach in which the observations all start in a single cluster and at each step they are removed from the cluster to create a new cluster or join other clusters. Tests on synthetic and real data showed that agglomerative and divisive methods have similar performances for our purpose. We selected agglomerative clustering because it makes the process tracking easier. A measure of dissimilarity is needed to merge the clusters. This is usually achieved by appropriate metric and linkage criterion. Euclidean distance is used here to determine the metric between two dispersion curves: (1) where D is the distance, ν i and ν j are phase velocity of the two dispersion curves at different frequencies. Linkage criterion determines the distance between sets of observations or clusters. Average linkage, in which the distance between clusters is calculated based on the average distance between each component of one cluster to each component of the other cluster, is used here to maximize the contribution of all DCs in the clusters. After the DCs are grouped, in each group (cluster) the DC with the highest quality and broad