Improvement of HVSR technique by self-organizing map (SOM) analysis

Data interpretation is one of the most important and thorny tasks in geosciences. Difficulties occur especially in non-invasive geophysical techniques and/or when the data that have to be analyzed are multidimensional, non-linear and highly noisy. Another important task is to ensure an efficient automatic data analysis, in order to allow a data interpretation as independent as possible from any a priori knowledge. This paper describes the post-processing application of a kind of neural network (self-organizing map, SOM) to the identification of the fundamental HVSR frequency of a given site. SOM results can be represented as two-dimensional maps, with a non-parametric mapping that projects the high dimensional original dataset in a fashion that provides both an unsupervised clustering and a highly visual representation of the data relationships. This innovative application of the SOM algorithm is presented with a case study related to the characterization of a mineral deposit.