Performance of Different ANNs in Microtremor H/V Peak Classification

Abstract

Abstract The microtremor horizontal-to-vertical (H/V) technique is extensively used to both assess the seismic amplification potential of soils and, in combination with other surface wave-based techniques, to reconstruct the near-surface seismic stratigraphy. The H/V peak frequencies are traditionally interpreted in terms of soil resonances, and, in this case, they are assigned a “stratigraphic origin.” However, not all H/V peaks mark subsoil resonances, and some of them have “anthropic” or “artefactual” origin. Recognizing the nature of H/V peaks should be mandatory before any stratigraphic interpretation. Nonetheless, this problem is not given the attention it deserves. Because this classification is not easy to achieve using standard statistical techniques, we decided to train two supervised neural networks: a traditional artificial neural network using a set of input values extracted from the individual (horizontal and vertical) microtremor spectra and a convolutional neural network working on images of the microtremor spectra. The nets were trained on an Italian dataset and tested on a U.S. dataset, collected by different operators and with different instruments. Both the nets achieved a classification accuracy of ∼90%; however, the convolutional one showed a greater generalization capability compared to the traditional one. Such machine learning algorithms can be useful tools to discriminate the origin of H/V peaks, complementing the traditional SESAME guidelines, which do not go into much detail on this topic.