N. Rabin1 , Y. Bregman2 , O. Lindenbaum3 , Y. Ben Horin2 , A. Averbuch3
1Afeka Academic College of Engineering, Tel Aviv, Israel
2Soreq Nuclear Research Center, Israel
3Tel Aviv University, School of Computer Science, Tel Aviv, Israel
Characterization of seismic events is an important component of the CTBT verification regime. Non-linear machine learning techniques are capable of compactly modeling complex datasets by using a local similarity metric. This process results in a low-dimensional representation of the dataset, in which each data item is characterized by a small number of intrinsic parameters. In this work, we apply a machine learning technique called diffusion maps for automatic earthquake-explosion discrimination and explosion classification. Diffusion maps construct a geometric representation of the seismograms that capture the intrinsic structure of the signal at each channel. As a pre-processing step, the seismograms are converted to normalized sonograms. In the obtained low-dimensional representation, seismic events with similar source mechanism from the same region have a similar representation. In addition, the single channel based classification method is extended to a multi-station one by introducing a kernel multiplication technique. This method extends the standard diffusion maps framework by providing a solution to handle multi-views and multi-source inputs. Our approach is demonstrated on several seismic data sets that are embedded and can also be visualized in a low-dimensional space. High accuracy discrimination results are achieved by using simple classification analysis methods in the low-dimensional space.