Production of a homogeneous seismic catalog based on machine learning for northeast Egypt

Abstract

This research presents a new approach which addresses the conversion of earthquake magnitude as a supervised machine-learning problem through a multistage approach. First, the moment magnitude (M w) calculations were extended to lower magnitude earthquakes using the spectral P-wave analyses of the vertical component seismograms to improve the scaling relation of M w and the local magnitude (M L) of 138 earthquakes in northeastern Egypt. Second, using unsupervised clustering and regression analysis, we applied the k-means clustering technique to subdivide the mapped area into multiple seismic activity zones. This clustering phase created five spatially close seismic areas for training regression algorithms. Supervised regression analysis of each seismic area was simpler and more accurate. Conversion relations between M w and M L were calculated by linear regression, general orthogonal regression (GOR), and random sample consensus (RANSAC) regression techniques. RANSAC and GOR produced better results than linear regression, which provides evidence for the effects of outliers on regression accuracy. Moreover, the overall multistage hybrid approach produced substantial improvements in the measured-predicted dataset residuals when individual seismic zones rather than all datasets were considered. In 90% of the analyzed cases, M w values could be regarded as M L values within 0.2 magnitude units. Moreover, predicted magnitude conversion relations in the current study corresponded well to magnitude conversion relations in other seismogenic areas of Egypt.