Automatic Detection, and P- and S-wave Picking Algorithm: an application to the 2009 L'Aquila (Central Italy) earthquake sequence

In order to process the enormous amount of digital waveforms continuously recorded at permanent and temporary seismic stations in Italy as quickly as possible, we implemented a semi-automatic procedure in order to identify local earthquakes and to provide consistently-weighted P- and S-wave arrival times. Local earthquake detection is obtained by a STA/LTA ratio-based algorithm applied to 3-component seismograms from individual stations. A minimum of 4 triggered stations are required to declare a seismic event. This setting proves to be extremely effective to detect a very large number of very low magnitude earthquakes (ML>1.5) with a small number of false alarms. The automatic picking system Mannekenpix (Aldersons, 2004), originally working on vertical component data, has been improved to tackle 3-component data. In order to increase the reliability of P-wave and S-wave picking, the system is now virtually capable of discriminating P-wave samples and S-wave samples, among noise samples. This Identification is performed by a C5 decision tree (Quinlan, 1993) derived from training data. Five groups of predicting variables are included: Energy, Polarization, Spectral Power, Skewness and Kurtosis. In addition, the SEDSL algorithm (Magotra et al., 1989) is also used as a predictor. The picking procedure requires a preliminary calibration derived from a reference subset of high-quality manual picks. After calibration, the picking system is statistically able to mimic the picking by a human analyst and to provide consistent uncertainty estimates translated into picking weights. We illustrate very satisfying results of the successful automatic procedure showing P- and S-phase automatic readings for the L’Aquila sequence. These readings are fully comparable to those of a good human analyst allowing high quality earthquake locations of many low-magnitude events in an extremely short space of time. Within a day of continuous recordings, we obtain around 2600 triggers, 75% of these are very high quality locatable events (mean RMS 0.1s) with at least 25 phase readings. References Aldersons, F. (2004) Toward a Three-Dimensional Crustal Structure of the Dead Sea region from Local Earthquake Tomography. Ph.D. thesis, Tel Aviv University. Magotra, N., Ahmed, N. and Chael, E. (1989) Single-Station Seismic Event Detection and Location IEEE Transactions on Geoscience and Remote Sensing, 27, 1, 15-23. Quinlan, J.R. (1993) C 4.5: Programs for machine learning. The Morgan Kaufmann Series in Machine Learning, San Mateo, CA.