AUTOMATIC PICKING OF SEISMIC ARRIVALS IN LOCAL EARTHQUAKE DATA USING AN ARTIFICIAL NEURAL-NETWORK

A preliminary study is performed to test the ability of an artificial neural network (ANN) to detect and pick seismic arrivals from local earthquake data, This is achieved using three-component recordings by utilizing the vector modulus of these seismic records as the network input. A discriminant function, F(t), determined from the output of the trained ANN, is then employed to define the arrival onset. 877 pre-triggered recordings from two stations in a local earthquake network are analysed by an ANN trained with only nine P waves and nine noise segments. The data have a range of magnitudes (M(L)) from -0.3 to 1.0, and signal-to-noise ratios from 1 to 200. Comparing the results with manual picks, the ANN can accurately detect 93.9 per cent of the P waves and also 90.3 per cent of the S waves with a F(t) threshold set at 0.6 (maximum is 1.0), These statistics do not include false alarms due to other non-seismic signals or unusable records due to excessive noise, In 17.2 per cent of the cases the ANN detected false alarms prior to the event. Determining the onset times by using the local maximum of F(t), we find that 75.4 per cent of the P-wave estimates and 66.7 per cent of the S-wave estimates are within one sample increment (10 ms) of the reference data picked manually. Only 7.7 per cent of the P-wave estimates and 11.8 per cent of the S-wave estimates are inaccurate by more than five sample increments (50 ms). The majority of these records have distinct local P and S waves. The ANN also works for seismograms with low signal-to-noise ratios, where visual examination is difficult, The examples show the adaptive nature of the ANN, and that its ability to pick may be improved by adding or adjusting the training data. The ANN has potential as a tool to pick arrivals automatically. This algorithm has been adopted as a component in the early stages of our development of an automated subsystem to analyse local earthquake data, Further potential applications for the neural network include editing of poor traces (before present algorithm) and rejection of false alarms (after this present algorithm).