Automatic tremor detection with a combined cross-correlation and neural network approach

Low-amplitude, long-duration, and ambiguous phase arrivals associated with crustal tremor make automatic detection difficult. We present a new detection method that combines cross-correlation with a neural network clustering algorithm. The approach is independent of any a priori assumptions regarding tremor event duration; instead, it examines frequency content, amplitude, and motion products of continuous data to distinguish tremor from earthquakes and background noise in an automated fashion. Because no assumptions regarding event duration are required, the clustering algorithm is therefore able to detect short, burst-like events which may be missed by many current methods. We detect roughly 130 seismic events occurring over 100 minutes, including earthquakes and tremor, in a three-week long test data set of waveforms recorded near Cholame, California. The detection has a success rate of over 90% when compared to visually selected events. We use continuous broadband data from 13 STS-2 seismometers deployed from May 2010 to July 2011 along the Cholame segment of the San Andreas Fault, as well as stations from the HRSN network. The large volume of waveforms requires first reducing the amount of data before applying the neural network algorithm. First, we filter the data between 2 Hz and 8 Hz, calculate envelopes, and decimate them to 0.2 Hz. We cross-correlate signals at each station with two master stations using a moving 520-second time window with a 5-sec time step. We calculate a mean cross-correlation coefficient value between all station pairs for each time window and each master station, and select the master station with the highest mean value. Time windows with mean coefficients exceeding 0.3 are used in the neural network approach, and windows separated by less than 300 seconds are grouped together. In the second step, we apply the neural network algorithm, i.e., Self Organized Map (SOM), to classify the reduced data set. We first calculate feature vectors describing the frequency content and motion in 0.5-sec windows as input vectors. Given a set of N stations, each vector consists of 6N values, including amplitude values for three frequency bands between 2 - 8Hz, 15 - 30Hz, 0.5 - 1.5Hz, and one value for P*abs(Q). P*abs(Q) is the horizontal-vertical component product, P multiplied by the product of its Hilbert transform, Q. Seismic signals of interest, e.g. earthquakes and tremor, have high P*abs(Q) values, whereas the value for noise tends to be much lower. Next, the SOM algorithm initializes a 2-D grid in the 6N-dimensional parameter space that is fitted to the data in an iterative training process. Data with similar features are linked to the nearest grid points, and automatically clustered into groups based on distance. The method detects tremor signals with amplitudes above 350 nm/s and a RMS-signal to noise ratio above 3 with a success rate of nearly 100%. The detection rate decreases to approximately 50% for signal to noise ratios <3.