A Non-Parametric Method for Automatic Determination of P-Wave and S-Wave Arrival Times: Application to Local Micro Earthquakes

rawles, christopher ; thurber, clifford;

We present a simple, fast, and robust method for automatic detection of P- and S-wave arrivals using a nearest neighbours-based approach. The nearest neighbour algorithm is one of the most popular time-series classification methods in the data mining community and has been applied to time-series problems in many different domains. Specifically, our method is based on the non-parametric time-series classification method developed by Nikolov. Instead of building a model by estimating parameters from the data, the method uses the data itself to define the model. Potential phase arrivals are identified based on their similarity to a set of reference data consisting of positive and negative sets, where the positive set contains examples of analyst
identified P- or S-wave onsets and the negative set contains examples that do not contain P waves or S waves. Similarity is defined as the square of the Euclidean distance between vectors representing the scaled absolute values of the amplitudes of the observed signal and a given reference example in time windows of the same length. For both P waves and S waves, a single pass is done through the bandpassed data, producing a score function defined as the ratio of the sum of similarity to positive examples over the sum of similarity to negative examples for each window. A phase arrival is chosen as the centre position of the window that maximizes the score function. The method is tested on two local earthquake data sets, consisting of 98 known events from the Parkfield region in central California and 32 known events from the Alpine Fault region on the South Island of New Zealand. For P-wave picks, using a reference set containing two picks from the Parkfield data set, 98 per cent of Parkfield and 94 per cent of Alpine Fault picks are determined within 0.1 s of the analyst pick. For S-wave picks, 94 per cent and 91 per cent of picks are determined within 0.2 s of the analyst picks for the Parkfield and Alpine Fault data set, respectively. For the Parkfield data set, our method picks 3520 P-wave picks and 3577 S-wave picks out of 4232 station–event pairs. For the Alpine Fault data set, the method picks 282 P-wave picks and 311 S-wave picks out of a total of 344 station–event pairs. For our testing, we note that the vast majority of station–event pairs have analyst picks, although some analyst picks are excluded based on an accuracy assessment. Finally, our tests suggest that the method is portable, allowing the use of a reference set from one region on data from a different region using relatively few reference picks.