Detection and Characterization of Transient Deformations in GPS Data

Recently there has been a growing awareness of the existence of transient deformation and slip events that occur over much longer times scales (from hours to months) than do earthquakes. Clearly identified instances of such events are rare, but include events in Mexico, Cascadia, Kantou, and Peru. We present a suite of methods for detecting and characterizing such events in GPS data. Two of the detection methods are based, respectively, on principle component analysis (PCA) and the related latent factor analysis (LFA), and rely on processing the time series data en masse to utilize the joint information in the time series. Another method, based on the use of covariance descriptors, can be used on time series individually to detect anomalous signals. This method is computationally efficient enough to be readily employed even on real-time (1Hz) GPS data and is readily extensible to incorporate a wide range of higher order features, including derivatives, moments, and wavelet or Fourier components. Furthermore, it is robust in the presence of missing values. The last detection method is a human-in-the-loop method that presents to the user a space-time visualization of calculated strain. Once a transient is detected, we use an inversion based on the downhill simplex method to characterize it, ascribing the transient to slip on modeled fault patches. For each fault patch the method adjusts the location, strike, dip, length, width, and slip to reach a locally optimum solution. In many cases, initial estimates of these parameters and the fault geometries can be well framed by the PCA and LFA detection methods, reducing the uncertainty of inversion results. We discuss the blind results from these methods on several sets of the synthetic GPS data offered as part of the Phase I and Phase II Southern California Earthquake Center (SCEC) transient detection excercises. In addition, we apply the methods to approximately 15 years of daily GPS field data collected from Southern California and compare the results with those from the synthetic data sets.