The time series analysis is performed by fitting hidden Markov models (HMMs) to individual GPS time series. The fitted HMMs allow us to segment the GPS time series into discrete modes; these modes are designated by different colors in the clickable plots for each time station. The geographical distribution of state changes among GPS stations provides insight into the geophysical processes of the region. We summarize these graphically by color coding the station markers in the map above to indicate current and recent state changes.
Navigation through time can be performed using the slider bar below the map or through the calendar to the left. A KML file that can be animated using Google Earth can be generated using the form under the calendar. The tab next to the map view display shows a plot of the total number of GPS stations that have changed state versus time, summarizing the total activity of the network.
The entire process is performed in the absence of labeled training data or other human supervision, and is entirely a data-driven approach. In general, fitting a hidden Markov model in the absence of a priori information using the standard expectation-maximization (EM) method is a difficult problem, due to the presence of numerous local maxima in the objective function. We address this problem through the use of the regularized deterministic annealing EM (RDAEM) algorithm, which produces stable, high-quality model fits. This algorithm has been implemented by the QuakeSim project in the RDAHMM software package.