Spatiotemporal filtering for regional GPS network in China using independent component analysis

Abstract

Removal of the common mode error (CME) is a routine procedure in postprocessing regional GPS network observations, which is commonly performed using principal component analysis (PCA). PCA decomposes a network time series into a group of modes, where each mode comprises a common temporal function and corresponding spatial response based on second-order statistics (variance and covariance). However, the probability distribution function of a GPS time series is non-Gaussian; therefore, the largest variances do not correspond to the meaningful axes, and the PCA-derived components may not have an obvious physical meaning. In this study, the CME was assumed statistically independent of other errors, and it was extracted using independent component analysis (ICA), which involves higher-order statistics. First, the ICA performance was tested using a simulated example and compared with PCA and stacking methods. The existence of strong local effects on some stations causes significant large spatial responses and, therefore, a strategy based on median and interquartile range statistics was proposed to identify abnormal sites. After discarding abnormal sites, two indices based on the analysis of the spatial responses of all sites in each independent component (east, north, and vertical) were used to define the CME quantitatively. Continuous GPS coordinate time series spanning $$\sim $$ 4.5 years obtained from 259 stations of the Tectonic and Environmental Observation Network of Mainland China (CMONOC II) were analyzed using both PCA and ICA methods and their results compared. The results suggest that PCA is susceptible to deriving an artificial spatial structure, whereas ICA separates the CME from other errors reliably. Our results demonstrate that the spatial characteristics of the CME for CMONOC II are not uniform for the east, north, and vertical components, but have an obvious north–south or east–west distribution. After discarding 84 abnormal sites and performing spatiotemporal filtering using ICA, an average reduction in scatter of 6.3% was achieved for all three components.