Sources Separations in Geodetic Time Series: Synthetic Tests and Results

Abstract:

Geodetic time series are the superimposition of a number of geophysical signals, errors, and noise. The objective of sources separation is to isolate and to identify each contribution. Indeed, when using geodetic data to study mass redistributions within the Earth system and the dynamics their reflect, it is a crucial step to separate the different geophysical signals in the time series.

Sources separation techniques generally estimate a separation matrix, which depends on assumptions on the nature of the sources. Here, several sources separation techniques were tested : PCA (Principal Component Analysis), rotated PCA, and two methods of ICA (Independant Component Analysis). PCA assumes that the sources are linearly independent, whereas ICA searches for sources that are statistically independent. Among the ICA techniques, different methods can be used to ensure this statistical independence. We tested two of them: non-gaussianity of the sources and minimization of the mutual information. First, we tested the adequacy between the hypothesis on the sources involved in each kind of separation technique, and the nature of and characteristics of these sources. Then, we built from these a priori sources synthetic geodetic time series and we tested the performances of each separation technique to retrieve the original sources. Results show great differences between the results obtained with the different separation techniques. In particular, if periodic signals are well retrieved by all separation techniques, PCA gives better results in order to retrieve interannual climatic signals and ICA performs better for some specific a priori sources like co-seismic signals.