Removing noise from MT data by using independent component analysis

We carried out a MT survey in the Boso peninsula (Chiba, Central Japan) to investigate the resistivity structure of the area where the slow slip event have occurred at least five times within 20 years. Large artificial noise contaminated in the MT data and the resistivity and phase showed near field effect at the frequency band below 1Hz. To avoid the local noise, we attempted to apply the independent component analysis (ICA). ICA is one of the multivariate analysis methods and in which complicated data sets can be separated into all underlying sources without knowing these sources or the way that they are mixed. It assume that the mixing is liner, and yields the relation x(t)=As(t) with input signals x(t), mixing matrix A and source signal s(t). ICA has the ability to compute the matrix W (=A-1). In this study, we used the ICA programs for complex signals to deal with phase part in frequency-domain data. This is extension of FastICA algorithm which was introduced by Aapo and Hyvärinen (2000) and is based on a fixed-point iteration scheme to complex valued signals. We applied the ICA method to improve horizontal magnetic components in MT data using both the data observed in Boso area and the noise free magnetic data observed in Memanbetsu Branch of Geomagnetic Observatory. The observatory is located in eastern Hokkaido, Northern Japan, and apart from approximately 800km distance from the Boso area. After applying ICA, each component is not defined intensity scale. To extract noise free data in original data scale, kept the noise free component, and other noise components set to 0 and applied inverse matrix of W to obtain original x, i.e. x(t)=W-1u'(t), where u'(t): components vector after ICA, x(t): the original data vector. Secondly, we tried to improve the electric components in accordance with the noise free magnetic components. Finally, we calculated the apparent resistivity and phases using the data processed as above. In comparison between before and after the ICA processing, the apparent resistivity take promising values and the phases take non-zero values. This result showed some parts of the noise components were removed. The apparent resistivity and phases improved by ICA were probably free from an influence of near-field phenomenon. These results revealed that ICA has a potential to handle noisy data.