09:45 〜 10:00
[SSS03-04] PCA-based noise reduction of seafloor pressure data to improve tectonic transient signal detection -Application to DONET long-term continuous data
キーワード：Ocean Bottom Pressure-gauge、Slow slip event、Seafloor geodesy、Oceanographic fluctuations
Ocean Bottom Pressure-gauges (OBPs) are essential sensors to continuously observe the vertical seafloor displacement and to understand the mechanism of the interplate seismological phenomena. On the other hand, OBP data also contain the non-tidal oceanographic fluctuations, which have a similar time constant (more extended than several days) with the unsteady tectonic signal and have 5-10 hPa amplitude. Therefore, it is necessary to correctly estimate and remove the oceanographic signal to observe the tectonic deformation. The spatial characteristics of oceanographic fluctuations have been evaluated using the observed data analysis or comparing the observation and numerical oceanographic models (e.g., Fredrickson et al., 2019; Dobashi and Inazu, 2021; Inoue et al., 2021). Inoue et al. (2021) evaluated the spatial similarity of OBP time series in the Hikurangi margin, focusing on the distance and sea-depth difference of the observation sites. They found that the standard deviation of OBP time series shows a sea-depth dependency. The similar characteristic also confirmed in DONET (Otsuka et al., 2021, JpGU). On the other hand, more detailed characteristics of the sea-depth-dependent oceanographic fluctuations have not been clarified. In this study, we focus on the DONET OBP time series, which contains long-term and wide-area observation data. The noise level of the OBP time series was reduced by efficiently classifying the oceanographic fluctuations. In addition, the ability to detect the transient crustal deformation by DONET time series was evaluated.
Since the oceanographic fluctuations are expected to have a spatial similarity, we utilized the Principal Component Analysis (PCA) to evaluate their characteristics. Before PCA application, we removed the tidal component and instrumental drift from the OBP time series from 2016 to 2019, using a low-pass filter and exponential approximate, respectively.
As the extracted time series contribution ratio by PCA, the first principal component (PC1) is approx. ~80%, PC2 is approx. ~10%, and other higher PCs are more minor than several percent. The spatial distribution of the eigenvectors for each PC shows that the PC1 is common to the entire observation network. The amplitude of the PC2 depends on the sea-depth of the observation sites. PC4 shows the hyperbola-like spatial feature with similar characteristics in shallow and deep regions. PC3 does not have a significant spatial dependency. We interpreted the PC1, PC2, and PC4 as oceanographic fluctuations and investigated the characteristic of residual time series by removing these components. The averaged standard deviation of the time series is clearly reduced when the PC1 is removed (averaged SD: 0.91hPa), PC1, 2, and 4 are removed (0.62hPa) compared to the observation (1.98hPa). Furthermore, when we compare distance and sea-depth dependence versus SD of the OBP time series with the removal of each PC, we found that the sea-depth dependency, which is apparent even with the removal of the PC1, can be well reduced with the removal of PC1, 2, and 4. These results clearly suggested that the PCA is useful to remove the common oceanographic fluctuations from OBP time series.
We also verified how the PCA classified the principal components when the transient crustal deformation occurred. We assumed the single rectangular fault models, synthesized the magnitude of earthquake-dependent vertical displacement distribution expected from the models, then added the synthetic ramp to the actual DONET OBP time series, and finally applied the PCA. The assumed crustal deformation can be classified in the PC1-PC3 depending on the magnitude, and the oceanographic fluctuations are classified into another PC as same as steady time period. These results suggest that we should classify the unsteady tectonic signal with noise such as oceanographic signal by PCA application to OBP time series. We will provide a quantitative evaluation of the accuracy of crustal deformation detection for OBP data by conducting a comprehensive numerical experiment in the DONET OBP network.
Since the oceanographic fluctuations are expected to have a spatial similarity, we utilized the Principal Component Analysis (PCA) to evaluate their characteristics. Before PCA application, we removed the tidal component and instrumental drift from the OBP time series from 2016 to 2019, using a low-pass filter and exponential approximate, respectively.
As the extracted time series contribution ratio by PCA, the first principal component (PC1) is approx. ~80%, PC2 is approx. ~10%, and other higher PCs are more minor than several percent. The spatial distribution of the eigenvectors for each PC shows that the PC1 is common to the entire observation network. The amplitude of the PC2 depends on the sea-depth of the observation sites. PC4 shows the hyperbola-like spatial feature with similar characteristics in shallow and deep regions. PC3 does not have a significant spatial dependency. We interpreted the PC1, PC2, and PC4 as oceanographic fluctuations and investigated the characteristic of residual time series by removing these components. The averaged standard deviation of the time series is clearly reduced when the PC1 is removed (averaged SD: 0.91hPa), PC1, 2, and 4 are removed (0.62hPa) compared to the observation (1.98hPa). Furthermore, when we compare distance and sea-depth dependence versus SD of the OBP time series with the removal of each PC, we found that the sea-depth dependency, which is apparent even with the removal of the PC1, can be well reduced with the removal of PC1, 2, and 4. These results clearly suggested that the PCA is useful to remove the common oceanographic fluctuations from OBP time series.
We also verified how the PCA classified the principal components when the transient crustal deformation occurred. We assumed the single rectangular fault models, synthesized the magnitude of earthquake-dependent vertical displacement distribution expected from the models, then added the synthetic ramp to the actual DONET OBP time series, and finally applied the PCA. The assumed crustal deformation can be classified in the PC1-PC3 depending on the magnitude, and the oceanographic fluctuations are classified into another PC as same as steady time period. These results suggest that we should classify the unsteady tectonic signal with noise such as oceanographic signal by PCA application to OBP time series. We will provide a quantitative evaluation of the accuracy of crustal deformation detection for OBP data by conducting a comprehensive numerical experiment in the DONET OBP network.