5:15 PM - 7:15 PM
[MTT37-P01] Proposal for a Correction Method for Displacement Steps Caused by Maintenance in SoftBank Original reference Sites Time Series
Keywords:GNSS, Dense Network, SoftBank Corp., Time Series Analysis
In crustal deformation analysis using GNSS data, coordinate changes unrelated to crustal movements can hinder the understanding of targeted geophysical phenomena. However, when these changes result from human-induced maintenance activities, their dates can be identified, making it easier to estimate the displacement magnitude. For GEONET, operated by the Geospatial Information Authority of Japan (GSI), a public list of maintenance activities allows researchers to apply corrections. In contrast, SoftBank proprietary reference points, increasingly utilized in recent years (e.g., Ohta and Ohzono, 2022), provide post-processing data through CSESS (https://csess.jp/), but maintenance dates remain undisclosed. Manual correction is feasible for short-term, small-scale analyses but impractical for long-term, large-scale studies. This study develops a semi-automatic method to generate candidate maintenance dates by analyzing SoftBank reference point time series.
GNSS data under CSESS are routinely processed by Tohoku University to estimate daily coordinate values using GipsyX Ver. 2.2's precise point positioning with ambiguity resolution (PPP-AR) method. Only GPS data from RINEX observations are used for daily coordinate estimation. The analysis period spans May 1, 2021, to December 31, 2024. Observation points missing data for over 50 consecutive days were excluded. Earthquake-related offsets were removed based on the "List of Earthquakes Observed by GEONET" by GSI.
Outlier removal was conducted as follows. First, a Smirnov-Grubbs test (99% confidence) was applied to coordinate estimation errors for each component, identifying Level 1 outliers. Next, Hampel filters with (1) a 50-day window and 4σ threshold and (2) a 10-day window and 3σ threshold were applied after removing Level 1 outliers. Days flagged by both were classified as Level 2 outliers, while those flagged by only one were Level 3 outliers. Level 1 and 2 outliers, representing extreme deviations, were removed. Level 3 outliers were retained for identifying potential maintenance-related offsets.
Following this, Level 1, 2, and 3 outliers were removed, and missing values were interpolated using the last available coordinate before the gap. A new time series was constructed by calculating the median coordinate difference over a 10-day moving window. The analysis proceeded through three processes: (i) A Hampel filter (30-day window, 3σ threshold) was applied to the difference time series to extract anomalous daily coordinate values. (ii) Displacement magnitudes in the difference time series were subjected to a Smirnov-Grubbs test (90% confidence) per component. (iii) Displacement magnitudes were again tested with Smirnov-Grubbs (99% confidence).
Thresholds were set for each observation point to extract days with significant displacement magnitudes. Days identified by both (i) and (iii) were classified as Maintenance Level A, those by (i) and (ii) as Maintenance Level B, and among days classified only by (iii), those exceeding the median displacement magnitude in horizontal components and the mean +1σ threshold in vertical components were classified as Maintenance Level C.
The average number of days classified as Level A was about seven, significantly higher than for GEONET, suggesting that the method may detect residual outliers or short-wavelength crustal deformation as maintenance candidates. Therefore, caution is required when applying this method.
Acknowledgments: The SoftBank's GNSS observation data used in this study was provided by SoftBank Corp. and ALES Corp. through the framework of the "Consortium to utilize the SoftBank original reference sites for Earth and Space Science".
GNSS data under CSESS are routinely processed by Tohoku University to estimate daily coordinate values using GipsyX Ver. 2.2's precise point positioning with ambiguity resolution (PPP-AR) method. Only GPS data from RINEX observations are used for daily coordinate estimation. The analysis period spans May 1, 2021, to December 31, 2024. Observation points missing data for over 50 consecutive days were excluded. Earthquake-related offsets were removed based on the "List of Earthquakes Observed by GEONET" by GSI.
Outlier removal was conducted as follows. First, a Smirnov-Grubbs test (99% confidence) was applied to coordinate estimation errors for each component, identifying Level 1 outliers. Next, Hampel filters with (1) a 50-day window and 4σ threshold and (2) a 10-day window and 3σ threshold were applied after removing Level 1 outliers. Days flagged by both were classified as Level 2 outliers, while those flagged by only one were Level 3 outliers. Level 1 and 2 outliers, representing extreme deviations, were removed. Level 3 outliers were retained for identifying potential maintenance-related offsets.
Following this, Level 1, 2, and 3 outliers were removed, and missing values were interpolated using the last available coordinate before the gap. A new time series was constructed by calculating the median coordinate difference over a 10-day moving window. The analysis proceeded through three processes: (i) A Hampel filter (30-day window, 3σ threshold) was applied to the difference time series to extract anomalous daily coordinate values. (ii) Displacement magnitudes in the difference time series were subjected to a Smirnov-Grubbs test (90% confidence) per component. (iii) Displacement magnitudes were again tested with Smirnov-Grubbs (99% confidence).
Thresholds were set for each observation point to extract days with significant displacement magnitudes. Days identified by both (i) and (iii) were classified as Maintenance Level A, those by (i) and (ii) as Maintenance Level B, and among days classified only by (iii), those exceeding the median displacement magnitude in horizontal components and the mean +1σ threshold in vertical components were classified as Maintenance Level C.
The average number of days classified as Level A was about seven, significantly higher than for GEONET, suggesting that the method may detect residual outliers or short-wavelength crustal deformation as maintenance candidates. Therefore, caution is required when applying this method.
Acknowledgments: The SoftBank's GNSS observation data used in this study was provided by SoftBank Corp. and ALES Corp. through the framework of the "Consortium to utilize the SoftBank original reference sites for Earth and Space Science".