2:00 PM - 2:15 PM
[SCG60-02] Developing anomaly detection method for InSAR time series with the aim of unknown slow-slip event detection
Keywords:InSAR, machine learning, anomaly detection, slow slip event
Slow slip events (SSE) are known as slow earthquakes mainly observed by geodetic techniques and have been thought to be linked to megathrust earthquakes (Obara and Kato, 2016). Global Satellite Navigation System (GNSS) is mainly used to observe SSE. GNSS can measure surface displacements in millimeters during SSE. However, because it requires ground observation points, SSE observations are limited to areas where GNSS observation points have been installed. In areas with few observation points, it is possible that SSE have been missing. InSAR (Interferometric Synthetic Aperture Radar) is a technology that can measure the distribution of surface displacements with high precision. InSAR is also capable of detecting small surface displacements of SSE. Cases of SSE detection by InSAR have been reported in Guerrero in 2006 (Cavalié et al., 2014) and off the Boso Peninsula in 2018 (Kinoshita and Furuta, 2024). However, the identification of the surface displacement signal ultimately relied on the visual judgment of the analyst. The purpose of this study is to develop anomaly detection method for InSAR time series to efficiently and objectively detect unsteady surface displacements caused by SSE. Taking advantage of InSAR, we aim to detect SSE in areas where GNSS observation networks are insufficient.
We simulated SSE displacement by using the rectangular fault model (Okada, 1992). Since InSAR observations contain atmospheric delay noise, the turbulent component of the atmospheric delay was reproduced by applying a Gaussian filter with random variables following a normal distribution with a mean of 3 cm, standard deviation of 1 cm. The actual InSAR time series also contains other noise, such as elevation correlation delay noise and ionospheric noise. Both noises are not included in this simulation because effective correction methods are available. In this simulation, it was considered that SAR observations were conducted 50 times every 14 days, and that the SSE displacement period was 2 months. 100 sets of the above simulation data were prepared to evaluate the signal detection capability. We used a machine learning method called the sliding window k-nearest neighbor method to detect anomalies of this simulated data. In this method, one time series data was grouped together with multiple adjacent observations and represented as a collection of multiple partial time series. Then, the multiple partial time series before SSE were used as a reference. The minimum distances between the partial time series for anomaly detection and reference partial time series were defined as the anomaly level. Fault displacement by SSE occurred between the 26th ~ 30th observations, and the collection of partial time series up to the 20th observation was used as the reference for this method.
The sliding window k-nearest neighbor method detected anomaly where the ratio of the maximum amplitude of displacement to the maximum amplitude of noise was 0.4, i.e., the signal amplitude was up to about 2.0 cm. Looking at the transition of anomaly in one time series, during the period when SSE was occurring, the anomaly level was increased. After the occurrence of SSE, the anomaly level remained high. When noise in the opposite direction of displacement occurred during the displacement period, the anomaly level remained low. But the anomaly level increased with a delay after the SSE ended, and in most cases the anomaly level was high after about three observations were made. When the moving average was applied to the simulation data, the increase in anomaly was suppressed for the noise-only case with no displacement, and the difference from the case with displacement was widened.
The model showed the possibility that anomaly detection from simulation data to InSAR time series data is effective for SSE detection. In the future, we will apply the model to InSAR time series data at locations where SSE has occurred and evaluate the utility of the model.
We simulated SSE displacement by using the rectangular fault model (Okada, 1992). Since InSAR observations contain atmospheric delay noise, the turbulent component of the atmospheric delay was reproduced by applying a Gaussian filter with random variables following a normal distribution with a mean of 3 cm, standard deviation of 1 cm. The actual InSAR time series also contains other noise, such as elevation correlation delay noise and ionospheric noise. Both noises are not included in this simulation because effective correction methods are available. In this simulation, it was considered that SAR observations were conducted 50 times every 14 days, and that the SSE displacement period was 2 months. 100 sets of the above simulation data were prepared to evaluate the signal detection capability. We used a machine learning method called the sliding window k-nearest neighbor method to detect anomalies of this simulated data. In this method, one time series data was grouped together with multiple adjacent observations and represented as a collection of multiple partial time series. Then, the multiple partial time series before SSE were used as a reference. The minimum distances between the partial time series for anomaly detection and reference partial time series were defined as the anomaly level. Fault displacement by SSE occurred between the 26th ~ 30th observations, and the collection of partial time series up to the 20th observation was used as the reference for this method.
The sliding window k-nearest neighbor method detected anomaly where the ratio of the maximum amplitude of displacement to the maximum amplitude of noise was 0.4, i.e., the signal amplitude was up to about 2.0 cm. Looking at the transition of anomaly in one time series, during the period when SSE was occurring, the anomaly level was increased. After the occurrence of SSE, the anomaly level remained high. When noise in the opposite direction of displacement occurred during the displacement period, the anomaly level remained low. But the anomaly level increased with a delay after the SSE ended, and in most cases the anomaly level was high after about three observations were made. When the moving average was applied to the simulation data, the increase in anomaly was suppressed for the noise-only case with no displacement, and the difference from the case with displacement was widened.
The model showed the possibility that anomaly detection from simulation data to InSAR time series data is effective for SSE detection. In the future, we will apply the model to InSAR time series data at locations where SSE has occurred and evaluate the utility of the model.