4:45 PM - 5:00 PM
[SCG60-10] Attempt to extract three-dimensional displacement fields due to fault slip by deep learning technique using InSAR and GNSS data
Keywords:InSAR, GNSS, CNN
Detecting displacements caused by fault slip is important to deepen understanding of the properties of asperities and their vicinity on a fault. Owing to the expansion of geodetic observation by satellites since the 1990s, including Synthetic Aperture Rader (SAR) and Global Navigation Satellite System (GNSS), large crustal deformation due to large earthquakes and slow slips have been clarified with their spatiotemporal extent. Nishimura et al. (2013) and Okada et al. (2022) extracted small displacements caused by slow slips using GNSS data removing common-mode errors by applying the spatial filter (Wdowinski et al., 1997). On the other hand, detecting millimeter-scale displacements is still challenging, using SAR interferometry (InSAR). Rouet-Leduc et al. (2021) extracted the cumulative displacement field of line-of-sight (LOS) component from InSAR time series by a convolutional neural network (CNN). Since InSAR and GNSS are complemental observations in terms of temporal and spatial observation intervals, joint usage of both datasets should improve the accuracy of the denoised displacement field. In this study, we attempt to develop a deep-learning model that extracts the three-dimensional displacement field due to fault slip using InSAR and GNSS data and to apply it to moderate earthquakes on megathrust faults.
We developed a CNN-based architecture following previous studies (e.g., Nakagawa et al., in prep.; Rouet-Leduc et al., 2021). In the first part of our model, we input InSAR and GNSS time series independently with ten and eleven convolutional layers, respectively. We add a digital elevation model (DEM; Farr et al., 2007) as a feature map during the processing of InSAR (Rouet-Leduc et al., 2021). In the second part, we connect the InSAR and GNSS feature maps created in the first part and perform the convolution with four layers. We trained our model in one hundred epochs and used the parameters at the final epoch.
We trained the model using synthetic InSAR and GNSS data. We created synthetic InSAR noises by summing the topography-correlated, long-wave-length, and spatially correlated random noises (Ghayournajarkar & Fukushima, 2022; Fujiwara et al., 1999). As for the GNSS noises, we made a 121-day time series of horizontal and vertical components with random walk and Gaussian noises(cf. Fukuda et al., 2008), where the noise amplitudes were assumed to be different for horizontal and vertical components. The synthetic signals are the displacements predicted by a rectangular elastic dislocation model assuming slip on the megathrust (Nakagawa et al., in prep.; Okada, 1992).
Our preliminary results show that the training was completed without overfitting and that the trained model could reproduce the ground truth. We evaluate the model performance using a mean of structure similarity index measurements (mSSIM) (Wang et al., 2004). The mSSIM curve against the maximum displacement of the ground truth shows that the curve for the east component is slightly better than that of the north and up components. We speculate that the constraints from both InSAR and GNSS data are maximized in the east component because of the InSAR LOS directions and the relatively high signal-to-noise ratio of horizontal GNSS data.
Acknowledgments
We utilized Sentinel-1 data provided by LiCSAR (Lazecký et al., 2020).
We developed a CNN-based architecture following previous studies (e.g., Nakagawa et al., in prep.; Rouet-Leduc et al., 2021). In the first part of our model, we input InSAR and GNSS time series independently with ten and eleven convolutional layers, respectively. We add a digital elevation model (DEM; Farr et al., 2007) as a feature map during the processing of InSAR (Rouet-Leduc et al., 2021). In the second part, we connect the InSAR and GNSS feature maps created in the first part and perform the convolution with four layers. We trained our model in one hundred epochs and used the parameters at the final epoch.
We trained the model using synthetic InSAR and GNSS data. We created synthetic InSAR noises by summing the topography-correlated, long-wave-length, and spatially correlated random noises (Ghayournajarkar & Fukushima, 2022; Fujiwara et al., 1999). As for the GNSS noises, we made a 121-day time series of horizontal and vertical components with random walk and Gaussian noises(cf. Fukuda et al., 2008), where the noise amplitudes were assumed to be different for horizontal and vertical components. The synthetic signals are the displacements predicted by a rectangular elastic dislocation model assuming slip on the megathrust (Nakagawa et al., in prep.; Okada, 1992).
Our preliminary results show that the training was completed without overfitting and that the trained model could reproduce the ground truth. We evaluate the model performance using a mean of structure similarity index measurements (mSSIM) (Wang et al., 2004). The mSSIM curve against the maximum displacement of the ground truth shows that the curve for the east component is slightly better than that of the north and up components. We speculate that the constraints from both InSAR and GNSS data are maximized in the east component because of the InSAR LOS directions and the relatively high signal-to-noise ratio of horizontal GNSS data.
Acknowledgments
We utilized Sentinel-1 data provided by LiCSAR (Lazecký et al., 2020).