17:15 〜 18:45
[SCG40-P11] Deep-leaning-based fault slip estimation method using GNSS displacements:
comparison with the least-squares-based method
There is a growing global trend that GNSS networks for measuring crustal deformation are expanding and densifying. In Japan, the national dense GNSS network GEONET is being complemented by the GNSS network operated by SoftBank Corp., which is composed of ~3,300 sites throughout Japan (Ohta and Ohzono, 2022). This trend gives rise to the increase of automated methods that are free from complicated data preprocessing and subjective processing such as setting hyperparameters for estimating the fault slip distribution. In this presentation, we report our investigation on the performance of a deep-learning-based method developed to estimate fault slip parameters of slow slip events (SSEs) using GNSS displacement data, by comparing the results of a conventional non-negative least squares method with smoothness constraints of slip distribution.
In our two-step deep-leaning-based method, we employed a set of two convolutional neural networks (CNNs), one for denoising and interpolating the displacement field, and the other for estimating the fault slip parameters, which are trained using synthetic data sets. The synthetic data are surface horizontal displacements with randomly added noise, which are produced assuming rectangular faults with uniform slip on the plate interface. We used the same kind of synthetic data sets to investigate the performance of the methods. We assumed the geometrical setting of the Nankai subduction zone in western Shikoku and the real GEONET station locations in the area.
We compared the detectability of our method to the conventional linear inversion method of slip distribution by conducting sets of inversions for the cases of fault slip equivalent to Mw 6.0 using the two methods. As shown in Figure, the results indicated that, for low signal-noise data, our method can suppress slip outside of assumed faults due to supervised learning, while the conventional one produces slip outside of assumed faults affected by data with large noises.
(Figure caption) Example of slip distribution comparisons. (a) Assigned true slip on the assumed subfaults. The slip amount is adjusted by considering its size so that its magnitude is equal to Mw6.0. (b) Theoretical displacements at GNSS stations are calculated for the assigned slip following Okada (1992). (c) Synthetic GNSS displacements. Each figure in a vertical row indicates the total displacements corresponding to the theoretical ones in (b) added with noise patterns #1, #2, and #3. (d) Slip distribution estimated from corresponding observed displacements in (c) with the least squares method. Dashed lines indicate iso-depth contours on the plate interface of 10km, 20km, 30km, 40km, 50km, 60km, and 80km, respectively. The value of variance reduction (VR) for each inference is also represented in the corresponding figure. In general, VR is a parameter that indicates how similar the estimated result is to the correct answer. VR of 100% means the inference is completely the same as the answer, and it would be 0 % if there is no inference. (e) Estimated slip distribution with our method.
In our two-step deep-leaning-based method, we employed a set of two convolutional neural networks (CNNs), one for denoising and interpolating the displacement field, and the other for estimating the fault slip parameters, which are trained using synthetic data sets. The synthetic data are surface horizontal displacements with randomly added noise, which are produced assuming rectangular faults with uniform slip on the plate interface. We used the same kind of synthetic data sets to investigate the performance of the methods. We assumed the geometrical setting of the Nankai subduction zone in western Shikoku and the real GEONET station locations in the area.
We compared the detectability of our method to the conventional linear inversion method of slip distribution by conducting sets of inversions for the cases of fault slip equivalent to Mw 6.0 using the two methods. As shown in Figure, the results indicated that, for low signal-noise data, our method can suppress slip outside of assumed faults due to supervised learning, while the conventional one produces slip outside of assumed faults affected by data with large noises.
(Figure caption) Example of slip distribution comparisons. (a) Assigned true slip on the assumed subfaults. The slip amount is adjusted by considering its size so that its magnitude is equal to Mw6.0. (b) Theoretical displacements at GNSS stations are calculated for the assigned slip following Okada (1992). (c) Synthetic GNSS displacements. Each figure in a vertical row indicates the total displacements corresponding to the theoretical ones in (b) added with noise patterns #1, #2, and #3. (d) Slip distribution estimated from corresponding observed displacements in (c) with the least squares method. Dashed lines indicate iso-depth contours on the plate interface of 10km, 20km, 30km, 40km, 50km, 60km, and 80km, respectively. The value of variance reduction (VR) for each inference is also represented in the corresponding figure. In general, VR is a parameter that indicates how similar the estimated result is to the correct answer. VR of 100% means the inference is completely the same as the answer, and it would be 0 % if there is no inference. (e) Estimated slip distribution with our method.