Deep-learning-based denoising and interpolation of small displacements caused by fault slip using InSAR and GNSS data: Application to the 2019 Mw 6.2 Hyuga-nada earthquake
Abstract
In recent years, deep learning techniques have become important tools in Solid Earth science, including geodesy. Previous studies have applied deep learning techniques to synthetic aperture radar interferometry (InSAR) or Global Navigation Satellite System (GNSS) data, which are two major observation techniques of satellite geodesy, to mitigate noise or detect small signals. Because InSAR and GNSS are complementary techniques, a deep-learning-based joint analysis can provide more detailed information. In this study, we developed a deep learning model that reproduces three-dimensional displacement fields from multiple interferograms and daily GNSS time series with gaps in the spatial and/or temporal domains. We employed a convolutional neural network-based architecture and trained the model using a synthetic dataset generated from signals from interplate slip and noise from multiple sources. Synthetic test results show that the developed model reproduces the ground truth with a variance reduction of greater than 60% when the maximum displacement amplitude exceeds 2 mm, although the performance depended on the noise level of the original GNSS data. We then applied our model to real InSAR and GNSS datasets of the coseismic displacements of the 2019 Mw 6.2 Hyuga-nada megathrust earthquake. The application successfully produced continuous displacement fields while suppressing incoherent random noise. The horizontal location of the rectangular fault model was consistent with the seismologically determined epicenter of the earthquake. Our results demonstrate the potential of deep learning techniques for obtaining denoised full displacement fields by combining multiple types of datasets.
Related
- Super-resolving Herschel: A deep-learning-based deconvolution and denoising technique: I. First application at 500 µm in the COSMOS field
- Earthquake Seismogram Denoising Across Time, Time‐Frequency, and Hybrid Domain Approaches
- An internally consistent framework for calculating cascading probabilistic earthquake risk and its application to a case study in New Zealand
- Comparison of machine learning and spatial interpolation methods for hourly rainfall data imputation in the Oliwski Stream catchment
- A New Amphibious Earthquake Catalog for the Alaska Peninsula: Training, Testing, and Applying Machine Learning Earthquake Picking in Offshore and Amphibious Environments
BibTeX
@misc{okada2026deeplearningbased,
title = {{Deep-learning-based denoising and interpolation of small displacements caused by fault slip using InSAR and GNSS data: Application to the 2019 Mw 6.2 Hyuga-nada earthquake}},
author = {Yutaro Okada and Yo Fukushima},
howpublished = {EarthArXiv},
year = {2026},
month = aug,
doi = {10.31223/x5jf7h},
url = {https://eartharxiv.org/repository/view/14728/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Deep-learning-based denoising and interpolation of small displacements caused by fault slip using InSAR and GNSS data: Application to the 2019 Mw 6.2 Hyuga-nada earthquake
AU - Yutaro Okada
AU - Yo Fukushima
PY - 2026
DA - 2026/08/31
PB - EarthArXiv
DO - 10.31223/x5jf7h
UR - https://eartharxiv.org/repository/view/14728/
AB - In recent years, deep learning techniques have become important tools in Solid Earth science, including geodesy. Previous studies have applied deep learning techniques to synthetic aperture radar interferometry (InSAR) or Global Navigation Satellite System (GNSS) data, which are two major observation techniques of satellite geodesy, to mitigate noise or detect small signals. Because InSAR and GNSS are complementary techniques, a deep-learning-based joint analysis can provide more detailed information. In this study, we developed a deep learning model that reproduces three-dimensional displacement fields from multiple interferograms and daily GNSS time series with gaps in the spatial and/or temporal domains. We employed a convolutional neural network-based architecture and trained the model using a synthetic dataset generated from signals from interplate slip and noise from multiple sources. Synthetic test results show that the developed model reproduces the ground truth with a variance reduction of greater than 60% when the maximum displacement amplitude exceeds 2 mm, although the performance depended on the noise level of the original GNSS data. We then applied our model to real InSAR and GNSS datasets of the coseismic displacements of the 2019 Mw 6.2 Hyuga-nada megathrust earthquake. The application successfully produced continuous displacement fields while suppressing incoherent random noise. The horizontal location of the rectangular fault model was consistent with the seismologically determined epicenter of the earthquake. Our results demonstrate the potential of deep learning techniques for obtaining denoised full displacement fields by combining multiple types of datasets.
ER -