SCEC Award Number 22077 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Detecting creep transients in InSAR timeseries using deep neural networks
Investigator(s)
Name Organization
Roland Bürgmann University of California, Berkeley Gareth Funning University of California, Riverside Yuexin Li University of California, Berkeley
Other Participants
SCEC Priorities 1d, 2a, 3a SCEC Groups Geodesy, CXM, SDOT
Report Due Date 03/15/2023 Date Report Submitted 03/14/2023
Project Abstract
 Variable creep behaviors contribute to regional strain and stress variations, and accelerations of creep rates may produce stress perturbations on faults and even trigger earthquakes. Due to limited observations of such low-amplitude near-fault signals, the governing factors for variable creep behaviors are still not well understood. Better quantification of the spatiotemporal fault creep will not only reveal its underlying mechanisms but also provide insight for seismic hazard assessment. In this project, we develop a neural network structure that is well designed for detecting shallow creep transients, based on the encoder-decoder structure proposed by Rouet-Leduc et al. (2021). Specifically, we evaluate a series of neural network structures trained with synthetic timeseries incorporating synthetic and real noise features, and finally apply the denoising method on creeping faults in Turkey, Pakistan and California to identify transient creep signals. The development of the InSAR denoising method helps refine our understanding of how strain and stress evolve on continental creeping faults over time.
Intellectual Merit This project is pertinent to one of the SCEC5 basic questions of earthquake science “Q1: How are faults loaded on different temporal and spatial scales”. The denoising neural network can be further incorporated into SCEC’s Community Geodetic Model (CGM) and provide a cross-comparison between InSAR timeseries products, especially at short temporal scales. This work will also benefit the analysis of data from past (e.g. ERS, Envisat, ALOS) and planned future SAR missions (e.g. NISAR) by mitigating the atmospheric delay in radar measurements and enabling transient detection in InSAR timeseries.
Broader Impacts This project provided support to a female graduate student at UC Berkeley in the final year of her PhD studies. The machine learning methodologies developed in this effort may ultimately benefit earthquake hazard research more broadly.
Exemplary Figure Figure 3: Performance of the DAE neural network model on the (a-c) North Anatolian Fault and the (d-f) Chaman Fault. (a) Input CSK InSAR timeseries from Rousset et al. (2016). (b) Input DEM, original and DAE recovered cumulative displacement at the last scene (2013-10-23). (c) Comparisons of profiles between original and recovered displacement. Locations of profiles are labeled in (b). (d)Input Sentinel-1 InSAR timeseries from Kang Wang (pers. commun.). (e) Input DEM, original and DAE recovered cumulative displacement at the last scene (2016-07-29). (f) Comparisons of profiles between original and recovered displacement. Locations of profiles are labeled in (b). Figure Credit: Yuexin Li