Reconstructing Horizontal Displacement through Deep Learning in
Multiple-pairwise Satellite Image Correlation
Abstract
Data collected from optical remote sensing systems can be used to
effectively measure ground horizontal displacements related for example
to earthquake ruptures, glacier flows, dune migrations, and slow
landslides. However, displacement of interest is often corrupted with
and even hidden under a considerable amount of unwanted noise. Most
traditional methods can eliminate such noise but may change unexpectedly
our concerned displacement pattern. If the clean displacement can be
recovered, with unwanted noise removed, we can more accurately interpret
how tectonic process contributes to surface displacement. With the
recently successful application in geodetic deformation denoising, deep
learning is also expected to be able to address the problem associated
with optical displacement. Here, we trained a deep-learning-based
autoencoder on a real noise dataset generated from a large amount of
Sentinel-2 images. We applied the autoencoder to synthetic and real test
datasets to successfully eliminate the unwanted noises in given MPICs,
accurately revealing noise-free displacement signals. Interestingly, we
even discovered detailed surface migrating information about a dune near
the Maduo earthquake in the Tibet Plateau. The finding highlights the
autoencoder is a robust method that can also improve horizontal
displacement caused by those tectonic processes other than just
earthquakes.