Development of satellite image registration/correlation techniques using Deep Learning

Title : Development of satellite image registration/correlation techniques using Deep Learning

Keywords : image processing, machine learning, registration, earth science, seismic faults, remote sensing

Advisors :
James Hollingsworth (ISTerre)
Sophie Giffard-Roisin (ISTerre)
Mauro Dalla Mura (GIPSA-Lab)

Summary :
Recent developments in optical geodesy, in which co-registered satellite/aerial images spanning a period of time can be cross- correlated (with sub-pixel precision) to retrieve dense measurements of ground deformation, has revolutionized how we characterize Earth surface processes (e.g. earthquakes, volcanoes, etc) from space.

Given the value of near-field fault displacements for understanding how active faults rupture the shallow crust in earthquakes, there is an increasing need to resolve finer details close to fault ruptures. However, phase correlation methods require large windows to capture enough robust long-wavelength frequencies for an accurate measurement, while spatial correlation methods are more robust with smaller windows, although require spatial regularization to minimize noise (which effectively imposes a smoothing of the result over a broader wavelength than the imposed window size). Furthermore, little work has been done to characterize which factors influence sub-pixel bias when correlating satellite images, although recent studies highlight the sensitivity to variable illumination conditions.

In this project, the student will address these problems by taking advantage of more sophisticated sub-pixel interpolation techniques attainable using deep learning approaches, which will be used to generate the non-linear mappings between LR and HR images. The novelty of this project comes from the use of a weakly-supervised super- resolution approach, in which only unpaired LR-HR images (i.e. from different sensors) are provided for the training. This avoids problems with SR models learning the reverse operation of the predefined downsampling process, when HR images are used to generate the LR dataset. If time allows, the student will also explore using synthetic satellite imagery (incorporating known ground displacements) for training purposes, thereby leading to faster and more accurate correlators (in addition to greater spatial resolution). Finally, the student will explore how these developments may be integrated into open- source correlation tools commonly used by the research community to investigate Earth surface dynamics.

The project will be undertaken jointly between GIPSA-Lab (Mauro Dalla- Mura) and ISTerre (James Hollingsworth, and Sophie Giffard-Roisin), both labs are in the Université Grenoble-Alpes (UGA).

Start date : whenever from now to Apr. 2021

Duration : from 4 to 6 months

Contact : james.hollingsworth univ-grenoble-alpes.fr