DeepVibes
Estimating and correcting micro-vibrations in satellite images acquired with pushbroom cameras, validated on Sentinel-2B data.
Pushbroom cameras build an image line by line as the satellite moves along its orbit. Tiny oscillations of the satellite’s attitude during acquisition, the micro-vibrations, make each line look in a slightly different direction: straight roads become wavy and the spectral bands, which sit at different places on the focal plane, no longer line up.
DeepVibes models the acquisition process with the satellite attitude as a parameter, then corrects the images in two stages: an identification network estimates the attitude, and a deep unrolled network uses it to reconstruct the image. We found that the geometry of the focal plane plays a key role in making the vibrations identifiable.
Interactive demo
Sentinel-2B scene. Drag the divider to compare the image affected by micro-vibrations (left) with the DeepVibes reconstruction (right); zoom in to see the colour fringes and wobbly edges disappear.