Injecting spatial priors in Earth observation with machine vision

Summary

Remote Sensing (RS) imagery with submeter resolution is becoming ubiquitous. Be it from satellites, aerial campaigns or Unmanned Aerial Vehicles, this spatial resolution allows to recognize individual objects and their parts from above.

This has driven, during the last few years, a big interest in the RS community on Computer Vision (CV) methods developed for the automated understanding of natural images.

A central element to the success of \CV is the use of prior information about the image generation process and the objects these images contain: neighboring pixels are likely to belong to the same object; objects of the same nature tend to look similar with independence of their location in the image; certain objects tend to occur in particular geometric configurations; etc.

When using RS imagery, additional prior knowledge exists on how the images were formed, since we know roughly the geographical location of the objects, the geospatial prior, and the direction they were observed from, the overhead-view prior.

This thesis explores ways of encoding these priors in CV models to improve their performance on RS imagery, with a focus on land-cover and land-use mapping.