Our CTO, Baptiste Coulange, did his PhD for CNES (the French Space Agency) and University of Paris Descartes on detecting of aliasing in satellite images.
Earth observation satellites are optimised to obtain the best image quality. The size of the sensors and the parameters of the optical chain are designed to obtain the best compromise between image resolution and acquisition artefacts. The higher the resolution the more artefacts are present.
Aliasing which is one of the artefacts in satellite images can lead to image misinterpretation. It is thus critical to detect it. By using the duality between spatial localisation in the image and aliasing relationships in the Fourier transform plane, it was possible to develop and validate a aliasing detection algorithm.
The three main acquisition artefacts are :
Aliasing causes many difficulties, including image interpretation and the identification of objects in satellite images. To facilitate the work of the image interpreters, the goal of this work was to develop a detector of zones in the satellite images presenting aliased structures.
Aliasing is a defect intimately linked to any numerical acquisition. Our work started by an analysis of its effect on simple image models, specifically periodic patterns that are highly impacted by aliasing.
This study lead to the definition of an aliasing spectrum relationship that characterises 2 frequencies or frequency zones that are linked by the aliasing pattern. In parallel, by extending the notion of "analytical signals" to images, the analytical parts of the images were obtained. The modulus of these images in the "complex domain" have the interesting property of having a very low sensitivity to bad sampling? They enable the determination of the localization in the original image of the energy with in a given frequency domain. This is interesting since it enables to localize a potentially aliased zone within the image.
It was experimentally showed that incorrectly sampled images present numerous frequency couples in an aliasing spectrum relationship and having the same localization within the original image. The identification of this property of aliased images enabled us to propose two aliasing detection strategies, both using "a contrario" methods.
Classical images used in image processing tests are not adapted to testing aliasing. A test data-base was thus first defined to have a better test of the proposed algorithms. The algorithms were also used to evaluate the existing CNES satellites on their capacity to create more or les aliasing in the acquired images.
Finally thanks to the detection algorithms, a first correction algorithm was suggested. Thanks to pre-informed algorithms, it was demonstrated that correction algorithms using the hypothesis of locally non-overlapping spectrums could enable an efficient aliasing correction.