Multi-sensor change detection

Last week, on Dec. 2nd, Tarek Habib made the defense of his PhD dissertation. The research subject was the detection of abrupt changes on multi-sensor remote sensing images. This work is a result of a cooperation between CNES, Thales Alenia Space and GIPSA Lab and was funded by the two first parties.

The PhD jury was composed by:

• Laure Blanc-Féraud (INRIA-CNRS, president)
• Lorenzo Bruzzone (Univ. Trento, reviewer)
• Cédric Richard (Univ. Troyes, reviewer)
• Marc Spigai (Thales Alenia Space)
• Grégoire Mercier (Télécom Bretagne, supervisor)
• Jocelyn Chanussot (GIPSA Lab, supervisor)

and myself.

The work presented by Tarek uses a supervised classification as a tool for aided image interpretation. The starting point is a framework which was developped in ORFEO Toolbox, where the 2 images on which changes are to be detected are used to compute different features

• change indicators, as for instance differences, ratios, correlations, etc.
• mono-date features as statistics, textures, etc.

In this way, a feature vector can be built for each pixel position and used for supervised classification. The algorithm used for the classification is the Support Vector Machine. Using a graphical user interface, an operator can select samples of the two classes of interest: change and no change.

Since the fetaures computed in the first step can be very rich, these 2 classes can directly correspond to something like damage and no damage in the case of risk applications. This is so, because the operator an select only the changes of interest.

As said above, this is a very simple and pragmatic approach which works. It is generic: there is no specificity about the type of data or the type of event. However, it has a main drawback which is the computation time. Indeed, for the approach to be generic, many features have to be computed. Also, the SVM classification time is proportional to the complexity of the separating surface.

The work of Tarek consisted in proposing approaches to speed up the whole system. He has proposed and assessed new approaches for feature selection, kernel optimization, and classification surface simplification which allow to speed things up, but loosing some accuracy in the classification.

One interesting thing in this approach is that the user can tune the parameters which have a direct effect on the time vs. accuracy trade-off.