PhD Defense of Avik Bhattacharya
Friday, December 14 at 14:30, Amphi B312
ENST -- 46, rue Barrault -- 75013 Paris

Indexing of Satellite Images Using Structural Information

Avik Bhattacharya
Friday, Decembre 14, 2007 at 14:30
ENST, Barrault, Amphi B312
Directeur de thèse
Jury members
  • Matthieu Cord, Professeur (Lab. LIP6, UPMC University, France),
  • Nicolas Paparoditis, Directeur de recherche (Lab. MATIS, IGN, France).
  • Henri Maître, Professeur (ENST Paris),
  • Georges Stamon, Professeur (Université René Descartes-Paris-5, France),
  • Ian Jermyn, Chargé de recherche (ARIANA Project, INRIA Sophia Antipolis, France).
  • Jordi Inglada, Ingénieur d'études (CNES, France).


From the advent of human civilization on our planet to modern urbanization, road networks have not only provided a means for transportation of logistics but have also helped us to cross cultural boundaries. The properties of road networks vary considerably from one geographical environment to another. The networks pertaining in a satellite image can therefore be used to classify and retrieve such environments. In this work, we have defined several such environments, and classified them using geometrical and topological features computed from the road networks occurring in them. Due to certain limitations of these extraction methods there was a relative failure of network extraction in some urban regions containing narrow and dense road structures. This loss of information was circumvented by segmenting the urban regions and computing a second set of geometrical and topological features from them.

The small images forming our database were extracted from images acquired from the SPOT5 satellite with 5m resolution (each image of size 512x512 pixels). The set of geometrical and topological features computed from the road networks and urban regions are used to classify the pre-defined geographical classes. In order to avoid the burden of feature dimensionality and reduce the classification performance, feature selection was performed using Fisher Linear Discriminant (FLD) analysis and an one-vs-rest linear Support Vector Machine (SVM) classification was performed on the selected feature set. The impact of spatial resolution and size of images on the feature set have been explored in this work. We took a closer look at the effect of spatial resolution and size of images on the discriminative power of the feature set to classify the images belonging to the pre-defined geographical classes. Tests were performed with feature selection by FLD and one-vs-rest linear SVM classification on a database with images of 10m resolution. Another test was performed with feature selection by FLD and one-vs-rest linear SVM classification on a database with 5m resolution images each of size 256x256 pixels.

With the above mentioned approaches, we developed a novel method to classify large satellite images with patches of images each of size 512x512 extracted from them acquired by SPOT5 satellite of 5m resolution. There has been a large amount of work dedicated to the classification of large satellite images at pixel level rather than considering image patches of different sizes. Classification of image patches of different sizes from a large satellite image is a novel idea in the sense that the patches considered contain significant coverage of a particular type of geographical environment.

Road networks and urban region features were computed from these image patches extracted from the large image. A one-vs-rest Gaussian kernel SVM classification method was used to classify this large image. The classification labels the image patches with the one having the maximum geographical coverage of the area associated in the large image. The large image was mapped into a "region matrix", where each element of the matrix corresponds to a geographical class. In certain cases, this produces anomalies, as a single patch may contain two or more different geographical coverages. In order to have an estimate of these partial coverages, the output of the SVM was mapped into probabilities. These probability measures were then studied to have a closer look at the classification accuracies. The results confirm that our method is able to classify a large image into various geographical classes with a mean error of less than 10%

Page maintenue par le webmaster - 3 février 2010
© Télécom ParisTech/TSI 1998-2010