Section: New Results

ACI Masses de Données QuerySat

Indexing of remote sensing images using road networks

Keywords : database retrieval, remote sensing, road network, skeleton, graph matching.

Participants : Avik Bhattacharya, Ian Jermyn, Xavier Descombes, Josiane Zerubia.

This Ph.D. is co-supervised by Henri Maître, deputy director of ENST, Paris, in collaboration with Michel Roux from TSI department, ENST, [http://www.tsi.enst.fr ]. QuerySat is partly funded by the French Space Agency (CNES).

Retrieval from remote sensing image archives relies on the extraction of pertinent information from the data about the entity of interest (e.g. land cover type), and on the robustness of this extraction to nuisance variables (e.g. illumination). Most image-based characterizations are not invariant to such variables. However, other semantic entities in the image may be strongly correlated with the entity of interest and their properties can therefore be used to characterize this entity. Road networks are one example: their properties vary considerably, for example, from urban to rural areas. We study the dependence of a number of network features on the class of the image (`urban' or `rural'). The chosen features include measures of the network density, connectedness, and `curviness'. The feature distributions of the two classes are well separated in feature space, thus providing a basis for retrieval. Classification using kernel k-means confirms this conclusion. A careful study of the features shows that the classes are quite well separated in many of the plots, making it reasonable to use these features for classification (see figure 25 ).

Figure 25. Scatter plots of selected pairs of features. Red stars correspond to rural areas, blue circles to urban areas. From left to right, top to bottom: junction density versus average curvature variance; length density versus average curvature variance; junction density versus variance of ratio of lengths; length density versus junction density; variance of ratio of lengths versus average curvature variance; variance of density of junction edges versus length density.

The clustering result, displayed in table 1 , shows that the two classes can be well partitioned using the above five features. 19 and 25 images from `rural' and `urban' classes respectively were correctly classified, while 1 and 7 images from `urban' and `rural' classes respectively were incorrectly classified.