Skip to Main content Skip to Navigation
Theses

Segmentation Spatio-temporelle d'une séquence d'images satellitaires à haute résolution.

Abstract : Image Time series represent an amount of information far greater than individual images. In fact, considering time increases significantly the number of possible states corresponding to a bigger amount of information (in an entropic sense). Thanks to the new generation satellites agility and to their use in onstellations, satellite image time series, SITS, will from now on be available at higher resolution. These data are extremely rich, but, in compensation, are complex and difficult to interpret manually. Automatic analysis methods are thus required. High resolution SITS (HRSITS) differ from the other types of existing sequences by the type of change they show. In fact, contrary to low resolution SITS, HRSITS present objects. In order to take this particularity into account, an object oriented analysis must be performed. Such approaches exist in the video domain. However, the HRSTIS' objets may undergo changes in radiometry (for example, due to the plants growth) whereas the video object's radiometry is generally assumed time invariant. We thus need to design an object oriented analysis adapted to HRSITS. Besides, the temporal sampling of the HRSITS is irregular and generally sub-sampled compared to visible phenomena at the considered spatial resolution. Moreover, from the satellite viewpoint, some radiometry changes due to the atmosphere thickness cumulate with the radiometry evolution of the scene objects. Nevertheless, despite slight geometrical registration errors, the HRSITS' objects shape is temporally redundant. A building is in fact generally sustainable whereas an agricultural zone is rarely modified. We thus propose an object based HRSITS analysis method exploiting on one hand the intra-objects radiometry redundancy and on the other hand the temporal redundancy of the shape of the objects different versions. A phenomenological study of the dynamic of the scene and of the HRSITS allows us to identify the characteristics of an adapted representation of the changes description : a graph the nodes of which are the spatial objects linked by edges representing their temporal dependency. However, the computation of this graph is a difficult problem, and we thus propose to compute an approximation in two steps. In the first step, a strong approximation is considered on the temporal dependencies in order to facilitate the determination of spatial regions. These are then extracted thanks to a segmentation algorithm exploiting jointly the two kind of available redundancies, namely the spatial homogeneity of the radiometry and the temporal geometric redundancy. The nodes of the graph are thus determined and the approximation on the temporal dependencies can then be relaxed in order to achieve a finer estimation of the graph. We then propose two utilisations of the graph. The first one exploits the structural characteristics such as the nodes degrees in order to detect and qualify geometric changes. The other uses an attributed version of the graph with radiometry properties attached to the nodes in order to extract relevant frequent behaviour.
Document type :
Theses
Complete list of metadatas

Cited literature [270 references]  Display  Hide  Download

https://pastel.archives-ouvertes.fr/pastel-00658159
Contributor : Camille Le Men <>
Submitted on : Tuesday, January 10, 2012 - 3:17:15 AM
Last modification on : Friday, July 31, 2020 - 10:44:06 AM
Long-term archiving on: : Wednesday, April 11, 2012 - 2:22:33 AM

Identifiers

  • HAL Id : pastel-00658159, version 1

Collections

Citation

Camille Le Men. Segmentation Spatio-temporelle d'une séquence d'images satellitaires à haute résolution.. Traitement des images [eess.IV]. Ecole nationale supérieure des telecommunications - ENST, 2009. Français. ⟨pastel-00658159⟩

Share

Metrics

Record views

1164

Files downloads

2190