Classification supervisée d'images d'observation de la Terre à haute résolution par utilisation de méthodes markoviennes

Abstract : The classification of remote sensing images including urban areas is relevant in the context of the management of natural disasters (earthquakes, floodings...), and allows to determine land-use and establish land cover maps, or to localise damaged areas. The supervised classification methods developed during this PhD thesis are essentially based on Markovian models. The first part of the study deals with the classification of single-polarized, mono-resolution synthetic aperture radar (SAR) amplitude images. The selected method consists in modeling the class-conditional statistics by resorting to finite mixture models, and then to plug these statistics into a Markov random field (MRF). To improve the classification results in urban areas, affected not only by speckle noise, but particularly difficult to process given their heterogeneity, we extracted a textural feature from the initial image which aims at discriminating the urban areas (e.g. Haralick's variance). The textural feature statistics are combined with those of SAR amplitude image by using bivariate copulas. Next, we extended the single-scale model to a hierarchical Markov random field integrated in a quad-tree structure. This model includes a prior update that experimentally leads to results less affected by speckle. The mono-resolution data are hierarchically decomposed using a wavelet transform. The main advantage of this approach is that multi-resolution and/or multi-sensor acquisitions can directly be integrated in the tree. We applied the proposed hierarchical classification to SAR COSMO-SkyMed multi-resolution acquisitions, and to coregistered optical/SAR data. At each tree level, the statistics are independently modeled by finite mixtures of Gaussian distributions when considering optical images, and by finite mixtures of generalized Gamma distributions when considering the SAR amplitude data. Such statistics are then combined by using multivariate copulas, and plugged in the hierarchical model. We tested and validated these methods on real SAR and optical data.
Document type :
Image Processing. Université Nice Sophia Antipolis, 2012. French
Contributor : Aurélie Voisin <>
Submitted on : Friday, November 2, 2012 - 5:22:49 PM
Last modification on : Saturday, November 3, 2012 - 7:31:53 AM
Document(s) archivé(s) le : Sunday, February 3, 2013 - 3:37:22 AM


  • HAL Id : tel-00747906, version 1



Aurélie Voisin. Classification supervisée d'images d'observation de la Terre à haute résolution par utilisation de méthodes markoviennes. Image Processing. Université Nice Sophia Antipolis, 2012. French. <tel-00747906>




Notice views


Document downloads