Multilevel fusion for classification of very high resolution remote sensing images. (Fusion multiniveau pour la classification d'images de télédétection à très haute résolution spatiale)

Remote sensing is a promising technology that finds as diverse applications as defence, urbanplanning, healthcare, and environmental management. Collecting countrywide statistics of cropyield is one of the main tasks of remote sensing. Acquiring and processing very high-resolution(VHR) satellite images are means accomplishing this task. Processing these remotely sensed(RS) images requires not only great computational power but also efficient algorithms for imagesegmentation and classification.This thesis aims at presenting the work carried out for applying computationally efficientspectral and textural analysis on very high-resolution RS images, and combining the resultsfrom the two analyses for improved classification of vegetation covers. The spectral analysispresented here adopts the unsupervised approach of classification, whereas the textural analysisadopts the supervised approach of classification. The fusion of the contour information from theunsupervised spectral analysis with the pixel class information from the supervised texturalanalysis yields successful classification results.The thesis takes as a test case, a site covered with orchards, truck crops, crop fields, vineyards,forest, and fallows from Nîmes region, France. The real contribution includes improved versionof the unsupervised classification method based on k-means clustering, a method of introducingrotation invariance into the texture features based on discrete Fourier transform, and a methodof fusing a supervised classification with an unsupervised classification. This thesis is all aboutdeveloping these algorithms.