Contextual classification of hyperspectral remote sensing images Application in vegetation monitoring

The goal of this thesis is the study of strategies for including contextual information in the classification of hyperspectral remote sensing images. The objectives are twofold.;The first objective is the development of new techniques for including contextual information. To this end, an important category of techniques, i.e. modelling the relationships between local pixel neighbourhoods as Markov Random Fields, is first considered. A strategy to extend the flexibility of this technique, by describing the classification problem at hand by an extended hierarchical tree, is introduced. The second technique under study, i.e. the state-of-the-art approach to extract contextual information in the form of attribute profiles, is extended to colour images. As a practical application, two images from the same scene, including a hyperspectral and a high spatial resolution colour image, are jointly classified by first extracting colour attribute profiles from the latter. In addition, a hybrid decision fusion approach is proposed to perform the classification. The third technique, developed in this work, is an approach for assessing the accuracy of contextual classification results, by introducing a new reference, and considering a new measure, based on the complexity of edges, i.e. transitions between classes.;The second objective of this thesis is the application of contextual classification techniques to the essential problem of assessing the conservation status of Natura 2000 habitats. The main challenge is in the structural complexity of most of the habitats under study, since these habitats display a high degree of heterogeneity and, in addition, cannot be simply identified by the presence of a single or a few dominant species. In order to handle this complexity, a contextual framework has been developed to reduce the problem to a number of more manageable sub-problems. First, the list of habitats is translated to a hierarchical scheme that includes the most important land cover types that these habitats are composed of. Second, the already developed hierarchical Markov Random Field approach is applied for classification. Finally, a contextual reclassification approach is applied to the resulting land cover map in order to construct a habitat map, suitable for determining the conservation status of the habitats.