Développement et automatisation de méthodes de classification à partir de séries temporelles d'images de télédétection - Application aux changements d'occupation des sols et à l'estimation du bilan carbone

摘要

As acquisition technology progresses, remote sensing data contains an ever increasing amount of information. Future projects in remote sensing like Copernicus will give a high temporal repeatability of acquisitions and will cover large geographical areas. As part of the Copernicus project, Sentinel-2 combines a large swath, frequent revisit (5 days), and systematic acquisition of all land surfaces at high-spatial resolution and with a large number of spectral bands.The context of my research activities has involved the automation and improvement of classification processes for land use and land cover mapping in application with new satellite characteristics. This research has been focused on four main axes: selection of the input data for the classification processes, improvement of classification systems with introduction of ancillary data, fusion of multi-sensors, multi-temporal and multi-spectral classification image results and classification without ground truth data. These new methodologies have been validated on a wide range of images available: various sensors (optical: Landsat 5/7, Worldview-2, Formosat-2, Spot 2/4/5, Pleiades; and radar: Radarsat, Terrasar-X), various spatial resolutions (30 meters to 0.5 meters), various time repeatability (up to 46 images per year) and various geographical areas (agricultural area in Toulouse, France, Pyrenean mountains and arid areas in Morocco and Algeria). These methodologies are applicable to a wide range of thematic applications like Land Cover mapping, carbon flux estimation and greenbelt mapping.