Evaluation of Sentinel-2: Case Study of Land Cover Classification and Ice Velocity

A new era of open-access satellite data is upon us and satellite sensors as the Sentinel-2A and its twin satellite Sentinel-2B, will together provide imagery with relative high temporal resolution, and improve state of art environmental monitoring and science studies, as on e.g., land cover, cryosphere and ocean. The team behind Sentinel-2 is reporting a high measured co-registration accuracy between repeat orbit, as this is crucial for users by the Sentinel-2. However, registered co-registration displacement between two different relative orbits is reported to be high, and limits the full potential of Sentinel-2 data. Firstly in this thesis, a study of the geometric performance was carried out, testing both repeat and different relative orbits for the Sentinel-2A sensor. The study was carried out on different topography in order to determine errors related to the digital surface model used in the ortho-rectification of Sentinel-2A scenes. The main result was, because of a poor representation of the true surface of the digital surface model used in the ortho-rectification process, that horizontal displacement between two correlated images from different relative orbits is larger in high elevated topography where the surface model does not cope the steep increase in mountainous topography. Secondly this thesis has measured the ice velocity of three glaciers in Norway and four calving glaciers at the Greenland Ice Sheet using the Sentinel-2A satellite, as it is vital to know ice velocity dynamics as a response to climate change. The measured ice velocity was found to have high accuracy, and possible synergy with other Sentinel satellites, as e.g., Sentinel-1 is proposed. Thirdly, a land cover classification was assessed in a typical agricultural landscape in southeast Norway and western Sweden, by using both Landsat-8, and Sentinel-2A satellites for comparison of spectral differences, processing levels and seasonal changes. Overall good agreement was found on all sensors and processing levels, but classes which shares similar spectral response were poorly mapped by all classifications parameters