The European Space Agency (ESA) through its Scientific Exploitation of Operational Missions (SEOM) element, are funding a project entitled Sentinel-2 Global Land Cover or S2GLC for short. This project will focus on the classification of Sentinel imagery for the purpose of producing a global land cover map. The project began on February 1, 2016 with a total duration of 2 years and is led by CBK PAN with three partner organisations. In order to maximise the output of this study, five test sites were chosen to test and validate the applied classification techniques: Italy, Germany, Namibia, Columbia, and China. Each test site covers an area approximately 200 000 km2 and are found in a variety of bio-geographical locations in order to maximise the types of land-cover to be classified. The choices were a balance between access to good validation data, landscape variability, and the technical realities of testing multiple classification algorithms and methodologies. The first part of this study is an extensive review of the currently available Global Land Cover (GLC) maps and databases. The review study is expected to be ready around the time of the conference and the team believes that it is a perfect venue to present our review to get constructive feedback from the users of GLC databases in attendance. This review, together with feedback from the community, will influence the choices in algorithms and image processing methodologies tested within the scope of this study. The second and third parts of the study are testing of the land-cover classification methodologies and validation of those methods respectively in order to produce not only the highest quality maps, e.g. accuracy >80%, but also harmonised with current GLC products. In order to achieve this complex goal, many different tests of object-oriented as well as pixel based classification approaches will be made. In parallel, advanced data collection strategies for training and validation will be investigated. While the majority of the applied land-cover classification techniques will be based on optical imagery acquired by Sentinel-2 (S2), the team understands that globally this challenge can be supported by the Sentinel-1 SAR data. The different approaches will be benchmarked in order to understand the influence of a variety of factors on the performance of the proposed methods. Factors will include feature relevancy, the impact of atmospheric correction, the selected minimal mapping unit, seasonal changes, the incompleteness of training data, image mosaicking, and multi-temporal S2 data. The final part of the project will be to make recommendations based on the research for future S2 based GLC products.
S2GLC legend has been created based on the comparison and analysis of existing Global Land Cover data bases legends. Moreover it can be modified in case of discovering a significant need for adding or removing certain classes. Any possible changes can be easily adopted due to hierarchical structure of S2GLC legend.
Below we present example of classification results computed for one of Implementation Sites in Germany.