Scientists at the CBK PAN have developed a fast new method to automatically classify land cover at high resolution. A prototype map of Europe has just been released under a contract with the European Space Agency.
The new map shows land cover in Europe in 2017. Thirteen of the most important classes are highlighted, including agricultural areas, forests (both deciduous and coniferous), built-up areas, marshes and peatbogs. Fifteen thousand images from Sentinel-2 satellites were processed. With a spatial resolution of 10 meters per pixel, the new dataset is 2,500 times more detailed than the existing CORINE land cover inventory. Moreover, as production was fully automated, it took six weeks to generate – compared to 18 months for CORINE.
The new method was developed by Earth Observation scientists from the CBK PAN (space research institute of the Polish Academy of Sciences). “We adopted a machine learning approach to deal with the huge number of observations. First, the algorithm was supplied with reference data from low-resolution databases. Next, it learned – on its own – how to discriminate between individual land cover types” said Professor Stanisław Lewiński, the project’s principal investigator. Researchers tested their approach on maps of China, Colombia, Namibia, Germany and Italy, before eventually applying it to the generation of a new land cover map for the whole of Europe.
The biggest challenge proved to be the reliability and applicability of reference data. Typically, the algorithm is trained with data that is at least as detailed as the resulting map. However, global databases provide information that is several dozen times less precise. “We were forced to rely on data that was much more uncertain than we anticipated, so we had to come up with a robust analytical scheme. Eventually, we reached 86–89% accuracy on the European scale”, said Lewiński. Accuracy is even higher for individual countries: 93% for Poland, 96% for Germany, 91% for Finland, and 93% for Croatia and Slovenia.
A key factor in the project’s success was the nature of satellite data. Images come from the two Sentinel-2 satellites launched in 2015/2017 under the European Copernicus program. Working in tandem, the pair fly over Europe every 5 days. This frequent revisit allowed the Polish team to classify land cover anywhere in Europe multiple times, based on the 20 best annual acquisitions. The aggregation of observations not only resulted in a highly-accurate map, but also solved the biggest problem in satellite imagery – cloud cover.
Work was carried out as part of the three-year Sentinel-2 Global Land Cover (S2GLC) project, funded by the European Space Agency under the Scientific Exploitation of Operational Missions initiative. The CBK PAN scientists developed the classification scheme and software. The consortium of international partners includes IABG, EOEXPLOR and the University of Jena in Germany. Detailed calculations were carried out by CloudFerro, operator of the CreoDIAS platform.
