S2GLC operating on CREODIAS
Sentinel-2 Global Land Cover (S2GLC) project, founded by ESA was implemented in 2016 - 2019 by a consortium led by CBK PAN together with three partners: IABG, EOXPLORE and the University of Jena. The project was divided into two phases: Development of LC classification on a global scale and LC classification of Europe.
The goal of Phase 1 was to develop the classification methodology ready to use for the global LC mapping based on Sentinel-2 images. According to the developed methodology each Sentinel-2 tile has been classified separately using a set of multi-temporal images. To maintain the 10 m resolution of Sentinel-2 imagery a pixel-based approach has been chosen. Multi-temporal images have been classified separately with Random Forest (RF) classifier applying training samples selected randomly from low resolution databases. The final result has been calculated using a dedicated aggregation procedure. This aggregation analyses, for each individual pixel, all multi-temporal results together with their prediction scores returned by RF classifier. The last step of the classification workflow was post-processing applied mostly to the pixels classified with low probability. The developed classification procedure was tested on five study areas: the whole territory of Germany and Italy and three other test sites in China, Columbia and Namibia.
The main objective of Phase 2 of the project was performing LC classification over a major area of the European continent in the CREODIAS infrastructure. For this purpose, the developed methodology was adjusted to the European landscape and restructured to full automation of all processes. The classification method and legend established in the initial part of S2GLC project has been also adapted to European conditions. Training samples have been derived from existing LC databases – CORINE LC and High Resolution Layers.
All the classifications were carried out on the CREODIAS platform. Over fifteen thousands of Sentinel-2 images representing 815 Sentinel-2 tiles have been processed using a dedicated software developed by CBK PAN. Validation of the final map was performed based on a large set of 52 000 randomly distributed samples. The overall accuracy (OA) of the complete map with 13 LC classes was estimated to be over 86% and 89% for the legend reduced to 10 classes.
More information can be found at the S2GLC project website: http://s2glc.cbk.waw.pl
Example of S2GLC, as seen in CREODIAS promotional materials