Accuracy assessment of global land cover maps: lessons learnt from the GlobCover and GlobCorine experiences

The validation of global land cover products becomes a critical and challenging issue as more global products are made available more regularly to the international community. The GlobCover 2005 product delivered in 2008 was the first global land cover product at 300 m resolution. Later on, the MODIS 500 m land cover product has been released in 2009 while the GlobCover 2009 is expected to be achieved as soon as early 2009. The GlobCover 2005 accuracy assessment was the first global exercise implemented according to the CEOS Land Product Validation group recommendations in order to assess the thematic accuracy of the 22 different land cover types described using the UN Land Cover Classification System (LCCS). The validation process consisted in three main steps: the reference data sources, the sampling strategy and the accuracy assessment. A specific working environment for on-screen collection of 'ground truth'data has been used by 16 international experts invited for 6 different 5-day workshops. Using web mapping capabilities like Google Earth, Virtual Earth and others as well as 8 years NDVI profiles, the experts were able to characterize more than 4000 validation points using the LCCS classifiers. This validation exercise has been repeated for the GlobCover 2009 product. The same set of experts were asked to confirm their land cover interpretation and possibly, to update their own interpretation from 2005 to 2009. This second set of validation points as well as those submitted to the interpretation of two different experts, allow studying the reference information consistency and quality over a large data set. Moreover, the GlobCorine 2005 product was also analysed with regards to the CORINE Land Cover product for further understanding. The research question of this paper is to investigate why the three last global land cover assessments, i.e. GLC2000, GlobCover, MODIS Collection 5 land cover product, varies from 68% to 74 % as overall accuracy figure. What is the level of legend providing the best accuracy? Is there a natural limit around 75 % of overall accuracy related to the quality of the reference data set? The results of this in-depth analysis are described and the lessons learnt from these global validation exercises are described to support any forthcoming global product validation plan. Furthermore, these results document the practical meaning of a global product overall accuracy figure from an end-user point of view