Incorporating Sub-Dominant Classes in the Accuracy Assessment of Large-Area Land Cover Products: Application to GlobCover, MODISLC, GLC2000 and CORINE in Spain

Various global land cover (LC) datasets have been produced from remote sensing data in response to the need for information about LC. Nevertheless, the potential use of global LC products is often very limited by the lack of detailed accuracy information at regional to national level. This paper proposes a methodology for performing accuracy assessment of large-area LC products, which takes into account a number of factors arising from intrinsic characteristics of LC, such as thematic uncertainty that results from the partial overlap in legend definitions and lack of homogeneity within reference and classification data. The approach compares the LC pixel label not only with the dominant reference label but also with sub-dominant LC types within the extent of the sampling unit. The methodology was illustrated in Spain using four LC datasets (GlobCover, MODIS land cover (MODISLC), GLC2000 and CORINE). The variety of reference label data offered by a detailed national database, namely SIOSE, supported several different fuzzy agreement definitions in order to derive unbiased estimates of accuracy measures. CORINE followed by GLC2000 showed the highest accuracy scores, whereas GlobCover and MODISLC showed the lowest scores. Nevertheless, the fuzzy approach revealed that a great amount of disagreement in GlobCover and MODISLC datasets does not actually correspond to classification errors, but it can be associated to legend ambiguity and mixed coverage pixels.