Retrieving sub-pixel land cover composition through an effective integration of the spatial, spectral and temporal dimensions of MERIS imagery

摘要

MERIS on Envisat delivers imaging spectroscopy data at 300m spatial resolution. MERIS has demonstrated its great potential for regional and global land cover mapping. This paper illustrates that the combination of the spatial, spectral, and temporal dimensions of MERIS has, in addition, the potential to retrieve sub-pixel land cover composition. Three MERIS FR Level 1b scenes acquired over The Netherlands in April, July and August 2003 were used in this study to derive fractional composition of 9 main land cover types. Linear spectral unmixing (with an optimized number of endmembers per pixel) was applied in both a mono- and multi-temporal fashion. A morphological eccentricity index (MEI) was used to explore the MERIS spatial dimension and, subsequently, to support the selection of the endmembers. The Dutch land use database (LGN5) was used as a reference in this study. Classification accuracy was assessed both at sub-pixel and per-pixel level. The best classification results were obtained for the combined image of April and July with a classification accuracy above 58%. In general, sub-pixel and per-pixel classification accuracies were similar. Spectral confusion was detected for several classes and dates indicating that the phenological status plays an important role in choosing the optimal acquisition date combination.