In this work we present a cloud mask for MERIS, developed as part of the EU NAOC project, which can discriminate between optically thick and thin clouds. The method is based on the expert selection of a small labelled data set of cloudy and cloud free pixels in MERIS observations taken over the ocean, guided by meteorological knowledge. This small labelled data set is augmented by a larger unlabelled data set randomly extracted from a number of MERIS scenes over the ocean. This unlabelled data set is used to characterise the structure of the MERIS spectra that are observed, using pattern recognition methods called the generative topographic mapping and Neuroscale. The generative topographic mapping constructs a density model for the 16 dimensional (i. e. the MERIS bands and a ratio between the radiances at the 11th and the 10 th bands) data in a lower (typically 2) dimensional latent space, which allows visualisation and understanding of the structure and distribution of the data. The Neuroscale algorithm is a distance preserving data projection algorithm without a density model. The lower dimensional structure is then used to define a non-linear projection, which retains information, but permits the construction of simpler classification models, something that will be especially important with future hyper-spectral instruments. We show the results of our cloud classification on several MERIS scenes and contrast our cloud mask with the standard MERIS cloud mask.
|Number of pages||6|
|Journal||European Space Agency (Special Publication)|
|Publication status||Published - 1 Dec 2004|