Multi-satellite cloud product generation over land and ocean using canonical coordinate features

Typical neural network approaches to cloud classification use labels that are arbitrary in interpretation. A new method of labeling is presented in order to provide more meaningful labels for clouds based on their compositions. These labels are derived from the cloud products of the MODerate resolution Imaging Spectroradiometer (MODIS) instrument. The products from the MODIS instrument are widely accepted as the state-of-the-art by the meteorological community, but are not available at a regular temporal frequency over one specific region due to the polar orbiting nature of the instrument. Moreover, the algorithms used to create these products cannot simply be applied to other more temporally regular satellite data. An innovative method is presented to estimate MODIS-like cloud products from the geostationary satellite Meteosat 8. This method exploits the common characteristics between MODIS and Meteosat 8 by using a canonical coordinate decomposition (CCD) to develop features of Meteosat 8 that are most coherent with MODIS for classification purposes. The CCD analysis is performed separately between the visible, infrared, and water vapor channels of both satellite instruments. The canonical coordinate features of Meteosat 8 are labelled with the MODIS product labels, when both data sets are present, to create the training set for a back-propagation neural network (BPNN) classification system. The quality of the generated cloud products has been demonstrated on several data sets from July 2004.