Radiance spectra classification from the Ocean Color and Temperature Scanner on ADEOS

Multispectral information from the ocean color sensors of remote sensing satellites can be used to classify the ocean surface waters into a number of classes. Nine areas distributed over the Pacific Ocean have been used to demonstrate this approach. Unsupervised neural networks were used to separate water pixels from land and cloud pixels and classify water into a variety of ocean colors. Self-organizing feature maps chose radiance spectra by minimizing least square differences amongst multichannel pixels. Pixels with similar radiance spectra were coded with similar colors. It has been shown that radiance spectra, after correction for the atmospheric absorption of a “standard atmosphere” for varying Sun and satellite viewing angles, could be classified into a single set of radiance spectra that apply over the whole ocean. No ground truth data were required to make this classification. Examinations of the classified images showed that the method could extract a large number of ocean color categories and provide a basis to separate case 1 waters from the case 2 and ocean radiances with a high influence of the atmosphere. Also, areas of high pigment, inappropriately masked out by the conventional routine, were correctly classified. This opens the possibility that in the future a robust global algorithm for chlorophyll estimation might be constructed