Analysis of structure and context for cloud classification

Human eye is inherently superior to computers in distinguishing land, water, ice, and clouds from satellite images, but is limited in operational applications. An ideal tool is a computer aided, at least partially, with the distinguishing qualities of the human eye. Computer analysis of satellite images in the past were mostly based on pixel-wise classification. Each pixel was attributed to one of the fixed classes, e.g. the class of cloud, water, or land. Today, however, it is possible to divide an image into segments instead of per-pixel classification. Then with the "Region Merging" method, pixels are combined until the internal inhomogeneity of a region exceeds a threshold. The classification and analysis is afterwards based on these segments, where the size of a segment varies between a pixel to the entire scene. A basic requirement is that a smaller segment cannot extend beyond the border of a larger segment. Each segment can be classified by using its features, e.g. the relationship between the length and the width. The classes in the neighbourhood of the segment can also be used in the classification. The technique allows an exact classification, e.g. to identify a lake as a lake and to distinguish it from a river. Using "Fuzzy Logic" it is possible to perform an extended differentiation between classes. In the present work, the method is extended to classify clouds based on context and structure analysis. It is applied on AVHRR data to distinguish different types of clouds, e.g. contrails are separated from cumulonimbus. Convective systems are also automatically identified. An objective analysis of clouds allows a long time evaluation of satellite images, e.g. from Meteosat. The results of such an analysis are beneficial to the studies of changes in water circulation.