Abstract
Recently, remotely sensed multispectral data have been proved to be very useful for many applications in the field of Earth surveys. For certain applications, however, limits in the spatial resolution of satellite sensors and variation in ground surface restrict the usefulness of the available data, since the observed spectral signature of the pixels is the result of a number of surface materials found in the area of the pixel. Two mixed pixel classification techniques which have shown high correlation with vegetation coverage of single pixels are described in this thesis: the vegetation indices and the linear mixing model. The two approaches are adjusted in order to deal with sets of pixels and not individual pixels. The sets of pixels are treated as statistical distributions and moments can be estimated. The vegetation indices and the linear mixing model can then be expressed in terms of these statistics. The illumination direction is an important factor that should be taken into account in mixed pixel classification, since it modifies the statistics of the distributions of pixels, and has received no attention until now. The effect of illumination on the relation between the vegetation indices and the proportion of sets of mixed pixels is examined. It is demonstrated that some vegetation indices, which are defined from the ratio of statistics in two spectral bands, can be considered relatively invariant to illumination changes. Finally, a new illumination invariant mixing model is proposed which is expressed in terms of some photometric invariant statistics. It is shown to perform very well and it can be used to un-mix accurately sets of pixels under many illumination angles. The newly introduced mixing model can be considered a suitable choice in the mixed pixel classification field. Key words: Mixed pixels, sets of pixels, vegetation index, illumination invariants.