Unsupervised Image Endmemeber Definition in Multisensor Image Data

In a recent studie (Greiwe and Ehlers, 2005), the author developed a decision based data fusion approach for the analysis of spatial and spectral high resolution image data of an urban scene. Image segments of high resolution orthophotos were classified with additional material information derived from hyperspectral image data. To retrieve this material information, the hyperspectral image data were classified by the Spectral Angle Mapper (SAM). Results of SAM were included into the classification of the high resolution orthophotos. The inclusion of hyperspectral image data significantly increased the classification accuracy of a high resolution orthophoto. One of the steps of the data fusion approach described above that more often requires the user input , is the SAM classification of hyperspectral data. One of the preconditions for the analysis of hyperspectral image data is the definition of a set of material reference spectra – often called image endmember. A manual selection of image pixel as a definition of image endmember leads to user dependent results. To avoid this, many approaches for unsupervised image endmember definitions, such asPPI or N-FINDR (Winter 1999) have been developed. These algorithms detect those pixels that define the convex hull of the n-dimensional feature space of hyperspectral image data. This produces endmembers that express ‘spectral extreme’ features. Although the resulting endmembers are useful for spectral unmixing approaches, they are not suitable for material detection approaches like SAM. Algorithms like SAM, determine the spectral similarity between a pixel’s spectra and a given endmember. In contrast to endmembers that are defined by ’spectral extremes’, SAM needs endmembers that represent mean spectra of material types in order to produce best possible results between spectrally similar material classes. The objective of this work is the development and implementation of a unsupervised image endmember definition approach for material detection methods like SAM. Information on the high spatial resolution orthophoto is used to detect homogeneous areas in the hyperspectral image data. Pixels of the hyperspectral image data in such homogenous areas are marked as endmember candidates. Then, a spectral correlation analysis (van der Meer and Bakker 1997) is used to calculate the spectral similarity between the candidates. At n given candidates, the n*n correlation matrix of all candidates is introduced as a new feature space that expresses spectral similarity between the candidates. Candidates with similar spectral behavior are grouped by a density based cluster algorithm. The mean spectrum of each cluster is stored in a spectral library for further processing. In an urban test site, several endmember for different materials could be defined by the proposed approach.