The Analysis of Multispectral Image Data with Self−organizing Feature Maps

The analysis of multispectral sceneries is still a challenging task although many different algorithms for a mathematical analysis exist. Most classification algorithms work in a supervised mode only, i.e. they need to know the possible land usage classes prior to the calculation. In this step many simplifications and assumptions have to be done which directly influence the result. KOHONEN ́s self−organizing feature maps, which are based on a biological model, provide a powerful tool to describe the multispectral scenery under consideration with a limited number of reference vectors, the so−called codebook. The resulting code−book, generated with an unsupervised learning scheme, is a compressed description of the multi−dimensional data in terms of non−linear principal components which thus overcomes the problems of a linear principal component analysis. Using this code−book as a basis for a lateral fully interconnected network and introducing a non−linear activity flow between radial− basis functions located at the positions of the reference vectors results in an unsupervised clustering scheme. This method has been successfully adapted to multispectral sceneries recorded by casi (compact airborne spectrographic imager).