Optimal Band Selection Methods Of Hyperspectral Remote Sensing Data

Compared with multispectral remote sensing, the characteristics of hyperspectral remote sensing data are more number of channels, higher spectral resolution, narrower bandwidth and larger amount of data.It is very important making optimal band combination in numerous bands for false color synthesis according to specific application purpose for further managements analysis and information extraction of hyperspectral remote sensing dada.With the example of the complex ground surface hyperspectral image of nearby the Huangpujiang in Shanghai obtained by the airborne Pushbroom Hyperspectral Imager (PHI) ,one the one hand, the information contents the correlation of different channels and spectral character of ground objects are analysed to extract band subsets with vast scale information content , smaller correlation, larger spectral difference ,then selected optimal band combination of hyperspectral remote sensing image by way of combine_entropy,covariance matrix, optimal index and band index based on information content;on the other hand,according to optic explain primary ground objects in the image are classed for nine classes by way of non_supervision,then caculated stastical distance(standard distance, measure of dispersions B_distance) and spectral correlation coefficient, similarity coefficients mixed distance and Eulerian distance to get max (mix for correlation coefficient) for the two ground objects differentiated effectively between arbitrary two classes on arbitrary three band combination,in the end ,the whole image is taked into account to select optimal band combination interpretated the whole image effectively.Band combinations by way of all kinds of methods for false color composition are analysed,in conlusion:theoretically speaking,combine entropy can reflect information content of image effectively but the result is not perfect. Covariance matrix can get bands with ample information content but it is a model educed under normal distribution condition,in addition its defect is large numbers of data, slower disposing speed. Optimal index cannot meet for hyperspectral data,in spite of considing standard deviation of image and correlation of between bands,it can make adjacent band combination,after considering spectral character of ground objects it can do better.Band index not only contains information content of image and correlation of between bands but also it is a model educed under grouping condition,so it is better than optimal index for hyperspectral data.When selecting bands by divisibility model between classes,sometimes inverse matrix of covariance matrix does not present,this confines application of measure of dispersion and_B_distance,whereas simpler standard distance between mean value contains mean value and standard deviation,its result do better than measure of dispersion and_B_distance,but it is a measuring of one dimension feature space not fit for the study of multi-variance.Spectral correlation coefficient based spectral dimensioncontains not only spectral itself but mean value of band combination,its result does better than SAM,whereas SAM reflected vector center pixel vector does better than spectral mixed_distance.