OPTIMUM BAND SELECTION OF HYPERSPECTRAL REMOTE SENSING DATA

Hyperspectral remote sensing data with waveband width of nm level has tens or even several hundreds channels and contains abundant spectral information. Different channels have their own properties and show the spectral characteristics of various objects. Selecting optimum bands from the varieties of channels is very important for the effective analysis and information extraction of hyperspectral data. This paper, taking Shunyi district of Beijing as a study area, comprehensively analyzed the spectral feature of hyperspectral data. On the basis of analyzing the information content of bands, correlation among different channels, band separability and spectral absorption characteristics of objects, a fundamental method of optimum band selection from hyperspectral remote sensing data was proposed. Three factors, the information amount of bands, correlation between bands and separability of objects in bands, are considered in selecting bands. The major steps of band selection are: (1) Compute the correlation matrix of hyperspectral data, analyze the correlation between bands, and then according to the correlation partition the complete data set into three band groups. The bands in same group are highly correlated and the different groups are relatively independent. (2) Considering that hyperspectral data has many channels which appear in groups, define the band index as P i, in which σ i is standard variance of band i, R w is absolute value of average correlation coefficient between band i and other bands in same group, R a is the sum of absolute value of correlation coefficient between band i and all other bands in different groups. It is evident that with higher R w and lower R a, the P i is higher and the corresponding band i is better in whole. Thus P i is an important parameter in selecting band. (3) Select several typical spectral classes as training samples, which are important objects to be classified in study area and have similar spectral feature, compute the separability of classes in different bands by Bhattacharyya distance. (4) On the basis of band comprehensive evaluation by band index and separability, select optimum bands bearing abundant information and high separability. The method derived from hyperspectral data of Shunyi District was also applied to different types of hyperspectral data of other region and similar conclusion was got. It shows that the proposed method in this paper is of general significance