Research On Filtering Noise In Spectral Domain Of Airborne Hyperspectral Remote Sensing Images

Hyperspectral remote sensing, as one of the most advanced techniques in the development of remote sensing, has been applied successfully to many domains. Hyperspectral remote sensing technique can provide detail spectra of different surface types, so it is applied widely. And the precesion improvement of hyperspectral applicaion depends on the improvement of hyperspectral image spectra's SNR, so, both noise in spatial domain and spectral domain need to be filtered before hyperspectral data are applied. Having PHI data(one kind of airborne hyperspectral remote sensing data) as the research object, the detection and analysis of noise in spectral domain, the calculation of SNR, the removal of noise in spectral domain and the effect of different noise filter methods on hyperspectral data are discussed. The dissertation can be categoried as following:1, Noise in airborne hyperpsectral data is analysed using standard variance, histogram, correlation coefficient among bands and image spectra; the methods on calculating SNR of hyperspectral data are summarized; a new method on calculating SNR of hyperspectral data is developed based on the calculating of SNR of the reflectant spectral curve; Tests and comparison of SNR calculating methods are performed using 2003 PHI data.2,DSGF(Derivative based Savitzky-Golay Filter) method is given in this dissertation. First, noise level of each especial band is determined based on second derivative of reflectance , then reflectance is filtered twice by Savitzky-Golay filter sizes, which results in the last, spectral noise in hyperspectral image removed, but most fine features of reflectance remained, after hyperspectral image is filtered pixel by pixel with DSGF filter.3,Wavelet transform is used to filter noise of spectral domain in hyperspectral images: first image spectra are decomposed to smooth components and noise components by multi-resolution anaylsis of wavelet transform; then the wavelet tranformed results of different surface image spectra are analysed and it is thought that wavelet inverse transform, with D1 and D2 components removed, can reach the best trade-off between noise reduction and the preservation of spectral features. 4,Methods on filtering noise of spectral domain are analysed and evaluated according to statistical tests, spectral comparability computation and class distance computation; the hyperspectral data before and after filtering is classfied using MLC method and SAM method; it is found that the effect of spectral noise filtering on precesion of MLC classfication is weak, but as to the result of SAM method, the spectral noise filtering not only improves the classification precesion but aslo reduce the number of unclassfied pixels in hyperspectral image.