In the past three decades, with the imaging spectroscopy technology unceasing development, the remote sensing images obtained by imaging spectrometer aboard the plane or satellite platforms contain more and more rich information about space, radiation and spectrum, and thus provide a powerful means for surface material information extraction and target detection. The most significant characteristic of hyperspectral remote sensing data is its fusion of the spatial information and spectral information, it enables us to acquire the ground surface spatial image and get the corresponding continuous spectral information at the same time, thus realize the information inversion and geophysics recognition based on information of object spectrum. Hyperspectral technology has shown superiority and importance in the fields such as crop yield estimation, resource survey and so on, but one problem it faces with is the mixed pixels. The spatial resolution and spectral resolution is always a pair of contradictory for a set of optical remote sensor. Hyperspectral unmixing is to decompose the mixed pixel into “basic components”, which can also be called “endmembers”. In this way, the remote sensing application could reach the sub-pixel level. In recent years the unsupervised algorithm has been applied to hyperspectral unmixing and been paid more and more attention. Among this kind of method, the one based on the model of nonnegative matrix factorization has shown great performance, and some constraints have been introduced into this algorithm to render better estimates. Most constraints in this respect only paid attention to the hyperspectral data cube’s macroscopic spatial and spectral properties and did not deeply excavate the inner characteristic of the spectrum itself and the physical mechanism of the spectrum's formation. This thesis takes the theoretical basis of material molecular spectroscopy and inner property of the spectrum as the starting point, and then introduces absorbance curve which is common in infrared spectroscopy and also presents the intrinsic properties of this curve. Based on the obvious transformation relation between spectral reflectance and its relevant absorbance curve, a characteristic of spectrum is introduced, that is the sparseness of the spectrum’s Gabor transform coefficients. We corporate this sparisity-constraint into the cost function and propose a novel algorithm. Results obtained with synthetic and real data are used to...
修改评论