Hyperspectral Imagery Sparse Unmixing Based On Spatial And Spectral Analysis

Hyperspectral imagery is of very high spectral resolution, containing hunderds of contiguous naorrow spectral bands. The spectra of an imagery pixel can form an almost continuous spectral signature curve, which supplies abundant spectral information of this pixel for more accurate analysis of material composition. Hence hyperspectral imagery has many important applications in a wide range of fields. However, as low spatial resolution, pixels of hyperspectral imagery are often not pure pixels but consisted of different materials, which are called mixed pixels. The mixed pixel phenomenon impedes hyperspectral data's application. Through the unmixing techniques of hyperspectral imagery, the spectral information and abundance could be quantitatively analyzed in subpixel level. Hence, pixels unmixing problem becomes an essential issue in hyperspectral data analysis, the performance of unmixing procedure plays a key role in the sequential application of hyperspectral data.The number of endmembers in a mixed pixel is commomly less than that of the entire imagery, and far awary less than that of the endmember library. This common sense is called the sparseness of mixed pixel. The unmixing method based on the constraint of sparseness could make the unmixing solution sparser and more accurate than the traditional unmixing methods. Just the same as the spectral information, the hyperspectral imagery also owns affluent spatial information. Traditional unmixing methods focus on the imagery analysis in spectral space, while the integration of spatial information would promote the unmixing performance.The major research of this thesis focuses on the sparse unmixing of hyperspectral remote sensing imagery based on spatial-spectral analysis, which contains the following content:(1) The sparse regression-based approach for hyperspectral imagery unmixing is introduced with statement of models and algorithms in details. In the sparse regression unmixing frame, the factors which influence the behavior of hyperspectral imagery unmixing procedure are analyzed with sufficient comparison experiments. Furthermore, the determination of sparse controlling parameter is discussed.(2) The assessment of algorithms' performance is systematically discussed. Besides the overall assessment of algorithms' performance, the evaluating indicators of sparsity of unmixed result and the precision of endmembers recognition are introduced since the sparseness character of mixtures is very important in the unmixing process.(3) The endmembers extracted from hyperspectral imagery own strong spectral variability which affect the unmixing accuracy of mixed pixels. Because of endmembers spectral variability, the sum-to-one constraint of fractional abundances fails to hold. The effect of endmembers spectral variability in unmixing process is studied in the sparse regression unmixing frame.(4) In the sparse regression unmixing frame, the distribution of unmixed fractional abundances in the endmember library is studied. As the analysis of fractional abundances distribution and the sparseness character of the representation of mixed pixels by the endmembers from endmember library, a sparser endmembers sublibrary is extracted from the whole endmembers library. The sparser endmembers sublibrary is closer to the endmembers subset which contains all the endmembers really exist in the mixed pixel. Then unmixing process would be restricted on the sparser endmembers sublibrary.(5) Integration of spatial information to the hyperspectral imagery unmixing process in pixel-wise level is studied. Hyperspectral imagery owns spatial smoothness in some degree. The smoothness of hyperspectral imagery reflects the smoothness of unmixed fractional abundances. Then a smoothness constraint of fractional abundances is added to the unmixing process under the sparse regression unmixing approach, where the smoothness is measured through the spatial and spectral analysis of hyperspectral data.Innovations of this thesis include the following points:(1) A hyperspectral imagery sparse unmixing algorithm based on endmember bundles is proposed. Due to endmembers spectral variability, the sum-to-one constraint of unmixed fractional abundances fails to hold. In the first step of this algorithm, hyperspectral imagery is recorrected by a band-dependent correction process in order to abate the endmember variability induced by the variation of atmospheric conditions. Then a representation model of endmember variability induced by the variation of illumination conditions and corresponding endmembers bundles are established. Endmembers bundles are instead of endmembers in the sparse unmixing procedure. Finally, the fractional abundances are recorrected by scaling factors according to the endmembers bundles. The recorrected fractional abundances would satisfy the sum-to-one constraint. This algorithm would reduce the endmember variability, which mainly caused by the variation of atmospheric conditions and illumination conditions, and promote unmixing performance.(2) A hyperspectral imagery unmixing algorithm based on sparser endmember subset is proposed. Through analyzing the distribution of unmixed fractional abundances after a coarse unmixing procedure in the sparse regression unmixing frame, the endmembers in the representation of mixed pixel induced by noise and endmembers' spectral similarity are kicked out. Then a sparser endmember subset is got and the unmixing process restricted on the sparser endmember subset would suppose sparser and more accurate results.(3) A hyperspectral imagery sparse unmixing algorithm based on nonuniform spatial smoothness constaint measured by spectral homogeneity index is proposed. A spatial homogeneity indicator is introduced in the spatial-spectral analysis of hyperspectral imagery to measure the smoothness of hyperspectral imagery. Then a spatial smoothness constraint, according to the spatial homogeneity indicator, is added to the unmixing process in the sparse regression frame. Compared with unmixing algorithms based on sparse regression, this algorithm would promote the unmixing accuracy and the smoothness of fractional abundances.In summary, this thesis focuses on the sparseness of representation of mixtures by endmember library, the smoothness of fractional abundances distribution in the spatial space, the endmembers spectral variability. In the sparse regression frame, this thesis proposes hyperspectral imagery unmixing algorithms based on sparser endmembers subset, spatial homogeneity analysis and considering endmembers spectral variability, which are benefited try to promote unmixing performance and get unmixed result more concident to the reality.