| Spectral unmixing is the process of decomposing the spectral signature of a mixed pixel into a set of endmembers and their corresponding abundances. Endmembers are spectra of the pure materials present in the image and abundances at each pixel represent the percentage of each endmember that is present in the pixel. Many spectral unmixing techniques treat a pixel as independent of its neighbors, therefore, only spectral characteristics of the image are used to address the spectral unmixing problem. However, a number of recent studies have found that spatial autocorrelation provides useful information for spectral unmixing. Combining spatial information with its spectral counterpart can lead to improvements in the unmixing results. In this dissertation, the unmixing methods that incorporate spatial information are termed spatial spectral unmixing, whereas those exploiting only spectral information are referred to as spectral-only unmixing. The available spatial spectral unmixing methods are thoroughly reviewed according to the following three categories: 1) endmember extraction, 2) selection of endmember combinations, and 3) abundance estimation. In addition, a suite of novel spatial spectral unmixing methods are proposed: (1) a spatial statistical preprocessing method is developed for the detection of homogenous regions prior to endmember extraction; (2) spatially-interpolated endmembers are investigated for spectral unmixing in order to account for endmember variability; (3) a linear spatial spectral mixture model is developed to incorporate spatial interactions arising from adjacency effect in abundance estimation. The experimental results, with both synthetic and real hyperspectral images, demonstrate the effectiveness of the proposed methods. |
