Automatic Spectral Unmixing From Hyperspectral Remote Sensing Imagery

Hyperspectral imaging spectrometer which provides images with very high spectral resolution, can provide diagnostic feature for identifying various surface materials. Due to the highly complicated scenes and the spatial resolution limitation in remote sensing, mixed pixels are inevitable, which are a mixture of more than one ground cover materials. Spectral unmixing is an efficient method for interpreting mixed pixels, which is a quantitative analysis procedure to decompose a mixed pixel into a collection of constituent spectra (endmember), and a set of corresponding fractions (abundances). Spectral unmixing, which breaks the spatial limitation of hyperspectral sensor, reveals the distribution of ground materials at sub-pixel level, and improves the accuracy of interpretation. It is critical for automatic processing and remote sensing quantitative applications. So the thesis focuses on the following problems in spectral unmixing with hyperspectral images:the negligence of spatial information, difficulty for estimation of endmembers’ number, nonexistence of pure pixels. And the research deeply surveys the incorporation of spatial information into automatic endmember and spectral unmixing.The main research work and the corresponding contributions of this dissertation are as following:Spectral mixture model provides the foundation for spectral unmixing. The model can be grouped into linear spectral unmixing model and nonlinear spectral unmixing model. The classical linear spectral unmixing model and its extended models, including weighted least squares model and Normal Composition Model (NCM), are studied. Also, the classical nonlinear spectral unmixing model and and extend nonlinear spectral unmixing model are studied. And based on linear spectral unmixing model, automatic endmember extraction and spectral unmixing algorithm are summarized. And the application of blind unmixing into spectral unmixing is studied. The drawbacks of above studies are analyzed. The accuracy of automatic endmember extraction and spectral unmixing can be improved from the following three aspects:(Ⅰ) Integration of spectral and spatial information to endmember extraction and spectral unmixing,(Ⅱ) accurate estimation or adaptive adjustment of endmembers’ number,(Ⅲ) application of blind unmixing into spectral unmixing when there are no pure pixels in the images.Most of current automatic endmember extraction and spectral unmixing algorithms rely on the exploitation of spectral information alone. To integrate the spatial and spectral information, a hybrid automatic endmember extraction algorithm (HEEA) is proposed in this thesis. Without dimensionality reduction, HEEA uses the spectral-information-divergence and spectral-angle-distance metric to measure the similarity, and the orthogonal subspace projection (OSP) method to search for the endmembers, which can decrease the correlation between extracted endmember spectra. Moreover, it is based on a local window which integrates both spatial and spectral aspects to extract endmembers. The combinated spectral information divergence and spectral-angle-distance metric can enlarge the separability between endmembers. Experiments show that the hybrid automatic endmember extraction algorithm with spatial information performs better than the endmember extraction algorithm without spatial information procedure.Current automatic endmember extraction and spectral unmixing should be predetermined the number of endmembers, which is a challenge problem. To adaptive adjust the number of endmembers, a new automatic sparse pruning endmember extraction algorithm with minimum volume and deviation constraint (SPEEVD) is proposed in the thesis. The proposed algorithm can adaptively determine the number of endmembers through a sparse pruning method, and can weaken the noise interference by a minimum volume deviation constraint. A non-negative matrix factorization solution based on the projection gradient is applied to solve the constrained optimization problem, which makes sure the convergence is steady and saves computation cost. The experimental results indicated that the proposed method manifested an improvement in both the root-mean-square error (RMSE) and the endmember spectra, compared to the other state-of-art methods, all of which need an accurate pre-estimation of endmember number.For better spectral unmixing when the non-existed pure pixels in the image, a new blind unmixing method with minimum volume and deviation constraint with spatial preprocessing is proposed in the thesis. To avoid non-unique of blind unmixing in spectral unmixing, the minimum volume and deviation constraint is added with the non-negative matrix factorization frame. And the spatial and spectral information is intergrated in the spatial preprocessing, rather than spectral information alone. The experimental results indicated that the proposed algorithm with the minimum volume and deviation manifested better performance in spectral unmixing than traditional blind unmixing, which also alleviate the interference of abnormal and noise. And the experimental results reveal integrated spatial information and spectral information can improve accuracy of the spectral unmixing.