Research On Spectral Unmixing Algorithms For Hyperspectral Remote Sensing Image

Hyperspectral remote sensing provides a powerful method for accurate identification and fine classification of materials by integrating the spectra representing to the radiant properties of materials with the images reflecting the spatial and geometric relations. With the expansion of application fields and escalation of application requirements, the development of hyperspectral remote sensing shows the trend of quantification analysis. However, there are lots of mixed pixels in remote sensing image, which not only affects the recognition and classification precision of materials, but also prevents the progress of quantification analysis for hyperspectral remote sensing technology. Spectral unmixing, as one of key technologies for solving the problem of mixed pixel, has become a research hotspot in hyperspectral remote sensing application fields. In this study, several spectral unmixing algorithms involving endmember extraction and abundance estimation are studied based on linear spectral mixture model. The main research works are described as follows:1. Endmember extraction plays an important role in spectral unmixing. Traditional endmember extraction algorithms (EEAs) only use the spectral information to extract endmember, ignoring the spatial characteristic of the remote sensing image, which make the algorithms are susceptible to the noise and anomaly pixel, resulting in reducing accuracy of endmember estimation. Focusing on this problem, a new EEA combining subspace projection and local spatial information is proposed. Based on the theory of convex simplex, the algorithm sequentially extracts the endmembers by combining the subspace projection and simplex volume analysis. During the extracting process, the local spectral similarity constraint is used to improve the robustness to the noise and anomaly pixel, which also avoids to the huge computational cost caused by global spatial information Furthermore, the simplex volume calculation is free of dimensionality reduction which may cause the possible loss of original information. The experimental results on synthetic and real hyperspectral image show that the proposed algorithm can improve the accuracy of the endmember extraction, and is more robust to the noise and anomaly pixels compared to the spectral-based EEAs.2. Fully constrained linear spectral unmixing is often formulated as a convex optimization problem that requires more advanced optimization technology, leading to excessive computational load. For the hyperspectral image, it usually covers a large geospatial area and thus the number of endmember is quite large, which increase the computational load further. To solve this problem, a subspace-projection-based geometric unmixing (SPGU) algorithm is proposed. SPGU expresses the abundance as the barycentric coordinates of pixel with respect to endmember simplex. Then, Laplace expansion is applied to the calculation of barycentric coordinates, which reduce the computational load of the SPGU. For the pixel violating the nonnegative constraint, it is iteratively projected onto the subspace projection spanned by translated endmember vertices until the one obtain the fully constrained abundance estimation. SPGU is in line with the least squares criterion. Experimental results demonstrate that the SPGU has almost the same results with the well-known fully constrained least squares spectral unmixing algorithm while has better computational performance when the hyperspectral scene has a larger number of endmembers.3. Endmember identification algorithm (EIA) has a defect of lower accuracy when it is applied to the hyperspectral image which does not completely satisfy the pure-pixel-assumption. To overcome this defect, a new two-stage spectral unmixing algorithm combining pure pixel identification and constrained nonnegative matrix factorization(NMF) is proposed. In the first stage, the EIA is used to obtain the initial pixel candidates, and then principal component analysis is performed on the homogenous pixel candidates to determine the pure endmembers. In the second stage, NMF is used to generate virtual endmembers. To overcome the local minimization, the sum of squared distance and abundance sum-to-one constraints are introduced into the NMF. Moreover, pure endmembers as prior knowledge is introduced into NMF to improve the convergence speed and decomposition accuracy. The experimental results show that the proposed algorithm has the potential of yielding accurate estimates of both endmember spectra and abundance fractions.