Hyperspectral Unmixing Based On Generalized Bilinear Model

Due to the space resolution of remote sensing data sets, spectral signatures collected by remote sensing imager in the natural environment must be a mixture of various substances. Therefore, spectral unmixing must be used for accurate estimation. Unmixing method in accordance with the employed decomposition model can be divided into method based on linear spectral mixture model and method based on nonlinear spectral model. This paper considers the sparse information of abundance, spatial information and modeling diversity, and improve the existing solutions of nonlinear mixing method, with details as follows:Firstly, correlation of hyperspectral data will lead to the sparsity of data, and each pixel does not contain all endmembers. Most existing nonlinear unmixing algorithm does not consider the sparse information of data. Due to the fact that nonlinear mixed algorithm does not consider the data sparse information, this paper proposes sparsity-constrained generalized bilinear model for hyperspectral unmixing. Generalized bilinear model?GBM? has been widely used for nonlinear hyperspectral image unmixing. However, Most existing nonlinear unmixing algorithms do not take the sparse information of the data into account, which is a significant characteristic resulting from the correlation of hyperspectral data. And most of the pixels may not contain all of the material in the scene. Recently regularization methods are usually applied to enforce the sparsity constraint on endmember abundance.This paper aims to extend the GBM by incorporating the sparsity constraint of abundance matrix with the semi-nonnegative matrix factorization, by dividing GBM into the linear part and the second-order part, which are optimized using an alternating optimization algorithm respectively. L_1/2-norm is used to explore the sparse characteristic, and the L_1/2-constrained semi-nonnegative matrix factorization(L_1/2-semi-NMF) algorithm is presented, which leads to better results on both synthetic and real data.Secondly, multiple scattering usually occurs between vegetation and soil in bilinear scene, and hyperspectral images containing substances such as vegetation and soil may occur bilinear mixing in the border areas. Taking into account the differences of the image area, a region adaptive segmentation based hyperspectral image unmixing process is proposed. The method uses K-means clustering method for hyperspectral data clustering. Then the image is divided into homogeneous regions and detail regions. Homogeneous region is assumed to be based on linear model, which can be solved by sparse constrained non-negative matrix factorization method. And the detail region is assumed to be based on generalized bilinear model, which can be solved by sparse constrained semi-nonnegative matrix factorization. The method can get more precise abundance and is good at keeping the edge information of bilinear abundance. Comparative experiments show that the proposed method is effective to improve the accuracy of unmixing hyperspectral remote sensing image.Thirdly, For the reason that rnon-linear unmixing algorithm does not consider the manifold structure of nonlinear unmixing,we proposed graph regularized bilinear model for unmixing. The existing semi-NMF algorithm for hyperspectral nonlinear unmixing considering only the Euclidean space structure. In fact, hyperspectral data is more likely to locate in a low-dimensional manifold in high-dimensional space. The addition of graph regularization can keep the original image and the abundance in close contact, the method is able to improve the performance of unmixing.