Multiple Endmember Spectral Mixture Analysis Algorithm Research

In remote sensing data exploitation, spectral mixture analysis technique is generally used to detect the land cover materials and their corresponding proportions present in the observed scene. Traditionally, a fixed endmember spectral signature set for each land cover material is used to estimate the abundance fraction of each potential mixed pixel by many spectral unmixing algorithms. However, among those approaches, illposed inverse problem for the abundance fraction of each land cover material because of model inaccuracies, observation noise, environment conditions, endmember variability,and data set size. In the literature, some scholars have proposed to perform the unmixing by taking the spectral variability into consideration. Among these spectral variability based unmixing approaches, multiple endmember spectral mixture analysis(MESMA)is probably the most widely used method. However, when the number of land cover materials is large, the computational load of the MESMA method would be very heavy and would not be applied for many complex scenes. For the problem, the major work of this thesis are as follows:(1) We propose a new sparse multiple endmember spectral mixture model(SMESMM)to improve spectral unmixing task. This model treats the spectral mixture procedure as a linear block sparse inverse problem. Although, land cover materials are generally complex in the whole scene. In a mixed pixel is usually less than the whole number of materials present in a given scene, which implies that the estimated solution of endmember model for mixed pixel is sparse.(2) We also propose a new MESMA approach based block sparse Bayesian learning(BSBL) algorithm frame. The approach is first solved using the block sparse Bayesian learning(BSBL) algorithm to obtain an initial block sparse solution. Then, MESMA is used to resolve the mixed pixel by using the selected land cover materials, which correspond to the non-zero blocks in the solution obtained by the BSBL algorithm. The block sparse solution obtained in the first step can help to determine how many and which land cover materials are involved in the concerned mixed pixel. This can largely decrease the number of possible candidate models for the MESMA method when the number of land cover materials is large.(3) Experimental results on simulated and real hyperspectral data demonstrate the validation of the proposed method, by comparing with other methods based a fix spectral endmember data set. Especially, in the cost of computing resource, the approach proposed is much less than the MESMA method.