| Hyperspectral remote sensing images has been an important research direction in the field of remote sensing images in recent years as they can provide rich feature information with hundreds of bands. The pixels in the image are usually mixed by different sets of endmembers because of the low spatial resolution, making it difficult for subsequent applications such as object classification and target detection. So the exact unmixing solution of mixed pixels is of great significance to provide detailed feature information. Improved algorithms are proposed based on the commonly used models from the aspect of improving the accuracy or efficiency of abundance estimation algorithms. In this paper, the main research work is as follows:1) The theoretical system of mixed pixel unmixing are introduced. According to the different assumptions of photon transport process, spectral mixing model can be divided into linear mixing model and nonlinear mixing model. Detailed explanation of these two models are made, following by a discussion of the characteristics and applicable situations. A brief overview of the algorithms based on different model is made.2) Based on the linear mixing model, this paper proposes an fast fully constrained linear unmixing algorithm according to the simplex volume abundance estimation and fully constraints. By introducing the determinant lemma, determinant calculation of simplex volume is avoided and only simple matrix operation is needed. The abundance estimation results meet the non-negativity constrain by combing with a newly proposed iterative projection method. This efficiency of the algorithm is demonstrated both on synthetic and real AVIRIS images.3) Particle swarm optimization is introduced to solve nonlinear mixing problem based on the generalized bilinear unmixing model. As time efficiency guaranteed, Algorithm shows the superiority in accuracy to other traditional algorithms and its robustness is verified by a series of simulation data experiments. |
PDF Full Text Request