چكيده به لاتين
Hyperspectral images are measured in hundreds or thousands spectral channels. These valuable data are used for recognition of the materials in the field of view. Spectral unmixing refers to the decomposition of a mixed pixel into a group of pure spectral signatures and their corresponding proportions to find the involved materials. The main aim of the thesis is to propose some algorithms that utilize the new statistical models to improve the performance of unmixing.
Due to the contribution of only a small number of endmembers of an extremely large library in each hyperspectral pixel, the abundance vector could be assumed sparse. Here, accordingly we consider the same case for estimating abundance vectors in a Bayesian sense. As our first novelty, we propose the Sparse Dirichlet Prior in a semisupervised manner for unmixing of hyperspectral images in Bayesian sense. The Markov Chain Monte Carlo (MCMC) sampler is used to generate samples based on the derived posterior. In the second proposed algorithm along with the sparse Dirichlet prior, we also make use of the Markov Random Fields to benefit from spatial correlation in the Bayesian based unmixing algorithm. Here, the classification and unmixing are performed simultaneously.
To model the more realistic conditions, in the remaining parts of the research different nonlinear mixing models have been employed. We apply the sparse Dirichlet prior to a polynomial postnonlinear mixing model in our third proposed algorithm. The results show that the nonlinear unmixing technique was more effective than linear ones. Next, in the fourth proposed algorithm, the MRF is similarly used to improve the unmixing procedure. This, as a result, leads to a major improvement in unmixing accuracy. Then this algorithm has been simplified for supervised applications. In the latter algorithm, a new model based on the linear quadratic mixing model is proposed. Using MRF to model the spatial correlation in a supervised scenario, results enhancement in unmixing accuracy.
