Classification And Unmixing Of Hyperspectral Remote Sensing Images Based On MCMC

In recent years,hyperspectral remote sensing plays an important role in precision agriculture,geography survey,resource investigation and so on.However,due to the technical facts,such as the constraints between spatial resolution and spectral resolution,and the presence of intimate mixtures,mixed pixels are prevalent.Therefore,how to achieve accurate classification and unmixing of hyperspectral remote sensing images has an important influence on the realization of quantitative remote sensing.In order to solve the related problems,this thesis proposes a classification and unmixing method for hyperspectral remote sensing images,which incorporates endmember variability and spatial information and thereby can achieve high-precision classification or unmixing results.In addition,with the purpose of batch processing and rapid analysis,the Flask-based hyperspectral remote sensing image analysis platform is built to provide a hyperspectral application module to meet a variety of user demands.Multinomial logistic regression is initially used to establish a prior model that only considers spectral information.Multinomial logistic regression is a pixelwise classification method,and its probability is to meet the nonnegative and sum-to-one requirements of abundance.Therefore,this study takes the initial probability of multinomial logistic regression as a priori abundance and introduces prior knowledge into the model by the equivalence relation between Bayesian and multinomial logistic regression under normal distribution assumption.The real data based computational results show that the multinomial logistic regression reflects the ground truth of land-cover to a certain extent.In order to take the variability of endmembers into account,the normal composition model is used to characterize the mixed pixels in the hyperspectral image by assuming that the spectrum of the mixed pixels is a linear combination of Gaussian random vectors with abundance as coefficients.Combined with the conditional independence of the spectrum,the set of spectral vectors is an independent normal mixture distribution.Considering the normality and non-negativity of abundance,the Dirichlet distribution is chosen as the prior distribution of abundance,and the neighborhood abundance information and granularity are introduced into the model by re-parameterization.Based on the joint distribution of spectral vector,granularity and abundance,a hierarchical Bayesian model is established to obtain the full posteriori distribution of granularity and abundance.The posterior distribution is estimated by Gibbs sampling,which is a kind of MCMC,in order to estimate abundance.At the same time,recursive feature elimination combined with cross validation is used for band selection.The classification experiments of Pavia University,MUUFL Gulfport and unmixing experiments of Indian Pines and Salinas datasets proves the high performance of our methods.Finally,according to the above theory and algorithm research,the hyperspectral remote sensing image analysis platform is established which adopts the lightweight web framework Flask as a development framework,MySQL and Flask-SQLAlchemy to manage the database,HTML / CSS / JavaScript,jQuery,Bootstrap to form front-end pages,and integrates user login / registration,features selection,classification(multinomial logistic regression,support vector machine,MCMC),unmixing(MCMC)and other functions.Through the test of each functional module,the completeness of platform function is verified.