Research On Unmixing Of Hyperspectral Remote Sensing Images Based On Spatial Spectral Structure Information Mining

Hyperspectral images(HSIs)of remote sensing have rich spectral information,and each pixel can extract a complete and high-resolution spectral curve,which have attracted widespread attention.However,due to the limited spatial resolution of the imager and the complex distribution of features,mixed pixels appears in HSIs,that is,the image unit is mixed by multiple pure features.The mixed pixels will affect the accuracy of the subsequent processing of HSIs,such as classification,target detection and so on.Therefore,the decomposition of mixed pixels,i.e.,unmixing,is very important for the refined processing of HSIs.Studies have shown that the rich spatial-spectral information in HSIs is particularly important for the study of unmixing tasks.How to effectively use the spatial-spectral information in the image as the priori and add it to the construction of model to further explore the potential structure of the image data,so as to achieve efficient research on unmixing,is the focus and dificulty for the processing of HSIs.This paper uses the nonnegative matrix factorization model as the basic research tool,and proposes the research on unmixing for HSIs based on spatial spectral structure information mining,by combining with the machine learning methods including the manifold learning,sparse representation,low-rank representation and so on,which aims at solving the problems of traditional algorithms and the application of unmixing.The main research content is summarized as follows:(1)Aiming at the problem of nonnegative matrix factorization model easily falling into local optimality,a linear unmixing algorithm for HSIs based on endmember independence and spatial similarity is proposed.Traditional unmixing algorithms often only impose a single constraint on endmembers or abundances.The proposed algorithm considers the characteristics of endmembers and abundance,and on the one hand,considering the independence of endmembers,the algorithm uses the autocorrelation matrix to construct independence constraint.On the other hand,the algorithm utilizes the spatial neighborhood information to construct the weight matrix of abundance,which is introduced into the sparse constraint.At the same time,it also adopts the correlation coefficient to measure the similarity between pixels to construct the manifold regularization to further explore the potential structure of hyperspectral data.The proposed linear unmixing algorithm not only considers the characteristics of endmembers and their abundances,but also fully utilizes the rich spatial-spectral information in the image.(2)Aiming at the problem that there are no constraints related to image characteristics in the objective function of the generalized bilinear model,a nonlinear unmixing model for HSIs based on semantic structure low-rank and self-representation learning is proposed.Most unmixing algorithms based on manifold learning usually uses the heat kernel function to construct the connection weight.Self-representation learning is adopted to mine the similarity of pixels in the image,which is introduced into the manifold learning.In addition,considering the correlation between pixels in the image,the semantic structure information is obtained by superpixel segmentation algorithm,which will be constrained by low-rank representation.The proposed nonlinear unmixing algorithm fully explores the similarity of the pixels in the image,and mines its potential structure information.(3)Aiming at the problem that a single spectral mixing model is difficult to completely reflect the complex interaction between different features in the real scene,a hybrid model based on structural region for HSIs is proposed to adaptively unmix.Traditional unmixing algorithms are often based on single spectral mixing model,but due to the complexity of surface features,a single mixing model usually does not accord with the spectral mixing in the real scene.Therefore,in order to study and process more accurately,a hybrid model is proposed to adaptively unmix for different regions of the image.The algorithm uses clustering algorithm to mine the spatial spectrum information of the image,and constructs a hybrid model for spectral unmixing by combining with manifold learning.The proposed algorithm fully considers the differences of spectral mixing patterns between regions,and realizes more refined scene modeling by mining the spatial-spectral information of the image.(4)Aiming at the problem that the corresponding abundances in different regions have different sparse characteristics,a sparse unmixing algorithm for HSIs based on visual computing and regional difference is proposed.The traditional sparse unmixing algorithm usually adopts a single sparse regularization,which ignores the differences of different regions in the image.Studies have shown that pure pixels tend to be located in the region with more uniform distribution in the image,while the region between homogeneous regions is more prone for spectral aliasing due to its special position,and its mixing is more serious.Inspired by visual computing,the algorithm applies the sketch map technology in visual computing to HSIs for the first time to mine the regions with different characteristics,and uses different sparse regularities to constrain,so as to construct a sparse unmixing hybrid model.The proposed algorithm fully considers the sparse differences of regions,and realizes the sparse unmixing of HSIs by combining the visual computing.(5)Aiming at the inherent connection between the tasks of unmixing and anomaly detection,an anomaly detection algorithm for HSIs based on spectral unmixing and anomaly characteristic analysis is proposed.Traditional detection algorithms often detect and process anomalies from a single point of view,and the detection effect is not good.Driven by anomaly detection task,a new unmixing model is constructed by exploring the spatial-spectrum information of the image.Through analyzing the local and global characteristics of anomaly,the anomaly in the image is detected more comprehensively from the perspective of pixel level and sub-pixel level,combining with the spectral unmixing technology.The proposed algorithm fully analyzes the local and global characteristics of the anomaly,and deeply organically integrates the tasks of unmixing and anomaly detection,which further improves the detection performance and makes the study of unmixing more meaningful.