Research Of Unmixing Algorithm For Hyperspectral Remote Sensing Image

With the development of spectral imaging technology,hyperspectral remote sensing images have been widely used in ground object detection tasks.However,due to the limitation of imaging principles,hyperspectral images generally have low spatial resolution,resulting in a large number of mixed pixels in the image.The goal of the hyperspectral unmixing is to decompose the mixed pixels in the remote sensing image into pure spectra(endmembers)and the proportion of fractions(abundances)in each pixel.It transfers hyperspectral images from pixel-level research to sub-pixel level,and provides more detailed spatial and spectral information for subsequent accurate classification,identification,observation,and analysis of ground targets.It is used in geological exploration and precision agriculture,and national defense security.Most of the existing unmixing methods use the idea of combining spatial-spectral information.However,with the emergence of more and more complex ground scenes,the internal structure of the obtained hyperspectral data is enhanced.Therefore,traditional data structure description methods have begun to be inapplicable,which brings new challenges to the unmixing task of hyperspectral remote sensing.The main problems are:1)the expression of effective spatial similar features;2)the problem of noise interference to the image;3)the effective modeling of complex data.Hence,this paper has carried out a series of research on the hyperspectral unmixing around the above key issues.The main research contents and innovations are as follows:(1)To address the local smoothing and non-local similarity of image space,this paper proposes a hyperspectral unmixing algorithm based on local smoothing and nonlocal spatial similarity.Traditional spatial information exploration methods usually use Euclidean distance to calculate the similarity between two pixels or regions to form a similarity matrix,thereby constraining the abundance matrix,which ignores the structural relationship existing in the data.In this paper,local regions and non-local similarity sets are established.The self-expression idea of data is used in each individual subspace to reconstruct pixels,which reduces the interference of noise on unmixing,and fully explores the spatial structure of the data.The proposed algorithm can improve the performance of understanding.(2)To address the shortcomings of spatial-spectral joint information expression,this paper proposes a hyperspectral unmixing algorithm based on subspace clustering constraint.Traditional spatial similarity features are usually measured by Euclidean distance.It is not only difficult to explore the subspace structure relationship of the data,but also very sensitive to noise.This paper uses the idea of subspace clustering and the idea of self-expression in the original data space to reconstruct the image inside each subspace,which can describe the multi-level interactive information contained in the internal space of the data.In addition,the use of self-expression to reconstruct the pure endmember matrix avoids the generation of artificial endmembers and ensures that the extracted endmember spectra have physical interpretability and practical significance.(3)To address the problem of strong noise interference in images,this paper proposes a hyperspectral unmixing method based on spatial-spectral information.The traditional similarity measurement method is greatly affected by noise,especially when there are unknown areas or serious pollution in the image,it will cause a large deviation in the similarity measurement.To select the most similar pixel set in the image,this paper designs a new similarity metric,which fully considers the spatial relationship and spectral correlation of the pixels,and can also analyze the data under high noise pollution conditions.The similarity is accurately measured,so as to carry out effective unmixing modeling,and has good adaptability to complex data types.(4)To address the three-dimensional characteristics for hyperspectral data,this paper proposes a hyperspectral unmixing method based on non-negative tensor decomposition.The traditional unmixing model based on non-negative matrix factorization destroys the original spatial structure of hyperspectral data,resulting in a certain amount of information loss in unmixing.This paper regards the three-dimensional hyperspectral data as a tensor,directly decomposes the tensor and learns a tensor low-rank expression to learn the low-rank attributes of the complete data.In order to maintain spatial smoothness in the unmixing process,this paper also uses a total variational deviation regularization to constrain the abundance tensor,and further explore the spatial relationship among the data.This paper uses the complete structure information of the data to assist the rich spectral information,and verifies the effectiveness of unmixing on a lot of synthetic data and real data.