Research On Multiple Endmember Hyperspectral Mixed Pixel Decomposition Algorithm

Hyperspectral image is visually represented as a cube that not only expresses the twodimensional spatial information of the hyperspectral image,but also has a spectral dimension that characterizes the physical properties of the pixel.Hyperspectral image has a very high spectral resolution and can solve many problems that cannot be solved under multispectral conditions.However,the spatial resolution of hyperspectral images is relatively limited,and a single pixel may contain many different features,which is a mixed pixel problem in hyperspectral images.The hybrid pixel problem is an important obstacle limiting the quantitative development of remote sensing technology.Therefore,it is very important to study and solve the mixed pixel problem in hyperspectral image.Spectral unmixing is an important way to solve mixed pixels in hyperspectral image and the purpose is to determine the types of features contained in the pixels and their corresponding proportions in the pixels.It is also a more accurate classification technique.The commonly used pixel mixing models are linear models and nonlinear models.Because the physical meaning of the linear hybrid model is relatively clear and the model is easy to establish,it is welcomed by researchers.In contrast,the establishment and solution of nonlinear models is difficult,so the research on spectral unmixing techniques of nonlinear models is relatively rare.Most of the traditional spectral unmixing techniques are based on the single endmember mode,and the spectral variation phenomenon in the hyperspectral image is not considered,resulting in unsatisfactory unmixing result.In order to deal with the influence of spectral variation on unmixing result,we use more than one spectral curve to represent a certain feature whether in simulative data experiment or in real data experiment.Hyperspectral images are affected by noise during the acquisition process and the presence of noise will affect the unmixing result of hyperspectral images.Especially when the pixels contain little abundance features,the noise may directly covered the spectral signals of these features,and result in a unmixing error.In order to deal with the influence of noise on little abundance features,a grouping optimization idea is used.The idea first uses all endmembers of a feature to perform unmixing processing and gain the corresponding coefficients of all endmembers in pixels.The purpose of this is to fully magnify the role of all endmembers of each feature in the pixels and reduce the phenomenon that little abundance features are covered by noise.The multiple endmember spectral mixture analysis has a high degree of unmixing accuracy,but it needs to exhaust different combinations of all endmembers of different feature for each pixel,so the amount of computation required is also considerable.The hierarchical multiple endmember hyperspectral unmixing algorithm has the characteristics of dividing the complex problem into several simple steps,the requiring relatively small amount calculation be relatively small and ensuring the high unmixing accuracy.However,hierarchical multiple endmember hyperspectral unmixing algorithm does not use all the endmembers of a feature during unmixing,so the result of unmixing can’t reflect the variation of a feature in all pixels,and the accuracy of unmixing is not very high.In order to cope with this situation,this paper integrates the ability of the extended linear mixing model to reflect variation of features in all pixels(ELMM)into the hierarchical multiple endmember hyperspectral unmixing algorithm and verifies it on the simulative data and the real data respectively.The results show that even if the number of variant endmembers contained in the hyperspectral image is large,this method can obtain higher unmixing accuracy.