Endmember Extraction Algorithm Research For Hyperspectral Remote Sensing Image Based On Iterative Error Analysis And Spatial Information

The mixed pixel has been a major obstacle in image processing, classification,recognition and understanding. It is also the main reason for the difficulty of the traditional pixel level remote sensing classification and area measurement accuracy.Mixed pixels exist in hyperspectral images which greatly increases the difficulty of accurate feature analysis and image classification of hyperspectral images. The process to solve the problem of mixed pixel is called mixed pixel decomposition, and it can be divided into two steps: endmember extraction and abundance inversion.Accurate and efficient extraction of endmember is the basis for improving the accuracy of abundance inverse. Therefore,endmember extraction is a key step in the process of mixed pixels decomposition. In this paper, we focus on the endmember extraction algorithm based on the decomposition of mixed pixels.Endmember is also called pure pixel, which refers to the pixels that contain only one kind of ground objects. At present, the endmember extraction algorithms can be roughly divided into two categories. The first category is the endmember identification algorithm which considers that there are all kinds of pure pixels in the image and finds the endmember from the image. The other is the endmember generation algorithm that does not think there are pure pixels in hyperspectral images,so it does not find endmembers from the image, but directly generate endmember spectral features. Usually, we can find pure pixel that respect various objects in nature,the iterative error analysis algorithm is an endmember identification algorithm,compared with the other endmember identification algorithm, it regard the error of unmixing error as the basis of determining endmembers, but it ignores the correlation and redundant information between endmembers in endmember extraction process.Moreover, it only considers the spectral information of the image and ignores the spatial information of the image.In view of the above problems, the work of this paper mainly includes:(1) A method of endmember Spectral optimization method based on image space information. In this method, the spectral angle and the amplitude difference between each pixels and potential difference are considered, and in order to unify the difference in numerical calculation,using the Gauss function to calculate the spectral weight of each pixel. And combined with the weight calculation method based on the distance of the pixel coordinates,we give a comprehensive weight to optimize the endmember spectra.(2) An optimized endmember extraction method based on iterative error analysis is proposed. In this method, iterative error analysis algorithm is used to extract candidate endmembers and Schimidt Orthogonalization is used to find vectors that orthogonal to existing endmembers. The projection length of the candidate endmembers in the orthogonal vectors is used to determine endmembers and the endmember spectral optimization is also carried out by means of endmember spectral optimization in (1). It can not only get a small inversion error but also reduce the correlation and redundancy between the endmembers.(3) The endmember extraction is carried out by using the endmember extraction method proposed in this paper,and then the endmember spectrum is optimized by using the endmember spectral optimization method proposed in this paper. Finally,the constrained least squares method is used to abundance inversion operation and obtains the unmixing error. And the extracted endmember spectrum is compared with the endmember spectrum in the USGS spectral library.Finally, experiments are carried out on the simulated data and real hyperspectral data. The experimental results show that compared with the traditional iterative error analysis algorithm (IEA), a hybrid endmember extraction algorithm (HEEA) and volume-invariant constrained geometric optimization model (VIC-GOM) algorithm and, the proposed endmember extraction algorithm in this paper has certain advantages in spectral angle distance and abundance inversion error.