A Study Of Hyperspectral Image Sparse Unmixing Based On The Learning Of End-members

Remote sensing technology is started in 1960's and its rapid development makes a lot of contribution to other technologies. Hyperspectral imaging is one behalf of these technologies.Hyperspectral data, collected by imaging spectrometer, has two obviously characteristics.The first one is that the pixels of hyperspectral image contain lots of band and the other one is high spectral resolution which is much higher than before. Basing on these properties, then hyperspectral image is used in identification of surface features. Unfortunately, low spatial resolution of imaging spectrometer and the uneven distribution of surface features cause bad quality of hyperspectral image in practical application that pure pixels can hardly exist and most of them are composed by several surface features, which lead to a result that the precision of identification of surface features is low. In order to better using Hyperspectral image, pixels should be decomposed into end-members and the corresponding abundances which is just the work of hyperspectral unmixing.For the sake of improving the result of hyperspectral unmixing, this article carry on a deep research on unmixing and the main work can be described as follows:1. Basing on the learning of end-members, a method of hyperspectral sparse unmixing from the point of view of space similarity is proposed. Integrating the thought of learning into hyperspectral unmixing, first learn the end-members combine with it's characteristic of smoothness and select out good quality of end-members with the help of digital spectral library. While solving abundances, constraint are added from the perspective of space that similar pixels mean that they have similar space structure too which can described as low-rankness. In addition, the number of end-members distribute in the whole image is relatively small compared with the number of pixels, leading to a phenomenon that just parts of end-members can be contained in a pixel which cause sparseness indirectly. So combine with low-rankness and sparseness, hyperspectral image is unmixed. Experiments show that the new solving mode can acquire higher precision in hyperspectral unmixing compared with traditional method.2.Basing on the learning of end-members, a method of hyperspectral sparse unmixing from the perspective of regular weighting is proposed. While learning end-members, real situation that some bands of end-members are eliminated which cause the property of piecewise smoothness are considered. Basing on this reason, additional constraint is added to the objective function. While solving abundances, problem is considered from the view of space distance and then low-rank constraint is replaced by constraint of regular. By virtue of several experiments, results prove that the new algorithm proposed here can do better than that of last algorithm.3.Basing on the learning of end-members, a method of hyperspectral sparse unmixing from the view of neighbor similarity is proposed. As to the learning process, constraints are still about smoothness and piecewise smoothness. While solving abundances, neighboring areas are considered that data is represented by its neighbor data with the same structure. This is an angle from which new relationship among neighbor data can get. Experiments indicate that the algorithm proposed here can have good results.