The Study On Hyperspectral Image Endmember Unmixing

Hyperspectral remote sensing is a kind of technology which is for feature remote sensing imaging by using very narrow and contiguous spectral channels. The images of the Earth’s surface acquired by this technology contains triple information in space radiation and spectral. Hyperspectral remote sensing for object detection and recognition is an important reason that why high-spectral image can be widely used in military and civilian fields. However, the limited spatial resolution sensor cause mixed pixels(a pixel within a variety of surface features, such as vegetation, rivers, roads, etc.) appear in remote sensing image. Mixed pixel problem affects not only the feature identification and classification accuracy, but also the development of remote sensing technology to quantitative. Therefore, how to effectively interpret the mixed pixel is a prominent problem of hyperspectral remote sensing applications.The papers introduced the the basic theory of the hyperspectral image data representation, the mixed model of the pixel, and classical algorithm of endmember extraction from hyperspectral images mixed pixel imaging principle. Then studyed traditional spectral unmixing and unsupervised spectral unmixing two method under the premise of a linear mixed model.AMEE and non-negative matrix factorization were selected as a typical algorithm, and proposed an improved algorithm based on the advantages and disadvantages of both the two algorithms. The computer simulation experiments found that the method does have fairly good results.