Research On Hyperspectral Spectral Signal Extraction Method Based On Nonlinear Mixed Model

As a foundation technology of hyperspectral image processing, hyperspectral spectrum signal extraction is widely applied in hyperspectral unmixing, target recognition, anomaly detection, and so on. Therefore, how to effectively obtain spectral information in hyperspectral image will have an impact on the applications of hyperspectral image.Currently, there are two methods to get spectral signal. One of them is that the spectral signal is obtained by experiment or selected from spectrum library. another is that the spectral signal is obtained from hyperspectral image, which is usually called Spectral Signal Extraction Algorithm(SSEA). However, because of many mixing pixels are widely present in hyperspectral image, spectral signal extraction algorithms is facing enormous challenge. At present, linear mixed model is usually used in SSEA,such as PPI(pure pixel index), VCA(Vertex Component Analysis), and so on. But,during the formation of the mixed pixel, it is easily affected by atmospheric transport,the instrument and the interaction between solar radiation and materials. Therefore,the mixed pixels should be described by non-linear mixed model. Meanwhile, with the improvement of spatial resolution, the nonlinear degree of mixed pixel becomes more and more serious. Therefore, in this paper, we study the spectral signal extraction algorithm based on nonlinear mixed model.Firstly, the cause of mixed pixels and mixed model is analyzed and discussed in this paper. Then, based on hyperspectral unmixing and similarity of vector, a framework of verifying the effectiveness of SSEA is proposed. we deeply study linear spectral signal extraction algorithm performance in different nonlinear mixed models for hyperspectral data, because the approximate linearization method is widely used to deal with the nonlinear components of mixed pixel. The experimental results show that this method lead to the increase of computational error, even unable to extract the spectral signals from the nonlinear mixed data. Finally, nonlinear SSEA and linear SSEA will be compared in the experiment. The results show that whether the simulated hyperspectral data or the actually measured hyperspectral data, the spectral signal extraction algorithm based on nonlinear mixed model can effectively extract the spectral signal. Meanwhile, the overall performance of nonlinear spectral signal extraction algorithm is better. Combined with the traditional target recognitionmethods, the extracted signal spectrum can effectively identify target in hyperspectral image.