Research On Hyperspectral Unmixing Techniques

Hyperspectral remote sensing images can provide more detailed information because of its high spectral resolution,so it can be better used for the classification and target recognition.At the same time,the remote sensing image contains a large number of mixed pixels because of the complex diversity of ground objects and the low spatial resolution of the sensor.Spectral unmixing is proposed to interpret hyperspectral remote sensing image quantitatively.This thesis investigates spectral unmixing algorithm of hyperspectral image,and is organized as follows:(1)This thesis introduces the background of the unmixing research,and summarizes the research status of the spectral imaging technology and spectral unmixing.The spectral unmixing algorithm which is based on the linear spectral mixture model is discussed from the two aspects of the endmember extraction and the abundance estimation.(2)Two improved Simplex Growing Algorithms(SGA)are proposed,and the proposed methods are applied to spectral unmixing.Although the Simplex Growing Algorithm is widely used,it has some flaws,and it is necessary to reduce the dimension of the data when the Simplex Growing Algorithm is used to extract endmembers.The Simplex Growing Algorithm can be improved by using a new formula to calculate the volume and the method of Orthogonal Subspace Projection(OSP)to acquire the first endmember.Another new Simplex Growing Algorithm is proposed to improve the unmixing performance by the N-FINDR method.The improved algorithms are tested by simulated data and real hyperspectral data,and experimental results indicate that the proposed methods can obtain more accurate results of endmember extraction and abundance than the Simplex Growing Algorithm.(3)An improved Automated Morphological Endmember Extraction(AMEE)algorithm is proposed.In order to mitigate the limitations of the Automated Morphological Endmember Extraction algorithm,an improved morphological operator is proposed after introducing the concept of a reference spectral vector,and a new method to calculate MEI is given.To avoid information loss,four candidate endmembers are chosen from each improved even-numbered structure element.The modified automated morphological endmember extraction algorithm is tested based on a hyperspectral data set.The experimental results show that the improved method can obtain correct candidate endmembers from scenario where more than two pure pixels are involved,and information loss in the procedure of dilation is also avoided.Finally,the proposed method produces more accurate results of endmember extraction and the spectral unmixing.