Linear Spectral Unmixing And Target Detection Of Hyperspectral Data Basedon Hyperion Image-a Case Study Of Land Cover Recognition

The hyperspectral remote sensingcontributes to moreprecisely characterizing the land cover type because of the detailed spectral information of its imaging system.But,due to thelow spatial resolution of the sensor,the mixed pixels existin urban and suburban images.In this dissertation,the linear spectral unmixing of hyperspectral mixed spectra is simulated and the abundance errorof different endmember group is analyzed;the land cover of urban and suburban area of Beijing is studied by spectral mixture analysis with Hyperion imagebased on V-I-S model;the experimental verification of mapping specific target via detection algorithm are conducted and the improved methods are proposed.This dissertation conducts a comprehensive analysis on the application of hyperspectral image with medium spatial resolution to urban and suburban remote sensing.The main contents and conclusions of this dissertation are thefollowing:(1)The study simulates mixed spectra with vegetation,impervious and soil spectra from ENVI spectral library,unmixes the simulated mixed spectra with different endmember combination,and thenpresents the plural endmember spectral unmixing which is applicable to hyperspectral data.The result shows that unmixing with plural endmember combination through a fully constrained linear spectral mixture model is conducive to error reduction and improves the correlation coefficient under the circumstance that spectra curveof the same surface differ in amplitude.(2)The abundance estimated through the fully constrained linear spectral unmixing with the commonly used endmember combination and the plural endmember combination are compared with actual abundance from Google Earth images by visual interpretation.The results indicate that the abundance estimated by the plural endmember combination have lower error.The validity of applying the plural endmember combination to large hyperspectral image unmixing through the fully constrained linear mixture model is verified with Hyperion image.(3)The feature of bands that are highly correlated to the vegetation abundance is analyzed by using the vegetation indexes.The NDVI computed by the reflectance near the red-edge wavelength is more significantly correlated to the vegetation abundance.Besides,the quadratic polynomial curve fitting is superior to linear fitting in characterizing the relationship between vegetation indexes and abundance.(4)Through unmixingthe Hyperion image acquired on August 2004 and June 2010 by linear spectral mixture model,the land cover changes of the northern and the southern part of Beijing are analyzed with abundance image.The land cover changes showed that vegetation and soil cover changed to impervious surface to a certain extent in both the northern and southern part of Beijing from 2004 to 2010.This change was more obvious within the area near the urban center and the impervious surface expanded outward on both sides of the artery in the remote area.(5)Small targets are identified through detection algorithm with hyperspectral image.Through experiment with Hyperion data,many detection algorithms are optimized and a multi-target detection algorithm is modified to be more sensitive to desired target and insensitive to undesired target in Hyperion image.