Mixed Pixel Decomposition And Its Applications Research

Hyperspectral remote sensing refers to use very narrow and continuous spectrum of remote sensing image features continuous channel of technology. The advent of hyperspectral imaging technology is a major leap of remote sensing technology application, it obtains the surface of the earth image contains rich space, radiation and spectral triple information, and widely used in geological exploration, marine remote sensing, vegetation, marine precision agriculture, atmosphere and environment monitoring, military reconnaissance and terrain mapping, etc.Hyperspectral remote sensing make remote sensing processing technology has been gradually deepening from qualitative analysis to quantitative analysis. However, due to limited spatial resolution and features of the complex diversity, mixed pixels exist in remote sensing image generally(that is a pixel in a variety of surface features), on the ground of the regional distribution of residential complex, especially. Not only the problem of mixed pixels effects identification and classification precision of ground, but also it is major barrier of the development of remote sensing technology to quantitative. Therefore, how to effectively interpret mixed pixels is an important problem of hyperspectral remote sensing applications.This thesis developed with the mixed pixels decomposition and its application, carrying on the following studies:①The research on AMEE endmember extraction.Because of the prior endmember extraction technolog only use spectral information, according to the linear spectral mixed model of mixed pixels decomposition, this article introduces the mathematical morphology theory to solve the problem of mixed pixels, and combines the extraction method organically, obtains an algorithm which can take account of the spatial correlation of spectral information and pixels to extract.②Hyperspectral image classification based on Endmember extraction.The theory of traditional classification algorithm basis on different categories of spectral exist differences, and ignore spectral information of mixed pixels doesn't on behalf of any single terrain types. Hyperspectral image classification is foundational technology, but only use traditional classification algorithms classify hyperspectral images can cause the classification accuracy reduce.Therefore, the article on the mixed pixel is decomposed, and then on hyperspectral image classification. This thesis extracts endmember by AMEE and PPI, then carrying on linear unmixing and using the fraction image to classify, obtains more satisfactory result.③The research on comparing with hyperspectral image classification.The thesis conducts the following group comparisons:1) Compare classification results with extraction endmember methods of AMEE and PPI, the result shows that after AMEE endmember unmixing solution fraction image by maximum likelihood classification accuracy higher than the PPI solution.2) The research on comparing classification results with support vector machine classification and maximum likelihood classification, the result shows that support vector machine classification accuracy higher than the traditional maximum likelihood classification.3) Compare the decomposition of mixed pixels to impact on classification. These four classification results mentioned above are compared, it also concludes that the classified precision of fraction image classification after linear unmixing is highest.