Research On Hyperspectral Database And Hyperspectral Data Mining

Hyperspectral remote sensing is a cutting-edge technology of acquiring land surface information, with a high resolution in spectrum. With the fast development of spaceborne and airborne sensors, hyperspectral data are more frequently and conveniently received. As a result, data storage, management and effective information extraction are becoming key challenges to hyperspectral remote sensing science and technology. Especially, automatic or semi-automatic information extraction is a predictabley trend for hyperspectral researches and applications.In this dissertation, the intension and extention of the concept hyperspectral database and spectral data mining has been brought forward after a background research on hyperspectral remote sensing and database. Based on this conception, ground-object spectra database, remote sensing image database, data mining technology have been studied, while spatial data mining and image data mining technology are focused. A conceptual design with data mining and application oriented is provided basetd upon project experience and international study. The hyperspectral database has been built up and more than 6,000 sets of spectral data have been uploaded. There are spectra of rocks, minerals, waters, concretes, trees, wheats, soils and hyperspectral images in this database. With this hyperspectral database, some researches on spectral data mining and hyperspectral image data mining have been carried out. In view of hyperspectral remote sensing technology and application, there are some advangates of this dissertation as follows:1. Based on hyperspectral database, new storage models of spectra and hyperspectral image has been put forward. A new design has combined data, methods and models in a whole database platform.2. To integrage different sources of spectra in one database platform, a new conceptual data structure, dualistic-core star structure, has been brought forward.3. Forward (from spectra to attributes) and backward researches have been impleted based on hyperspectral database. Automatic optimizations have been applied in band combination and band selection.4. From a totally different view of hyperspectral image, a projection from hyperspectral images to relation tables has been built up to improve analysis and information extraction. The following key techniques are implemented in hyeprspectral database: Band selection based on Minium Length Model, Feature extraction based on Non-negative Matrix Fraction, Data compression and decompression based on Non-negative Matrix Fraction, Object Decection based on Support Vector Machine.