| Summary: | 碩士 === 國立中山大學 === 機械與機電工程學系研究所 === 107 === With the advancement of remote sensing technology, the applications of hyperspectral imagery (HSI) are more and more popular. Despite of many success achieved by HSI techniques, there are still some problems to be solved. For instance, HSI data provides huge information which usually contains a lot of redundancy because of its inherent nature. It imposes difficulties for classification because of the curse of dimensionality issue. The enormous data volume of HSI also causes the issues of data storage and long processing time. Therefore, how to select the representative bands from the original image cube without significant loss of information is one of the most important topics in remote sensing society. We call it Band Selection (BS).
In this thesis, we combine a new concept, Band Subset Selection (BSS), with self-sparse representation (SSR) model as the objective function, to create an effective BS method for data dimensionality reduction. This method is called self-sparse representation based BSS (SpaBSS). In order to efficiently implement SpaBSS, two iterative algorithms are developed: successive SpaBSS (SC-SpaBSS) and sequential SpaBSS (SQ-SpaBSS). Unlike many existing BS approaches which may only find the locally optimal solution by a single path, the proposed SpaBSS can obtain the nearly globally optimal solution by continuously updating the band subset based on minimizing the reconstruction error of SSR model. The experiments conducted on three real hyperspectral datasets demonstrate that both SpaBSS methods can find appropriate band subsets for effective hyperspectral image classification and endmember extraction.
|
|---|
