Wang, Z; (2017) Essays on hyperspectral image analysis: classification and target detection. Doctoral thesis , UCL (University College London).

Preview

Text Thesis_Ziyu.pdf Download (6MB) | Preview

Abstract

Over the past a few decades, hyperspectral imaging has drawn significant attention and become an important scientific tool for various fields of real-world applications. Among the research topics of hyperspectral image (HSI) analysis, two major topics -- HSI classification and HSI target detection have been intensively studied. Statistical learning has played a pivotal role in promoting the development of algorithms and methodologies for the two topics. Among the existing methods for HSI classification, sparse representation classification (SRC) has been widely investigated, which is based on the assumption that a signal can be represented by a linear combination of a small number of redundant bases (so called dictionary atoms). By virtue of the signal coherence in HSIs, a joint sparse model (JSM) has been successfully developed for HSI classification and has achieved promising performance. However, the JSM-based dictionary learning for HSIs is barely discussed. In addition, the non-negativity properties of coefficients in the JSM are also little touched. HSI target detection can be regarded as a special case of classification, i.e. a binary classification, but faces more challenges. Traditional statistical methods regard a test HSI pixel as a linear combination of several endmembers with corresponding fractions, i.e. based on the linear mixing model (LMM). However, due to the complicated environments in real-world problems, complex mixing effects may exist in HSIs and make the detection of targets more difficult. As a consequence, the performance of traditional LMM is limited. In this thesis, we focus on the topics of HSI classification and HSI target detection and propose five new methods to tackle the aforementioned issues in the two tasks. For the HSI classification, two new methods are proposed based on the JSM. The first proposed method focuses on the dictionary learning, which incorporates the JSM in the discriminative K-SVD learning algorithm, in order to learn a quality dictionary with rich information for improving the classification performance. The second proposed method focuses on developing the convex cone-based JSM, i.e. by incorporating the non-negativity constraints in the coefficients in the JSM. For the HSI target detection, three approaches are proposed based on the linear mixing model (LMM). The first approach takes account of interaction effects to tackle the mixing problems in HSI target detection. The second approach called matched shrunken subspace detector (MSSD) and the third approach, called matched cone shrunken detector (MSCD), both offer on Bayesian derivatives of regularisation constrained LMM. Specifically, the proposed MSSD is a regularised subspace-representation of LMM, while the proposed MSCD is a regularised cone-representation of LMM.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item