Research On Dimensionality Reduction Algorithm For Hyperspectral Data

Hyperspectral data dimensionality reduction is the first step of the analysis andapplication of land resources by using remote sensing data. At the same time, it is animportant means for people to obtain remote sensing information. In order to meet therequirements of accuracy, speed and generalization, aiming at the characteristic ofhigh-dimensionality, nonlinear, small sample size, large amount data, and expensivelabelling, the problem of hyperspectral data dimensionality reduction is researched byusing machine learning, pattern recognition and remote sensing scienceinterdisciplinary theories and methods. Based on non-negative sparse representation,tensor space and transfer learning, the following aspects are studied in thisdissertation.1. Dimensionality reduction of hyperspectral data based on (block) non-negativesparse representation. At first, in order to keep the sparse structure between samplesand the inner manifold structure in each sample unchanged, a non-negative sparseembedding projection based on sample-dependent repulsion graph algorithm isproposed for dimensionality reduction, according to the non-negative sparserepresentation and the sample-dependent repulsion graph building adjacency graph.Then, for the high calculation complexity, low reconstruction accuracy and othercharacteristics of non-negative sparse representation, an over-complete blockdictionary to design a block non-negative sparse representation is introduced to solvethe deficiency. Three block non-negative sparse representation algorithms areproposed, i.e. dimensionality reduction of hyperspectral data based on unsupervisedblock non-negative sparsity reconstruction embedding, supervised non-negativesparsity graph, and semi-supervised non-negative sparse semi-supervised algorithm.2. Tensor-based hyperspectral data dimensionality reduction. According to thetensor characteristic of hyperspectral data, a dimensionality reduction algorithm basedon high-quality tensor neighborhood graph and patch alignment is proposed.Hyperspectral data is represented as tensor form through a window field for keepingthe spatial information of each pixel. According to the patch alignment frameworkextended to the tensor space, which is introduced for the purpose of representing thespatial information between local tensor samples and achieving global optimal.Additionally, considering the traditional KNN or-ball neighborhood algorithms areboth based on Euclidean distance, but the traditional graph-building method cannot really reflect the distance between tensor data, so the high-quality tensor distance isdesigned to construct a high-quality tensor neighborhood graph containingdiscriminating information.3. Feature-based transfer learning for hyperspectral data dimensionality reduction.Many machine-learning-based dimensionality reduction algorithms performs poorwhen the source and target hyperspectral data are drawn from the differentdistribution. Therefore, a dimensionality reduction algorithm based on pairwiseconstraints discriminative analysis and non-negative sparse divergence forhyperspectral data is proposed. The algorithm can automatically obtain the pairwiseconstraints samples which contained discriminating information, a bridge for thedifferent distributions between source and target domain hyperspectral data is builtthrough a non-negative sparse divergence criterion, which can achieve the goal ofknowledge transferring from source to target hyperspectral data.The research results can not only provide a new analysis and design methods andtechnical reserves for hyperspectral data dimensionality reduction, but also furtherdeepen and enrich the existing machine learning, pattern recognition, remote sensingscientific and other interdisciplinary theory, which is of great theoretical significanceand practical value.