Nonlinear Feature Extraction Of Hyperspectral Remote Sensing Data Based On A Fast Manifold Learning Strategy

Hyperspectral Remote Sensing plays an important role in Earth Observation and is a cuting edge technique among all the Remote Sensing applications. The development of the Hyperspectral Sensors accelerates the growth of spacial and spectral resolution of the Hyperspectral Remote Sensed dataset, which burdens the workload of the Hyperspectral data analysis and processes, and helps the Feature Extraction become an indispensible part of the Hyperspectral dataset preprocessing. This thesis proposes a new kind of feature extraction algorithms, as known as Manifold Learning (ML). ML is a kind of nonlinear feature extraction techniques with simple structure, few parameters, and guaranteed globally optimal solution. But its high computational complexisity impedes its development in the Hyperspectral applications where the large scale data size is beyond the processing power of ML. To overcome this problem, an unified framework of the ML algorithm was established after two kinds of ML algorithms were systematically analyzed, globally structure preserved methods and locally structure preserved methods respectively. Spectral Angle Sensitive Hashing Forest (SASHF) and Nystrom’ algorithm were introduced to improve the computational efficiency of nearest neighbor searching and low dimensional coordinates calculating which are the most complicated parts of the ML algorithms. In the thesis, these two algorithms were proved theoretically and experimentally that they have the advantages in computing efficiency against other similar algorithms. An improved framwork of ML was built based on SASHF and Nystrom’ algorithm. In the experiments of clustering and classification on three sets of benchmark dataset, this new ML algorithm can not only maintain the information of the original dataset, but also help to improve the performance of the clustering/classification algorithm. The features extracted by the new ML algorithm were proved to outperform the features extracted by Principle Component Analysis and even the original dataset itself.