Unsupervised target detection and classification for hyperspectral imagery

Hyperspectral imaging has become an emerging technique in remote sensing analysis. With high spectral/spatial resolution many unknown material substances can be revealed by hyperspectral imaging sensors for data exploitation. While it significantly improved the capability for target detection it also threw great challenges in designing and developing effective methods to process data and to get the needed information for image analysis. The primary focus of this dissertation is the development of unsupervised approaches to target detection and classification for hyperspectral imagery. The research represents a significant departure from supervised approaches where the target information is assumed to be given or can be obtained a priori. Three major types of spectral targets are of particular interest in this dissertation. One is endmembers whose spectral signatures are idealistically pure. Another is anomaly which usually occurs in small size with signatures significantly different from image background. The third is human-made objects. All these types of spectral targets are usually appear in small population and occur with low probabilities, e.g., special spices in agriculture and ecology, toxic wastes in environmental monitoring, rare minerals in geology, drug/smuggler trafficking in law enforcement, combat vehicles in the battlefield, landmines in war zones, chemical/biological agents in bioterrorism, weapon concealment and mass graves. These spectral targets are generally considered as insignificant objects because of their very limited spatial information but they are actually critical and crucial for defense and intelligence analysis since they are insignificant compared to targets with large sample pools and generally hard to be identified by visual inspection. From a statistical point of view, the spectral information statistics of such special targets cannot be captured by 2nd order statistics as variances but rather by high-order statistics (HOS) as skewness, kurtosis and etc. In light of this interpretation we categorize the image pixels into two classes. One is background (BKG) with pixel spectral signature characterized by 2nd statistics. The other is target with pixel signature characterized by high-order statistics. Once image pixel vectors are categorized into BKG and target classes according to spectral properties, the follow-up task is how to find them. One is how many of them. The other is how to extract them. The key discovery that led to the techniques developed in the remaining chapters of this dissertation is the recognition that a data normalization technique called sphering can retain high order statistics but remove the 1st and 2nd order statistics of a data set. Three least-squares based unsupervised virtual endmember finding algorithms (LS-VEFA) and a component-analysis based unsupervised virtual endmember finding algorithms (CA-VEFA) are developed in this dissertation to extract target and background signatures for image analysis based on linear spectral mixture analysis (LSMA) model. In addition, to address the drawbacks of the current virtual dimensionality (VD) techniques which is originally designed to estimate the number of distinct signatures in a hyperspectral image, an orthogonal subspace projection approach has developed to find the number of signal resources in hyperspectral imagery with the implementation of two separately developed algorithms, unsupervised target sample generation (UTSG) algorithms and unsupervised background sample generation (UBSG) algorithms. The mathematical basis of these new processing approaches is the concept of subspace projection. Projection of each pixel in a hyperspectral image sequence onto an appropriate subspace provides least squares optimal rejection of interfering spectral signatures, maximization of signal-to-noise ratio for the spectral signature(s) of interest.