A Research Of Feature Extraction And Classification Techniques For Target Detection In Hyperspectral Image

Hyperspectral remote sensing effectively integrate the spectral feature with geometric characters of targets. Theoretically speaking, hyperspectral data is greatly propitious to explore target' s physical and chemical characters deeply or to classify different targets precisely. But conventional data processing methods in panchromatic and multispectral remote sensing can not satisfy demands of hyperspectral data' s information extraction. In case of so much spectral bands and such huge quantities of data, the key problem is how to extract the interest information.This dissertation explored the theories and methods for feature extraction and classification of hyperspectral data target detection, which are parts of important research contents of National 863-708 Hi-Tech, and concentrated on hyperspectral data target detection, studied de-noising methods of hyperspectral data, small targets detecting methods of hyperspectral data, small training samples classifying of hyperspectral data, nonlinear feature extracting of hyperspectral data mainly. In general, the major works and contribution of this paper are as follows.1 Hyperspectral data' s characteristic and its effect on target detection are systematically summarized and analyzed. Factor and problem that should be noticed are explored.2 The de-noising techniques of hyperspectral data are studied deeply. Two kinds of methods that based on the cubic smooth spline and stationary discrete wavelet transform respectively are researched, and a new improved threshold de-noising method is brought forward. From noise filtering, signal-to-noise calculating, targets classifying experiments of PHI data, it can be concluded that two methods al 1 can filter the noise in hyperspectral data effectively, which change with wavelengh, and improve data' s qualities on signal-to-noise and classification.3 The feature extraction techniques for hyperspectral data small targets detection are studied deeply. Three kinds of methods that based on Fast Independent Component Analysis (FICA), Real Coding Genetic Algorithm Projection Pursuit(RCGAPP), Multiscale 1-D Wavelet Transform respectively are brought forward. We settle the problem that ICA is not feet to hyperspectral data with FICA. On the grounds of this, a small targets feature extracting method that based on FICA is brought forward. From small targets detecting experiments of AVIRIS and OMIS data, it can be concluded that feature extraction method, which based on fast independent component analysis (FICA), can extract thenon-gauss structure of hyperspectral data effectively, and is greatly propitious to detect small targets from the background objects that distribute uniformly.? We optimize the projection index with Real Coding Genetic Algorithm, and brought forward a small targets feature extracting method that based on RCGAPP. From small targets detecting experiments of AVIRIS and OMIS data, it can be concluded that projection pursuit can mine the small targets' information from hyperspectral data. Compared with ICA, it can extract different characteristic structure by set projection index more flexible.? Multiscale 1-D wavelet feature extracting is a method for target' s spectral characters. It is derived from theories of information structure, can mine the absorbing features on different scales, and reduce quantities of data. From sub. pixel targets detecting experiment of AVIRIS data, it can be concluded that different scale' s features can describe targets' spectral characteristic excellent, and improve target detection performance.4> Feature extraction and classification method that based on small training samples learning are studied deeply. A projection index for small training samples learning and a feature extraction and classification method based on SVM projection pursuit is brought forward, and are generalized for multi-class problem by Error Correcting Output Code. From targets classifying experiment of Man-made and AVIRIS data, it can be concluded that this method has an nice ability on small training samples classifying, and can extract the classification feature exactly by little pre-information.5> Training samples kernel-based techniques, which extract nonlinear feature for hyperspectral data, are studied. A feature extraction method that based on kernel Bhattacharyya projection pursuit is researched, and are generalized for multi-class problem by Error Correcting Output Code. From targets classifying experiment of Man-made and AVIRIS data, it can be concluded that kernel-based methods can mine nonlinear feature from hyperspectral data with better performance and simpler calculation.