Target Detection And Classification For Hyperspectral Imagery

Target detection and classification is one of primary tasks of hyperspectral imaging. In terms of the method of spectral expression, the style of unsupervised cluster, probability in the data, the geometrical construction of hyperspectral imaging in the band space and the it's continuity in the image space, the dissertation draws some conclusion on feature extraction, unsupervised classification, endmember selection, linear unmixing, target detection and anomaly detection as follows:1. A method for spectral feature extraction was developed based on spectral recomposition. By arranging the spectra by the sort of their reflectance or DN, the spectral curves that are originally difficult to be extracted features from will usually produce some obvious features. It is helpful to feature extraction and father analysis and process.2. The concept of spatial continuity was proposed and successfully applied to image classification, spectral optimization, redundancy reduction and real-time endmember determination.3. A unsupervised classification method was proposed based on universal gravitation. Each pixel that was taken as a star in the universe would move with the gravitation of all the other pixels. The last formed galaxy is corresponding to the result of classification.4. Two approaches of autonomous spectral endmember determination were developed. Based on the convex nature of hyperspectral data in its characteristic space, Gram-Schmidt Orthonormalization process, high dimensional analytic geometry and distance between pixels were introduce to find a unique set of purest pixels in an image; A new volume formula of simplex which was independent of dimension of the data was introduced to find all the endmembers which are larger than any other volume formed from any other combination of pixels. The concept of endmember constrution function was first proposed, so the weightiness of each endmember is depended on it's influence to the simplex contraction, but not it's information magnitude. It's significant to the small target extraction.5. A method of target extraction based on endmember projection vector was developed. Based on the convex nature of hyperspectral data in its band space, a series of vectors named endmember projection vector are produced for use of object extraction. The technique is based on the fact that in band space, any endmember is the farthest point from the hyperplane consisted of all the other endmembers.6. A new theorem about the nature of simplex was proposed and applied to spectral linear unmixing. Once all the endmembers were found, the image cube can be "unmixed" into fractional abundances of each material in each pixel by a simple ratio of volume.