Methods Of Mineral Feature Extraction And Classification By Hyperspectral Remote Sensing Images

Space, radiation and spectral information of hyperspectral remote sensing imagesare rich, whose spectral resolution is at λ/100orders of magnitude. A number of verynarrow and contiguous spectral image data can obtained in the electromagneticspectrum of ultraviolet, visible, near infrared and mid-infrared region. Hyperspectralremote sensing images have lots of bands, can describe surface features’ spectralproperties in detail which bring some advantages in surface features identification.But the bands are strongly correlated and information is redundant which also bringsome challenges. Dimension disaster(Hughes phenomenon) is usually occurred inhyperspectral remote sensing image classification.When traditional supervisedclassification methods are used for hyperspectral image classification, trainingsamples is increased dramatically with the increase of band number. Lots oftraining samples are difficult to obtained in hyperspectral remote sensing images.Hyperspectral remote sensing feature extraction can alleviate Hughes phenomenon,it can compress data and remove noise of data on the one hand; on the other hand,spectral characteristics of the target is more obvious through feature extraction, it ismore conductive to the subsequent classification and information extraction. Linearfeature transformation methods such as principal component analysis(PCA), min/maxautocorrelation factors(MAF), minimum noise fraction(MNF) are used widely inhyperspectral remote sensing feature extraction. But they are likely to result indistortion and loss of data information for non-linear hyperspectral remote sensingimage data. Scholkopf expanded principal component analysis to nonlinear kernelprincipal component analysis (KPCA) in1998, from then on KPCA is studied homeand abroad, but few researchers studied kernel min/max autocorrelationfactors(KMAF) and kernel minimum noise fraction(KMNF).Hyperspectral remote sensing is a new area, majority of the data are charged. Thedissertation focuses on theory and methods, so the free AVIRIS hyperspectral remotesensing images in Cuprite, Nevada, United States are used as data source. The paperstudies hyperspectral mineral feature extraction based on kernel methods (KPCA,KMAF, KMNF); extracts end-member after feature extraction; studies hyperspectral image classification and information extraction based on spectral features and supportvector machine (SVM).The dissertation has the following achievements and innovations:(1) Design and implementation of hyperspectral image feature extraction based onkernel methods (KPCA, KMAF, KMNF). We introduced kernel methods in PCA,MAF, MNF algorithms, developed KPCA, KMAF, KMNF nonlinear featureextraction algorithms and did experiments to study their parameters, compared PCAand KPCA, KPCA and KMAF/KMNF. It showes that KPCA, KMAF and KMNFparameterσhas little influence on algorithms’ time efficency; with the increase ofsample number running time increases rapidly, but hyperspectral image featureextraction based on KPCA、KMAF、KMNF can get good results with small samples;PCA, KPCA methods reduct data dimension quickly, but images after it are notstrictly sorted by their quality; KMAF, KMNF reduct data slower than PCA,KPCA,but images after it are strictly sorted by their quality, this is very usefull for bandselection, image classification and information extraction.(2) Design and implementation of PPI endmember extraction based on PCA,KPCA, KMAF/KMNF. Based on different images after feature extraction, use PPIand N-dimensional visualizer to extract endmembers. We find that PPI endmemberextraction based on PCA depends more on operators’ skills and may lost someendmembers because PCA compress data quickly and may lost image information;PPI endmember extraction based on KPCA,KMAF/KMNF can get more endmembersbecause they consider image non-linear characteristics; PPI endmember extractionbased on KMAF/KMNF are easy and efficent because images are strictly sorted bytheir quality.(3)Implementation of hyperspectral image classification based on spectralcharacteristics. Binary encoding, spectral angle mapper (SAM), matchedfiltering(MF), mixture-tuned matched filtering(MTMF）were used in hyperspectralimage classification, accuracy assessments were done. We find that binary encoding issuitable for hyperspectral image roughly classification; MF is a fast classificationmethod and can produce more false signals; SFF is an absorption feature basedclassification method, it can identify minerals which have obvious absorption featureefficently; SAM angle values have nothing to do with spectrum vector modes, itsimply compares the spectrum shape similarity; MTMF is the combination of MF andlinear mixture theory, it reduces the false signal appeared in MF and can get highclassification accuracy.(4) Hyperspectral image classification and information extraction based onimproved support vector machine (SVM). The theoretical basis of support vectormachines and its classification principles are studied. One-against-one method is usedto extend the two-class support vector machine to multi-class in hyperspectral imageclassification, shringking and caching techniques are used to improve its efficiency. SVM-based hyperspectral image classification experiments are done and influence ofdata dimension, kenel function and sample number are analyzed. We find that datadimension has little influence on SVM-based classification, and it has some noiseimmunity; different kernel functions may get similar classification accuracy. Withreasonable parameters, we can get high classification accuracy when sample numberis small. All of this shows superiority of SVM-based hyperspectral imageclassification.（5）Initially forms a set of scientific and practical mineral weak informationextraction methods and technology based on hyperspectral remote sensing images.Fist get hyperspectral remote sensing reflectance data by radiometric correction; thenuse kernel methods for feature extraction, so data was reducted; extract endmemberbased on reducted data; finally based on spectral characteristics and SVM for imageclassification and information extraction.