| Hyperspectral Imagery(HSI)can obtain ground-objects information from hundreds of narrow and adjacent spectral channels with the advantage of high spectral resolution and strong spectral distinction.HSI has been widely used for detailed ground-objects classification and lithologic mapping.The large number of bands,high redundancy and great amount bring many challenges to these studies.Therefore it is necessary to reduce the dimension of HSI.Band(feature)selection is a common technique for dimensionality reduction,i.e.select an optimal subset from all original bands and can reserve most information without changing the physical meaning of original data,in order to achieve dimensionality reduction.The process of HIS band selection is a complex optimization of band combination.The criterion function for band selection should has validity,i.e.,the selected bands subset should with strong discrimination,high class separability and less correlation,and the search strategy should has a certain timeliness.Therefore,on the basis of the results of previous studies,this thesis takes advantage of priori-information to construct the feature metric(FM)and use affinity propagation(AP)to select the optimal bands subset,then two efficient band selection algorithm have been proposed for ground-objects classification and lithologic mapping.The main contents of this paper are as follows.1.The research of HSI band selection based on discriminative component analysis(DFM-AP).for the difficult of label sample obtains in HSI,chunklet information is formed use the pairwise constraints,then a criterion function is defined combine a effect distance metric learning method,i.e.,discriminative component analysis(DCA),for minimizing the total variance of data points within the same chunklets and maximizing the total variance between the discriminative chunklets,then the discriminative feature metric(DFM)is constructed.Finally,AP is used as the search strategy to select the optimal subset which has high class separability in ground-objects classification and low correlation between bands.In the experiments,the AVIRIS and Hyperion imagery are used,comparing with seven widely used methods,i.e.,information divergence(ID),Fisher’s linear discriminant analysis(FLDA),traditional AP and adaptive affinity propagation(AAP),etc.The experimental results consistently prove the effectiveness of the DFM-AP in the stability time-consuming and the sensitivity analysis.2.The research of HSI band selection based on SLIC superpixel feature metric(SFM-AP).According to the complex geologic aspect of research area(Nevada Cuprite,USA),the SLIC superpixel segmentation is combined with the spectral angle distance(SAD)for dividing the HSI into superpixel with different size,which have good homogeneity.A SLIC superpixel chunklets(SSC)and SLIC superpixel-based feature metric(SFM)are setablished.The AP is also used as the search strategy to select the optimal bands.The classification and mapping of 11 lithologic units in 9 lithologic class of research area are summarized with spectral angle mapping(SAM)and spectral feature fitting(SFF)analysis from differences of spectral characteristics and spectrum characteristics of the lithologic units.From the results of the proposed algorithm,the selected optimal bands subset can clearly and precisely express the distribution of lithologic units of research are compared with an original bands and can provide a new idea in lithologic classification and mapping study. |
