Hyperspectral Remote Sensing Image Classification Based On Grouped Band Selection And Compact Dictionary

Hyperspectral remote sensing image contains the rich spectral information which can effectively reflect the potential differences of the various land covers.Consequently,it has become an important medium for people to perceive the surface information from a macro perspective and has been widely used in the fields of military,national defense,environmental monitoring and so on.Although hyperspectral images can provide rich spectral information and bring convenience to hyperspectral image processing,there are still many challenges:1)Due to its large data volume and high redundancy,it requires a lot of data storage space and computing resources during processing.2)The phenomena of same object with different spectrum,different objects with same spectrum still exist in hyperspectral remote sensing images,and the collection of labeled samples is difficult,which impedes the effective use of hyperspectral data for the analysis of ground objects.Band selection can find the most representative band subset from the original hyperspectral data and also can effectively reduce data volume while maintaining the physical chara-cteristics of the original data,thus,it is an important preprocessing step for hyperspectral image processing.Meanwhile,classification is the basis of many applications of hyperspectral remote sensing images.Therefore,this paper launches the research on hyperspectral remote sensing image processing from two aspects:band selection and classification.The main contributions are summarized as follows:1)A novel band selection method based on grouped neighborhood is proposed for hyperspectral remote sensing images.First,the first level non-decimated wavelet decomposition is applied to each band of hyperspectral image.Then,the local iterative clustering algorithm is used to divide the low-frequency components into many neighborhood groups.At last,a band selection criterion,which considers both intra group correlation and inter group difference and takes into account the influence of each band noise,is designed to effectively select the representative and low noise bands in each group.2)A hyperspectral remote sensing image classification method based on compact dictionary sparse representation is proposed,which constructs a compact dictionary for each test sample based on the neighboring known labels and the spectral similarity between test sample and each subdictionary.Furthermore,a spatial location expansion strategy is used during the classification,which makes full use of the labels of classified test samples.3)The proposed band selection and classification methods are respectively compared with several state-of-art methods in three commonly used hyperspectral datasets.Experiments show that the proposed band selection method can effectively remove the noise band and select a subset of bands that are representative in the overall data and conducive to the classification identification.The proposed classification method not only achieves high classification accuracy,but also obtains faster classification speed than similar methods.