Ranking-based-clustering And Superpixel Segmentation For Hyperspectral Remote Imagery Dimensionality Reduction

Through imaging the same spatial area by hyperspectral sensors at different spectral wavelengths simultaneously, the acquired hyperspectral imagery often contains abundant semantic information, which makes hyperspectral remote sensing imagery in the field of surface objects classification and identification attract widespread attention. Currently, hyperspectral remote sensing technology has played an important role in many areas such as military and civilian. However, due to the difficulty of obtaining sufficient labeled training samples in practice, the high dimensional features of spectral bands unavoidably leads to the problem of “dimensionality disaster”(also called Hughes phenomenon). Therefore, dimensionality reduction should be applied for hyperspectral imagery. Band(or feature) selection and feature extraction are as important dimensionality reduction means and frequently applied for hyperspectral imagery processing. Conventional band selection methods are two main categories: ranking-based band selection methods and clustering-based band selection methods. The former methods choose the representative bands by ranking the bands with defined metrics(such as non-Gaussianity) and then assigned a sorted weight for every band to select a specified number of bands. The latter formulate the band selection problem as a clustering procedure and select the cluster centers as representative bands. Although the idea of two methods are different, but they have complementary advantages. Therefore, it is beneficial to use both methods together to accomplish the band selection task. This paper proposed a ranking-based clustering band selection method called E-FDPC. Besides, a method which combines entropy rate superpixel segmentation information is studied for feature extraction. The main contributions of this paper contain the following aspects:Firstly, the original fast density peak clustering(FDPC) method has been enhanced to be suitable for hyperspectral imagery band selection from two folds: On the one aspect, the local density and intra-cluster distance in the procedure of clustering have been modified for subsequent band analysis by weighting these two parameters. On the other aspect, an exponential-based learning rule is employed to adjust the cutoff threshold for different number of selected bands, which makes the selected bands more representative.Secondly, an effective strategy, which is called the isolated-point-stopping strategy is developed to automatically determine the appropriate number of bands to be selected. That is, the clustering process will be stopped by the emergence of an isolated point(the only point in one cluster), which can be as a rational indicator for choosing an appropriate number of bands and achieve the balance between classification accuracy and dimensionality reduction.Finally, a method that is based the entropy rate superpixel segmentation information in feature extraction for hyperspectral imagery dimensionality reduction is proposed. Compared with the origin pixel-based feature and simple linear iterative clustering(SLIC) superpixel segmentation method, it can obtain more distinctive classification features and improve the ability of feature set for hyperspectral surface identification.