A Method Of Segmentation For Hyperspectral Image Based On Manifold Learning Pixel Distribution Flow

Hyperspectral remote sensing is the current frontier remote sensing technology, which uses a lot of narrow band of electromagnetic waves from objects of interest to obtain the relevant data, which contains abundance information of spatial, radiation and spectral.In dealing with the image data with more spectral bands, an inevitable question arises, hundreds or even thousands of bands presented in the dimensions of the new requirements on the traditional image processing methods. It has become a common problem of analysis of hyperspectral image that is how to use the abundance spatial and spectral information of hyperspectral image data to carry out a more precise classification, recognition and the segmentation of hyperspectral image.Manifold learning methods results effectively apply to nonlinear dimensionality problem. Recently, there are a lot of related researches carried on hyperspectral image processing using manifold learning. For the problem that normally using manifold learning algorithms applied to hyperspectral data which can hardly given label information, a method based on manifold learning algorithm which the spatial context within a 2-dimensional array is introduced for classification of hyperspectral image is proposed in this paper. We weighted combination of the spatial information and the information of spectral reflection coefficient, a regular variation of stability from the spatial information to the information of spectral reflection coefficient, through adjusting the weights of the two characteristics, is acquired. Using this variation the pixels of the boundaries for classification is found, and labeled the hyperspectral data with high accuracy.This paper mainly introduced the spatial characteristics which describes the relation of position of two pixels on the image. Propose the joint Gassian distance measurement to improved Laplacian Eigenmap algorithm, constructed the Pixel Distribution-Flow and apply to the Hyperspectral image data. The main innovative points are as follows:(1) Introduce the spatial characteristics to the spectral characteristics of each pixel, we combined these two kind of characteristics which used for the selection of neighbors via multiplied Gaussian distances, including how to combine the data of two input spaces, and the selection of parameters. The construction of Pixel Distribution Flow is proposed in this paper. The actual physical meaning of Pixel Distribution Flow is explained through the simplifying the parameters.(2) Apply Laplacian Eigenmap Pixel Distribution-Flow on Hyperspectral image data. According to the actual characteristics of the data set and the intrinsic problem of the result of the mapping, we proposed two kind necessary adjustments. Including nonlinear geometry adjustment which can obtain stable mapping result via fix the original mapping through the prior knowledge of the spatial characteristics. And the adjustment for aggregation of the marginal part of the image. This adjustment is for the algorithm introduced the spatial characteristics lead to the unreasonable aggregation of marginal part of the image compare to the internal part of the image. Then we select the boundary points via hard-threshold and obtain the result of segmentation of hyperspectral data with high accuracy.(3) For large scale problem applying Laplacian Eigenmap Pixel Distribution-Flow, the aggregation of the marginal part of the image will be more serious, and this will lead the hard-threshold do not work anymore, in this paper we proposed mulit-thresholds method which use several useful thresholds to select the boundary points simultaneously to solve this problem.