Homogeneous-Region Analysis Of Hyperspectral Image Based On HDA And MRF

Hyperspectral remote sensing is a new technology, and this technology brings people strongly method of accurately-identify surface feature by means of acutely ability of spectral feature detecting.Recent analysis methods are commenly based on pixels, which thinking way confirm land surface typethrough every pixel's spectral feature. In this way, many methods appeared which were represented by spectral-match and multi-dimension spectral feature classification. Many worthy research results were accumulated and pushed forward the development of hyperspectral remote sensing applied technology.But from the angle of people's investigation of the world, the targets'feature is multi-attributive and the targets'implication is often subjective. In this condition, the way based on pixel is hardly afford satisfaction in target recognition, and homogeneous-regions give a more rich expression to the targets, so that, the image apprehension based on homogeneous-region is advantageous.As the basis of latter work, image segmentation is the key step in producing homogeneous-regions (blocks), and hyperspectral image segmentation is a brandnew research subject, which was what this paper focus on and delve into. This paper firstly bring forth the frame of hyperspectral image classify based on homogeneous-regions, and exausted the pisition and effect of homogeneous-region analysis in the frame. Based on this, this paper bring forth a new type of Successive Projection Pursuit algorism, which steply select effective feature component by means of multi-time one dimension projection.And then, this paper take a measure of Density Analysis to segmente hyperspectral remote sensing image. Operated against the complexity of remote sensing image spectral space, this paper present an algorithm of HDA (Hierarchy Density Analysis) to analyze the feature-space in order to get self-adaptive segmentation, and get good result. Nextly, this paper described the mapping model of spectral vector into 2-dimension space using Markov Random Field (MRF), established a texture model of multispectral remote sensing image based on MRF, and analyzed the calculations of Gibbs potentials and Gibbs parameters. Based on this, this paper presents an iterating algorism of multi-band image texture segmentation. Further, this paper and analyzed the limitation of traditional Gibbs model, and present a Non-parameters Gibbs model and taken a measure of multi-level segmentation. Laterly, this paper present an new algorism of hyperspectral remote sensing image texture segmentation.Based on this, this paper presents a Binary Markov Random Field model for multi-scale texture segmentation of multi-dimension remote sensing image. This paper converts a one-time multi-class segmentation problem into a continual binary segmentation problem. At the same time, The new B-MRF model synthetically considered scale of 2-d image space and scale of spatial space. Associated with Spectral Code Pyramid, new method got multi-scale texture segmentation for multispectral or hyperspectral image.Lastly, in order to demonstrating the utility of homogeneous-regions, this paper discusses the image classification based on homogeneous-regions. Aimed at the problem of optimum scale selecting, this paper take a measure of crossover-scale supervised classification based on blocks, and inderectly solved this problem during the sample training.