A Study Of Spectral-spatial Based Hyperspectral Image Classification

In recent years, the hyperspectral image processing has attracted wide attention. Through hyperspectral sensors in different space platforms, in the ultraviolet spectrum, visible spectrum, near infrared spectrum and infrared spectrum of the electromagnetic spectrum, take image of target area with the continuous spectrum band of dozens to hundreds at the same time. When obtaining the spatial information from target area, also obtain its spectral information. Realize the fusion of spatial information and spectral information.Compared with the multispectral, hyperspectral technology has improved highly in the information abundance of target object. So hyperspectral technology makes it possible for more reasonable and effective analysis and processing of remote sensing image processing. A major application of hyperspectral technology is the classification. For a period of time before the hyperspectral image classification mainly relys on the spectral information, but it loses the information in the spatial domain, graph size or structure etc. Recently literature shows that the classification of hyperspectral image takes attention of combination spatial information with spectral information, called spectral-spatial classification.This dissertation is mainly concerned with the design and implementation of spectral-spatial classification of hyperspectral image. The author’s major contributions are outlined as follows:1. Some existing spectral-spatial classification algorithm of hyperspectral image have been studied, also carry some simulation experiment. Further, some existing spectral-spatial classification method are combined to give new classification result by decision fusion. The experiment result show that spectral-spatial strategy is able to effectively promote the result, not only the classification precision, also the visual representation in classification image. It’s clear that the improved method can further improve the classification result.2. A new spectral-spatial classification algorithm of hyperspectral image on the basis of the hierarchical probabilistic model has been proposed. The hierarchical probabilisticmodel is a mathematical model in computer vision. It can learn and construct the structure information in the feature. We introduce it to construct a new spectral-spatial classification algorithm on the basis of the hierarchical probabilistic model, also achieve good result by the classification experiment.3. A new spectral-spatial classification algorithm of hyperspectral image on the basis of the three-dimensional morphological profile has been proposed. Based on the three-dimensional characteristic in the hyperspectral image, we propose a three-dimensional morphological processing derived from the exiting two-dimensional morphological processing, and construct the three-dimensional morphological profile for feature extraction used for hyperspectral image. Different from spatial information extraction in an image by the traditional two-dimensional morphological processing, the proposed new method takes consideration of three-dimensional data characteristics in hyperspectral image, and extracts its three-dimensional spectral-spatial feature. Simulation experiments are carried on three different hyperspectral images. Compared with the existing method, our proposed new method shows its availability in spectral-spatial classification of hyperspectral image.