The Research Of Hyperspectral Image Precise Analysis Based On Multi-Feature Fusion

With the development of remote sensing technology and the widespread application, the people’s ability of knowing the world can be enhanced by the extension of spectrum and image. The advent of hyperspectral remote sensing technology, it is cleverly combined with the features of the physical and chemical properties of spectral information and the spatial pattern reflects the image feature space information. Hyperspectral image data through image mode and spectrum mode unity, visualization thinking and logical thinking can get together in the epistemology, it makes people’s cognition to the objective world has improved greatly.Every hyperspectral image pixel has an accurate and continuous spectra contain a wealth of information of features. Although, many classic remote sensing data processing methods are still used in hyperspectral image processing and analysis, however, due to the characteristics of hyperspectral image data and the realization of the aim of detailed analysis, making these methods are often difficult to meet the requirements of hyperspectral imagery interpretation. Therefore, it needs to develop some innovative and different from the traditional image processing methods to achieve the precise analysis of hyperspectral image. Over the past decade, aiming at the problems in hyperspectral imagery interpretation, experts and scholars in domestic and abroad made many fruitful scientific researches. From the perspective of research and application, it can be roughly divided into two pieces of content, one is hyperspectral image precise classification, and the other is hyperspectral image target recognition.Spectral curve reflects the different characteristics of surface features. Hyperspectral image reflected in the precise spectral differences make precise classification and small target recognition become a possible things. Many hyperspectral classification methods have been proposed in recent years, with mainly two classes:spectral matching technology and multidimensional spectrum characteristic classification technology. Spectral matching technology is relatively mature in hyperspectral image precise classification, by comparing the spectral similarity degree between spectral curve of pixels and standard spectral in database to classify. Multidimensional spectrum characteristic classification technology is based on the statistical characteristic of data, according to the pattern recognition method based on the statistical characteristics to classify. For hyperspectral images small target recognition, the small target recognition algorithm is used to detect hyperspectral images in particular the existence of weak signal or distribution problems, due to the dimensions of the hyperspectral image is numerous, it can better reflect the nuances of ground objects, thus, results of hyperspectral images of small target recognition has a wide range of practical significance and application value. Taking into account the two content of the hyperspectral image interpretation, this paper to solve these problems from three aspects respectively:one is a multi-feature conversion adaptive classification of hyperspectral remote sensing image, second is classification of fusion of spatial and spectral information in hyperspectral image based on image segment, third is a novel algorithm for hyperspectral image target recognition.First, this paper introduces the spectral similarity measure and the role in the classification of spectral matching, studied the existing spectrum matching classification method using spectral similarity measure. And this paper proposes a multi-feature conversion adaptive classification of hyperspectral remote sensing image. The experimental results demonstrate that this method has a good classification accuracy and stronger adaptability. The different similarity measures values vector versions performed better than only one spectral similarity measures versions.Then, classification of fusion of spatial and spectral information in hyperspectral image based on image segment. It is using the software "eCongnition" to make a multi-scale segmentation of image. And classify by a classifier. Next, the initial result will use the class co-occurrence matrix to optimize the initial result of classification. Experiments show that, by combining spatial information, it can further improve the classification accuracy, and it can eliminate noise or specklelike. And reduce the impact of the mixed-pixels, effectively.Finally, in view of the problem of small targets in hyperspectral image, this paper proposes a weighted independent component analysis orthogonal subspace projection algorithm (ICA-OSP). By independent component analysis method to extract target features and background features. By using the weighted orthogonal subspace projection is to realize separation of target and background. The experimental results show that under the same situation of known image target spectral information and unknown background spectrum information. The ICA-OSP algorithm will improve the detection probability and reduce the false alarm probability. The ICA-OSP outperforms the other traditional target recognition algorithm.