Research On Lunar Soil Analysis Based On Multi-spectral Image Fusion And Target Classification Of Neural Network

The Chang'e Project, China's first lunar exploration mission, is one of the most important scientific projects in China, and has raised great attention. The project also creates great opportunities for hi-tech scientific application and research such as space science research. Because research of lunar soil is a significant part of the lunar exploration mission, this essay includes research on multi-spectral information analysis and processing based on lunar soil using the technologies including wavelet analysis and neural network. The research contains two parts. One is multi-spectral image fusion research while the other is multi-spectral target classification research of the lunar soil based on neural network classifier.In the research of the multi-spectral image fusion, this essay introduces the definition, procedure and key technologies of the multi-spectral image fusion. It analyzes the characteristics of image local feature, directional contrast of wavelet coefficients and high-frequency components of images. Then, the essay puts forward a multi-spectral image fusion algorithm based on relativity of directional contrast and weight of wavelet coefficients. After the fusion experiments of the multi-spectral image of the earth surface, the subjective and objective analyses of the fusion results show that this algorithm is effective in improving fusion quality of the multi-spectral image. The result of the fusion experiment of multi-focus images indicates that the algorithm is good to spread.In the research of the multi-spectral target classification research of the lunar soil based on neural network classifier, this essay analyzes the features of neural network after introducing the theory and classic methods of the multi-spectral target classification and the basic knowledge of neural network. Having studied the multi-spectral information of the lunar soil using the wavelet theory and relativity analysis, the conclusions are drawn that the problem of the multi-spectral information dimensionality reduction can be solved by using the low-frequency wavelet coefficients, and that the problem of the decorrelation can be solved by using the first level high frequency wavelet coefficients. Then, network weights and bias are figured out after a three-level feedforward neural network classifier is designed and receives trainings. The multi-spectral test samples of four kinds of materials including lunar soil are classified by using the network classifier. This essay makes a meaningful attempt in using the neural network classifier in the multi-spectral target classification.This essay also gives brief introduction of the physical and chemical components and property of lunar soil in the appendix.