Research On Independent Component Analysis Algorithm And Its Applications In Hyperspectral Remote Sensing

Independent Component Analysis,referred to as ICA,is a method for blind source separation,which carry out the independency and non-Gauss from the multi-dimensional data.The biggest advantage of ICA is that it can well figure out the optimal estimations of source signals and then separate them from the mixed signals by assuming the statistical independence,when any information of source signals and mixing matrix is unknown.Because ICA reflected higher-order statistics characteristic of image data brilliantly,it had successful application in many fields of hyperspectral remote sensing image processing.The fast independent component analysis algorithm was studied deeply in this paper,and an improved algorithm based on negative entropy was proposed,the experimental results show its better robustness.The improved algorithm is applied in dimensionality reduction and small detection under the uniform background in hyperspectral images,which has Significant value.The main work and innovation of this paper as follows:(1)The foundations of mathematics,basic principles and solving constraints of ICA were briefly summarized,and the general solving process was discussed in detail,meanwhile the popular ICA algorithms were analyzed and compared.(2)The fast independent component analysis(FastICA)was researched emphatically.The FastICA algorithm based on negative entropy is Sensitive to The initial weights and has slower speed of iterative calculation,as to this issue,an advanced algorithm named M-Fast ICA was presented to improve the convergence performance.M-FastICA added a searching agent in the direction of Newton iterative,to improve its dependence on the initial weights,meanwhile combined with the fifth-order convergence of Newton iterative method to improve the convergence speed.(3)Analyzing the data characteristics of hyperspectral remote sensing images,we considered the bands images of hyperspectral remote sensing were mixed by the spectral features of diverse surface feature randomly.Comparing with dimensionality reduction method by principal component analysis and minimum noise fraction transform based on the second-order statistics,the M-Fast ICA algorithm was applied to the dimensionality reduction and de-noising of hyperspectral image data,and obtained the statistically independent band component,which were more conducive to subsequent processing and application.(4)The feature extraction techniques for hyperspectral image data small targets detection are studied.The M-Fast ICA was applied to feature analysis of hyperspectral images,which can extract the non-gauss structure of hyperspectral data,and small target in image can be detected effectively in homogeneous background.