Independent Component Analysis For Mineralizing Alteration Information Extraction Of Hyperspectral Remote Sensing Data

Hyperspectral remote sensing obtains information of continuous spectrum of ground objects by aerial or space imaging spectrometer. The spectral resolution of hyperspectral images reach up to 10nm, making undetectable surface features in traditional remote sensing be able to be detected in hyperspectral remote sensing. Currently, hyperspectral remote sensing has been widely used in geological mapping, vegetation investigation, marine remote sensing, agriculture remote sensing, atmospheric research, environmental monitoring and so forth. Geological prospecting is one of the most successful areas of hyperspectral remote sensing application, including mineral identification, geological mapping and alteration or mineralization information extraction. Hyperspectral mineral identification is based on the theory of the stability of chemical composition and physical structure of rock minerals with stable characteristics of the intrinsic absorption. Therefore, minerals can be effectively identified by appropriate identification methods after analyzing the characteristics and structure of hyperspectral image.Hyperspectral remote sensing data has the property of high spectral resolution and low spatial resolution, which makes it possible to achieve the goal of fine identification of surface features, brings some problems at the same time. One is the high dimension of hyperspectral data makes the processing, storage and transmission difficult; the other is the phenomenon of mixed pixel of hyperspectral image because of low spatial resolution, which reduces the accuracy of remote sensing classification and information extraction. Therefore, how to extract useful information from hyperspectral image and decompose the mixed pixels are the two major challenges we are facing.At present, a variety of pattern recognition methods based on statistical and data processing theory have been applied to the process of classification and information extraction of hyperspectral image. Independent component analysis (ICA) is a analysis method based on higher order statistical properties of signal. Observed signal has been linearly decomposed into statistically independent and non-Gaussian components. ICA uses higher order statistical properties of the data, which can achieve dimensionality reduction of hyperspectral data and hyperspectral unmixing. The surface features will be represented as separate components by ICA, making separation between surface features to maximize. Now research ideas can be divided into two kinds: Firstly, from the point of view of data reduction and effective band selection, ICA was used to transform the spectral band to find the spectral effective representation; secondly, ICA can be used in spectral unmixing, completing remote sensing classification and information extraction.In this paper, the characteristics of hyperspectral remote sensing data were analyzed. Using USGS Mineral Spectral Library to construct hyperspectral simulated images to explore the application of ICA in hyperspectral image processing, trying to find a better method for classification of hyperspectral image and information extraction of mineral alteration.The major works of this paper are as follows:1. The basic principle of ICA technology, algorithms, and criteria were reviewed and summarized;2. The characteristics of hyperspectral remote sensing data were analyzed, the traditional method of feature extraction and classification in hyperspectral image were reviewed. A feature extraction model and spectral unmixed model based on ICA were explained combined with the property of ICA;3. Constructing simulated images using USGS Mineral Spectral Library, experimenting ICA in feature extraction and spectral unmixing. Spectral angle mapping (SAM) method was used to classify the simulated data before and after feature extraction. The classification accuracy of feature extraction remains more than 90% after feature extraction, which indicates that the feature extraction based on ICA is effective. For ICA-based decomposition of mixed pixel, the endmember extraction error was limited in 10? 3.Experimental results showed that the ICA technique can be used to effectively extract features of hyperspectral data and can also be used in linear unmixing model, which obtained both endmember and abundance information at the same time.