Hyperspectral Remote Sensing Data Mining Using Multiple Classifiers Combination

Since the advent of remote sensing in the second half of 20th century, nowadays there have been great changes in theory and technology. The advent of hyperspectral was one of the most significant breakthroughs in remote sensing. Hyperspectral remote sensing has higher spectral resolution as the same time retain higher spatial resolution, so its capability of distinguishing the different and describing the same ground objects in details enhanced greatly. It acquires image in a large number (typically over 40), narrow (typically 10 to 20 nm in width) and contiguous spectral bands to enable the extraction of reflectance spectra at a pixel scale, so it can produce data with sufficient resolution for the direct identification of those materials with diagnostic spectral features (Goetz et al., 1985). The objective of hyperspectral remote sensing is to measure quantitatively the components of the Earth System from calibrated spectra acquired as images for scientific research and applications (Vane & Goetz, 1988). The rationale behind this technology for geological applications is that mineral species have diagnostic absorption features from 20 to 40 nm wide in electromagnetic wavelength ranges which is larger than hyperspectral spectral resolution (van der Meer & Bakker, 1997). Goetz demonstrated firstly that direct identification of carbonates and hydroxyl-bearing minerals is possible by remote measurement from Earth orbit (Goetz et al., 1982). There are two main categories of extracting information method from hyperspectral remote sensing image: based on feature space and based on spectral space. Many statistics-based classification methods based on feature space have been successfully applied to multispectral remote sensing data in the past years (Pal & Mather, 2003, Wen et al., 2008b). However, they are not effective for hyperspectral remote sensing data. The problem is caused by curse of dimensionality and Hughes phenomenon (Hsu, 2007), which refer to the fact that the sample size required for training a specific classifier grows exponentially with the number of spectral bands. Usually simple but sometimes effective ways to overcome this problem is to increase sample numbers or to reduce the dimensionality of hyperspectral remote sensing data. The former needs a lot of sample numbers, so it will cost many human and material resources; the latter will lead to some useful information lost. The Matched Filtering methods based on spectral space are successfully used in hyperspectral data. These O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg