Extraction of Absorption Features from Hyperspectral Data Using Gaussian Model with Parameter Initialization

• Main Authors: Jenn-Yii Wu, 吳震乙
• Other Authors: Wei-Song Lin
• Format: Others
• Language: en_US
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/66585933405633603349

碩士 === 國立臺灣大學 === 電機工程學研究所 === 87 === Hyperspectral image can provide sufficient spectral sampling to define unique spectral signatures on a per-pixel basis and can be applied broadly to variety fields. With the enormous volume and abundant information, techniques to extract information from of the hyperspectral data sets need to be developed. Gaussian model is commonly used to model reflectance spectra as a series of Gaussian curves with each curve characterized by a central wavelength position, width, and strength. It can potentially provide additional information concerning the physical mechanisms that are responsible for absorption in the region of the spectra. In this thesis, the Levenberg-Marquardt (L-M) iterative method instead of the Hessian method is applied to fit the spectrum curve because of its robustness to initial parameters. A method to estimate the initial parameters, including the widths of absorption bands, of the Gaussian model is proposed to improve the efficiency in extracting the absorption features. Applying derivative analysis, Huguenin’s criteria are used to estimate the center positions of the absorption features. Then by subtracting the amount contributed by adjacent absorption bands from the overall spectra, the strength of each absorption band can be estimated easily. Finite approximation is responsible for calculating derivatives and mean filter is applied to remove the effect of noise. An experimental system named ‘Hyperspectral Imaging Lab” (HIL) has been built to acquire the experimental data. Results of simulations and experiments show that the proposed method can extract absorption features efficiently and accurately.