关键词: 高光谱图像分类, 加权指数函数模型, 最小二乘法, 模糊C均值, 光谱曲线, 曲线拟合, 主成分分析, 最小噪声分离

Abstract:

In recent years, the hyperspectral remote sensing technology has developed rapidly. Hyperspectral images obtained by hyperspectral sensors contain abundant spectral information of ground objects, such that they are good for fine spectral recognition. In hyperspectral image processing, its accurate classification is a solid foundation for the subsequent interpretation tasks. However, the numerous bands in hyperspectral imagery not only provide sufficient characteristics information for classification, but also bring the problem of how to use these characteristics effectively. In this paper, to make full use of the spectral information of hyperspectral images so to achieve accurate classification, a hyperspectral image classification method based on the Weighted Exponential Function (WEF) was proposed that considers the multi-peak characteristics of the spectral response curve of pixels. Firstly, the WEF model was used to build an ideal model of the spectral response curve of pixels, composed of several exponential functions with different weights. Because there are many parameters in the model (including weight, peak position and peak width), it is difficult to solve them. Therefore, the WEF model with fixed peak positions and number of exponential functions was used to model the spectral response curve of all pixels. Then, the parameters of the WEF model were determined according to the least square principle to fit the spectral response curve. Finally, the parameter set was used to replace the spectral measure vector, and the WEF model parameter vector of the pixel was used as its feature. Fuzzy C-means (FCM) algorithm was used for image classification. To validate the feasibility and effectiveness of the proposed method, the classification experiments of Salinas and PaviaU hyperspectral images were conducted by using respectively the proposed method, Principal Component Analysis (PCA) based classification method, Minimum Noise Fraction (MNF) based classification method, and FCM method with the spectral measure vector as the classification feature. The user accuracy, product accuracy, overall accuracy and Kappa coefficient of the results from these classification methods were calculated, and the experimental results were evaluated qualitatively and quantitatively. Compared with other methods, the classification accuracy of the proposed classification method for the Salinas image increased from 51% to 60%, and for the PaviaU image from 43% to 51%. In addition, the proposed classification method reduced the amount of hyperspectral image data while preserving the rich spectral information of hyperspectral images.

Key words: hyperspectral image classification, weighted exponential function model, least square method, Fuzzy C-means, spectral curve, curve fitting, principal component analysis, Minimum Noise Fraction