新型高光谱图像的超分辨率制图方法

A novel Super-resolution mapping method for hyperspectral imagery

摘要:

由于高光谱图像的应用在很大程度上受限于其较低的空间分辨率,为此提出了一种结合支持向量机和小波变换的高光谱图像超分辨率制图方法．先对高光谱图像进行光谱解混得到分量图,然后对分量图进行一级小波分解．各局域窗内中心像元的3个高频系数与邻域像元低频系数之间的对应关系表示为训练样本,用于支持向量机的学习．训练好的模型用来对低分辨率图像即分量图进行超分辨率制图．实验表明,这种借助小波变换来获取训练样本的学习方法无需先验信息,相比采用BP神经网络学习的方法,支持向量机的超分辨率制图效果更佳．

Abstract:

The application of hyperspectral imagery (HSI) is quite limited to its low spatial resolution. A new method for super-resolution mapping of HSI is proposed using support vector machine(SVM) and wavelet transform. Firstly, spectral unmixing is processed for HSI and the fraction images are obtained. Then the wavelet decomposition is processed on these fraction images. In the local window, the relation between the three high-frequency coefficients of the center pixel and low-frequency coefficients of neighbour pixels are described by training samples, which are used for the learning process of SVM. The trained SVM models are utilized to predict the super-resolution mapping results of the coarse resolution images, i.e., the fraction images. Experiment results show that using wavelet transform can eliminate the dependence on prior information, and compared with the learning method based on BP neural network, SVM can produce higher accuracy super-resolution mapping results.