小样本的高光谱图像降噪与分类

发布时间：2018-05-04 11:41

本文选题：高光谱图像分类 + 特征提取　；参考：《遥感学报》2017年05期

【摘要】：在样本数目稀少情况下实现高光谱图像精细分类是个挑战性的问题。高光谱图像信噪比提高比较困难,噪声大小对分类结果有最直接的影响。利用高光谱图像相邻波段之间的相关性和相邻像素之间的相关性,提出多级降噪滤波的高光谱图像分类方法,通过改进的两阶段稀疏与低秩矩阵分解方法,去除高光谱图像中能量较高的噪声,利用主成分分析方法去除高光谱图像中能量较低的噪声,引导滤波方法去除分类结果图中的"椒盐噪声"。选取两幅真实高光谱图像进行实验,结果表明,两阶段稀疏与低秩矩阵分解法和主成分分析法两种降噪方法具有较强的互补性;引导滤波方法使得分类图更加平滑且分类精度更高。与其他光谱空间分类方法相比,本文方法分类精度更高,且在样本极少时能获得很高的分类精度。
[Abstract]:It is a challenging problem to realize hyperspectral image fine classification when the number of samples is scarce. It is difficult to improve the signal-to-noise ratio of hyperspectral images, and the noise level has the most direct influence on the classification results. Based on the correlation between adjacent bands and adjacent pixels of hyperspectral images, a multilevel de-noising filtering method for hyperspectral image classification is proposed. The improved two-stage sparse matrix decomposition method and low-rank matrix decomposition method are used to classify hyperspectral images. The noise with high energy in hyperspectral image is removed, the noise with low energy in hyperspectral image is removed by principal component analysis, and the "salt and pepper noise" is removed by guided filtering method. Two real hyperspectral images are selected for experiments. The results show that the two methods of sparse matrix decomposition and low rank matrix decomposition and principal component analysis are complementary to each other. The guided filtering method makes the classification map smoother and the classification accuracy higher. Compared with other spectral spatial classification methods, this method has higher classification accuracy and can obtain high classification accuracy when few samples are available.
【作者单位】：山东科技大学计算机科学与工程学院;
【基金】：国家自然科学基金(编号:41406200) 山东省自然科学基金(编号:ZR2014DQ030)~~
【分类号】：TP751

【相似文献】