空谱特征分层融合的高光谱图像特征提取

Spatial and spectral feature hierarchical fusion for hyperspectral image feature extraction

利用基于光谱维的特征提取方法将原始高光谱图像数据降到一定维数,对降维后的数据采用多尺度自适应加权滤波器(adaptive weighted filters,AWF)进行滤波,将在所有尺度上得到的滤波结果分层融合为新的图像,设计了分层融合框架,有效提取出了高光谱图像中重要的空谱特征,从而提高了分类精度。又将主成分分析(principal component analysis,PCA)算法融入到该框架中,提出了分层融合-主成分分析(hierarchical fusion principal component analysis,HF-PCA)算法。该方法不仅降低了波段间的冗余性,而且削弱了样本的类内差异性,提高了高光谱图像的分类精度。在Indian Pines和Salinas数据库上的实验结果表明,即使在训练样本数量较少的情况下,由HF- PCA算法得到的分类精度明显高于其他算法,2种数据总体分类精度的最大值分别为86.73%和95.01%,有效提高了高光谱图像的分类精度。

In this paper, the multi-dimensional adaptive weighted filter (AWF) is used to filter the hyperspectral image with a certain dimension which are reduced by the feature extraction method based on spectral dimension. Then, the filter results obtained on all scales are hierarchical fusion into a new image, and the hierarchical fusion framework is designed. These treatments make the essential spatial and spectral features in hyperspectral images extracted effectively, so the classification accuracy is improved. The principal component analysis (PCA) algorithm is integrated into the framework, and a hierarchical fusion-principal component analysis (HF-PCA) algorithm is proposed. This method not only reduces the redundancy between bands, but also weakens the internal differences of the samples and improves the classification accuracy of hyperspectral images. Experimental results on the Indian Pines and Salinas databases demonstrate that the classification accuracy obtained by the HF-PCA algorithm is significantly higher than that of other algorithms, even when the number of training samples is small, and the maximum value of the overall classification accuracy is 86.73% and 95.01%, respectively. The classification accuracy of hyperspectral images is improved effectively.