基于半监督稀疏流形嵌入的高光谱影像特征提取

摘要

Hyperspectral image contains the properties of much bands and high redundancy, and the research of hyperspectral image classification focuses on feature extraction. To overcome this problem, a Semi-Supervised Sparse Manifold Embedding (S3ME) algorithm is proposed in this paper. The S3ME method makes full use of labeled and unlabeled samples to adaptively reveal the similarity relationship between data with the sparse representation of tangent space. It constructs a semi-supervised similarity graph via the sparse coefficients and enhances the weight between labeled samples from the same class. In a low-dimensional embedding space, S3ME preserves the similarity of graph to minimize the sum of the weighted distance. Then, it obtains a projection matrix for feature extraction. S3ME not only reveals the sparse manifold structure of data but also enhances the compactness of the same class data, which can effectively extract the discriminating feature and improve the classification performance. The overall classification accuracies of the proposed S3ME method respectively reach 84.62%and 88.07%on the PaviaU and Salinas hyperspectral data sets, and the classification performance of land cover is improved compared with the traditional feature extraction methods.%高光谱影像具有波段数多、冗余度高的特点，因此特征提取成为高光谱影像分类的研究热点。针对此问题，该文提出一种半监督稀疏流形嵌入(S3ME)算法，该方法充分利用标记样本和无标记样本，通过基于切空间的稀疏流形表示来自适应地揭示数据间的相似关系，并利用稀疏系数构建一个半监督相似图。在此基础上，增加了图中同类标记样本的权重，然后在低维空间中保持图的相似关系不变，并最小化加权距离和，获得投影矩阵实现特征提取。S3ME方法不仅能揭示数据间的稀疏流形结构，而且增强了同类数据的集聚性，能有效提取出鉴别特征，改善分类效果。该文提出的S3ME方法在PaviaU和Salinas高光谱数据集上的总体分类精度分别达到84.62%和88.07%，相比传统特征提取方法提升了地物分类性能。