针对传统图嵌入方法仅采用单一图结构无法有效表征高维数据中复杂本征结构, 本文提出了一种半监督多图嵌入(SSMGE)方法, 并应用于高光谱影像特征提取。该方法首先利用标记样本的类内、类间近邻点来构建类内超图、类间超图、类内普通图、类间普通图, 然后通过无标记样本的近邻点和远离点构建无监督本征超图和惩罚超图, 并以多图协同方式来表征高维数据间的复杂几何关系, 实现鉴别特征提取。本文提出的SSMGE方法不仅能有效揭示数据点间超图和普通图的结构, 而且在低维嵌入空间中增强同类数据聚集性和非同类数据的远离性, 提取的鉴别特征可改善地物分类精度。在PaviaU和Urban高光谱数据集上进行了分类实验, 本文方法的总体分类精度分别可达到85.92%和79.74%。相比普通图嵌入和超图方法, 该算法明显提升了地物的分类性能。

Traditional graph embedding methods often use single graph structures for feature extraction (FE).However, these methods cannot effectively represent the complex intrinsic structuresof high-dimensional data. To address this problem, a Semi-Supervised Multi-Graph Embedding (SSMGE) algorithm was proposed for FE of Hyper Spectral Images (HSIs). First, the SSMGE method constructed one of each of intra-and inter-class hypergraphs and intra-and inter-class graphs through intra-and inter-class neighbors of labeled samples.In addition, it constructs an unsupervised intrinsic hypergraph and a penalty hypergraph using unlabeled samples. The fusion of graphs and hypergraphs can effectively characterize the complex relationships in high-dimensional data. The SSMGE method not only effectively reveals the intrinsic structure of an HSI by exploring the collaboration of graphs and hypergraphs but also enhances the discriminative ability of extracted features in low-dimensional embedding space.This enables improved classification performance of HSI data. Experimental results on the PaviaU and Urban hyperspectral datasets show that the overall accuracies of the proposed method reached 85.92% and 79.74%, respectively. The SSMGE method can significantly improve classification performance as compared with some state-of-the-art FE methods.