关键词: 高光谱遥感影像, 分类, 特征提取, 图嵌入, 多流形学习

Abstract: Manifold learning method can find the low-dimensional manifold structures embedded in high-dimensional data.However,the traditional manifold learning algorithms assume that all samples lie on a single manifold,while the samples in different subsets may belong to different sub-manifolds.To solve the above problem,a new dimensionality reduction (DR) method termed supervised multi-manifold discriminant embedding (SMMDE) is proposed for classification of hyperspectral remote sensing image.At first,the proposed method explore the labels of HSI data to divide samples into different sub-manifolds.Based on the graph embedding framework,the intra-manifold and inter-manifold graphs are constructed to represent the multi-manifold structure of HSI data,and the intra-class aggregation and inter-class separation are enhanced by minimizing the intra-manifold distance and maximizing the inter-manifold distance simultaneously.Therefore,low-dimensional discriminant features are obtained to improve the performance of HSI classification.Experimental results on the PaviaU and KSC hyperspectral data sets show that the overall classification accuracies respectively reach 88.04% and 84.53% when 2% training samples are randomly selected for training.The proposed SMMDE method can effectively improve classification performance compared with many state-of-art DR algorithms.

Key words: hyperspectral remote sensing image, classification, feature extraction, graph embedding, multi-manifold learning