使用深度对抗子空间聚类实现高光谱波段选择

高光谱图像（HSI）由数百个波段组成，波段之间的相关性强且具有较高的冗余度，导致出现维度灾难并且分类的复杂性很高。为此，使用深度对抗子空间聚类（DASC）网络进行高光谱的波段选择，并引入拉普拉斯正则化使网络更优，在保证分类精度的前提下降低分类的复杂度。该网络通过在编码器和解码器中引入自表达层来模仿传统子空间聚类的“自表达”属性，充分运用光谱信息和非线性特征转换得到波段之间的相互关系，解决传统波段选择方法无法同时考虑光谱和空间信息的问题。同时，引入对抗学习来监督自编码器的样本表示和子空间聚类，使得子空间聚类具有更好的自表达性能。为了使网络性能更优，加入拉普拉斯正则化来考虑反映图像几何信息的局部流形结构。实验在两个公开的高光谱数据集上进行，所提出的方法和几种主流的波段选择方法进行对比的结果表明，DASC方法在分类精度上优于对比方法，其选出的波段子集可以满足应用需求。

Hyperspectral band selection based on deep adversarial subspace clustering

HyperSpectral Image （HSI） consists of hundreds of bands with strong intra-band correlations between bands and high redundancy， resulting in dimensional disaster and increased classification complexity. Therefore， a Deep Adversarial Subspace Clustering （DASC） method was used for hyperspectral band selection， and Laplacian regularization was introduced to make the network performance more robust， which reduces the classification complexity under the premise of ensuring classification accuracy. A self-expressive layer was introduced between the encoder and the decoder to imitate the “self-expression” attribute of traditional subspace clustering， making full use of the spectral information and nonlinear feature transformation to obtain the relationships between the bands， and solving the problem that traditional band selection methods cannot consider spectral-spatial information simultaneously. At the same time， adversarial learning was introduced to supervise the sample representation of the auto-encoder and subspace clustering， so that the subspace clustering has better self-expression performance. In order to make the network performance more robust， Laplacian regularization was added to consider the manifold structure reflecting geometric information. Experimental results on two public hyperspectral datasets show that compared with several mainstream band selection methods， DASC method has higher accuracy， and the selected band subset of the method can satisfy application requirements.