基于孪生网络架构和图卷积的多尺度高光谱图像分类

Siamese Network Architecture and Graph Convolution-Based Multi-Scale Hyperspectral Image Classification

大量的工作表明,融合谱-空信息的高光谱图像分类方法可以获得令人较为满意的结果.如何充分利用高光谱图像中所包含的空间信息,并将其与谱特征进行恰当地融合,从而获得更理想的分类结果,是遥感领域一个挑战性的问题.文章提出了一种基于孪生网络架构和图卷积的多尺度高光谱图像分类方法.该方法首先将原始高光谱图像分为三幅相同大小但具有不同谱特征的图像,然后,分别对三幅图像采用不同的尺度进行超像素分割且进行合并.合并后的超像素不仅极大地缩减了图的规模,提高了计算效率,而且进一步增强了空间信息在分类中的作用.接下来,基于扩展的孪生网络架构,分别对三幅图像进行图卷积操作,提取三幅图像的主要谱特征.最后,采用注意力机制对提取到的谱特征进行融合,并输入到全连接网络进行分类.在Indian Pines和Salinas两个公开数据集上的实验结果和比较结果表明,所提出方法的分类完成性优于其他几个竞争性的方法.

Extensive work has shown that the hyperspectral image classification method based on the fusion of spectral-spatial information can obtain satisfactory results.It is a challenging problem in the field of remote sensing that how to make full use of the spatial information contained in hyperspectral images and properly integrate it with spectral features to obtain better classification results.In this paper,a multi-scale hyperspectral image classification method based on siamese network architecture and graph convolution is proposed.First,the original hyperspectral image is divided into three images of the same size but with different spectral features.Then,three images are segmented separately into superpixels using different scales and merged them.The merged superpixels not only greatly reduce the size of the graph and improve the computational efficiency,but also further enhance the role of spatial information in classification.Next,based on the extended siamese network architecture,the main spectral features of the three images are extracted using graph convolution,respectively.Finally,the extracted spectral features are fused using the attention mechanism and input into the fully connected network for classification.Experimental and comparative results on two public datasets,Indian Pines and Salinas,show that the proposed method performs better than several competing methods in classification completion.