基于光谱注意力图卷积网络的高光谱图像分类

近年来，图卷积网络因其特征聚合的机制，能够同时对单个节点以及近邻节点的特征进行表示，被广泛应用于高光谱图像的分类任务。然而，高光谱图像(HSI)中常存在波段冗余、同物异谱等问题，使得直接利用原始光谱特征构建的初始图可靠性不足，从而导致高光谱图像的分类精度低。为此，该文提出一种基于光谱注意力图卷积网络(SAGCN)的高光谱图像半监督分类方法。首先，利用注意力模块对光谱的局部与全局信息进行交互，以增加重要光谱的权重、减小冗余波段以及噪声波段的权重，从而实现光谱的自适应加权；然后，针对光谱加权处理后的高光谱图像，通过空间-光谱相似性度量构建更为准确的近邻矩阵；最后，通过图卷积对标记和无标记样本进行有效的特征聚合，并使用标记样本的聚合特征训练网络。在Indian Pines, Kennedy Space Center和Botswana 3个真实高光谱图像数据集上的实验结果验证了所提方法的有效性。

HyperSpectral Image Classification Based on Spectral Attention Graph Convolutional Network

In recent years, graph convolutional network has been widely used in hyperspectral image classification because of its feature aggregation mechanism, which can simultaneously represent the features of a single node and neighboring nodes. However, there are many problems in HyperSpectral Images(HSI), such as band redundancy and different spectrum of the same object, which results in the inadequate reliability of the initial graph constructed by directly using the original spectral features, thus leading to the low classification accuracy of hyperspectral images. Therefore, a semi-supervised classification method for hyperspectral images based on Spectral Attention Graph Convolutional Network (SAGCN) is proposed. Firstly, the attention module is used to interact with the local and global information of the spectrum, and realize the adaptive weighting of the spectrum. Then, for the hyperspectral images after spectral weighting, a more accurate nearest neighbor matrix is constructed by using spatial-spectral similarity. Finally, effective feature aggregation of labeled and unlabeled samples is carried out by graph convolution, and the network is trained with the features of labeled samples. Experimental results on three real hyperspectral image datasets including Indian Pines, Kennedy Space Center and Botswana demonstrate the effectiveness of the proposed method.