卷积神经网络因为其可以自动提取低阶特征并组合在高维空间中，以此提高了分类性能，使得其广泛的应用在高光谱图像的分类研究中，并取得了非常重大的进展，无论是仅使用光谱特征还是综合了空谱特征进行联合分类，相较于早期的使用图像降维和手工提取特征的机器学习方法来说，有着十分长足的进步。然而目前应用在高光谱图像分类的卷积核因为其十分规则的形状，使得在对光谱特征亦或是空谱特征进行提取时，会忽略空间像素点之间的内在联系，这会导致分类的结果中出现不是很好的轮廓，以及非常小块的虚假预测，因此如何有效的综合利用高光谱图像的空间信息和光谱信息来提升最终的分类精度，并减少小区域的虚假预测和不良好的预测边界便成为了一个亟需解决的问题。同时高光谱数据的标签需要专业人员目视以及实地勘察才可以确定下来，耗费人力物力，而传统卷积神经网络在小样本下会出现性能下降的现象，导致分类精度降低，那么如何在利用少量的标签数据来训练网络的同时网络模型的精度还可以维持在一个较高水平便成为另一个需要解决的问题。
最近几年出现的图神经网络因为其灵活的卷积方式，可以充分的利用高光谱图像丰富的空间和光谱信息。因此针对上述提到的两个问题：（1）规则的卷积方式会导致缺少像素之间内在的信息；（2）小样本问题导致分类精度的降低。本文以提高高光谱图像分类精度为目标，并减少对标记样本的依赖，针对传统卷积神经网络因为规则的卷积方式难以有效提取空间信息的问题，提出基于图神经网络的高光谱图像分类算法，通过图所固有的点和边的特点，融合高光谱图像的空间信息和光谱信息，同时充分利用未标记样本的特征参与训练，提高分类模型的效果；而高光谱图像数据量往往比较大，构建较大的图数据会导致模型训练难的问题，针对这一问题，本文提出符合高光谱图像特点的超像素分割算法，通过综合计算像素点之间的光谱距离和空间距离将相似特征的相邻像素聚合到一起，组成超像素，充分利用未标记样本的信息，以减少对标记样本的依赖。主要创新有（1）将图网络引入到高光谱图像分类的研究中；（2）引入超像素分割算法处理高光谱图像来得到图数据。具体的研究内容如下：
（1）提出一种新的高光谱遥感图像分类方法，即基于图神经网络的HSGACN模型，该方法由一层有三个抽头的图注意力层和一层图卷积层组成，根据高光谱图像中像素之间的关系得到三种不同领域范围的邻接矩阵，通过计算交叉熵损失函数并进行反向传播来对网络进行训练。
（2）针对高光谱数据以像素为单位所构成的图数据过大的问题，提出基于SLIC算法的高光谱图像超像素分割算法，通过计算像素点之间的空间距离和光谱距离，并平衡权重，迭代的更新超像素聚类中心和范围边界，在新的聚类中心和旧的聚类中心之间的误差小于一定范围时停止迭代，最终得到一个由超像素构成的高光谱图像数据，减少图数据的大小和HSGACN的计算量。
本文提出的基于图神经网络的HSGACN网络模型对构建的图数据进行训练，可以利用图神经网络的特征传播特性改变未标记样本的特征，使得相同类别的超像素的特征更为相似，利用图注意力网络可以改变边的权重的特性，来逐渐优化图数据的结构，利用图卷积网络可以对全局运算的特性，充分利用图数据的特征和结构信息。本文为验证方法的有效性，在三个公开数据集Indian pines、Pavia University和Kennedy Space Center上实验，在每个类别只有30个标记样本的情况下分别达到91.84%、95.69%和98.42%的精度。

The convolutional neural network can automatically extract low-level features and combine them in high-dimensional space to improve the classification performance, making it widely used in the classification research of hyperspectral images, and has made very significant progress, regardless of Compared to the early machine learning methods that use image dimensionality reduction and manual feature extraction, whether they use only spectral features or integrate spatial spectrum features for joint classification has made great progress. However, the convolution kernel currently used in hyperspectral image classification because of its very regular shape makes it possible to ignore the internal relationship between spatial pixels when extracting spectral features or spatial spectral features, which will cause classification The results are not very good contours, and very small blocks of false predictions, so how to effectively use the spatial information and spectral information of hyperspectral images to improve the final classification accuracy, and reduce false predictions and small bad areas Predicting the boundary becomes an urgent problem. At the same time, the labels of hyperspectral data can be determined by professional visual inspection and field survey, which consumes manpower and material resources, and the traditional convolutional neural network will have a performance degradation phenomenon under a small sample, resulting in reduced classification accuracy, so how to use a small amount To train the network with the label data, the accuracy of the network model can also be maintained at a high level, which becomes another problem to be solved.
Graph neural networks, which have appeared in recent years, can make full use of the rich spatial and spectral information of hyperspectral images because of their flexible convolution methods. Therefore, for the two problems mentioned above: (1) the regular convolution method will lead to the lack of inherent information between pixels; (2) the problem of small samples leads to a reduction in classification accuracy. In this paper, we aim to improve the classification accuracy of hyperspectral images and reduce the dependence on labeled samples. Aiming at the problem that traditional convolutional neural networks are difficult to effectively extract spatial information because of the regular convolution method, a hyperspectral image classification algorithm based on graph neural network is proposed. Through the inherent point and edge characteristics of the graph, the spatial information and spectral information of the hyperspectral image are fused, and the characteristics of the unlabeled samples are fully utilized to participate in the training to improve the effect of the classification model; while the amount of data in the hyperspectral image is often relatively large, The construction of larger graph data will lead to the problem of difficulty in model training. For this problem, this paper proposes a superpixel segmentation algorithm that conforms to the characteristics of hyperspectral images. By comprehensively calculating the spectral distance and spatial distance between pixels, the phases of similar features are compared. Adjacent pixels are grouped together to form superpixels, making full use of the information of unlabeled samples to reduce dependence on labeled samples. The main innovations are (1) the introduction of graph networks into the study of hyperspectral image classification; (2) the introduction of superpixel segmentation algorithms to process hyperspectral images to obtain graph data. The specific research content is as follows:
(1) A new hyperspectral remote sensing image classification method is proposed, namely the HSGACN model based on graph neural network. This method consists of a three-tap graph attention layer and a graph convolution layer. According to the hyperspectral image The relationship between the pixels in the three kinds of adjacency matrices in different domains is obtained, and the network is trained by calculating the cross-entropy loss function and performing back propagation.
(2) Aiming at the problem that the hyperspectral data is composed of pixels whose data is too large, a hyperspectral image superpixel segmentation algorithm based on SLIC algorithm is proposed, by calculating the spatial distance and spectral distance between pixels, and balancing the weights , Iteratively update the superpixel clustering center and range boundary, stop iterating when the error between the new clustering center and the old clustering center is less than a certain range, and finally get a hyperspectral image data composed of superpixels, reducing The size of the graph data and the amount of calculation of HSGACN.
The HSGACN network model based on the graph neural network proposed in this paper trains the constructed graph data. The feature propagation characteristics of the graph neural network can be used to change the characteristics of unlabeled samples, making the characteristics of superpixels of the same category more similar. Force networks can change the characteristics of edge weights to gradually optimize the structure of graph data. Using graph convolution networks can make full use of the characteristics and structural information of graph data for the characteristics of global operations. In order to verify the effectiveness of the method, this paper tested on three public data sets, Indian pines, Pavia University and Kennedy Space Center, and achieved accuracy of 91.84%, 95.69% and 98.42% with only 30 labeled samples in each category.