高光谱遥感图像是典型的三维立方体数据，其包含数十甚至上千个光谱波段，不仅可以反映地物的空间特征，而且丰富的光谱信息可以反映地物内在的属性特征，在农业、环境、海洋、气象、地质、军事等多个领域得到了广泛的应用。高光谱遥感图像分类是高光谱图像进行分析和研究的基础，一直以来是遥感领域的研究热点。尽管基于深度学习的方法在此领域取得了突破性进展，但是由于高光谱遥感图像数据量大、波段冗余、特征维度高、标记样本难以获取等原因，分类中存在边界像素易分错、模型训练耗时、标记样本不足等问题，高光谱遥感图像分类仍然面临着严峻的挑战。
      近年来，注意力机制作为深度学习的重要技术之一，可以自动筛选输入数据的关键特征，抑制不重要的特征，得到越来越多国内外研究者的关注。因此，本文根据高光谱遥感数据图谱合一的特点，结合深度学习中的注意力机制，针对高光谱遥感图像分类中边界像素易分错、模型训练耗时、标记样本不足等问题展开研究。本文的主要研究内容、成果和创新如下：
      (1) 针对高光谱遥感图像中边界像素干扰信息多、易分错的问题，提出了一种基于空谱联合中心注意力网络的高光谱遥感图像分类方法。首先，设计了一个自动打分函数，该函数基于输入特征和目标特征之间的差异，计算所有空谱联合特征与目标特征之间的相关性得分，得分结果与相关性大小成正比；其次，提出了中心注意力模块，该模块基于输入特征的相关性得分，生成对应特征的注意力系数，得分越高系数越大，得分越低系数越小，并将输入特征与对应的注意力系数加权求和，生成新的空谱联合特征。由于新生成的空谱联合特征与目标更相关，这些特征更具区分性和判别性。实验结果表明，提出的空谱联合中心注意力网络能够提取与目标更相关的空谱联合特征，能有效减少边界像素的错分，提升了高光谱遥感图像分类算法的性能。
      (2) 针对高光谱遥感图像特征维度高、模型训练耗时的问题，提出了一种基于快速空谱联合中心注意力网络的高光谱遥感图像分类方法。首先，设计了一种光谱联合降维的方法，该方法有效结合主成分分析和线性判别分析，综合利用所有数据的主成分信息和已知样本的标签信息，实现高光谱遥感数据的特征降维；其次，提出了卷积前馈模块，通过采用二维卷积和特征前馈层替代三维卷积，分离提取样本的空间特征和光谱特征，降低特征图的维度。由于输入数据维度的降低和特征图计算量的减少，模型的训练效率得到提升。实验结果表明，基于快速空谱联合中心注意力网络的分类方法不仅可以提升模型的训练效率，而且可以兼顾模型的分类性能。
      (3) 针对高光谱遥感图像分类中标记样本难以获取、标记样本不足的问题，提出了一种基于半监督空谱联合图注意力网络的高光谱遥感图像分类方法。首先，设计了一种高光谱遥感图像超像素分割的方法，本文选取图像的前三个主成分数据进行局部区域聚类，将特征相似且距离相近的多个像素分割成超像素块；其次，基于超像素块和空间邻域关系构建节点（样本）的邻接矩阵，采用图注意力网络自动学习节点间的连接权重，进行特征提取和特征聚合。实验结果表明，在标记样本有限的情况下，半监督的空谱联合图注意力网络能够有效利用标记样本和未标记样本的空间信息和光谱信息，提高分类性能。
      (4) 为验证本文提出的三种高光谱遥感图像分类方法在实际场景中的有效性，本文选择新疆某地高光谱遥感数据作为研究对象，针对区域内十三类岩性进行分类。实验结果表明，相比现有的基于深度学习的分类算法，本文提出的方法具有明显的优势，可以高效快速地完成不同岩性的分类，为矿产勘查和地质填图提供了强有力的技术支持。
      综上，本文针对高光谱遥感图像分类，结合深度学习中的注意力机制，提出了空谱联合中心注意力网络、快速的空谱联合中心注意力网络和半监督的空谱联合图注意力网络等多种分类方法，并在新疆某地的实际场景中进行了不同岩性分类的验证。本文的研究一方面将为高光谱遥感图像的分类提供新思路和新方法，满足高光谱遥感图像数据分析和处理的现实需求；另一方面也将丰富深度学习领域的相关理论。

      Hyperspectral remote sensing images are typical three-dimensional (3D) cube data, containing dozens or thousands of spectral bands, which can reflect the spatial characteristics and the internal attribute characteristics of objects. They are used in many fields such as meteorology, agriculture, geology, environment and military. The classification of hyperspectral remote sensing images is the basis for the analysis and research of hyperspectral data, and is a research hotspot all along in the remote sensing field. Although the methods based on deep learning have made a breakthrough in the remote sensing field, there are still some problems in the classification such as boundary pixels being easily misclassified, time-consuming on model training and insufficient labeled samples, due to the large amount of hyperspectral data, redundant bands, high feature dimension, and difficulty in collecting labeled samples. Therefore, the classification of hyperspectral remote sensing images still faces severe challenges.
      Recently, as an important technology in deep learning, attention mechanism can automatically screen key features of input data and suppress unimportant features, which increasingly more attention from domestic and foreign researchers. Therefore, this study combines the characteristics of hyperspectral remote sensing images with the attention mechanism in deep learning, and takes an in-depth research on the problems concerning the classification of hyperspectral remote sensing images, such as boundary pixels being easily misclassified, time-consuming on model training, and insufficient labeled samples. The research contents, achievements and innovations of this study are mainly as follows:
      (1) This study proposes a classification method on the basis of spatial-spectral center attention network with reference to the problems that boundary pixels have much interference information and are easy to be misclassified in hyperspectral remote sensing images. First, an automatic scoring function is designed to calculate the correlation score between all the spatial-spectral features and the target features based on the difference that is mechanized from the difference between the input features and the target features. The scoring outcomes are in proportion to the correlation scores. Second, the center attention module, deriving from the correlation scores of the input features, is proposed to generate the attention coefficients with corresponding features and obtain new spatial-spectral features. Since the newly generated spatial-spectral features are more relevant to the target, these features are more discriminative and recognizable. The experimental results show that the spatial-spectral center attention network can extract spatial-spectral features that are more relevant to the target, effectively reduce the misclassification of boundary pixels, and improve the performance of hyperspectral remote sensing image classification methods.
      (2) This study proposes a classification method of hyperspectral remote sensing images on the basis of spatial-spectral center attention network with reference to the problems of high feature dimension of hyperspectral remote sensing images and time-consuming on model training. First, a dimension reduction method for spectrum is designed. It effectively combines principal component analysis and linear discriminant analysis, and utilizes the principal component information of all data and the label information of training samples to achieve feature dimension reduction. Secondly, a convolution feedforward module is proposed to reduce the dimension of the feature maps by extracting the spatial features and the spectral features of samples separately by 2D convolution and feedforward layers. Because of the reduction of the input data dimension and the reduction of feature computation, the training efficiency of the model could be improved. The experimental results show that the classification method based on the fast spatial-spectral center attention network can not only improve the training efficiency of the model, but also reconcile the classification performance of the model.
      (3) This study proposes a method based on semi-supervised spatial-spectral graph attention network with reference to the problems of difficulty in collecting labeled samples and insufficient labeled samples in hyperspectral remote sensing images. First, an algorithm for superpixel segmentation of hyperspectral remote sensing images is designed to perform local area clustering by using the first three principal component data of the images, and to divide some pixels with similar features and adjacent distances into superpixel blocks. Secondly, the method constructs the adjacency matrix of nodes (samples) based on superpixel blocks and neighborhood relationships, and uses graph attention network to automatically learn the connection weights among the nodes for feature extraction and feature aggregation. The experimental results show that in the case of limited labeled samples, the semi-supervised spatial-spectral graph attention network can effectively utilize the spatial and spectral information of labeled and unlabeled samples for feature extraction and feature aggregation, and improving the classification performance.
      (4) In order to verify the three classification methods with regard to hyperspectral remote sensing images in the actual scene, this study selectes the hyperspectral remote sensing data in Xinjiang as the study area, and classifies thirteen types of lithology in the area. The experimental results show that the method proposed has obvious advantages with the comparison of the excellent classification algorithms existing in deep learning, can complete the task of lithologic classification efficiently and quickly, and provide strong technical support for mineral exploration and geological mapping.
      To sum up, this study combines the classification of hyperspectral remote sensing images with the attention mechanism, proposes several models consisting of a spatial-spectral center attention network, a fast spatial-spectral center attention network, and a semi-supervised spatial-spectral graph attention network, and carries out the lithologic classification in an actual scene in Xinjiang. The research in this thesis, on the one hand, provides new ideas and new methods for the classification of hyperspectral remote sensing images to meet the practical needs of hyperspectral remote sensing image analysis and processing; on the other hand, enriches the relevant theories in the field of deep learning.