高光谱图像分类方法可以同时利用空间特征和光谱特征对地物进行精细分类，已成为遥感领域的基础应用技术和重要研究方向。传统高光谱图像分类以高光谱图像降维和光谱特征与空间特征提取为核心内容，积累了大量研究成果。而近年来，深度学习因其高效的特征学习能力而对实现精细高光谱图像分类具有重要价值，迅速成为高光谱遥感分析领域的研究热点。但是，基于深度学习的高光谱图像分类仍面临如下问题：1）高光谱图像数据维度高、冗余信息多，易降低分类性能、增加计算负荷，需要在分类之前选择合适的波段组合进行数据降维，而传统的有监督波段选择方法在所选波段发生变化时通常需要重复训练模型，采用深度模型时计算压力更大；2）高光谱图像标记样本获取困难、数量较少，易导致分类模型过度拟合，如何在少量标记样本空间下进行有效的特征学习；3）高光谱图像具有空谱合一的特点，同时具备空间与光谱特征，如何有效整合空谱特征以提高最终的分类效果。 针对以上问题，本文以高光谱遥感图像地物分类为目标，以深度学习为基础，结合高光谱图像维度高、标记样本少和空谱合一等特点，分别从波段选择、半监督光谱分类和空谱联合分类三个方面开展高光谱图像分类的深入研究，提出了基于深度卷积神经网络的波段选择方法、基于光谱角损失函数的生成对抗网络高光谱图像半监督分类模型和基于交互引导图滤波与深度学习的高光谱图像空谱联合分类方法等一系列新模型和新方法。本论文的主要创新点如下： （1）针对高光谱图像数据维度高、冗余信息多带来的分类性能降低与计算负荷过高等问题，本文提出了一种基于深度卷积神经网络的波段选择算法在高光谱图像分类之前进行数据降维。传统的监督方式波段选择模型在波段组合的内部波段数量和组成波段发生变化时，往往需要重新训练模型，计算负荷较高；再加上深度模型本身的训练时间较长，直接采用深度学习模型进行波段选择的计算压力更大。为解决这一问题，本文结合高光谱图像一维光谱特征，设计了一维深度卷积神经网络作为波段组合评价模型进行波段选择；提出一种非选择波段置零方法，可以实现模型一次训练、多次选择波段的效果，大幅降低了计算量。另一方面，针对随机选择波段组合效率低下的问题，本文提出了以距离密度引导的波段选择数量自适应配置方法，进一步提高波段选择的效率。最后，根据距离密度算法对传统的数据增强方法进行了改进。实验表明，相比传统的波段选择方法，本文提出的方法能在保证具有较高分类精度的要求下高效地选择波段组合。 （2）针对高光谱标记样本少、未标记样本多导致分类系统容易过拟合、不易训练等问题，本文提出了基于光谱角损失函数的生成对抗网络高光谱图像半监督学习模型。1）首先，针对传统的生成对抗网络模型无法直接应用一维光谱进行训练的问题，本文提出了一种基于一维生成对抗网络的高光谱图像半监督学习算法。通过生成器生成数据增强高光谱图像样本空间，利用判别器获取样本基本特征，生成对抗网络可以有效地提高少量标记样本空间下高光谱图像的分类效果。2）在此基础上，结合高光谱图像光谱特征，通过改进生成对抗网络的目标函数，提出一种基于光谱角距离损失函数的生成对抗网络模型，实现了高效、稳定的生成对抗网络训练过程。然后通过构建一个新的以多层卷积融合特征为输入的卷积神经网络，实现了高光谱图像半监督分类模型。最后实验表明，本文提出的方法充分利用了未标记和已标记样本的信息，相比传统机器学习方法和基于卷积神经网络光谱分类方法能够大幅提高光谱分类精度，并优于目前流行的一些方法。 （3）针对高光谱数据空谱合一的特点，为了有效利用空间与光谱特征，本文提出了基于交互引导图滤波与深度学习的高光谱图像空谱联合分类方法。首先，针对以往引导图滤波无法有效处理引导信息与目标信息结构不一致的问题，提出一种采用交互引导图滤波获取空间特征的方法。然后针对传统的基于引导图滤波的空谱联合分类方法采用单一主成分作为引导图导致的信息失真问题，提出一种基于波段距离分组主成分为引导图的空间特征提取框架。该方法根据高光谱图像波段距离对波段进行分组，分别提取每组波段子集的主成分作为当前波段子集的引导图对高光谱图像进行滤波。对滤波之后的高光谱图像建立深度卷积神经网络模型和生成对抗网络模型，实现了对高光谱图像的空谱联合分类。实验表明，相比传统方法以及最近流行的一些基于滤波的空谱联合分类方法，本文提出的基于交互引导图滤波的方法可以有效地提取高光谱图像的空谱联合特征，提高高光谱图像的分类精度。 （4）最后，我们将上述提出的基于深度学习的高光谱图像分类算法应用到新疆某地的高光谱图像地质体分类实践中。首先使用基于深度卷积神经网络的波段选择方法选择波段组合，然后应用基于生成对抗网络的半监督分类方法与基于交互引导图滤波的空谱联合分类方法进行地质体分类。实验表明，本文提出的基于深度学习的高光谱图像分类算法，在光谱分类和空谱联合分类上相比传统方法和部分最近流行的方法都具有很大的优势，可以显著地提高高光谱遥感图像地物分类的精度，为地质制图提供了一种高效的辅助方法。

Hyperspectral images (HSIs) classification methods become the basic application and important research field of remote sensing because they can classify ground targets finely using spectral features and spatial features. Traditional HSIs classification technology has accumulated a large number of research results. In recent years, the deep-learning-based application has gradually become a research hotspot in the field of hyperspectral remote sensing because of its great value for fine HSIs classification that its efficient representation learning ability brings. However, the current classification technology of HSIs still faces the following problems: 1) HSIs have high dimensional and redundant information, which is easy to reduce classification performance and increase computational load; Therefore, it is necessary to select the band set for data dimensionality reduction before classification. However, the traditional supervised band selection methods usually need to train the model repeatedly when the selected bands are changed, and the calculation load is larger when the deep model is used; 2) It is difficult to acquire HSIs labeled samples, and the number of the HSIs labeled samples is small, which is easy to cause over-fitting of the classification model; 3) HSIs have both spatial and spectral features, and how to effectively integrate the spatial and spectral features to improve the final classification effect. In order to overcome the above problems, taking the HSIs classification as the target, the in-depth researches for HSIs data band selection, semi-supervised spectral classification and spectral-spatial classification based on deep learning, which is combined with the characteristics of image-spectrum merging, are discussed in this dissertation. A series of new models and methods about the HSIs band selection method based on deep convolutional neural network (CNN), HSI semi-supervised classification model based on generative adversarial networks (GAN) with the spectral angle distance (SAD) loss function and HSIs spectral-spatial classification methods based on mutually guided image filter (muGIF) and deep learning are proposed separately in this dissertation. The main innovation points in this dissertation are presented as follows: (1) Aiming at the problems with poor generalization ability and high computational load brought by high dimensional and redundant information of HSIs, this dissertation proposes a new band selection method based on deep convolutional network. The traditional supervised band selection model often needs to be retrained when the number of bands or the bands themselves in the band component are changed, which leads to higher calculation load. Combined with the 1-D spectral features of HSIs, a 1-D deep CNN model is designed as a band combination evaluation model for band selection. A non-selected bands zero-filling method is proposed, which can realize the effect of one-time training and multi-time band selection of the model. On the other hand, in order to solve the problem of low efficiency of randomly band selection, this dissertation proposes the concept of distance density to reflect the features of the spectral curve objectively. Different numbers of bands can be selected in different regions according to the distance density, so that the selected band component can retain more spectral features as much as possible and can increase the efficiency of band selection. Finally, the traditional data augmentation method is improved according to the distance density algorithm. The experiments show that compared with the traditional band selection methods, the proposed method can efficiently select the band combination under the requirement of ensuring high classification accuracy. (2) As for the problems that the classification system is easy to over-fit and difficult to be trained caused by the few labeled samples and many unlabeled samples, this dissertation proposes semi-supervised learning methods of HSIs based on GAN. 1) Firstly, in order to solve the problem of the traditional GAN model can not be trained by the 1-D spectrum directly, this dissertation proposes a semi-supervised learning algorithm of HSIs based on 1-D GAN. Because the generator generates data can enhance the hyperspectral image samples, and the discriminator can obtain the base features of samples, the GAN can effectively improve the classification effect of HSIs in a small amount of labeled samples. 2) Secondly, combining with the spectral features of HSIs, this dissertation further proposes a GAN model based on the SAD loss function. HSIs have their own 1-D spectral features that distinguish them from other common images. However, current methods based on GAN do not consider these features well, resulting in some unstable training and slow speed problems when performing the classification task. In order to solve these problems, this dissertation improve the objective function of GAN using spectral angle distance loss function based on the spectral features of HSIs, which can achieve an efficient and stable GAN training process. On this basis, a semi-supervised learning model of HSIs with better classification effect is realized by constructing a new CNN with multi-convolution fusion features as input. Finally, the experimental results show that the proposed methods can use the information of unlabeled and labeled samples. Compared with traditional machine learning and CNN-based spectral classification methods, the spectral classification accuracy of the proposed methods can be greatly improved，and is better than several popular methods. (3) In view of the characteristics of hyperspectral data with both spatial and spectral features, in order to effectively utilize spatial and spectral features, this dissertation proposes a method for HSIs spatial feature extraction based on muGIF, and combined with deep CNN and GAN, the method can classify HSIs using spectral-spatial. Firstly, aiming at the problem that the previous guided image filtering (GIF) cannot effectively deal with the inconsistent information structure between the guided information and the target information, a method for extracting spatial features by using muGIF is proposed. Then, aiming at the problem that the information loss caused by a single principal component as a guided image in the traditional GIF-based spectral-spatial classification, a spatial feature extraction framework based on band-distance-grouped principal component is proposed. The method groups the bands according to the band distance, and extracts the principal components of each set of band subsets as the guide map of the current band subset to filter the HSIs. A deep CNN model and GAN model for the filtered HSIs are constructed and then trained by samples for HSIs spectral-spatial classification. Experiments show that compared with the traditional methods and several popular spectral-spatial HSIs classification methods based on filter, the proposed methods based on muGIF can effectively extract the spectral-spatial features, and improve the classification accuracy of HSIs. (4) Finally, we apply the proposed HSIs classification methods based on deep learning to the HSIs geologic body classification located in a certain area in Xinjiang. Firstly, the band component is selected by using the band selection method based on the deep CNN and the distance density, and then the spectral classification method based GAN and the spectral-spatial classification method based on muGIF are applied to the data classification of the study area. Experiments by applying the above methods to the study area data show that the proposed HSIs classification methods based on deep learning have great advantages over traditional methods and several popular methods in spectral classification and spectral-spatial classification, which can significantly improve the accuracy of classification of HSIs, and provides an efficient auxiliary method for geological mapping.