数字高程模型（Digital Elevation Model，DEM）是地球表面地形高程的数字表示，通常为栅格格式。高分辨率 DEM 可以提供更详细准确的地球表面信息，使其在地质学、水文学和农业等领域广泛应用。然而，通过硬件设备获取高分辨率 DEM 需要花费较高的成本，于是基于深度学习的方法在 DEM 超分辨率领域得到快速的发展。目前 DEM 超分辨率重建大多使用的是基于单图像超分辨率（Single Image Super-Resolution，SISR）方法，基于参考的图像超分辨率方法（Reference-Based Image Super-Resolution，RefSR）研究甚少。这是因为目前还没有公开的DEM 参考数据集，而通过人工制作 DEM 参考数据集是一项成本非常高的任务。在DEM 超分辨率重建任务中，地形表面复杂区域的细节信息较难恢复，这对基于生成对抗网络（Generative Adversarial Network，GAN）方法中的判别器提出更高要求，同时基于 GAN 的方法在训练过程中容易发生训练不稳定的问题。针对上述问题，本文的主要工作为：

（1）提出了一种基于自相似和 Transformer（Self-Similarity Transformer，SSTrans）的 DEM 超分辨率重建方法，利用地形自相似性的特点构造 DEM 参考数据集，解决了没有公开的 DEM 参考数据集从而无法开展 RefSR 工作的问题。 SSTrans 使用深度残差网络作为主干网络提取低分辨率 DEM 数据特征；使用残差特征提取模块以获取高质量的参考图像特征；使用特征融合模块通过注意力机制计算低分辨率 DEM 与参考 DEM 之间深层对应关系，将获取到的参考信息融合到低分辨率 DEM 数据中。

（2）提出了以 SSTrans 为生成器， U-Net 网络为判别器的基于Transformer 和生成对抗网络（Transformer U-Net GAN，TUGAN）的 DEM 超分辨率重建方法，解决了基于 GAN 的 DEM 超分辨率方法对判别器更高的要求和容易发生训练不稳定的问题。TUGAN 的判别器由编码器和解码器组成，通过在编码器和解码器之间加入跳跃连接，能够更好地向生成器反馈全局与局部信息，进一步提升网络精准分割细节信息的能力，实现高程点级别的误差传输。同时，在判别器中加入谱归一化，在损失函数中使用 WGAN-GP 以实现梯度惩罚，达到模型训练稳定的目的。

实验结果表明，SSTrans 方法在四个实验区域均重建出高质量的 DEM 数据和 DEM 地形属性，尤其在地形表面复杂的区域中，重建质量远超其他方法。这是因为参考 DEM 数据提供了丰富的细节信息，面对复杂的区域仍能维持高质量的重建效果。TUGAN 为 SSTrans 的改进方法，通过判别器向生成器反馈全局与局部信息，引导生成器生成更高质量的 DEM 数据。TUGAN 方法的重建质量在 SSTrans 的基础上取得进一步提升，DEM 数据和 DEM 地形属性的重建结果均达到所有方法中最好的效果，在视觉上也恢复了比 SSTrans 更丰富的细节信息。

Digital Elevation Model(DEM) are digital representations of the Earth's surface topography, typically in raster format. High-resolution DEMs provide more detailed and accurate information about the Earth's surface, making them widely used in geology, hydrology, and agriculture. However, obtaining high-resolution DEMs through hardware equipment is costly, so deep learning methods have rapidly developed in the field of DEM super-resolution. Currently, most DEM super-resolution reconstruction methods are based on Single Image Super-Resolution (SISR) techniques, while research on Reference-Based Image Super-Resolution (RefSR) methods is scarce. This is because there are no publicly available DEM reference datasets, and creating DEM reference datasets manually is a very expensive task. In the DEM super-resolution reconstruction task, it is difficult to recover the detailed information of complex terrain surface areas, which poses higher requirements for the discriminator in Generative Adversarial Network (GAN) methods. Furthermore, GAN-based methods are prone to training instability issues. To address these issues, this thesis presents two main contributions:

(1) A DEM super-resolution reconstruction method based on Self-Similarity and Transformer (SSTrans) is proposed, which utilizes the self-similarity characteristics of terrain to construct a DEM reference dataset, solving the problem of not having a publicly available DEM reference dataset for RefSR work. SSTrans uses a deep residual network as the backbone to extract low-resolution DEM data features; employs residual feature extraction modules to obtain high-quality reference image features; and uses a feature fusion module to calculate the deep correspondence between low-resolution DEMs and reference DEMs through an attention mechanism, integrating the obtained reference information into the low-resolution DEM data.

(2) A DEM super-resolution reconstruction method based on Transformer and Generative Adversarial Network (Transformer U-Net GAN, TUGAN) is proposed, with SSTrans as the generator and U-Net network as the discriminator. This addresses the higher requirements for the discriminator and the training instability issues in GAN-based DEM super-resolution methods. TUGAN's discriminator consists of an encoder and a decoder, and by adding skip connections between the encoder and decoder, it can better provide global and local information feedback to the generator, further enhancing the network's ability to precisely segment detailed information and achieve elevation point-level error transmission. Additionally, spectral normalization is added to the discriminator, and WGAN-GP is used in the loss function to implement gradient penalties, ensuring stable model training.

Experimental results show that the SSTrans method reconstructs high-quality DEM data and DEM terrain attributes in all four experimental areas, especially in areas with complex terrain surfaces, where the reconstruction quality far exceeds other methods. This is because the reference DEM data provides rich detail information, which can maintain high-quality reconstruction results even in complex areas. TUGAN is an improved method of SSTrans, with the discriminator providing global and local information feedback to the generator, guiding it to generate higher-quality DEM data. The reconstruction quality of the TUGAN method further improves upon SSTrans, achieving the best results among all methods in terms of DEM data and DEM terrain attribute reconstruction, and visually recovering richer detail information compared to SSTrans.