High-resolution Digital Elevation Models (DEMs) are essential for precise geographic analysis. However, obtaining high-resolution DEMs in regions with dense vegetation, complex terrain, or satellite imagery voids presents substantial challenges. This study introduces a deep learning approach using three-dimensional (3D) terrain features combined with Conditional Generative Adversarial Networks (CGANs) to reconstruct DEMs. The 3D terrain features, such as valley and ridge lines, exhibit topographic relief patterns and provide constraints for CGANs to reconstruct DEMs. Experiments conducted in the Loess Plateau of Shaanxi confirmed the performance of the proposed method, demonstrating marked improvements in the accuracy of DEM reconstruction compared to models based on two-dimensional (2D) terrain features. The elevation accuracy of the reconstructed DEMs by the proposed method is 5.30 m, which is higher than that of the 2D terrain features method (18.90 m) by 71.96 %. Meanwhile, the proposed method shows a 15.78 % and 17.64 % improvement in elevation accuracy and slope accuracy, respectively, when reconstructing a 5 m high-resolution DEM from a 30 m low-resolution DEM. The proposed method can be flexibly used for reconstructing, repairing, and filling voids in DEM data.

中文翻译：

---

使用条件生成对抗网络从 3D 地形特征重建高分辨率 DEM


高分辨率数字高程模型 (DEM) 对于精确的地理分析至关重要。然而，在植被茂密、地形复杂或卫星图像空白的地区获得高分辨率 DEM 面临着巨大的挑战。本研究介绍了一种深度学习方法，使用三维 (3D) 地形特征与条件生成对抗网络 (CGAN) 相结合来重建 DEM。 3D 地形特征（例如山谷和山脊线）表现出地形起伏图案，并为 CGAN 重建 DEM 提供约束。在陕西黄土高原进行的实验证实了该方法的性能，与基于二维（2D）地形特征的模型相比，DEM重建的精度显着提高。该方法重建DEM的高程精度为5.30 m，比二维地形特征方法的高程精度（18.90 m）提高了71.96%。同时，当从 30 m 低分辨率 DEM 重建 5 m 高分辨率 DEM 时，该方法的高程精度和坡度精度分别提高了 15.78 % 和 17.64 %。所提出的方法可以灵活地用于DEM数据中的重建、修复和填充空隙。