Digital Elevation Models (DEMs) are essential for comprehending the three-dimensional (3D) structure of the Earth’s surface. When the resolution of airborne or satellite sensors fail to meet the requirements for high-precision terrain observation, super-resolution methods offer an alternative solution, which infers the high-resolution DEMs from coarse-resolution DEMs. This study aims to create an unsupervised DEM super-resolution method that utilizes untrained neural networks, integrating both deep prior from network structure and terrain knowledge, namely UnTDIP. Specifically, the untrained neural network iteratively processes a single DEM, and the deep prior from network structure learns low-order statistical information from the DEM that is then applied to the super-resolution task. Furthermore, the terrain knowledge for topographic consistency is integrated into the iterative process and synergies with the deep prior to reconstructing the DEM with the correct 3D structure. Most neural network-based DEM super-resolution methods are data-driven, while the proposed method does not require any training datasets or pre-trained network parameters, making it broadly applicable. The UnTDIP method was evaluated in three study areas with different spatial resolutions of 5 m, 10 m, and 30 m. The method performs well in multiple scale factors and spatial resolutions, even surpassing the fully supervised DEM super-resolution method in some evaluation metrics. Moreover, the integration of terrain knowledge reduces the Mean Absolute Error (MAE) by 6.08 % to 34.65 %. Overall, the proposed method is well-suited for DEM data and can be adapted for super-resolution tasks for other geospatial data with the integration of domain knowledge.