Learning DEM Completion with Pretrained Spatial Representations

Digital Elevation Models (DEMs) are a fundamental representation of terrain surfaces and are widely used in geomorphological analysis and terrain modeling. Traditional interpolation-based techniques are commonly used to address missing or unobserved regions in DEMs. However, these methods can struggle to recover coherent terrain structures when the gaps are large or irregular.

In this work, we investigate DEM completion by adapting model representations pretrained on large-scale image datasets to elevation data, which provides an effective initialization for Transformer models without requiring training from scratch. A Transformer-based architecture is employed to capture long-range spatial dependencies and global terrain structure, which are essential for reconstructing missing elevation regions. Geomorphological constraints are incorporated into the learning process to guide the reconstruction toward structurally consistent and physically plausible terrain surfaces.

Experimental results on a subset of the swisstopo DEM dataset demonstrate that the proposed approach achieves a mean absolute error (MAE) of approximately 6 m and a root mean squared error (RMSE) of approximately 11 m on the validation. Given that the average per-sample local topographic amplitude is approximately 256 m, an MAE of 6 m corresponds to less than 3% of this scale. The proposed approach leads to more coherent and structurally consistent DEM reconstructions compared to traditional interpolation-based methods, particularly in regions with missing or sparsely sampled elevation data. While the framework is developed for DEM completion, similar modeling principles could potentially be explored for other spatial surface reconstruction problems involving incomplete geospatial data.