Reconstructing Digital Elevation Models From Single Synthetic Aperture Radar Images

Abstract Traditional radar‐based methods of generating topography involve complex dual‐antenna or formation‐flying satellite configurations, interferometric processing, and specialized expertise. We present an approach for reconstructing digital elevation models (DEMs) from single Synthetic Aperture Radar (SAR) images using deep learning. SAR‐to‐DEM inversion is an inherently ill‐posed problem: the geometric distortions that encode topography also cause multiple ground points to contribute to each radar pixel, creating a many‐to‐one mapping. To establish whether this inverse problem is learnable across diverse global terrain, we adapt architectures from computer vision and find that a Dense Prediction Transformer effectively interprets the geometric distortions and textural patterns in SAR imagery to recover multi‐scale topography. We evaluate three calibration approaches with varying ground truth requirements: global scaling (no ground truth), point calibration, and coarse DEM calibration. The latter shows particular promise by combining fine‐scale feature detection with regional accuracy from existing low‐resolution elevation models. Despite challenges in resolving the position of topographic boundaries, this approach reduces the processing and technical barriers associated with mapping topography and offers substantial advantages in accessibility, processing speed, and flexibility. It represents a step forward for generating topography in time‐critical humanitarian efforts, disaster response, and applications in data‐scarce regions.