2025 CSDMS meeting-136
From CSDMS
(if you haven't already)
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu
Browse abstracts
Constructing Digital Elevation Models from Single Synthetic Aperture Radar Images
Hamish Mitchell,
MIT Cambridge Massachusetts, United States. whamitch@mit.edu
Taylor Perron, MIT CAMBRIDGE Massachusetts, United States. perron@mit.edu
Numerous humanitarian and disaster relief missions require updated topography to provide time-critical support following events such as earthquakes, wildfires, tsunamis, or armed conflicts. Radar techniques are particularly advantageous over other methods (e.g., LiDAR) in these scenarios because they are insensitive to weather and lighting conditions, allowing data collection through clouds and smoke, or at night. Despite these advantages, current radar-based methods for generating topography face significant challenges related to data acquisition logistics, processing complexity, and the specialized expertise needed.
Inspired by recent advances in computer vision and monocular depth estimation, we present a novel approach to generate Digital Elevation Models (DEMs) from single Synthetic Aperture Radar (SAR) images using deep learning. Our method leverages a global dataset of open-source SAR-DEM image pairs to train multiple architectures, including Vision Transformers (ViTs) and fully convolutional networks. We evaluate various supervised and adversarial training strategies across a diverse range of Earth's landscapes. Our approach streamlines topographic reconstruction by working directly in ground coordinates and eliminating specialized pre-processing, making DEM generation more accessible. By utilizing open-source satellite radar data with a 6-day revisit time, our method enables topographic reconstruction at a significantly improved temporal resolution.
