Martian DEM Void Filling With Orientation-Guided Diffusion Transformer and Topographic Constraints

Abstract

Mars exploration plays a crucial role in deep-space studies and requires high-quality digital elevation models (DEMs) for terrain analysis, landing site evaluation, and rover navigation. However, Martian DEMs frequently contain voids due to sparse observations and stereo matching errors, which significantly limit their usability. In this study, we propose a diffusion-based framework for filling Martian DEM voids. A diffusion transformer (DiT) is adopted to capture global terrain structures, while an orientation-encoding module introduces directional constraints from low-resolution DEMs to guide terrain reconstruction. In addition, a consistent sampling strategy is taken in the known region to preserve the observed terrain and improve the continuity of the boundaries during the reverse diffusion process. Experiments on HiRISE DEM datasets demonstrate that the proposed method effectively reconstructs complex Martian terrain. Compared with the generative adversarial network (GAN)-based method, the proposed method improves mean absolute error (MAE) from 11.737 to 4.858 and multiscale structural similarity index (MS-SSIM) from 0.648 to 0.923.