Orbit-Augmented Geometric Consistency for Lunar Neural Elevation Model

Minseok Song, Suwan Lee, Seokju Lee

Keywords: digital elevation model, multi-view 3d reconstruction, remote sensing

Abstract: Reconstructing high-resolution lunar digital elevation models (DEMs) from sparse orbital imagery remains challenging due to limited viewpoint diversity and strong illumination variations. We present an extension of the Lunar Neural Elevation Model (LNEM) that incorporates orbit-aware data augmentation and multi-orbit geometric consistency, improving reconstruction robustness under sparse-view conditions. Our approach synthesizes additional training views via orbit-conditioned sampling while preserving pushbroom imaging geometry. To enforce cross-orbit consistency, we introduce a geometric constraint based on bidirectional reprojection between orbit-derived views, enabling stable multi-view alignment despite limited parallax. The proposed framework integrates orbit augmentation and geometric consistency into a unified training pipeline without requiring dense image coverage. Experiments on LROC NAC data show that the proposed constraint improves geometric consistency and reduces bias-corrected LOLA RMSE relative to the LNEM baseline without geometric consistency. We further report NAC DTM and SLDEM as contextual production DEM references, rather than as direct baselines to be surpassed. These results demonstrate the effectiveness of orbit-aware augmentation and multi-orbit geometric constraints for sparse-view pushbroom lunar reconstruction.

Submission Number: 10