Neural Radiance Methods for Lunar Terrain Modeling

Recent advancements in scene representation, such as Neural Radiance Fields (NeRFs), provide continuous representations of objects and scenes using coordinate-based neural networks. These methods are effective for Digital Elevation Model (DEM) reconstruction but struggle with accurately modeling largely shadowed regions due to their lack of depth-supervision and shadow-awareness. To address this, we aim to develop Lunar Neural Radiance Methods (LunarNRM), a novel neural surface reconstruction method that incorporates shadow-aware and depth-aware techniques into a NeRF pipeline. LunarNRM will generate shadow-controlled DEMs of the lunar surface, allowing for the relighting of topographic depressions such as craters. By incorporating multi-sensor data from the Lunar Reconnaissance Orbiter (LRO) satellite, specifically optical data from the Narrow Angle Camera (NAC) and altimeter data from the Lunar Orbiter Laser Altimeter (LOLA), we aim to demonstrate that LunarNRM can reconstruct largely shadowed regions in the lunar South Pole, which are critical targets for the NASA Artemis campaign.