Files
Abstract
This research aims to explore modern neural rendering methods for 3D surface reconstruction using remote sensing imagery from Earth observation satellites. Traditional multi-view stereo (MVS) photogrammetric methods rely on rigid surface textures to accurately reconstruct a scene. However, polar regions across Earth exhibit uniform color and texture due to snow and ice coverage, making it difficult for traditional methods to create accurate surface models of the terrain. Neural Radiance Fields (NeRFs) overcome this problem by using neural volumetric rendering techniques to continuously learn the color and geometry of a scene. In this research, we test NeRF variants, Shadow NeRF (S-NeRF) and Satellite NeRF (Sat-NeRF), using WorldView-2 satellite imagery of a scene in Anchorage, Alaska, and Mount Doran, located in Chugach, Alaska. Our goal is to assess the feasibility and performance of NeRF methods in generating digital surface models (DSMs) of terrain from polar regions on Earth using multi-view satellite imagery. Furthermore, we present a comparative analysis of S-NeRF and Sat-NeRF using high-density point clouds derived from LiDAR ground-truth to determine which method performs better with sparsely-textured geo-spatial data. The results from this research will ultimately reveal whether NeRFs could serve as a plausible contender to the current state-of-the-art methods for accurate 3D scene reconstruction from satellite remote sensing data in sparse-textured terrains.