Neural Radiance Field Methods for Satellite Imagery of Polar and Arid Climates

Abstract

This research aims to explore modern neural rendering methods for 3D surface reconstruction using remote sensing imagery from satellites. Traditional multi-view stereo photogrammetric methods rely on rigid surface textures to accurately reconstruct a scene. However, many arid and polar regions across Earth exhibit uniform color and texture, 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 geometry and color radiance of a scene. In this research, we test NeRF variants Shadow NeRF (S-NeRF) and Satellite NeRF (Sat-NeRF), using satellite imagery of the Eureka Sand Dune located in Death Valley, California and Mount Doran located in Chugach, Alaska. Our goal is to assess the feasibility of NeRF methods at generating surface meshes of terrain from arid and polar regions on Earth, from multi-view satellite imagery. Furthermore, we present a comparative analysis on S-NeRF and Sat-NeRF to determine which method performs better with sparsely-textured geo-spatial data. The results from this research will ultimately reveal whether NeRFs serve as a plausible contender to current state-of-the-art for 3D scene reconstruction with remote sensing data, furthering fields in computer vision, photogrammetry, and geographic sciences.