Photogrammetry for Unconstrained Optical Satellite Imagery With Combined Neural Radiance Fields

We propose a novel generic method to address the challenge of handling unconstrained multi-view optical satellite photogrammetry under time-varying conditions of illumination and reflection. For one thing, we innovatively represent the surface radiance and albedo produced by extensive lights with continuous radiance fields based on the radiometry principle and then combine the static and transient components for satellite photogrammetry. For another, a novel self-supervised mechanism is introduced to optimize the learning process which leverages dark regions accentuation, transient and static composition, as well as occlusion and shadow suppression. We evaluate the proposed framework via real-world multi-date WorldView-3 images and demonstrate that our proposed model consistently outperforms the existing state-of-the-art methods.