Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ShadowGS: Shadow-Aware 3D Gaussian Splatting for Satellite Imagery

Feng Luo, Hongbo Pan, Xiang Yang, Baoyu Jiang, Fengqing Liu, Tao Huang

TL;DR

ShadowGS tackles shadow variability in multi-temporal satellite imagery to enable accurate 3D reconstruction. It extends 3D Gaussian Splatting with a remote-sensing physics-based rendering equation, per-Gaussian SH-based appearance, a global skylight SH, and ray-marched, geometry-aware shadows, alongside depth–normal and shadow-consistency constraints. It introduces a shadow map prior for sparse views and optimizes a multi-term loss to separate geometry, albedo, and illumination. Evaluations on DFC2019 and IARPA datasets show improved shadow disentanglement, DSM accuracy, and novel-view synthesis with minutes of training per area, across RGB, pan-sharpened, and sparse-view inputs.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a novel paradigm for 3D reconstruction from satellite imagery. However, in multi-temporal satellite images, prevalent shadows exhibit significant inconsistencies due to varying illumination conditions. To address this, we propose ShadowGS, a novel framework based on 3DGS. It leverages a physics-based rendering equation from remote sensing, combined with an efficient ray marching technique, to precisely model geometrically consistent shadows while maintaining efficient rendering. Additionally, it effectively disentangles different illumination components and apparent attributes in the scene. Furthermore, we introduce a shadow consistency constraint that significantly enhances the geometric accuracy of 3D reconstruction. We also incorporate a novel shadow map prior to improve performance with sparse-view inputs. Extensive experiments demonstrate that ShadowGS outperforms current state-of-the-art methods in shadow decoupling accuracy, 3D reconstruction precision, and novel view synthesis quality, with only a few minutes of training. ShadowGS exhibits robust performance across various settings, including RGB, pansharpened, and sparse-view satellite inputs.

ShadowGS: Shadow-Aware 3D Gaussian Splatting for Satellite Imagery

TL;DR

ShadowGS tackles shadow variability in multi-temporal satellite imagery to enable accurate 3D reconstruction. It extends 3D Gaussian Splatting with a remote-sensing physics-based rendering equation, per-Gaussian SH-based appearance, a global skylight SH, and ray-marched, geometry-aware shadows, alongside depth–normal and shadow-consistency constraints. It introduces a shadow map prior for sparse views and optimizes a multi-term loss to separate geometry, albedo, and illumination. Evaluations on DFC2019 and IARPA datasets show improved shadow disentanglement, DSM accuracy, and novel-view synthesis with minutes of training per area, across RGB, pan-sharpened, and sparse-view inputs.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a novel paradigm for 3D reconstruction from satellite imagery. However, in multi-temporal satellite images, prevalent shadows exhibit significant inconsistencies due to varying illumination conditions. To address this, we propose ShadowGS, a novel framework based on 3DGS. It leverages a physics-based rendering equation from remote sensing, combined with an efficient ray marching technique, to precisely model geometrically consistent shadows while maintaining efficient rendering. Additionally, it effectively disentangles different illumination components and apparent attributes in the scene. Furthermore, we introduce a shadow consistency constraint that significantly enhances the geometric accuracy of 3D reconstruction. We also incorporate a novel shadow map prior to improve performance with sparse-view inputs. Extensive experiments demonstrate that ShadowGS outperforms current state-of-the-art methods in shadow decoupling accuracy, 3D reconstruction precision, and novel view synthesis quality, with only a few minutes of training. ShadowGS exhibits robust performance across various settings, including RGB, pansharpened, and sparse-view satellite inputs.
Paper Structure (16 sections, 19 equations, 22 figures, 5 tables)

This paper contains 16 sections, 19 equations, 22 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Shadow in Remote Sensing
  4. NeRF for Satellite Images
  5. 3DGS for Satellite Images
  6. Method
  7. Camera Model and Geometry
  8. Physics-based Rendering Equation
  9. Shadow Consistency Constraint
  10. Shadow Prior for Sparse View
  11. Training Strategy and Total Loss
  12. Experiment
  13. Multi-View Experimental Results
  14. Sparse-View Experimental Results
  15. Ablation Studies
  16. ...and 1 more sections

Figures (22)

  • Figure 1: ShadowGS reconstructs 3D geometry with consistent shadow modeling from multi-temporal satellite imagery. The top row displays reconstructed DSMs while the bottom row shows shadow decomposition results. Compared to EO-NeRF ref8 and EOGS ref11, our method produces superior reconstruction quality with sharper edges, richer details, smoother surfaces, and shadows that align precisely with scene geometry.
  • Figure 2: The overall pipeline of ShadowGS.
  • Figure 3: Shadow consistency constraint. Shadows are visible when the satellite's viewing direction differs from the solar direction (position A), but become self-occluded by the object's geometry when the two directions align (position B).
  • Figure 4: Geometric reconstruction visualization on the JAX 068 dataset. The fourth column shows the error map between each method's reconstructed DSM and the ground truth (GT), where red indicates overestimation and blue indicates underestimation of height values.
  • Figure 5: Albedo Decomposition and Shadow Modeling Results. (a) Input image; (b-c) Albedo and shadow maps decoupled by ShadowGS; (d-f) Shadow maps under different solar elevation angles; (g-i) Shadow maps under different solar azimuth angles.
  • ...and 17 more figures