EOGS++: Earth Observation Gaussian Splatting with Internal Camera Refinement and Direct Panchromatic Rendering

Pierrick Bournez1, Luca Savant Aira 2, Thibaud Ehret3, Gabriele Facciolo1,
1Universite Paris-Saclay, CNRS, ENS Paris-Saclay, Centre Borelli, 91190, Gif-sur-Yvette, France 2 Politecnico di Torino, Corso Duca degli Abruzzi, 24, 10129 Torino TO, Italia 3AMIAD, Pôle Recherche, France
Paper arXiv Code Data
Training panchromatic image
Rendered panchromatic image
GT Lidar DSM
Rendered DSM
Qualitative result of the EOGS++ pipeline.

Abstract

Recently, 3D Gaussian Splatting has emerged as an efficient alternative to NeRF for Earth observation, delivering competitive reconstruction quality with substantially reduced training times. In this work, we introduce EOGS++, an enhanced extension of the Earth Observation Gaussian Splatting framework designed to operate directly on raw high-resolution panchromatic imagery, thereby removing the need for external preprocessing.

By leveraging optical flow, we embed bundle adjustment directly into the training process, avoiding dependence on external optimization tools while improving camera pose estimation. We further incorporate several refinements, such as early stopping and TSDF-based post-processing—that jointly yield sharper reconstructions and improved geometric accuracy.

Experiments on the IARPA 2016 and DFC2019 datasets demonstrate that EOGS++ achieves state-of-the-art performance in both reconstruction quality and computational efficiency. It surpasses the original EOGS and NeRF-based baselines while preserving the speed advantages of Gaussian Splatting. Notably, our method reduces the mean MAE on building structures from 1.33 m to 1.19 m, highlighting its effectiveness for satellite-based 3D reconstruction.

EOGS
EOGS ++
compared DSM between EOGS and EOGS ++ on scene IARPA 002

Method

Method Overview
Schematic overview of the proposed training pipeline. From the 3D Gaussian primitives, an image is rendered. The rendered image is then aligned with the training observation using an optical flow algorithm, after which the model is trained accordingly.

Visual Results

IARPA 002 IARPA 003
None
Learn wv
Optical flow
External B.A.
GT Lidar DSM

Qualitative comparison of the different methods for correcting camera-pointing errors on the IARPA 002 and IARPA 003 scenes.

Video comparison

EOGS
EOGS ++
.

Related links

Several recent works and repositories explore complementary approaches to satellite-based 3D reconstruction.

EoNerf proposes an Earth-observation–oriented adaptation of NeRF, focusing on improved geometric consistency across multi-view satellite imagery.

EOGS introduces a Gaussian Splatting framework tailored to satellite data, providing efficient reconstruction pipelines for large-scale scenes.

BibTeX

@article{bournez2025eogsearthobservationgaussian,
  title={EOGS++: Earth Observation Gaussian Splatting with Internal Camera Refinement and Direct Panchromatic Rendering},
  author={Bournez, Pierrick and Savant Aira, Luca and Ehret, Thibaud and Facciolo, Gabriele},
  journal={arXiv preprint arXiv:2511.16542},
  year={2025},
  url={https://arxiv.org/abs/2511.16542}
}