Earth Observation Gaussian Splatting Boosts 3D Reconstruction

December 1, 2025
|
Updated February 9, 2026
|

2 min read

Earth viewed from space, illuminated by sunlight, highlighting land and water features.

Earth Observation Gaussian Splatting (EOGS++) Enhances 3D Reconstruction

Earth Observation Gaussian Splatting (EOGS++) is transforming 3D reconstruction from satellite imagery, streamlining workflows and improving accuracy. Developed by a research team including Pierrick Bournez and Luca Savant Aira, this new method tackles the traditional headaches of satellite-based modeling. Previous workflows relied on complex pre-processing steps like pansharpening and external optimization tools, which often slowed down the pipeline. EOGS++ changes the game by ingesting raw, high-resolution panchromatic data directly. It skips the tedious prep work and streamlines the entire reconstruction process, making it a powerful tool for anyone interested in large-scale 3D mapping and lidar applications.

Earth Observation Gaussian Splatting

The real innovation lies in how the system handles camera refinement. Instead of relying on separate tools to estimate where the satellite was positioned, the team embedded bundle adjustment directly into the training loop. By leveraging optical flow techniques, the system refines camera pose estimation during the process. This internal correction minimizes reprojection errors and ensures high geometric consistency without the need for external software. For professionals familiar with lidar datasets, this integrated approach results in cleaner models and sharper details without the computational overhead typically associated with Neural Radiance Fields (NeRFs).

The results are impressive. Tested on the IARPA 2016 and DFC2019 benchmarks, EOGS++ delivered state-of-the-art performance. It reduced mean absolute errors in building reconstruction from 1.33 to 1.19 compared to original models. The addition of early stopping mechanisms and post-processing further sharpens the final output. This ability to generate precise 3D representations of the Earth’s surface from raw data marks a significant step forward for the industry. It proves that Gaussian Splatting is a superior choice for efficient, high-quality earth observation.

Read More: https://quantumzeitgeist.com/earth-observation-gaussian-splatting-eogs-refines-camera-pose/

Written by Adam Clark. Adam has spent the past 13 years exploring the world from above by using drones, satellites, and mapping tools to better understand our landscapes. Connect with him on LinkedIn: Adam Clark

More Recent News: Lidar Coastal Digital Models Show Stable Beaches

Get Lidar News in Your Inbox

Weekly updates on lidar tech, geospatial industry news, case studies, and product reviews.

About The Author

Nathan Roe of Lidar News

SAM Managed geospatial services

Recent 3D Modeling Posts

Autonomous Navigation Lidar for ESA Moon Mission

Editor’s Note: Autonomous navigation lidar is enabling machines to understand and navigate complex environments without…

July 28, 2026

Lidar Supports Vivid Sydney Projection Mapping

For Vivid Sydney’s Creative Technology After Dark, lidar scanning was used to create accurate digital…

July 7, 2026

Preserving an Icon: Liberty Bell Digital Twin

The Liberty Bell Digital Twin Project recently created the first free, high-resolution public digital twin…

July 2, 2026
The four stages of 3D Gaussian Splatting reconstruction, visualised as a single progression. A sparse point cloud provides the initial spatial skeleton (left); large overlapping Gaussian primitives are seeded from those points and begin to occupy volume (second); the optimiser densifies and refines those primitives, packing them tighter as the form converges (third); finally, alpha-compositing from the correct viewpoint produces a photorealistic render with no visible primitives remaining (right). The primitives never disappear, they simply become coherent. This is not a surface mesh, not a voxel grid, not a neural network at render time. It is millions of learned ellipsoids, each carrying position, shape, rotation, colour and opacity, composited in depth order at real-time frame rates. Image: Mike Devers / Cr8ive Media

Gaussian Splatting and LiDAR: A Practitioner’s Field Guide

3D Gaussian Splatting (3DGS) is one of the most talked-about technologies in reality capture right…

June 23, 2026
New Coherent Lidar Expands What 3D Sensing Can See

New Coherent Lidar Expands What 3D Sensing Can See

A new lidar system developed by researchers at the University of Toronto, the Vector Institute,…

June 23, 2026

Cyprus History With 3D Heritage Modeling

An island-based French artist is blending traditional physical sculpting with modern digital workflows to preserve…

June 19, 2026

Popular Posts

Phoenix Lidar Systems

Get Lidar News in Your Inbox

Weekly updates on lidar tech, geospatial industry news, case studies, and product reviews.

Frontier Precision Emesent STX