GeoRay

Gauge-Aware Feed-Forward Satellite 3D Reconstruction in the Geodetic Frame

Paper · Code · Issues

GeoRay is a feed-forward framework for reconstructing dense satellite surface height directly in the absolute geodetic frame under rational polynomial cameras (RPCs). It combines RPC height-ray reasoning, gauge-aware relief–datum separation, and uncertainty-calibrated monocular/multi-view fusion.

Model details

• Architecture: frozen VGGT-1B and MoGe-2 backbones with lightweight GeoRay adaptation, relief, datum, calibration, and uncertainty modules
• Inputs: three-view satellite image patches and forward RPC coefficients
• Outputs: absolute surface height and predictive uncertainty; optional longitude/latitude/height back-projection
• Framework: PyTorch 2.7.x
• Code license: Apache-2.0
• Checkpoint file: georay_ckpt.pt
• Checkpoint SHA-256: 7bbe1fc5f04d471d1ecd843be94b73571d8e6aded9a407965277cf37a1a3f330

The released checkpoint contains GeoRay's trainable modules. Frozen backbone weights are obtained separately and remain subject to their upstream licenses.

Results

On 26 held-out US3D tiles, GeoRay reports:

Metric	Result
Absolute MAE ↓	2.99 m
Coverage ↑	91.9%
Completeness-aware accuracy gain ↑	+46.4 points over the strongest compliant feed-forward baseline
Model-forward time ↓	24.0 s / tile in the reported setting

See the paper for the complete evaluation protocol, cross-dataset and cross-city results, and baseline definitions.

Installation

git clone https://github.com/HIT-SIRS/GeoRay.git
cd GeoRay
conda create -n georay python=3.11 -y
conda activate georay

# Install the CUDA-enabled PyTorch 2.7.x build appropriate for your system first.
pip install -r requirements.txt
pip install -e .

Follow the repository installation guide to prepare the pinned VGGT code and external frozen weights.

Download

After this model repository is published, download the checkpoint with:

hf download hit-sirs/GeoRay georay_ckpt.pt --local-dir checkpoints/GeoRay
export GEORAY_CKPT="$PWD/checkpoints/GeoRay/georay_ckpt.pt"

Inference

python scripts/infer.py \
  --ckpt "$GEORAY_CKPT" \
  --config configs/train.yaml \
  --input /path/to/tile.tar \
  --out_dir outputs/tile

python scripts/export_dsm.py \
  --infer_out outputs/tile \
  --out_dir outputs/tile \
  --input /path/to/tile.tar

The input WebDataset format and RPC requirements are documented in the GeoRay README.

Intended use

GeoRay is intended for research on satellite multi-view 3D reconstruction, absolute-frame surface-height estimation, uncertainty-aware photogrammetry, and related geospatial studies.

Limitations

• The released model expects the preprocessing and three-view WebDataset structure documented by the project.
• Accuracy is not guaranteed for unseen sensors, geographic regions, RPC conventions, or height datums; validate outputs before downstream use.
• The exported product is a predicted digital surface model, not a surveyed digital terrain model.
• Predictive uncertainty is a model estimate and must not be treated as a formal safety guarantee.
• Inference requires an NVIDIA GPU and separately obtained frozen backbone weights.

Third-party components

GeoRay source code and its lightweight checkpoint are released under Apache-2.0. VGGT-1B, MoGe-2, and other third-party components are governed by their own licenses. In particular, users must review the applicable VGGT weight license before use or redistribution.

Citation

@article{dong2026georay,
  title   = {GeoRay: Gauge-Aware Feed-Forward Satellite 3D Reconstruction in the Geodetic Frame},
  author  = {Dong, Zhe and Wu, Wanqing and Sun, Yuzhe and Jiang, Haochen and Ma, Yuchen and Ren, Lecheng and Liu, Tianzhu and Gu, Yanfeng},
  journal = {arXiv preprint arXiv:2608.29680},
  year    = {2026}
}