THOR: Transformer based foundation model for Heterogeneous Observation and Resolution

January 30, 2026 · View on GitHub

A Versatile Foundation Model for Earth Observation Climate and Society Applications

THOR is a compute-adaptive geospatial foundation model developed by Norwegian Computing Center (NR), UiT The Arctic University of Norway and ESA Φ-lab.

Key Features

Multi-sensor support: Sentinel-1 (SAR), Sentinel-2 (MSI), Sentinel-3 OLCI & SLSTR
Flexible resolution: 10 m to 1000 m native resolutions
Compute-adaptive: Flexible patch sizes and ground covers (1000 m to +100,000 m)
Data-efficient: State-of-the-art performance in data-limited regimes
Model type: Vision Transformer (FlexiViT)

THOR unifies data from Copernicus Sentinel-1, -2, and -3 (OLCI & SLSTR) satellites, processing their native 10 m to 1000 m resolutions in a single model. THOR is pre-trained with a novel randomized patch and input image size strategy, allowing deployment at inference with any patch size for dynamic trade-offs between computational cost and feature resolution without retraining.

Setup

Using uv to set up the environment:

uv sync

Optional dependencies for development:

uv sync --group dev --extra scripts --extra test

Usage

For downstream applications, we recommend using the terratorch framework with our THOR terratorch extension.

Terratorch backbone loading example

# Example usage of THOR ViT backbone with terratorch

# Import our custom thor_terratorch_ext module to register THOR backbones
import thor_terratorch_ext  # noqa: F401

# Load the backbone registry
from terratorch import BACKBONE_REGISTRY

# List available THOR backbones
print([b for b in list(BACKBONE_REGISTRY) if "thor" in b])

# Build a THOR ViT model with specific bands
model = BACKBONE_REGISTRY.build(
    "thor_v1_tiny",
    pretrained=True,
    model_bands=["BLUE", "GREEN", "RED", "VV", "VH"],
    input_params=dict(  # Optional input parameters to customize
        ground_covers=[
            2880
        ],  # Ground cover in meters (typically input image size [px] * input image resolution)
        flexivit_patch_size_seqs=[8],  # Patch size in pixels
    ),
)

Model Training

Pretrain THOR architecture on THOR-Pretrain data from scratch

uv run main.py train thor/config/pretrain/final/thor-base.yaml

Evaluation

THOR demonstrates highly competitive performance on the PANGAEA benchmark, particularly in data-limited regimes. With only 10% training data, THOR-Base achieves the best average rank across all datasets.

Model	HLS Burns	MADOS	PASTIS	Sen1Floods11	FBP	DynEarthNet	CropMap	SN7	AI4Farms
CROMA	76.44	32.44	32.80	87.22	37.39	36.08	36.77	42.15	38.48
DOFA	71.98	23.77	27.68	82.84	27.82	39.15	29.91	46.10	27.74
Prithvi	77.73	21.24	33.56	86.28	29.98	32.28	27.71	36.78	35.04
SpectralGPT	83.35	20.29	34.53	83.12	39.51	35.33	31.06	36.31	37.35
Terramind-B	77.39	44.06	39.96	84.43	54.00	37.35	35.65	43.21	38.59
UNet Baseline	79.46	24.30	29.53	88.55	52.58	35.59	13.88	46.08	34.84
ViT Baseline	75.92	10.18	38.44	81.85	56.53	35.39	27.76	36.01	39.20
THOR-B	76.90	40.67	38.93	86.29	42.80	35.21	42.23	55.94	38.90
THOR-T	75.98	41.65	36.26	82.70	42.81	34.03	37.82	58.52	38.56

Results in mIoU on PANGAEA benchmark with 10% training data. Bold = best, italic = second-best.

Attribution

The development of THOR was funded and supported by European Space Agency (ESA) Φ-lab (FM4CS project, contract no. 4000143489/24/I-DT), and the Research Council of Norway (KnowEarth project no. 337481).

Citation

If you use THOR in your research, please cite the paper:

@article{forgaard2026thor,
      title={THOR: A Versatile Foundation Model for Earth Observation Climate and Society Applications}, 
      author={Theodor Forgaard and Jarle H. Reksten and Anders U. Waldeland and Valerio Marsocci and Nicolas Longépé and Michael Kampffmeyer and Arnt-Børre Salberg},
      year={2026},
      eprint={2601.16011},
      archivePrefix={arXiv},
      primaryClass={eess.IV},
      url={https://arxiv.org/abs/2601.16011}, 
}

License

THOR is released under the MIT License.

This project is based in part on code from the Stanford Machine Learning Group, also licensed under MIT.