Zhu X Lab

weights

open(ungated CC-BY-4.0 DOFA and DOFA v2 checkpoints on Hugging Face)

data

components-listed(paper section 6 gives the composition — ~11.5M images across Sentinel-1, Sentinel-2, NAIP, Gaofen and EnMAP, with per-modality channel counts — and a download link per subset, but no assembled DOFA corpus and no script that rebuilds it end to end)

code

open(pretraining entry point, masked-autoencoder engine, samplers, and launch script under pretraining/)

license

CC-BY-4.0(attribution only, not OSI-approved)

The pretraining pipeline is public, the checkpoints are ungated, and the paper links a download for each of the five training subsets, enough to audit the mixture. But those subsets are third-party datasets nothing in the repository assembles into one release, so the corpus is not fully open, and the weights carry CC-BY-4.0 rather than the repository's own MIT license.

• https://api.github.com/repos/zhu-xlab/DOFA/contents/pretraining recorded 2026-09-10

  Lists main_pretrain_ofa.py, engine_pretrain.py, models_base_ofa_mae.py, train_mae_all.sh, samplers, and a datasets directory holding only a loader and wavelength tables.

• https://arxiv.org/abs/2403.15356 recorded 2026-09-10

  Abstract states that the code and pretrained weights are publicly available. It says nothing either way about the corpus, so it establishes the artifacts and not the data dimension.

• https://arxiv.org/pdf/2403.15356v3 recorded 2026-09-10

  Section 4.5 — "DOFA is pretrained on a large-scale multimodal EO dataset comprising approximately 11.5 million images from five different modalities: Sentinel-1, Sentinel-2, NAIP, Gaofen, and EnMAP". Section 6 Data availability statements gives a download link per subset (SatlasPretrain for Sentinel-1 and Sentinel-2, a named Drive folder for Gaofen, hyspecnet.rsim.berlin for EnMAP).

• https://huggingface.co/api/models/earthflow/DOFA recorded 2026-09-10

  gated: false; private: false; license cc-by-4.0; DOFA_ViT_base_e100.pth, DOFA_ViT_large_e100.pth, dofav2_vit_base_e150.pth, and dofav2_vit_large_e150.pth present.

• https://huggingface.co/earthflow/DOFA/raw/main/README.md recorded 2026-09-10

  Card front matter and the model details both state License: CC-BY-4.0 for the released DOFA weights.

The Hugging Face mirror reports zero downloads because the checkpoints are pulled as raw files rather than through the download API, so GitHub stars are used as the last-resort reach signal instead.

• https://api.github.com/repos/zhu-xlab/DOFA recorded 2026-09-10

  stargazers_count: 211; archived: false; fork: false.

DOFA earns two capabilities directly: an adaptation pathway through linear-probing and segmentation fine-tuning, and a single checkpoint that accepts five different satellite sensors. Masked autoencoding, its pretraining objective, is not counted as generation, and that same sensor diversity is not counted a second time as multiple scales.

• https://arxiv.org/abs/2403.15356 recorded 2026-09-10

  Describes a wavelength-conditioned dynamic hypernetwork processing inputs from five distinct satellite sensors and continual pretraining on five EO modalities.

• https://raw.githubusercontent.com/zhu-xlab/DOFA/master/README.md recorded 2026-09-10

  Documents GeoBench classification linear probing and UPerNet segmentation fine-tuning under downstream_tasks, and shows one loaded model run over Sentinel-1 SAR, Sentinel-2, and NAIP RGB inputs with arbitrary channel counts.