OneScience-Group/DOFA

DOFA

Model Introduction

DOFA is a dynamic single-backbone foundation model for multisensor remote-sensing imagery. It generates band-adaptive weights through a wavelength-conditioned hypernetwork, enabling one model to process observations with different channel counts and spectral responses.

Paper: Neural Plasticity-Inspired Multimodal Foundation Model for Earth Observation
https://arxiv.org/abs/2403.15356

Model Description

DOFA was proposed by a research team from institutions including Wuhan University. The model uses five modalities, Sentinel-1, Sentinel-2, NAIP, EnMAP, and Gaofen, for multisensor masked pretraining, with 4-channel Gaofen inputs. It is suitable for multimodal remote-sensing representation learning, image reconstruction, and cross-sensor downstream task adaptation.

Use Cases

Scenario	Description
Multisensor pretraining	Use central wavelengths to drive dynamic patch-embedding and decoding weights.
Cross-modal representation learning	Use one checkpoint to process five remote-sensing modalities with different channel counts.
Land-cover and scene classification	Transfer and fine-tune shared representations for land-cover and remote-sensing scene classification across different sensors.
Semantic segmentation	Transfer wavelength-aware features to pixel-level remote-sensing interpretation tasks such as flood and land-cover segmentation and fine-tune them.
Local engineering validation	Use a small amount of synthetic data to check training, inference, and evaluation workflows.
Multi-accelerator training	Launch distributed training with torchrun.

Usage Guide

1.OneCode

Experience intelligent one-click AI4S programming through the OneCode online environment:

Click to Experience Intelligent One-Click AI4S Programming

2.Download and Installation

hf download OneScience-Group/DOFA --local-dir ./DOFA
cd DOFA

Environment Dependencies

Hardware Requirements

• A GPU or DCU is recommended.
• CPU can be used for connectivity validation with a small configuration; full training and inference are slower.
• DCU users must install DTK in advance. DTK 25.04.2 or above, or the OneScience-recommended version matching the current cluster, is recommended.

DCU Environment

# Please activate DTK and CONDA first
conda create -n onescience311 python=3.11 -y
conda activate onescience311
# uv installation is supported
pip install onescience[earth-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

GPU Environment

# Please activate CONDA first
conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12
conda activate onescience311
# uv installation is supported
pip install onescience[earth-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai

Training Data Introduction

By default, one 224x224 synthetic sample per data split is used for each modality to validate the engineering workflow. The five modalities are Sentinel-1 with 2 channels, Sentinel-2 with 9 channels, NAIP with 3 channels, EnMAP with 202 channels, and Gaofen with 4 channels. The synthetic EnMAP wavelengths are evenly spaced approximations used for protocol validation.

The synthetic data retain the channel counts, 224x224 spatial size, and per-channel wavelength input specifications of the authors' official pretraining configuration for the five sensors.

Real data must be preprocessed and converted to the following NPZ training protocol. This protocol is consistent with the model input specification but is not equivalent to the original datasets' download format.

images: float32 [N,C,224,224]
wavelengths: float32 [C]
modality: string scalar
data_range: float scalar

fake_data.py automatically writes the protocol, data_source, and wavelength_mode protocol metadata. These fields must be retained when using real data.

python scripts/fake_data.py

Training

python scripts/train.py

Multi-accelerator training can use:

torchrun --nproc_per_node=8 scripts/train.py

Training shares the dynamic backbone across five modalities and saves a checkpoint and overall training metrics. The default configuration is intended for quick workflow validation. Formal experiments should use the multimodal data scale, model configuration, and training schedule corresponding to the paper.

result/checkpoints/dofa.pt
result/training/metrics.json

Training Weights

This repository will provide DOFA training weights in the weight/ folder. The weight files will be uploaded soon and are expected to be available in the near future.

Inference

python scripts/inference.py

Inference loads the training checkpoint, performs masked reconstruction for each modality in the configuration, and saves the results to:

result/output/

Evaluation and Visualization

python scripts/result.py

Evaluation reports overall MSE, MAE, PSNR, and mask ratio on masked regions, and generates reconstruction comparison figures for each modality. Synthetic-data results are only for validating the engineering workflow and do not represent the full performance reported in the paper.

result/evaluation/metrics.json
result/evaluation/

Official OneScience Resources

Platform	OneScience Main Repository	Skills Repository
Gitee	https://gitee.com/onescience-ai/onescience	https://gitee.com/onescience-ai/oneskills
GitHub	https://github.com/onescience-ai/OneScience	https://github.com/onescience-ai/oneskills

Citation and License

This repository is a reproduction of the original DOFA paper.

The use of the code and data in this repository remains subject to the licenses and terms of use of their respective projects.