SatDiFuser

SatDiFuser: Can Generative Geospatial Diffusion Models Excel as Discriminative Geospatial Foundation Models?

ICCV 2025 [ArXiv]

By Yuru Jia, Valerio Marsocci, Ziyang Gong, Xue Yang, Maarten Vergauwen, Andrea Nascetti

SatDiFuser explores self-supervised learning (SSL) with diffusion models as geospatial foundation models (GFMs) for remote sensing. While most GFMs use contrastive learning or masked image modeling, SatDiFuser shows that diffusion-based generative models can also learn powerful discriminative representations. By analyzing multi-stage, noise-dependent features and introducing three fusion strategies, SatDiFuser achieves state-of-the-art results on remote sensing benchmarks, highlighting the untapped potential of diffusion models for SSL in RS.

<img src="figs/method-overview.png" width="85%" height="60%">

Please follow the instructions below to set up the experiments.

⚙️ Setup

• Environment: python 3.11.5, CUDA 12.4.0
• Clone:

  git clone https://github.com/yurujaja/SatDiFuser.git
  cd SatDiFuser

• Install dependencies:

pip install -r requirements.txt
pip install --no-deps geobench

• Download DiffusionSat checkpoints: Our approach builds on DiffusionSat (paper).
  • Download 512×512 and 256×256 checkpoints from 512×512 and 256×256
  • Place them under diffusionsat/pretrained/resolution512 and diffusionsat/pretrained/resolution256
  • Point pretrained_model_name_or_path in configs/_base_/diffusionsat.yaml to your root checkpoints directory

📊 Downstream Tasks

• GEO-Bench: We evaluate on GEO-Bench (paper).
• Download and setup:
  • See GEO-Bench downloading guide
  • See dataset implementation in PANGAEA
  • Export data root: export GEO_BENCH_DIR=/path/to/geobench

📎 Configuration

Configs are composed from multiple YAML files. An experiment file (e.g., configs/eurosat_exp.yaml) lists a base: array that includes:

• configs/_base_/diffusionsat.yaml — DiffusionSat backbone settings
• configs/_base_/satdifuser.yaml — SatDiFuser extraction/fusion settings
• Task-specific decoder base (e.g., configs/_base_/upernet.yaml for segmentation)
• Dataset/task config under configs/tasks/ (e.g., configs/tasks/meurosat.yaml)

Tips

• configs/_base_/diffusionsat.yaml

  • pretrained_model_name_or_path: root folder for checkpoints
  • use_metadata / num_metadata / low_cpu_mem_usage / revision: backbone loading flags

• configs/_base_/satdifuser.yaml

  • Fusion variant: Set fuser to gw, lw, or moe in configs/_base_/satdifuser.yaml.
  • fuser: fusion variant — gw (Global-Weighted), lw (Localized-Weighted), moe (Mixture-of-Experts)
  • layer_idxs: which UNet blocks/layers to extract from
  • projection_dim: bottleneck output channel dimension per feature stream
  • save_timesteps: diffusion timesteps to extract
  • num_timesteps / diffusion_mode: diffusion settings
  • MoE-only: num_experts, top_k

• configs/_base_/upernet.yaml (segmentation only)

  • rescales: per-scale resize factors into the feature pyramid
  • pool_scales: PPM pooling scales (e.g., 1, 2, 3, 6)
  • channels: decoder channels for UPerNet head (must be a positive int)

Dataset/task configs (configs/tasks/*.yaml)

• Classification example configs/tasks/meurosat.yaml:

  • task: classification
  • num_classes, original_img_size, img_size, num_channels, class_list
  • prompt, resize_outputs

• Segmentation example configs/tasks/mnz_cattle.yaml:

  • task: segmentation
  • num_classes, class_names, original_img_size, img_size, num_channels
  • prompt, resize_outputs

Experiment files (configs/*_exp.yaml)

Running

Update your experiment file to include the right dataset task file in its base: list.

Example: Eurosat classification, with experiment file: configs/eurosat_exp.yaml with configs/tasks/meurosat.yaml in base:

python run.py --config configs/eurosat_exp.yaml

Example: New Zealand cattle segmentation

Experiment file: configs/nzcattle_exp.yaml with configs/tasks/mnz_cattle.yaml in base:

python run.py --config configs/nzcattle_exp.yaml

✏️ Acknowledgements

We thank prior work and codebases: Diffusion Hyperfeatures, SLiMe, DiffSeg, DiffCut, PANGAEA, GEO-Bench. We also acknowledge the National Academic Infrastructure for Supercomputing in Sweden (NAISS, Grant No. 2022-06725) for supporting the computations and data handling.

🎓 Citation

@inproceedings{jia2025satdifuser,
  title={Can Generative Geospatial Diffusion Models Excel as Discriminative Geospatial Foundation Models?},
  author={Jia, Yuru and Marsocci, Valerio and Gong, Ziyang and Yang, Xue and Vergauwen, Maarten and Nascetti, Andrea},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2025}
}