---

license: cc-by-4.0 library_name: pytorch tags: - sar - remote-sensing - self-supervised-learning - foundation-model - masked-image-modeling - object-detection - image-classification - semantic-segmentation pipeline_tag: image-feature-extraction

SARATR-X-v2

Scale-Aware Structural Pre-Training for SAR Foundation Models

Weijie Li, Yafei Song, Yongxiang Liu, Bowen Peng, Jie Zhou, Jingyuan Xia, Wei Yang, Tianpeng Liu, Zhen Liu, and Li Liu

Paper (arXiv) · Code (GitHub) · BaiduYun · Zhihu

Note: This paper is currently under review. Data and checkpoints on this page / BaiduYun are available upon request — please email lwj2150508321@sina.com.

Model Description

SARATR-X-v2 is a vision foundation model for SAR target recognition, built upon scale-aware structural masked pre-training. Instead of reconstructing raw pixels, it reconstructs a multi-scale structural target that is robust to speckle and semantically consistent across scales, then transfers to classification, detection, and segmentation.

Key ideas:

• Multi-scale structural target (S1–S6) with learnable cross-scale fusion (multi)
• Pre-trained on 500K unlabeled SAR target chips
• Strong transfer on 12 SAR benchmarks (classification / detection / segmentation)

Files in This Repository

Pre-trained checkpoints and related data for SARATR-X-v2. Typical layout (see the GitHub weights/README.md for details):

Path	Description
base/jiaquan_simple/checkpoint-1200.pth	Main model, iTPN-Base
large/jiaquan_simple/checkpoint-1200.pth	Main model, iTPN-Large
base/hivit/checkpoint-1200.pth	HiViT comparison backbone
base/pixel, single_s*, multi, …	Target / fusion ablations

Default checkpoint used in the paper: checkpoint-1200.pth.

How to Use

1. Email lwj2150508321@sina.com to request access (gated repository).
2. Download checkpoints from this Hugging Face repo (or BaiduYun).
3. Place them under the GitHub repo’s weights/ directory as documented there.
4. Run downstream scripts from waterdisappear/SARATR-X-v2 (classification/, detection/, segmentation/).

Example (classification linear probe):

export SARATRX_PRETRAIN_CKPT=/path/to/weights/base/jiaquan_simple/checkpoint-1200.pth
cd classification/linear_eval
python SOC_50.py --data_path ../../dataset/classification/SOC_50classes

Citation

@article{saratrxv2,
  title={SARATR-X-v2: Scale-Aware Structural Pre-Training for SAR Foundation Models},
  author={Li, Weijie and Song, Yafei and Liu, Yongxiang and Peng, Bowen and Zhou, Jie and
          Xia, Jingyuan and Yang, Wei and Liu, Tianpeng and Liu, Zhen and Liu, Li},
  journal={IEEE Geoscience and Remote Sensing Magazine},
  year={2026},
  eprint={2607.23238},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

Contact

Questions or access requests: lwj2150508321@sina.com

License

CC BY 4.0 (this Hugging Face repository). Code on GitHub is released under MIT — see the GitHub repository for details.