---
license: cc-by-nc-4.0
language: en
tags:
  - vision
  - image-feature-extraction
  - sar
  - remote-sensing
  - synthetic-aperture-radar
  - masked-autoencoder
  - transformers
library_name: transformers
pipeline_tag: image-feature-extraction
datasets:
  - Wenquandan777/SAR-1M
arxiv: 2512.16635
base_model: Wenquandan777/SARMAE
model-index:
  - name: sarmae-vit-base-patch16-pretrain
    results: []
---

# sarmae-vit-base-patch16-pretrain

SARMAE (ViT-B, patch 16) encoder checkpoint converted to native Hugging Face Transformers format.

SARMAE is a Noise-Aware Masked Autoencoder for self-supervised SAR representation learning, pretrained on [SAR-1M](https://huggingface.co/datasets/Wenquandan777/SAR-1M) with Speckle-Aware Representation Enhancement (SARE) and Semantic Anchor Representation Constraint (SARC).

- **Paper:** [2512.16635](https://arxiv.org/abs/2512.16635)
- **Legacy weights:** [Wenquandan777/SARMAE](https://huggingface.co/Wenquandan777/SARMAE)
- **Stage:** `pretrain`
- **Input:** 3 x 224 x 224 (single-channel SAR is repeated to 3 channels)
- **Architecture:** 12 layers, hidden size 768, 12 heads

## Model specifications

| Property | Value |
|----------|-------|
| Model type | `sarmae` |
| Backbone | ViT-B |
| Patch size | 16 |
| Image size | 224 |
| Hidden size | 768 |
| Layers | 12 |
| Attention heads | 12 |
| Global pooling | `True` |
| Normalization mean | `[0.485, 0.456, 0.406]` |
| Normalization std | `[0.229, 0.224, 0.225]` |

## Intended use

- SAR image feature extraction for downstream classification, detection, and segmentation
- Initializing OpenMMLab backbones (`mmrotate`, `mmseg`) after weight porting
- Research and non-commercial use under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)

## Quick start

Install dependencies:

```bash
pip install transformers timm torch torchvision safetensors
```

### Feature extraction with `transformers.pipeline`

```python
from transformers import pipeline

pipe = pipeline(
    task="image-feature-extraction",
    model="BiliSakura/SARMAE-transformers", subfolder="vit-base-patch16-pretrain",
    trust_remote_code=True,
)
features = pipe(sar_image, pool=True, return_tensors=True)
print(features.shape)
```

### Direct model loading

```python
from transformers import AutoModel, AutoImageProcessor

model = AutoModel.from_pretrained("BiliSakura/SARMAE-transformers", subfolder="vit-base-patch16-pretrain", trust_remote_code=True)
processor = AutoImageProcessor.from_pretrained("BiliSakura/SARMAE-transformers", subfolder="vit-base-patch16-pretrain", trust_remote_code=True)

inputs = processor(images=sar_image, return_tensors="pt")
outputs = model(**inputs)
pooled_features = outputs.pooler_output
```

### Local checkout

```python
pipe = pipeline(
    task="image-feature-extraction",
    model="./sarmae-vit-base-patch16-pretrain",
    trust_remote_code=True,
)
```

## Preprocessing

- Resize to 224x224
- Scale pixel values to `[0, 1]` (`rescale_factor=1/255`)
- Repeat grayscale SAR to 3 channels when `repeat_grayscale_channels=true`
- Normalize with ImageNet mean/std (same as SARMAE fine-tuning code)

## Training data

Pretrained on **SAR-1M**, a million-scale SAR dataset with paired optical anchors for a subset of samples.

## Citation

```bibtex
@misc{liu2025sarmaemaskedautoencodersar,
  title={SARMAE: Masked Autoencoder for SAR Representation Learning},
  author={Danxu Liu and Di Wang and Hebaixu Wang and Haoyang Chen and Wentao Jiang and Yilin Cheng and Haonan Guo and Wei Cui and Jing Zhang},
  year={2025},
  eprint={2512.16635},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.16635},
}
```

## License

This model is released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).