DINOSAR

Self-supervised SAR image encoders with content-aware view construction.

DINOSAR learns transferable representations from single-channel synthetic aperture radar imagery. Its Content-Aware Multi-Crop (CAMC) strategy constructs informative and diverse views for teacher–student self-distillation.

Paper: DINOSAR: Large-Scale SAR Self-Supervised Pretraining with Content-Aware View Construction
Status: Submitted to the International Journal of Digital Earth (IJDE).

Project homepage · UniSAR-7M dataset · Pretrained checkpoints · Citation

DINOSAR pretraining framework. CAMC generates two global views and local views from a SAR structural content map and content anchors. The teacher processes global views, while the student processes all views. The teacher is updated by an exponential moving average of the student. Click the figure for full resolution.

Method

CAMC derives structural content maps from multi-scale ratio-gradient responses and samples views by balancing informative scattering content, anchor proximity, and coverage diversity. Pretraining combines image-level DINO distillation, patch-level iBOT distillation, and KoLeo regularization. Teacher targets use Sinkhorn–Knopp normalization; student predictions use softmax.

The ViT backbones use single-channel inputs, rotary position embeddings (RoPE), and four register tokens. They follow the DINOv3-compatible architecture with a single-stage self-distillation recipe. The exported teacher encoders provide global image representations and spatial patch features for downstream learning.

Released checkpoints

File	Architecture	Pretraining data	Epochs	Final step	CLS dimension	Training config	Training log
dinosar_s16_unisar7m_60e.pth	ViT-S/16	UniSAR-7M	60	330360	384	YAML	JSONL
dinosar_b16_unisar7m_60e.pth	ViT-B/16	UniSAR-7M	60	275340	768	YAML	JSONL

Both models are pretrained for 60 epochs on UniSAR-7M, a corpus of 7,047,666 SAR image samples, using CAMC. The Dataset Card documents the source composition, curation, download instructions, and usage considerations. Different effective batch sizes produce different step counts: 1280 for ViT-S and 1536 for ViT-B.

Files contain teacher backbone-only PyTorch tensor state dictionaries. They exclude projection heads and optimizer states and are intended for feature extraction and downstream initialization, not exact training resumption. See manifest.json and SHA256SUMS for file metadata and integrity verification.

Training logs cover the full 60-epoch runs at 20-step intervals. Training configurations preserve the recorded settings, with the dataset path replaced by /path/to/UniSAR-7M.

Loading

Install the DINOSAR Python package from the companion source repository (DINOSAR_official/DINOSAR, using its installation instructions), and install huggingface_hub and pyyaml. These are custom PyTorch encoders; Transformers AutoModel is not required.

from pathlib import Path
import torch
import yaml
from huggingface_hub import hf_hub_download
from dinosar.model import load_backbone_from_config

name = "dinosar_b16_unisar7m_60e"  # or dinosar_s16_unisar7m_60e
repo = "YTang/DINOSAR"
# Fixed revision containing the released weights and model configurations.
revision = "159c352055e64dc5703f59254920bc8473870272"
weights = hf_hub_download(repo, name + ".pth", revision=revision)
config_path = hf_hub_download(repo, "configs/" + name + ".yaml", revision=revision)
config = yaml.safe_load(Path(config_path).read_text())["model"]
model = load_backbone_from_config(config, weights, strict=True).eval()

# Shape-only smoke example; replace with a properly preprocessed SAR image.
x = torch.zeros(1, 1, 224, 224)
with torch.inference_mode():
    features = model.forward_features(x)
    cls = features["x_norm_clstoken"]
    patches = features["x_norm_patchtokens"]
print(cls.shape, patches.shape)

Input convention

The encoders use one input channel, patch size 16, and four register tokens. For the classification evaluation pipeline, use grayscale imagery, aspect-ratio-preserving resize and zero padding to 224 x 224, conversion to a [0, 1] tensor, and normalization with mean 0.219 and standard deviation 0.220. Use the companion code's ResizeAndPad implementation to match evaluation behavior. These statistics describe the prepared image representation, not arbitrary raw complex-valued SAR measurements. Dense prediction tasks use their own documented preprocessing protocols.

Intended use

The encoders support frozen feature extraction and initialization for downstream SAR classification, detection, segmentation, and image–text retrieval. Task-specific heads, annotations, and training protocols are required for those applications. The pretrained backbones alone do not constitute complete detectors, segmenters, or image–text models.

Licensing

The weight license statement is pending provenance review. Code, upstream dependencies, and training imagery have separate usage terms; the companion software license should not be interpreted as a blanket license for the weights or dataset. Training-data attribution and usage information is provided in the Dataset Card.

Citation

If DINOSAR contributes to your research, please cite the associated manuscript:

@unpublished{tang2026dinosar,
  title = {{DINOSAR}: Large-Scale {SAR} Self-Supervised Pretraining with Content-Aware View Construction},
  author = {Tang, Yan and Jin, Yifeng and Miao, Zekai and Ji, Min and Duan, Yu and Zhang, Shaoming and Wang, Jianmei},
  year = {2026},
  note = {Submitted to the International Journal of Digital Earth}
}