SAR-UMBRA-x4

SAR-UMBRA-x4 is a paired SAR image dataset for 4× super-resolution. Each sample contains one high-resolution target and two versions of the corresponding low-resolution input.

Dataset size

Split	Paired samples	HR images	LR oversampled images	LR crop images
Train	128,502	128,502	128,502	128,502
Validation	7,753	7,753	7,753	7,753
Total	136,255	136,255	136,255	136,255

The dataset contains 408,765 PNG files in total.

Sample components

Directory	Image size	Description
hr	1024 × 1024	High-resolution target image
lr_x4_crop	256 × 256	Native-size 4× low-resolution input
lr_x4_oversampled	1024 × 1024	The same 4× low-resolution input, Fourier-oversampled to the HR spatial size

lr_x4_crop and lr_x4_oversampled contain the same low-resolution information. The oversampled version is provided for models that require the input and target to have identical spatial dimensions.

Directory structure

After extracting the archives, the dataset is organized as follows:

SAR-UMBRA-x4/
├── train/
│   ├── hr/
│   ├── lr_x4_crop/
│   └── lr_x4_oversampled/
└── val/
    ├── hr/
    ├── lr_x4_crop/
    └── lr_x4_oversampled/

The train and validation splits are distributed separately as train.tar and val.tar.

Download and extract

hf download bw0826/SAR-UMBRA-x4 \
  --repo-type dataset \
  --local-dir SAR-UMBRA-x4

cd SAR-UMBRA-x4
tar -xf train.tar
tar -xf val.tar

Pairing HR and LR images

Corresponding images share the same scene, patch index, and viewing-angle information. HR filenames contain _hr_gt_, while both LR directories use the corresponding filename with _lr_x4_.

For example:

train/hr/..._METADATA_hr_gt_p000_angle_42.png
train/lr_x4_crop/..._METADATA_lr_x4_p000_angle_42.png
train/lr_x4_oversampled/..._METADATA_lr_x4_p000_angle_42.png

The corresponding LR filename can therefore be obtained directly from the HR filename:

from pathlib import Path
from PIL import Image

root = Path("SAR-UMBRA-x4/train")

for hr_path in sorted((root / "hr").glob("*.png")):
    lr_name = hr_path.name.replace("_hr_gt_", "_lr_x4_")

    crop_path = root / "lr_x4_crop" / lr_name
    oversampled_path = root / "lr_x4_oversampled" / lr_name

    hr = Image.open(hr_path).convert("L")
    lr_crop = Image.open(crop_path).convert("L")
    lr_oversampled = Image.open(oversampled_path).convert("L")

Recommended usage

• Use lr_x4_crop → hr for conventional 4× super-resolution training.
• Use lr_x4_oversampled → hr when the model expects equal-sized input and target images.
• Use the same pairing rule for both the train and validation splits.
• The PNG files contain grayscale SAR amplitude images; complex phase data is not included.

For details about dataset construction and preprocessing, please refer to the paper "http://arxiv.org/abs/2609.02377".

License

The dataset organization and processed outputs in this repository are released under CC BY-NC 4.0.

The underlying Umbra SAR imagery originates from the Umbra Open Data Program and remains licensed under CC BY 4.0. Users must provide appropriate attribution to Umbra and comply with the applicable source-data license.