UniSAR-7M

A large-scale, multi-source synthetic aperture radar image corpus for self-supervised representation learning.

UniSAR-7M contains 7,047,666 single-channel SAR image samples assembled from public SAR datasets and openly available imagery from commercial satellite constellations. It provides the pretraining corpus for DINOSAR, a self-supervised learning framework that uses Content-Aware Multi-Crop (CAMC) to construct informative views of SAR imagery.

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

Project homepage · Pretrained models · Google Drive distribution

Dataset available: The complete distribution is now available on Hugging Face: all 18 archive parts (000–017), totaling approximately 344.29 GiB (369.68 GB) of compressed data. Download all parts before reconstructing the archive.

Overview of UniSAR-7M. Left: locations of 2,274 commercial source scenes; filled and open markers indicate land- and ocean-centered scenes. Right: commercial-image processing and corpus statistics. Corpus-size comparisons count processed SAR pretraining samples and exclude auxiliary optical images. Click the figure to view it at full resolution.

Dataset overview

SAR imagery exhibits sensor-dependent appearance, speckle, and spatially uneven scattering responses. UniSAR-7M combines imagery from multiple sources to support the study of transferable SAR representations at scale. The corpus is used without task-specific labels during DINOSAR pretraining; downstream classification, detection, segmentation, and retrieval require their respective benchmark annotations and evaluation protocols.

Property	Description
Modality	Synthetic aperture radar imagery
Input representation	Single-channel images
Corpus size	7,047,666 image samples
Commercial source scenes	2,274 scenes across Capella, ICEYE, and Umbra
Commercial-scene tiling	Non-overlapping 256 × 256 pixel tiles
Primary purpose	Self-supervised pretraining and SAR representation learning
Distribution	One gzip-compressed TAR archive split into 18 binary parts

The corpus size counts image samples, including tiles extracted from larger scenes. It should not be interpreted as the number of independent satellite acquisitions or geographic locations.

Data sources and composition

The public-dataset component incorporates training imagery from SARDet-100K, ATRNet-STAR, FAIR-CSAR, FUSAR-Ship, and AIR-PolSAR-Seg-2.0. The commercial-imagery component is derived from Capella, ICEYE, and Umbra, whose contributing constellations operate in X-band.

Source	Number of image samples
SARDet-100K	138,240
ATRNet-STAR	104,249
FAIR-CSAR	479,440
FUSAR-Ship	5,243
AIR-PolSAR-Seg-2.0	6,168
Capella	3,564,112
ICEYE	791,379
Umbra	1,958,835
Total	7,047,666

These counts describe the complete curated corpus distributed in this repository.

Corpus construction

Commercial SAR scenes are partitioned into non-overlapping 256 × 256 pixel tiles. The curation pipeline applies NoData filtering and background filtering, followed by feature-based deduplication. The commercial component is reduced from 26,187,909 initial tiles to 11,846,546 after NoData filtering, 6,727,620 after background filtering, and 6,314,326 after deduplication.

Deduplication uses representations extracted by a ViT-S model pretrained on the SAR imagery of SAR-1M, with a cosine-similarity threshold of 0.8. One representative is retained from each duplicate group. This procedure is applied separately within each commercial imagery source, after background filtering. It is not a cross-source deduplication procedure and is not applied to the public-dataset component.

The public component follows source-specific preparation: SARDet-100K and FAIR-CSAR use 256 × 256 chips with 20% overlap and annotation-based foreground filtering; ATRNet-STAR contributes deduplicated SOC-40/SOC-50 chips; FUSAR-Ship and AIR-PolSAR-Seg-2.0 are prepared at 480 × 480 pixels. These public-source procedures are distinct from the commercial-scene pipeline.

Researchers evaluating representations on any source benchmark should retain that benchmark's prescribed validation and test partitions and document the relationship between their pretraining data and evaluation samples.

Download and reconstruction

This repository preserves the original multipart archive distribution. The parts are fragments of a single compressed archive, not independently extractable TAR files or WebDataset shards.

1. Download the archive parts

Using the Hugging Face CLI:

hf download YTang/UniSAR-7M \
  --repo-type dataset \
  --include 'UniSAR-7M.tar.gz.part_*' \
  --local-dir UniSAR-7M-download

cd UniSAR-7M-download

Parts 000–016 are 20 GiB each; part 017 is approximately 4.29 GiB. All 18 parts are required.

2. Check completeness and reconstruct the archive

Run the following in Bash. The explicit file list ensures that the 18 parts are concatenated in the correct order and prevents reconstruction when a required part is missing or empty.

set -euo pipefail

parts=()
for i in {000..017}; do
  part="UniSAR-7M.tar.gz.part_${i}"
  if [[ ! -s "$part" ]]; then
    echo "Missing or empty archive part: $part" >&2
    exit 1
  fi
  parts+=("$part")
done

cat "${parts[@]}" > UniSAR-7M.tar.gz
gzip -t UniSAR-7M.tar.gz
tar -xzf UniSAR-7M.tar.gz

gzip -t checks the integrity of the reconstructed compressed stream. Reserve disk space for the downloaded parts, the reconstructed archive, and the extracted images when using these commands.

3. Use with DINOSAR

Set data.data_path in the DINOSAR pretraining configuration to the extracted image root. The companion image-folder loader scans the root recursively and creates an index.txt when needed.

The released DINOSAR pretraining configurations use single-channel inputs with normalization mean 0.219 and standard deviation 0.220. Follow the companion code for image decoding, intensity scaling, and augmentations. These values refer to the prepared image representation and should not be applied directly to arbitrary raw complex-valued SAR measurements.

Intended uses and limitations

UniSAR-7M supports research on self-supervised SAR pretraining, view construction, representation transfer, and the effect of pretraining data composition. It is a pretraining corpus rather than a unified labeled downstream benchmark.

The source mixture is unequal, and the commercial component is drawn from X-band imagery. Dataset size alone does not establish balanced coverage of geographic regions, land-cover categories, acquisition geometries, frequency bands, or polarization settings. Tiles derived from the same scene can remain correlated despite non-overlapping extraction and within-source deduplication. Cross-source redundancy may also remain.

When reporting experiments, specify the dataset version, any additional filtering, and the downstream split protocol. Report results across relevant acquisition conditions rather than assuming that performance on one benchmark establishes generalization to all SAR domains.

Source attribution and usage terms

UniSAR-7M aggregates imagery from multiple providers and public datasets. Users should acknowledge the original sources in addition to the associated DINOSAR manuscript and consult the applicable source-specific usage and redistribution terms. The license of the companion software should not be interpreted as a blanket license for the underlying imagery.

Citation and contact

If UniSAR-7M contributes to your research, please acknowledge the dataset and the following manuscript:

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

A complete bibliographic citation and persistent paper link will be provided when available. For dataset questions or reports of archive issues, use the repository discussion page.