SAR Ship Detection Dataset

A dataset of Sentinel-1 GRD SAR image chips with corresponding bounding-box annotations for ship detection. This dataset accompanies our paper OSSDD – a New Open Dataset for Sentinel-1 Ship Detection, which describes the dataset creation and baseline detection results. This repository contains the imagery, annotation files, metadata, and a ready-to-use PyTorch Dataset class (ShipDataset) for training object detection models.

The dataset is available in two formats:

---

Table of Contents

• Dataset Overview
• Directory Structure
• SAR Image Files
• Annotation Format
  • Axis-Aligned Bounding Boxes (Metadata_aabb)
  • Rotated Bounding Boxes (Metadata_rbb)
• Metadata File (metadata.csv)
• Usage: ZIP + PyTorch ShipDataset
  • Parameters
  • Additional Notes on ShipDataset
  • Usage Example
• Usage: WebDataset
  • Shard Structure
  • Minimal Example
• Requirements
• Citation
• License

---

Dataset Overview

The dataset consists of Sentinel-1 SAR image chips (in VV and VH polarization) together with ship annotations. The underlying Sentinel-1 products are GRD (Ground Range Detected) images. Each image is provided as a single-band 32-bit GeoTIFF and comes with text-based annotation files describing both axis-aligned and rotated bounding boxes. Additionally, rotated bounding boxes are provided as binary masks (8-bit GeoTIFF).

---

Directory Structure

dataset_root/
├── Chip_VV/
│   ├── Chip_VV_<file_name>.tif
│   └── ...
├── Chip_VH/
│   ├── Chip_VH_<file_name>.tif
│   └── ...
├── Metadata_aabb/
│   ├── Metadata_aabb_<file_name>.txt
│   └── ...
├── Metadata_rbb/
│   ├── Metadata_rbb_<file_name>.txt
│   └── ...
├── Rotated_Masks/
│   ├── Rotated_Mask_<file_name>.tif
│   └── ...
└── metadata.csv

---

SAR Image Files

• All images in the chip folders are 32-bit TIFF files and represent GRD SAR amplitude images.
• Each image is single-channel (one polarization band per file).
• The patch size is 700 × 700 pixels for training data and 512 × 512 for validation and test data.
• File naming convention: Chip_<POL>_<file_name>.tif, where <POL> is either VV or VH.

The Sentinel-1 products used are GRD (Ground Range Detected) images. These are generated from SLC data by projection into ground geometry and multilooking. Due to this resampling, some images contain negative pixel values in the vicinity of very strong reflectors. These negative values are already present in the GRD products downloaded from the ESA Copernicus Hub.

This must be taken into account — for example, when applying a logarithmic transform to the images. The ShipDataset class handles this via optional clipping of extreme outliers and non-positive values before log10 scaling.

---

Annotation Format

Annotations are stored as plain text files inside Metadata_aabb (axis-aligned boxes) and Metadata_rbb (rotated boxes). All coordinates are pixel coordinates in the corresponding image patch, using a coordinate system whose origin is at the top-left.

Axis-Aligned Bounding Boxes (Metadata_aabb)

• The first entry is the number of bounding boxes contained in the mask corresponding to the file name.
• Each following line describes one bounding box:

Index || Min-x || Min-y || Max-x || Max-y || Center-x || Center-y

• The center values are rounded mean values.

Rotated Bounding Boxes (Metadata_rbb)

• The first entry is the number of rotated bounding boxes (RBB) contained in the mask corresponding to the file name.
• Each following line describes one rotated bounding box:

Index || Corner1-x || Corner1-y || ... || Corner4-y || Center-x || Center-y

• Again, coordinates refer to a top-left origin coordinate system.
• The center values are rounded mean values.

---

Metadata File (metadata.csv)

metadata.csv contains one row per image chip and defines the training, validation and test set. Additionally, it provides further information about the land coverage per image derived from a water-land mask (computed with SNAP). Relevant columns include:

---

Usage: ZIP + PyTorch ShipDataset

Download

Download ossdd.zip from the repository root and extract it:

from huggingface_hub import hf_hub_download
import zipfile

zip_path = hf_hub_download(
    repo_id="sylviaHoch/OpenSARShip-Detection-Dataset",
    filename="ossdd.zip",
    repo_type="dataset",
    local_dir="./data",
)

with zipfile.ZipFile(zip_path, "r") as z:
    z.extractall("./data/ossdd")

Then place ship_dataset.py in your working directory and use the ShipDataset class as shown below.

Parameters

Additional Notes on ShipDataset

• Multi-polarization handling: each polarization of an image is treated as a separate sample (dataset length is scaled by the number of polarizations).
• Random / fixed cropping: random crops are applied during training; fixed-offset crops are used for validation and testing.
• Outlier handling: optional clipping of values > 20 000 and <= 0 — important for Sentinel-1 GRD data (see Important Notes on Sentinel-1 GRD Data).
• dB scaling: optional log10 transform of pixel values.
• Multiple normalization schemes: per-patch normalization (zscore, robust_zscore, percentile_clip, min_max) and per-scene normalization (scene_min_max, scene_min_p99). Scene-based scaling parameters are read from / written to scaling_params.csv.
• Land augmentation: duplication of land-dominated scenes via augment_landmasks.
• Stratified subsampling: subsample the training set by class (water / land) via training_fraction. This option can be used to evaluate model performance with limited training data.
• Bounding-box clipping: for cropped image patches, the corresponding boxes are shifted to the crop coordinate system, clipped to the patch borders, and boxes outside the patch are removed.

Each item returned by __getitem__ is a tuple (patch, target):

• patch: torch.FloatTensor of shape (1, patch_size, patch_size).
• target: dict with
  • boxes: torch.FloatTensor of shape (N, 4) in [xmin, ymin, xmax, ymax] format,
  • labels: torch.Int64Tensor of shape (N,) (single class → label 1),
  • image_id: torch.Tensor([idx]).

Usage Example

from torch.utils.data import DataLoader
from ship_dataset import ShipDataset


def collate_fn(batch):
    return tuple(zip(*batch))


dataset = ShipDataset(
    data_dir="path/to/dataset_root",
    label_dir="path/to/dataset_root",
    normalization="robust_zscore",
    split="train",
    patch_size=512,
    polarization=["VV", "VH"],
    db_scaling=True,
    clip=True,
    seed=42,
)

loader = DataLoader(
    dataset,
    batch_size=4,
    shuffle=True,
    collate_fn=collate_fn,
)

for images, targets in loader:
    # images: tuple of tensors of shape (1, 512, 512)
    # targets: tuple of dicts with keys 'boxes', 'labels', 'image_id'
    pass

---

Usage: WebDataset

The WebDataset format enables efficient streaming of the dataset — either from local storage or directly from the HuggingFace Hub — without requiring the full dataset to be downloaded upfront.

Shard Structure

Shards are stored under webdataset/{split}/ as .tar files (e.g. train-000000.tar, val-000000.tar, test-000000.tar, …). Each sample inside a shard contains the following fields:

Minimal Example

import io
import json
import numpy as np
import rasterio
import webdataset as wds


def decode_tif(data: bytes) -> np.ndarray:
    """Read a GeoTIFF from a byte buffer and return band 1 as float32 array."""
    with rasterio.open(io.BytesIO(data)) as src:
        return src.read(1).astype(np.float32)


def decode_aabb(data: bytes) -> list:
    """Parse aabb.txt and return a list of [xmin, ymin, xmax, ymax] boxes."""
    boxes = []
    lines = data.decode("utf-8").strip().splitlines()
    for line in lines[1:]:          # first line contains the box count
        parts = line.split()
        if len(parts) >= 5:
            boxes.append([
                float(parts[1]), float(parts[2]),
                float(parts[3]), float(parts[4]),
            ])
    return boxes


def preprocess(sample: dict) -> dict:
    return {
        "vv":    decode_tif(sample["vv.tif"]),
        "vh":    decode_tif(sample["vh.tif"]),
        "boxes": decode_aabb(sample["aabb.txt"]),
        "meta":  json.loads(sample["meta.json"]),
    }


### --- Stream directly from the HuggingFace Hub ---

from huggingface_hub import hf_hub_url

shard_pattern = [
  hf_hub_url(
      repo_id="<sylviaHoch/OpenSARShip-Detection-Dataset",
      filename=f"webdataset/train/train-{i:06d}.tar",
      repo_type="dataset",
    ) for i in range(43)   # adjust range to the actual number of shards
]

### --- Or load shards from local storage ---

# shard_pattern = "path/to/webdataset/train/train-{000000..000042}.tar"

dataset = (
    wds.WebDataset(shard_pattern, shardshuffle=True)
    .shuffle(500)
    .map(preprocess)
)

for sample in dataset:
    vv    = sample["vv"]     # np.ndarray, shape (H, W), float32
    vh    = sample["vh"]     # np.ndarray, shape (H, W), float32
    boxes = sample["boxes"]  # list of [xmin, ymin, xmax, ymax]
    meta  = sample["meta"]   # dict: file_name, text, land_fraction, split
    break

---

Requirements

rasterio
numpy
pandas
torch
webdataset        # required for WebDataset usage only
huggingface_hub   # required for Hub download

---

Citation

If you use this dataset in your research, please cite the corresponding publication.

@misc{hammer2026ossddnewopen,
      title={OSSDD - a New Open Dataset for Sentinel-1 Ship Detection}, 
      author={Horst Hammer and Sylvia Hochstuhl and Antje Thiele and Tobias Brosch and Padraig Davidson and Tim Remiger and Michael Teutsch},
      year={2026},
      eprint={2608.01963},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
}

---

License and Attribution

Dataset Annotations and Code

The annotations, metadata, and accompanying code created by the dataset authors are released under the Creative Commons Attribution 4.0 International (CC-BY-NC-SA-4.0) license.

Underlying Sentinel-1 Imagery

The image data is derived from Copernicus Sentinel-1 products and remains subject to the Copernicus Sentinel Data Terms and Conditions.

This dataset contains modified Copernicus Sentinel data (the Sentinel-1 GRD scenes were cropped, preprocessed, and annotated).

Disclaimer: Copernicus Sentinel data is provided without any express or implied warranty, including as regards quality and suitability for any purpose. By using this data, you acknowledge the Copernicus Sentinel Data Terms and renounce any claims for damages against the European Union and the providers of the said data and information.

How to Attribute

When using or redistributing this dataset, please include the following notice:

Contains modified Copernicus Sentinel data, processed and annotated by the dataset authors. Annotations licensed under CC-BY-NC-SA-4.0.