dronefreak/MARIDA

MARIDA: Marine Debris Detection from Sentinel-2 (Object Detection)

<p align="center"> <img src="marida_banner.jpg" alt="MARIDA Dataset Banner"/> </p>

Unofficial redistribution of MARIDA (Marine Debris Archive), converted from its native per-pixel classification masks into a standardized YOLO/COCO object-detection layout via connected-component extraction.

Dataset Description

• —Homepage: https://marine-debris.github.io/
• —Repository: https://github.com/marine-debris/marine-debris.github.io
• —Paper: https://doi.org/10.1371/journal.pone.0262247
• —Point of Contact: akikakh@central.ntua.gr

Disclaimer

This repository is not an official release of MARIDA.

MARIDA was created by Katerina Kikaki, Ioannis Kakogeorgiou, Paraskevi Mikeli, Dionysios E. Raitsos, and Konstantinos Karantzalos (National Technical University of Athens), who retain all copyright. This repository does not claim ownership of any imagery, annotations, or metadata.

This repository exists for two purposes:

1. 1.To convert MARIDA's native task (weakly-supervised pixel-level semantic segmentation) into bounding-box object detection, since the source ships classification masks, not boxes.
2. 2.To render the raw 11-band Sentinel-2 reflectance patches into standard 3-channel images usable by ordinary YOLO/COCO-based training pipelines, which don't natively handle multi-band GeoTIFFs.

Two-hop provenance. This redistribution is sourced from the official MARIDA release on Source Cooperative (the Radiant MLHub successor host; originally published on Zenodo). Both the original creators and the redistribution host are credited below.

Dataset Overview

MARIDA (Marine Debris Archive) is a Sentinel-2 satellite-imagery dataset for marine debris and related ocean-surface phenomena: 1,381 256x256 patches (11-band surface reflectance) with per-pixel classification masks across 15 classes -- Marine Debris, Dense Sargassum, Sparse Sargassum, Natural Organic Material, Ship, Clouds, Marine Water, Sediment-Laden Water, Foam, Turbid Water, Shallow Water, Waves, Cloud Shadows, Wakes, and Mixed Water. Splits: train 694 / valid 328 / test 359 images, following the official patch-id lists.

MARIDA is natively a weakly-supervised segmentation dataset, not an exhaustively-annotated detection benchmark. Only a sparse subset of each patch's pixels are confidently labeled -- large, visually-obvious features (a cloud shadow, an unlabeled water-color boundary) are frequently left unlabeled entirely. A box in this repository means "a confidently-labeled contiguous region of this class", not "every visible instance is boxed". Several of the 15 classes (Marine Water, Clouds, Waves, Turbid Water, Shallow Water, Cloud Shadows, Sediment-Laden Water, Mixed Water) describe extended surface phenomena rather than compact objects, and correspondingly produce larger, sparser boxes than the foreground classes (Marine Debris, Sargassum, Ship, Foam, Wakes).

This repository preserves every officially-labeled region while converting the task representation (see Changes from the Official Release below).

Changes from the Official Release

Unlike a simple directory reorganization, this conversion involves two genuine representation changes -- the raw release ships neither bounding boxes nor natural RGB images:

1. Segmentation mask &rarr; bounding boxes

Each patch's classification mask (*_cl.tif, integer values 1-15, 0 = unlabeled) is converted to boxes via connected-component extraction per class value (8-connectivity): every contiguous labeled region becomes one box, tightly fit to its pixel extent. Components smaller than 4 pixels are dropped as extraction noise. No class was merged, split, or renamed -- all 15 official classes are preserved in their original order (id 1 &rarr; index 0, ... id 15 &rarr; index 14).

2. 11-band reflectance &rarr; RGB image

MARIDA's raw patches are 11-band Sentinel-2 surface reflectance (bands B01-B08, B8A, B11, B12; float32, roughly 0-0.3 range), not natural-color imagery. This repository renders a true-color-ish RGB image per patch (bands B04/B03/B02 as R/G/B, a per-patch 1st-99th percentile contrast stretch applied jointly across all three channels to preserve color balance, then a 1/1.5 gamma correction), saved as JPEG. A handful of patches (7 of 1,381) contain scene-edge no-data pixels as NaN in the raw reflectance; these are imputed with the patch's own per-channel mean before rendering, matching the approach the dataset's own reference PyTorch dataloader uses. This is a derived visualization/training input, not a radiometrically calibrated product -- the box coordinates themselves come straight from the classification mask, independent of this rendering choice.

No official split, class taxonomy, or label was altered otherwise.

Dataset Structure

<repo>/
├── README.md
├── marida_banner.jpg
└── data/
    ├── data.yaml
    ├── images/
    │   ├── train/   (694 *.jpg)
    │   ├── valid/   (328 *.jpg)
    │   └── test/    (359 *.jpg)
    └── labels/
        ├── train/   (694 *.txt)
        ├── valid/
        └── test/

where:

• —data/images/<split>/ contains the derived RGB renders for each split.
• —data/labels/<split>/ contains one YOLO-format .txt annotation file per image (class x_center y_center width height, normalized; empty for the rare patch with no labeled region).
• —data/data.yaml is the Ultralytics dataset configuration file (class names, split paths, relative to data/).
• —Splits: train 694 images / 1,533 boxes &middot; valid 328 images / 713 boxes &middot; test 359 images / 746 boxes (1,381 images / 2,992 boxes total).

Classes (15)

Full per-class instance counts and a class-distribution chart: see the dataset statistics report.

Dataset Sources

Original Paper

MARIDA: A benchmark for Marine Debris detection from Sentinel-2 remote sensing data

Katerina Kikaki, Ioannis Kakogeorgiou, Paraskevi Mikeli, Dionysios E. Raitsos, Konstantinos Karantzalos

PLOS ONE, 17(1):e0262247, 2022.

Official Resources

• —Project website: https://marine-debris.github.io/
• —GitHub Repository: https://github.com/marine-debris/marine-debris.github.io
• —Official data host: https://source.coop/ntua/marida
• —Zenodo archive: https://doi.org/10.5281/zenodo.5151941

Attribution

All credit for the dataset belongs entirely to the original MARIDA authors: Katerina Kikaki, Ioannis Kakogeorgiou, Paraskevi Mikeli, Dionysios E. Raitsos, and Konstantinos Karantzalos (National Technical University of Athens).

This repository only converts the official release's segmentation masks and multi-band reflectance into a bounding-box/RGB layout for improved usability with standard object-detection training pipelines.

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

License

MARIDA is distributed under Creative Commons Attribution 4.0 International (CC BY 4.0), as stated explicitly on the official Source Cooperative listing and in the dataset's own README.

Accordingly:

• —Attribution to the original creators is required.
• —Commercial and non-commercial use are both permitted.
• —No share-alike obligation.

The underlying imagery is Sentinel-2, distributed by the European Space Agency under the Copernicus open-data programme (free and open for any use), so there is no second-order sensor-rights restriction on top of the dataset's own CC BY 4.0 terms.

This repository is distributed under the same terms as its source (CC BY 4.0).

Citation

If you use this dataset, please cite:

@article{kikaki2022marida,
  title={MARIDA: A benchmark for Marine Debris detection from Sentinel-2 remote sensing data},
  author={Kikaki, Katerina and Kakogeorgiou, Ioannis and Mikeli, Paraskevi and Raitsos, Dionysios E. and Karantzalos, Konstantinos},
  journal={PLOS ONE},
  volume={17},
  number={1},
  pages={e0262247},
  year={2022}
}

Acknowledgements

We sincerely thank Katerina Kikaki, Ioannis Kakogeorgiou, and their co-authors for creating and publicly releasing this valuable marine-debris monitoring benchmark, and Source Cooperative / Radiant Earth for hosting the official data.