reefsupport/marine-data

Reef Support Marine Data

Underwater imagery with segmentation masks and bounding boxes for coral-reef computer vision, with the licence and attribution of every row recorded.

Dataset summary

Reef Support Marine Data v1.0 (October 2026) collects eight openly licensed sources into four ready-to-train configurations: benthic and scene semantic segmentation, instance segmentation and fish detection. Two of the sources contain Reef Support's own annotations, and one of them is Reef Support's own reef-survey imagery. Every row carries its licence and attribution, so provenance stays attached to the pixels. The release is built with the marine-data registry, which records the licence of every source before it is admitted.

All sources in this release are distributed by their authors under open licences (Apache-2.0, MIT, CC BY 4.0 or CC BY 3.0), and each requires attribution. Sources without a licence that permits redistribution are not part of it. The licences differ per source, so read Licensing before you share or publish results. That section records the licences as Reef Support read them from the primary sources and is not legal advice.

Configurations	4: coral-masks, scene-masks, instance-masks, fish-boxes
Rows	58,803 (train 44,142, validation 6,489, test 8,172). instance-masks and fish-boxes share images.
Tasks	semantic segmentation, instance segmentation, object detection
Format	Parquet shards (up to about 0.5 GB each) with images and masks embedded, streamable
Size	15.5 GB
Sources	8: Coralscapes, SUIM, UIIS, UIIS10K, USIS10K, Roboflow Aquarium Dataset, Seaview labels, Reef Support benthic
Version	v1.0 (October 2026), built 2026-10-06
Code	github.com/reefsupport/marine-data

Supported tasks

• Semantic segmentation of coral-reef benthos (coral-masks) and of underwater scenes (scene-masks).
• Instance segmentation of underwater objects and animals (instance-masks).
• Object detection of fish and other fauna (fish-boxes).

Labels are in English. Where a source and its labels differ in vocabulary, the native labels are kept.

Quick start

Load a configuration with the datasets library, in one of three ways. Every row carries its licence and attribution; keep them when you redistribute.

Quick look in a notebook: stream the split and read the first row.

from datasets import load_dataset

stream = load_dataset("reefsupport/marine-data", "coral-masks", split="train", streaming=True)
row = next(iter(stream))
print(row["licence"], row["attribution"])

Scripts and training: download the config once, cached afterwards. coral-masks is 12.6 GB (all splits are fetched); to try the pipeline first, use scene-masks (0.2 GB) or the one-source form below.

from datasets import load_dataset

ds = load_dataset("reefsupport/marine-data", "coral-masks", split="train")

One source only: download just that source's shards (file names are <split>-<source>-<n>-of-<total>.parquet).

from datasets import load_dataset

ds = load_dataset(
    "reefsupport/marine-data",
    "coral-masks",
    split="train",
    data_files={"train": "data/coral-masks/train-reef-support-seaview-labels-*.parquet"},
)

In a plain Python script, stopping a stream after a few rows can keep the process from exiting (an upstream issue in pyarrow's dataset scanner). Download the config or a single source instead.

Decode a mask and look up class names (semantic masks are uint8 index PNGs, 255 = ignore):

import numpy as np
from datasets import load_dataset

stream = load_dataset("reefsupport/marine-data", "coral-masks", split="train", streaming=True)
row = next(iter(stream))
mask = np.array(row["mask"])                                    # (H, W) class ids
names = {c["id"]: c["label_native"] for c in row["class_map"]}  # per-image id -> the source's own label

Boxes in fish-boxes are normalised to 0-1:

from datasets import load_dataset

stream = load_dataset("reefsupport/marine-data", "fish-boxes", split="train", streaming=True)
row = next(iter(stream))
print(row["boxes"][0])  # {"x_min": .., "y_min": .., "x_max": .., "y_max": .., "label": "fish"}

Select one source inside a configuration: ds.filter(lambda s: s == "coralscapes", input_columns="source").

Label mapping

The masks keep the native ids of each source, and those ids differ between sources, so marinedata.labels maps every source onto one of three shared schemes with fixed class ids. 255 is the ignore value in every scheme and, for partially annotated sources, it means "not annotated" rather than background, so mask the loss to labels.supervised_classes(source, scheme) for those sources.

Requires Python 3.10 or newer and git. The package is not on PyPI; install the tagged release. The hf extra adds datasets and Pillow:

pip install "marinedata[hf] @ git+https://github.com/reefsupport/marine-data@v1.0.0"

Stream a row and remap it with remap_row, which adds label, an HxW uint8 array:

from datasets import load_dataset
from marinedata import labels

stream = load_dataset("reefsupport/marine-data", "coral-masks", split="train", streaming=True)
row = labels.remap_row(next(iter(stream)), "benthic-coarse")  # adds `label`, an HxW uint8 array
names = labels.class_names("benthic-coarse")
present = sorted({int(v) for v in row["label"].ravel()} - {labels.IGNORE_INDEX})
print(row["source"], [names[i] for i in present])

In a script, download the config or one source as described under Quick start and pass its rows to remap_row the same way:

from datasets import load_dataset
from marinedata import labels

ds = load_dataset("reefsupport/marine-data", "scene-masks", split="train")
row = labels.remap_row(ds[0], "scene")

from datasets import load_dataset
from marinedata import labels

ds = load_dataset(
    "reefsupport/marine-data",
    "coral-masks",
    split="train",
    data_files={"train": "data/coral-masks/train-reef-support-seaview-labels-*.parquet"},
)
row = labels.remap_row(ds[0], "benthic-coarse")

The other configurations use the scene scheme. scene-masks rows work with remap_row as above:

from datasets import load_dataset
from marinedata import labels

stream = load_dataset("reefsupport/marine-data", "scene-masks", split="train", streaming=True)
row = labels.remap_row(next(iter(stream)), "scene")
names = labels.class_names("scene")
present = sorted({int(v) for v in row["label"].ravel()} - {labels.IGNORE_INDEX})
print(row["source"], [names[i] for i in present])

instance-masks rows carry a list of instances instead of one semantic mask, so map each instance label with labels.map_label (it returns None for a label the scheme does not cover):

from datasets import load_dataset
from marinedata import labels

stream = load_dataset("reefsupport/marine-data", "instance-masks", split="train", streaming=True)
row = next(iter(stream))
ids = [labels.map_label(row["source"], inst["label_native"], "scene") for inst in row["instances"]]
print(row["source"], ids)

fish-boxes rows have boxes and name their source in source_id:

from datasets import load_dataset
from marinedata import labels

stream = load_dataset("reefsupport/marine-data", "fish-boxes", split="train", streaming=True)
row = next(iter(stream))
ids = [labels.map_label(row["source_id"], box["label"], "scene") for box in row["boxes"]]
print(row["source_id"], ids)

Ids are the position in the class list, and 255 (ignore) applies to every scheme.

Scheme	Classes (id, name, meaning)	Configurations
benthic-coarse	0 HC hard coral; 1 MIL fire coral; 2 SC soft coral; 3 ALGAE algae; 4 ABIOTIC abiotic; 5 OTHER_FAUNA other fauna	coral-masks
coral-binary	0 NOT_CORAL every other mapped label; 1 CORAL hard, soft and fire coral	coral-masks
scene	0 BW background waterbody; 1 HD human divers; 2 PF plants and sea-grass; 3 WR wrecks and ruins; 4 RO robots and instruments; 5 RI reefs and invertebrates; 6 FV fish and vertebrates; 7 SR sea-floor and rocks	scene-masks, instance-masks, fish-boxes

Label policy. The schemes are set up for training on several sources at once. Dead coral is ignored by default: in benthic-coarse and coral-binary, a label whose condition is dead goes to 255 (the five dead coral classes of Coralscapes, and dead clam), because Coralscapes separates alive, dead and bleached coral while the Reef Support and Seaview masks record no condition. Bleached coral is alive and stays HC / CORAL. The Coralscapes class unknown hard substrate mixes rock with octocorals and is ignored in both schemes. A few Coralscapes labels the registry leaves open are mapped: sponge and anemone to OTHER_FAUNA, algae covered substrate to ALGAE. To opt out of the dead-coral default, pass exclude_conditions=(); ("bleached",) ignores bleached coral but keeps dead coral:

from marinedata import labels

table = labels.lut("coralscapes", "benthic-coarse", exclude_conditions=())  # dead coral is HC again

Partially annotated sources label a subset of the classes. For reef-support-benthic-own and reef-support-seaview-labels (HC and SC, or CORAL in coral-binary) everything else in the image, including real coral of other kinds, is 255. In coral-binary they supply positives and no NOT_CORAL examples. The instance and box sources (uiis, uiis10k, usis10k, roboflow-aquarium) are partial too: pixels outside the annotated objects are unannotated, not background. labels.is_dense(source) returns the flag.

Seaview masks include whole-frame hard coral. About 8 percent of the Seaview rows label the whole frame as hard coral. These frames are close-ups of dense coral, so their masks carry little boundary information. Weight or sample the source accordingly when you mix it with sources that outline individual colonies.

The class lists, the mapping of every native label and the full label policy are in docs/LABELS.md.

Dataset structure

Configurations

config	task	train	validation	test	total	GB
coral-masks	Benthic semantic segmentation	4,585	569	843	5,997	12.6
scene-masks	Underwater scene semantic segmentation	1,181	163	254	1,598	0.2
instance-masks	Underwater instance segmentation	19,409	2,284	3,603	25,296	1.4
fish-boxes	Object detection (fish and fauna boxes)	18,967	3,473	3,472	25,912	1.4

Splits are the upstream splits where the source provides them, otherwise a deterministic 80/10/10 hash of split_group.

coral-masks

Benthic semantic segmentation of coral-reef imagery: image, a uint8 index mask (255 = ignore) and a per-image class_map. The 5,997 images come from three sources with different label vocabularies. Native labels are kept and mapped to a shared taxon node (taxon_node) and a coarse group (coarse) where a mapping exists.

source	rows	train / val / test	licence
coralscapes	2,075	1,517 / 166 / 392	Apache-2.0
reef-support-benthic-own	1,226	930 / 129 / 167	CC BY 4.0
reef-support-seaview-labels	2,696	2,138 / 274 / 284	images CC BY 3.0, masks CC BY 4.0

scene-masks

Underwater scene segmentation from SUIM (8 classes including plants and seagrass, fish, reef, wrecks and divers): 1,598 images at 640x480 with uint8 index masks. Source: suim, MIT.

instance-masks

Underwater instance segmentation: image, a uint16 instance-id mask (0 = background) and instances, a list with the label and a normalised box for every instance id. 25,296 images.

source	rows	train / val / test	licence
uiis	4,622	3,932 / 690 / 0	Apache-2.0
uiis10k	10,042	8,035 / 0 / 2,007	Apache-2.0
usis10k	10,632	7,442 / 1,594 / 1,596	Apache-2.0

Images in this configuration originate from UIIS, UIIS10K and USIS10K, distributed by their authors under the Apache License 2.0. The original publications state that the images were gathered from public sources and earlier datasets, so copyright in individual images may rest with third parties. Rights holders can request removal via the issue tracker or the contact page.

fish-boxes

Fish and fauna photos with bounding boxes (x_min, y_min, x_max, y_max normalised to 0-1, native label). 25,912 images (mostly 640x480; sizes vary) with about 3.6 boxes per image (1.8 for usis10k, 7.6 for roboflow-aquarium). For the UIIS-family sources the boxes are derived from the upstream instance annotations, one box per instance. Labels are the upstream vocabularies (for example fish, corals, reefs, jellyfish, human divers). The build ran a deduplication gate: no image spans two splits, and no upstream test image is in train.

source	images	licence
uiis10k	10,042	Apache-2.0
usis10k	10,611	Apache-2.0
uiis	4,622	Apache-2.0
roboflow-aquarium	637	CC BY 4.0

Images from uiis, uiis10k and usis10k originate from UIIS, UIIS10K and USIS10K, distributed by their authors under the Apache License 2.0. The original publications state that the images were gathered from public sources and earlier datasets, so copyright in individual images may rest with third parties. Rights holders can request removal via the issue tracker or the contact page.

Data fields

coral-masks, scene-masks, instance-masks:

field	type	description
image_sha256	string	SHA-256 of the image bytes. Stable id and join key.
image	Image	RGB photo.
mask	Image (PNG)	Semantic: uint8 class id, 255 = ignore. Instance: uint16 instance id, 0 = background.
class_map	list of struct	coral-masks and scene-masks. Per image: id (value in the mask), label_native, taxon_node, coarse (may be null).
instances	list of struct	instance-masks only. id, label_native, taxon_node, coarse, bbox_xyxy_norm [x_min, y_min, x_max, y_max] in 0-1.
source	string	Source id, see Licensing.
licence	string	Licence of the row's data under its source.
annotation_licence	string	Licence of the annotations.
attribution	string	Text to reproduce when you use or share the row.
split, split_group	string	Split, and the group that is never divided across splits.
annotator_type	string	human for every configuration in this release.
width, height, lat, lon	int, float	Pixel size; WGS84 degrees (null unless the source provides them).

fish-boxes (this configuration uses the US spelling license):

field	type	description
image_sha256, image	string, Image	As above.
source_id	string	Source id.
license, licence_class, attribution	string	Licence, its class (open for every row) and the attribution text.
boxes	list of struct	x_min, y_min, x_max, y_max (0-1 of width and height), label.
points	list	Point annotations; empty in this release.
split, split_group	string	Split and non-splittable group.
upstream_id, upstream_url, lat, lon, depth_m, meow_realm, habitat	mixed	Optional context; null unless the source provides it.
width, height	int	Pixel size.

Note on the Seaview rows: the licence column of reef-support-seaview-labels rows reads exactly CC-BY-3.0-AU (images); CC-BY-4.0 (masks, Reef Support), and the attribution column cites CC-BY-3.0-AU for the images. The images are CC BY 3.0 (unported), as stated by the University of Queensland, and the attribution in the Licensing table is the one to reproduce.

Dataset creation

Sources

Eight sources contribute images and annotations; Licensing lists them with their licences and attribution. Coralscapes, the Seaview labels and Reef Support's benthic set cover coral-reef benthos. SUIM provides general underwater scenes. UIIS, UIIS10K, USIS10K and the Roboflow Aquarium Dataset provide fish, fauna and other underwater objects.

Annotation process

Annotations are the sources' own except where noted. Reef Support drew the masks for the Seaview images, and the benthic set is Reef Support's own imagery and annotation. The Seaview masks are decoded from stitched RGB renders with a fixed 5-colour palette (Hard Coral = 1, Soft Coral = 2, 0 = unlabelled). Boxes for the UIIS-family sources are derived from the upstream instance annotations. Every annotator_type in the mask configurations is human.

Build and verification

1. Source check. Each source is fetched and its licence is checked against the registry before it is admitted.
2. Conversion. The release builder writes sharded Parquet with images and masks embedded, per-row licence and attribution, class maps and splits. Each configuration ends with a CHECKSUMS.sha256.
3. Upload. Every file is checked (size and sha256 against CHECKSUMS.sha256) before and after upload, and row counts are compared with the build.

Each configuration has release notes in docs/<config>/: README.md, LICENSE, NOTICE and CHECKSUMS.sha256. To verify a local snapshot: grep ' data/' docs/<config>/CHECKSUMS.sha256 | sha256sum -c - (macOS: shasum -a 256 -c). The code is in the GitHub repository (src/marinedata, release builder in hf_export.py).

Personal and sensitive information

Some images show divers or aquarium visitors, and fish-boxes has human and human divers labels. Location fields (lat, lon) are null unless the source provides them.

Considerations and limitations

• Dropped rows. 35 coral-masks images were dropped because no class map could be built (reef-support-benthic-own 24 of 1,250, reef-support-seaview-labels 11 of 2,707). 12 instance-masks rows were dropped (uiis 6 of 4,628, uiis10k 6 of 10,048).
• Splits are not independently audited. The mask configurations were not checked for near-duplicates, so near-identical frames can fall on both sides of a split. fish-boxes ran a deduplication gate. The Seaview split_group (site partition plus the first 5 characters of the image id) is provisional.
• One image, several configurations. instance-masks and fish-boxes are built from the same UIIS, UIIS10K and USIS10K images and are split per configuration. Join on image_sha256 before evaluating across configurations.
• Vocabularies differ per source. label_native is the source's own label. The shared taxon_node and coarse mappings are partial and coarse is often null.
• Image provenance. For the UIIS-family sources, image rights are described under Licensing.
• Geographic and habitat coverage follows the sources (for example, the Coralscapes images come from Red Sea dive sites) and is uneven. Models trained here may not transfer to other regions or imaging conditions.

Licensing

Each source keeps its own licence. The table lists what is included from each source, the licence, and the attribution to reproduce. The full notices are in docs/<config>/LICENSE and docs/<config>/NOTICE. The per-row licence and attribution columns record the same information at row level. This is a statement of the licences Reef Support recorded, not legal advice.

Source	What is included	Licence	Required attribution
coralscapes	Images and masks, coral-masks	Apache-2.0	Sauder et al. 2025, The Coralscapes Dataset, ICCV 2025 Marine Vision Workshop
reef-support-benthic-own	Images and masks, coral-masks	CC BY 4.0	Reef Support, CC BY 4.0, https://reef.support
reef-support-seaview-labels	Images, coral-masks	CC BY 3.0	Gonzalez-Rivero et al., XL Catlin Seaview Survey, Seaview Survey Photo-quadrat and Image Classification Dataset, The University of Queensland, doi:10.14264/UQL.2019.930, CC BY 3.0. Changes: re-encoded into Parquet
reef-support-seaview-labels	Masks drawn by Reef Support, coral-masks	CC BY 4.0	Masks: Reef Support, CC BY 4.0
suim	Images and masks, scene-masks	MIT	Islam et al. 2020, SUIM. Copyright (c) 2020 Md Jahidul Islam. The MIT notice is reproduced in docs/scene-masks/LICENSE
uiis	Images and instance masks, instance-masks; images and derived boxes, fish-boxes	Apache-2.0	Lian et al., WaterMask, ICCV 2023, UIIS
uiis10k	as uiis	Apache-2.0	Li et al. 2025, UWSAM, UIIS10K
usis10k	as uiis	Apache-2.0	Lian et al., ICML 2024, USIS10K
roboflow-aquarium	Images and boxes, fish-boxes	CC BY 4.0	Roboflow, Aquarium Dataset, 2020, CC BY 4.0, https://public.roboflow.com/object-detection/aquarium

Changes made by Reef Support to every source: images and masks were re-encoded into Parquet, class_map or instances were added, and splits and split_group were assigned. Where a row's annotations are Reef Support's own, they are licensed CC BY 4.0.

UIIS, UIIS10K and USIS10K. Images in these configurations originate from UIIS, UIIS10K and USIS10K, distributed by their authors under the Apache License 2.0. The original publications state that the images were gathered from public sources and earlier datasets, so copyright in individual images may rest with third parties. Rights holders can request removal via the issue tracker or the contact page.

Seaview images. The University of Queensland record for the Seaview Survey dataset (UQ eSpace UQ:734799) states the licence as Creative Commons Attribution 3.0 International (CC BY 3.0). The Seaview rows of this release read CC-BY-3.0-AU (images); CC-BY-4.0 (masks, Reef Support) in their licence column, and CC-BY-3.0-AU in their attribution column for the images. The images are CC BY 3.0 (unported) as stated by the University of Queensland, and the attribution in the table above is the one to reproduce.

Citation

If you use this dataset, cite the release and the sources you use.

@misc{reefsupport_marine_data_v1,
  title        = {Reef Support Marine Data, v1.0},
  author       = {{Reef Support B.V.}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/reefsupport/marine-data}},
  note         = {Version 1.0 (October 2026). Licences differ per source; see the dataset card.}
}

BibTeX for the upstream sources

@inproceedings{sauder2025coralscapes,
  title     = {The Coralscapes Dataset: Semantic Scene Understanding in Coral Reefs},
  author    = {Sauder, Jonathan and Domazetoski, Viktor and Banc-Prandi, Guilhem and Perna, Gabriela and Meibom, Anders and Tuia, Devis},
  booktitle = {Proceedings of the International Conference on Computer Vision Joint Workshop on Marine Vision},
  year      = {2025},
  eprint    = {2503.20000},
  archivePrefix = {arXiv}
}

@article{gonzalezrivero2020monitoring,
  title   = {Monitoring of Coral Reefs Using Artificial Intelligence: A Feasible and Cost-Effective Approach},
  author  = {Gonz{\'a}lez-Rivero, Manuel and Beijbom, Oscar and Rodriguez-Ramirez, Alberto and others},
  journal = {Remote Sensing},
  volume  = {12},
  number  = {3},
  pages   = {489},
  year    = {2020},
  doi     = {10.3390/rs12030489}
}

@misc{gonzalezrivero2019seaview,
  title        = {Seaview Survey Photo-quadrat and Image Classification Dataset},
  author       = {Gonz{\'a}lez-Rivero, Manuel and others},
  year         = {2019},
  publisher    = {The University of Queensland},
  doi          = {10.14264/UQL.2019.930},
  note         = {XL Catlin Seaview Survey, UQ eSpace UQ:734799. CC BY 3.0}
}

@inproceedings{islam2020suim,
  title     = {Semantic Segmentation of Underwater Imagery: Dataset and Benchmark},
  author    = {Islam, Md Jahidul and Edge, Chelsey and Xiao, Yuyang and Luo, Peigen and Mehtaz, Muntaqim and Morse, Christopher and Enan, Sadman Sakib and Sattar, Junaed},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2020},
  eprint    = {2004.01241},
  archivePrefix = {arXiv}
}

@inproceedings{lian2023watermask,
  title     = {WaterMask: Instance Segmentation for Underwater Imagery},
  author    = {Lian, Shijie and others},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {1305--1315},
  year      = {2023}
}

@misc{li2025uwsam,
  title         = {Advancing Marine Research: UWSAM Framework and UIIS10K Dataset for Precise Underwater Instance Segmentation},
  author        = {Li, Hua and Lian, Shijie and Li, Zhiyuan and Cong, Runmin and Li, Chongyi and Yang, Laurence T. and Zhang, Weidong and Kwong, Sam},
  year          = {2025},
  eprint        = {2505.15581},
  archivePrefix = {arXiv}
}

@inproceedings{lian2024usis10k,
  title     = {Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset},
  author    = {Lian, Shijie and Zhang, Ziyi and Li, Hua and Li, Wenjie and Yang, Laurence Tianruo and Kwong, Sam and Cong, Runmin},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning (ICML)},
  series    = {PMLR},
  volume    = {235},
  pages     = {29545--29559},
  year      = {2024}
}

@misc{roboflow2020aquarium,
  title        = {Aquarium Dataset},
  author       = {{Roboflow}},
  year         = {2020},
  howpublished = {\url{https://public.roboflow.com/object-detection/aquarium}},
  note         = {CC BY 4.0}
}

Contact

• Questions, corrections and rights-holder requests: GitHub issues or the Reef Support contact page.
• Website: reef.support | Hugging Face: reefsupport
• Code, registry and release builder: github.com/reefsupport/marine-data

Reef Support builds tools to protect and restore marine ecosystems.