hotosm/fAIr-guru-td-seagrass

fAIr Open Mapping Gurus: seagrass

Seagrass segmentation training data from the Open Mapping Gurus programme, built from four Timor-Leste projects. Drone imagery from OpenAerialMap is paired with seagrass areas mapped and validated in HOT Tasking Manager.

At a glance

Tiles	4,497
Splits	train 3,600, validation 479, test 418
Split method	10 contiguous regions per project, 1 region to test, 1 region to validation; tiles at a region edge border tiles of another split; a feature crossing an edge has parts in both splits
Tiles with seagrass areas	4,121
Tiles without seagrass areas	376
Seagrass areas	27,420
Ground pixel size	0.02 m (zoom 23)
Area covered	0.10 sq km
Countries	Timor-Leste
Version	0.1.1, built 2026-10-26

Sample tiles

Each tile is shown beside its labels, coloured by class, with the main class and its count above.

Classes

Labels are features tagged seamark:type=seagrass, and a label's class is its seamark:type value. The wiki describes how each tag is mapped: Key:seamark:type, Tag:seamark:type=seagrass.

Class	Labels	Share
seagrass	27,420	100.0%

Source projects

Project	Country	Tiles	Seagrass areas	Ground pixel size
55908: Seagrass Mapping in Timor-Leste 2026 [Team Harish]	Timor-Leste	1,154	9,146	0.02 m
55909: Seagrass Mapping in Timor-Leste 2026 [Team Azhar]	Timor-Leste	1,127	3,638	0.02 m
55910: Seagrass Mapping in Timor-Leste 2026 [Team Mirna]	Timor-Leste	1,093	8,788	0.02 m
55911: Seagrass Mapping in Timor-Leste 2026 [Team Kiki]	Timor-Leste	1,123	5,848	0.02 m

How it was labelled

The labels are OSM Sandbox features mapped in HOT Tasking Manager projects (wiki). Tasking Manager divides a project area into small tasks, so many volunteers can map it together following the project instructions. A mapper traces the features in a task, and a validator then checks the task and marks it validated.

The imagery comes from OpenAerialMap (wiki), where people and organisations share openly licensed satellite, aerial and drone imagery through the Open Imagery Network. Each project sets this OpenAerialMap imagery for its mappers, and the tiles come from it.

The four projects cover one drone image and list LAND_USE and OTHER as mapping types. Mappers drew seagrass as areas tagged area=yes and seamark:type=seagrass and saved them to the OSM Sandbox, as the project instructions set out, so the labels are not part of OpenStreetMap. They come from the per-chip labels of vhr-seagrass-segmentation, joined back into whole areas by their sandbox id, and are published under label_files/. Tiles come only from task squares marked VALIDATED that lie inside the area covered by that dataset's chips. Masks are binary rasters of the areas.

Splits

The splits are spatial. Each project's tiles are grouped into contiguous regions, and each region goes whole to train, validation or test, as listed under Split method above. For model evaluation we recommend making your own split, as described in Make your own split.

Load

import json

from datasets import load_dataset
from rasterio.features import rasterize
from rasterio.transform import from_bounds

dataset = load_dataset("hotosm/fAIr-guru-td-seagrass", split="train")
labelled = dataset.filter(lambda count: count > 0, input_columns="num_features")
row = labelled[0]
image = row["image"]
labels = json.loads(row["label_geojson"])

bounds = (row["bbox_west"], row["bbox_south"], row["bbox_east"], row["bbox_north"])
transform = from_bounds(*bounds, image.width, image.height)
shapes = [feature["geometry"] for feature in labels["features"]]
mask = rasterize(shapes, out_shape=(image.height, image.width), transform=transform)

image is the 256x256 tile as served by OpenAerialMap. The bbox_* columns hold its bounds, which place label_geojson on its pixels.

Column	Description
image	RGB tile
mask	Binary raster of the labels, 0 background and 255 feature, in datasets built with masks
label_geojson	Labels in the tile with their OpenStreetMap tags, EPSG:4326
num_features, class_counts	Label count, and the count per class as JSON
tile_id, tile_x, tile_y, tile_z	XYZ tile identity
bbox_west, bbox_south, bbox_east, bbox_north	Tile bounds, EPSG:4326
project_id, project_name, project_url, country, organisation	Source Tasking Manager project
imagery_url, feature_type	OpenAerialMap tile URL, and the OpenStreetMap key of the classes

Browse the STAC catalogue

Each project is one item holding a Cloud Optimized GeoTIFF of its tiles, the labels and the tile footprints with their split. Open in STAC Browser to see the labels over the imagery, or follow the QGIS guide to open them in QGIS.

Known limitations

• The labels live in the OSM Sandbox, so they cannot be compared with or updated from OpenStreetMap.
• 9 to 13% of each project's validated area has no labels in vhr-seagrass-segmentation and is left out.
• All tiles come from one image, captured on 30 March 2026 according to its OpenAerialMap record, so validation and test measure new areas of that same image.
• 496 of the 16,876 seagrass areas (3%) carry a mapper note such as Unsure or check if seagrass. They are kept as labels, and the note is in their tags.

Sources and licensing

Component	Source	License
Imagery	OpenAerialMap	CC-BY 4.0, attribute contributors of Open Imagery Network
Labels	OSM Sandbox	ODbL 1.0, attribute HOT OpenStreetMap sandbox contributors

Built with fAIr-ds-builder.