Datasets¶

fAIr's models are trained on open datasets, and validated predictions are published as open data. This page explains how training data is created and lists the datasets HOT publishes on Hugging Face and the Humanitarian Data Exchange.

Most datasets combine imagery under CC-BY-4.0 with labels derived from OpenStreetMap under the ODbL (the Open Database License).

How training data is created¶

Training labels come from mapping in the Tasking Manager and from MapSwipe, and community mappers in the Open Mapping Gurus program create datasets that are released openly. Their deployments are listed below, and fAIrSwipe projects are on the Field projects page.

Open Mapping Gurus¶

Community mappers in the Open Mapping Gurus program, facilitated by HOT regional hubs (WNA: West and Northern Africa; ESA: Eastern and Southern Africa; LAC: Latin America and the Caribbean; AP: Asia-Pacific), map training data on the Tasking Manager. Each completed deployment links its released training datasets, described under Open Mapping Gurus datasets. Deployment status as of 13 October 2026.

#	Location	Hub	Features	Status	Tasking Manager	Training datasets
1	Timor-Leste	AP	Seagrass	Complete	55909, 55908, 55911, 55910	seagrass
2	Freetown, Sierra Leone	WNA	Trees, solar panels, buildings (roof type and colour), highways	Complete	50661, 60363	buildings, highways, road surface, trees, solar
3	Nairobi, Kenya	ESA	Buildings (roof type and colour), highways (surface type)	Complete	50793	buildings, highways, road surface, roof type
4	Stellenbosch, South Africa	ESA	Buildings (roof type and colour), highways (surface type)	Complete	40695	buildings, highways, road surface, roof type
5	Bogotá, Colombia	LAC	Buildings (roof type and colour), highways (surface type), water tanks, sports facilities, trees	Ongoing	66831
6	Payatas, Quezon City, Philippines	AP	Buildings (roof type and colour), highways (surface type), streetlights, electric posts, coconut trees, solar panels	Ongoing	65381, 65412, 65413, 65414
7	Borongan, Eastern Samar, Philippines	AP	Buildings (roof type and colour), highways (surface type), zebra crossings, solar panels, swimming pools	Ongoing	49805, 49839, 49838, 49837
8	Monkayo, Davao de Oro, Philippines	AP	Buildings (roof type and colour), highways (surface type), zebra crossings, streetlights, electric posts, coconut trees, solar panels	Ongoing	66864
9	Zimbabwe	ESA	Buildings (roof type and colour), highways (surface type)	Ongoing	66897
10	Malawi	ESA	Buildings (roof type and colour), highways (surface type)	Ongoing	67128
11	Sanur, Bali, Indonesia	AP	Parking lots, solar panels, swimming pools, trees	Planned
12	Parang Island, Mindoro, Philippines	AP	To be confirmed	Planned

Training datasets¶

Datasets used to train and evaluate geo-AI models, hosted on Hugging Face. Download totals are on Milestones.

Dataset	Feature	Task	License
hot-building-segmentation	Buildings	Image segmentation	CC-BY-4.0, ODbL
vhr-building-segmentation	Buildings	Image segmentation	CC-BY-4.0, ODbL
streetlevel-poles	Poles and towers	Object detection	CC-BY-SA-4.0
vhr-highway-segmentation	Highways	Image segmentation	CC-BY-4.0, ODbL
vhr-seagrass-segmentation	Seagrass	Image segmentation	CC-BY-4.0, ODbL
fAIr-guru-td-buildings	Buildings	Image segmentation	CC-BY-4.0, ODbL
fAIr-guru-td-highways	Highways	Image segmentation	CC-BY-4.0, ODbL
fAIr-guru-td-road-surface	Road surface	Image segmentation, image classification	CC-BY-4.0, ODbL
fAIr-guru-td-roof-type	Building roof type	Image segmentation, image classification	CC-BY-4.0, ODbL
fAIr-guru-td-solar	Rooftop solar panels	Image segmentation	CC-BY-4.0, ODbL
fAIr-guru-td-trees	Trees	Object detection	CC-BY-4.0, ODbL
fAIr-guru-td-seagrass	Seagrass	Image segmentation	CC-BY-4.0, ODbL

The fAIr-guru-* datasets come from the Open Mapping Gurus program; vhr-seagrass-segmentation holds the Timor-Leste seagrass labels that fAIr-guru-td-seagrass is built from. The streetlevel-poles dataset is an early experiment with street-level imagery.

Open Mapping Gurus datasets¶

Each Guru dataset pairs drone and aerial imagery from OpenAerialMap with the features mapped in the deployment's Tasking Manager projects. Tiles come only from task squares marked as validated, and every tile is assigned to a train, validation, or test split. Each release stores its labels with the tiles, so later map edits leave a released dataset unchanged.

Sample tiles beside their labels, from each dataset on Hugging Face:

All Guru datasets are listed in one STAC catalogue. It opens in STAC Browser, and in the QGIS STAC panel with this URL:

https://huggingface.co/datasets/hotosm/fAIr-guru-stac/raw/main/catalog.json

The buildings dataset in STAC Browser, with one item per Tasking Manager project: Stellenbosch, Nairobi, and Freetown.

The Stellenbosch buildings in QGIS: imagery tiles outlined by split, with the building labels on top.

The datasets are built with fAIr-ds-builder, whose guides cover loading, viewing in STAC Browser or QGIS, custom splits, and building from other Tasking Manager projects.

Quality issues in a dataset, for example found while reviewing a project in QGIS, can be reported on the fAIr-ds-builder issue tracker.

Prediction and crisis outputs¶

Validated predictions and crisis-response data produced with fAIr, published as open datasets.

On the Humanitarian Data Exchange:

Dataset	Description	License
Nepal Flood 2026: fAIr Damage Assessment, Upper Trishuli and Bhote Koshi	Building damage assessment for the flood, produced with fAIr	CC BY (as listed on HDX)
Venezuela - M 7.5 Earthquake - Damage Assessment Area	Damage areas on an H3 grid from fAIr and Microsoft's AI for Good Lab detections, validated in MapSwipe	CC BY-NC-SA 4.0
Venezuela - M 7.5 Earthquake - June 2026 - OSM & Overture Data	OSM and Overture layers, with fAIr building predictions and building density	ODbL; Overture places CDLA Permissive 2.0

On Hugging Face:

Dataset	Description	License
venezuela_eq_2026	Venezuela M 7.5 earthquake, June 2026	Not specified
colombia_eq_2026	Colombia earthquake, August 2026	Not specified
nepal_flood_2026	Nepal Flood 2026 (Upper Trishuli and Bhote Koshi)	CC-BY-4.0