Skip to content

Models

fAIr's models are published in an open catalog. A base model is a pretrained model that anyone can run. An advanced user can fine-tune a base model on their own area to produce a local model, which is versioned and published back to the collection. Fine-tuning works through the API today; the screens for it are planned.

Models are contributed to the fAIr-models repository. Registered models and their metadata are listed in the base-models STAC collection, and can be run at dev.ai.hotosm.org; see What is fAIr for the versions.

Base models

The models available in fAIr today and those planned. "Available" names the version that runs the model: the current version (v3) or the earlier one (v2).

# Model Feature Task type Contributor Status
1 Buildings (DINOv3) Buildings Segmentation HOT Available, v3
2 Building damage (DINOv3) Building damage Segmentation HOT Available, v3
3 Trees (DeepForest) Trees Object detection HOT Available, v3
4 Trees (UNet, Freetown) Trees Segmentation HOT West and Northern Africa hub Available, v3
5 Solid waste Solid waste Segmentation HeiGIT Available, v3
6 Buildings (RAMP, YOLOv8) Buildings Segmentation HOT Available, v2
7 Trees (Omdena challenge) Trees Object detection Omdena In development
8 Solar panels Solar panels Object detection Omdena In development
9 Bridges Bridges Object detection Omdena In development
10 Roof type classification Buildings Classification Omdena In development
11 Seagrass Seagrass Object detection Omdena In development
12 Highway segmentation Roads Segmentation Zindi In development
13 Road surface (paved / unpaved) Roads Classification HeiGIT In development
14 Road surface damage Roads Object detection Community member In development
15 Water storage tanks Water tanks Object detection Togglecorp In development
16 Breeding containers Disease vectors Object detection HeiGIT In development
17 Buildings (U-Net) Buildings Instance segmentation Community member In development
18 Open ground Open spaces Segmentation NAXA In development
19 Roof material Buildings Segmentation NAXA In development
20 Tent detection Tents Object detection Community members In development
21 Swimming pools Swimming pools Object detection HOT In development
22 Rural roads Roads Segmentation Community member Planned
23 Surface water Water Segmentation Community member Planned
24 Parking spaces Parking spaces Object detection HOT Planned

The DeepForest trees model uses weights from Weecology, packaged for fAIr by HOT. The Nepal flood damage model on Activations is a local model tuned from the building damage model, so it is not listed here. The swimming pools model builds on Mozilla AI's pretrained swimming pool detector; it is under review in pull request 65, and its item is already in the base-models STAC collection. Progress across the model pipeline is tracked on the Milestones page.

Open geo-AI challenges

Several models are being built through two open geo-AI challenges in 2026, which open model development to a wider community:

  • Omdena (around 50 participants): trees, solar panels, bridges, roof type classification, and seagrass.
  • Zindi: highway segmentation, in a challenge planned to run from 13 November to 11 December 2026.

Contributions also come through the open call from HeiGIT (solid waste, road surface, breeding containers), NAXA (open ground, roof material), Togglecorp (water storage tanks) and community members, alongside HOT's own models.

Through a fellowship with the Lund AI Society, a student-run nonprofit, a team of students works from September 2026 to January 2027 on cleaner building outlines from the building segmentation model.

Need a different feature? Ask the team in #fair-coord on Slack, or contribute a model (see below).

Contributing a model

New geo-AI models are contributed through a pull request to the fAIr-models repository, following its contribution guide. After review and approval, a model is registered into fAIr and becomes available to every kind of user. The registration flow is described in Register a base model, and the architecture in ML pipeline.

What kind of models

The collection is made of base models: models pre-trained on a specific feature, such as buildings, roads, or flood damage, so they can be reused across locations. Contributions can be classical machine learning models (for example random forest or k-means applied to geodata) or models built on foundation models.

Models coupled to a specific task work best. For example, a DINOv3 trained for buildings, roads, or damage fits well and can be run on a mapper's area directly, which is the form the collection is built around.

Existing open-source geo-AI models can also be integrated through an open call; see Vision.

If you build geo-AI models and are interested in contributing, apply through the open call. fAIr is open to expansion in the following categories:

  • Residential areas
  • Critical public infrastructure
  • WASH infrastructure
  • Livelihood infrastructure
  • Transportation infrastructure
  • Essential utilities
  • Emergency and public service facilities

Other humanitarian features beyond this list are also welcome.