| CI/CD |   |
|
|---|---|---|
| Tech Stack |       |
|
| Code Style |   |
|
| Community |   |
|
| Submodules |   |
|
| Other Info |     |
fAIr is an open AI-assisted mapping service developed by the Humanitarian OpenStreetMap Team (HOT) that aims to improve the efficiency and accuracy of mapping efforts for humanitarian and development purposes. The service uses AI models, specifically computer vision techniques, to detect objects such as buildings, roads, waterways, and trees from satellite and UAV imagery.
The name fAIr is derived from the following terms:
- f: for freedom and free and open-source software
- AI: for Artificial Intelligence
- r: for resilience and our responsibility for our communities and the role we play within humanitarian mapping
For mappers/project managers leading mapping projects
- Map object detection using open-source ML models, open imagery and open community-led training and validation data
- Use of open-source satellite and UAV imagery shared on HOT's OpenAerialMap (OAM) to detect map features
- Trainimg of base ML models using localised datasets for the creation of localised AI models that better fit the local context
- Constant feedback loop to eliminate model biases and ensure models are relevant to local communities
For model developers
- Sharing open ML models and making them available to the open mapping community
Unlike other GeoAI data platforms, fAIr is a free and open-source AI service that allows open mapping community members to access open AI models shared by model developer on the platform, and create and train their own AI models for mapping in their region of interest and/or humanitarian need. The goal of fAIr is to provide access to AI-assisted mapping across mobile and in-browser editors, using open ML models to ensure that the models are relevant to the places where communities aim to generate map data.
To eliminate model biases, fAIr is built to work with the local communities and receive constant feedback on the models, which will result in the progressive intelligence of computer vision models. Whenever an mapper uses the AI models for assisted mapping, fAIr can take those corrections as feedback to enhance the AI model’s accuracy.
The ML models suggest detected features to be added to OpenStreetMap (OSM) or other open datalayers.
| Status | Feature | Detailed Description | Release |
|---|---|---|---|
| ✅ | AI Community Contribution | AI developers can add their own models to fAIr (HOT, HeiGIT and Mozilla AI models available so far) | |
| ✅ | fAIrSwipe training datasets | Model creators can ask the MapSwipe community to help create training data (24 field projects in 15 countries) | |
| ✅ | fAIrSwipe validation of AI predictions | The community validates AI predictions in MapSwipe (47 projects, 850k+ results) | |
| ✅ | Documentation site | New user guides and project progress at hotosm.github.io/fAIr | |
| 🔄 | Start Mapping ("Try fAIr") | Pick any model, apply it to any imagery, and see the results | |
| 🔄 | Map larger areas | Request predictions over larger areas in one go | |
| 🔄 | Public prediction results | Share prediction results in a public list | |
| 🔄 | Private/Public Datasets/Models | Choose whether datasets and models are private or public | |
| 📅 | 15+ ready-to-use models | Production release with at least 15 different models to choose from | |
| 📅 | Open training data | Publish manually verified training data from fAIrSwipe and Open Mapping Gurus | |
| 📅 | Integrate Tasking Manager | Use validated Tasking Manager tasks as training data for building detection models | |
| 📅 | Clone model | Copy a model and its dataset to improve it further | |
| 📅 | Community discussion | Comment and give feedback on public models |
See the high level roadmap and milestones for additional details.
| Status | Feature | Detailed Description | Release |
|---|---|---|---|
| ✅ | Adopting YOLOv8 model | Improvements to the prediction algorithm | v2.0.1+ |
| ✅ | New UI/UX | redesign to enhance the user experience | v2.0.10+ |
| ✅ | fAIr evaluation | detailed research with Masaryk University & Missing Maps Czechia and Slovakia, welcome to join the efforts, here is the final report | no release |
| ✅ | Handling User Profile | Enable users to log in easily and have insights in their user activity, their own models/datasets and submitted trainings | v2.1.0 |
| ✅ | Notifications features | Training status change would trigger a notification on the web/email to let user know training is finished successfully or with a failure | v2.1.3 |
| ✅ | Replicable Models | Enable users to run a pre-trained model on new imagery/on a different area of their choice and using different satellite imagery | v2.2.0 |
| ✅ | Offline AI Prediction | Enable users to submit requests for prediction using any pre-trained model and any imagery and process it in the background and provide the results back to user. | v2.2.3 |
| ✅ | Post Processing Enhancement | Users would get enhanced geometry features (points/polygons) based on the need of the mapping process | v2.2.4 |
| ✅ | fAIrSwipe | Enable users to validate fAIR generated features and push them into OSM by integrating fAIr with MapSwipe, more details | v2.2.15 |
|👀| You can follow here the details and scope of each of the above features. and you can see and follow the Figma design progress for current in development 🔄 features
A higher level roadmap for 2025 can be found on Github.
- First We expect there should be a fully mapped and validated task in project Area where model will be trained on
- fAIr uses OSM features as labels which are fetched from [Raw Data API] (https://github.com/hotosm/raw-data-api) and Tiles from OpenAerialMap (https://map.openaerialmap.org/)
- Once data is ready fAIr supports creation of local model with the input area provided , Publishes model for that area which can be implemented on the rest of the similar area
- Feedback is important aspect , If mappers is not satisfied with the prediction that fAIr is making they can submit their feedback and community manager can apply feedback to model so that model will learn
The backend is using library we call it fAIr utilities to handle:
1. Data preparation for the models
2. Models trainings
3. Inference process
4. Post processing (converting the predicted features to geo data)
Checkout Docker Installation docs
- Start by reading our Code of conduct
- Get familiar with our contributor guidelines explaining the different ways in which you can support this project! We need your help!
By submitting imagery links to fAIr for model creation, you:
- Grant fAIr permission to download tiles covering your specified area of interest.
- Authorize fAIr to use these tiles for training and inference.
- Allow fAIr to redistribute the downloaded tiles to anyone who wishes to view or reproduce the dataset used for model training.
- Training: Image tiles will be downloaded, stored, and will be published as part of the training dataset for the specified area only (Not the full image).
- Inference: Image tiles will not be stored. They are only cached temporarily during prediction and removed after. Only the results, the prediction area, and the source imagery reference are stored.
- The original copyright remains with the imagery’s source or rights holder.
- You grant fAIr the right to license the downloaded tiles under CC BY 4.0.
- If you are using a commercial TMS (Tile Map Service) with your own token, fAIr will download, store, and derive information from the tiles for training in your specified area.
- For inference, tiles are cached only for prediction and removed afterward.
You must verify that your imagery provider’s license is compatible with fAIr’s intended use.
- When submitting imagery to fAIr, ensure you are not violating the license of the TMS or imagery provider.
- If you are using imagery from OpenAerialMap, review the legal page for applicable terms.
- If you plan to use the API or imagery services beyond the scope of this license, reach out to info@hotosm.org for guidance.
By submitting labels to fAIr for model creation, you:
- Grant fAIr permission to use the labels you provide for training.
- Allow fAIr to redistribute these labels under the same license applied to the training area you specify.
- Acknowledge responsibility for the content of the labels you share.
- All labels used and shared by fAIr are licensed under the ODbL 1.0.
- This is the same license used by OpenStreetMap.
- If labels are pulled from OpenStreetMap, no user account information is included.
- Only the attributes and geometry are extracted for use.
- If you upload your own labels, you must comply with ODbL.
- You are fully responsible for the data you submit and what information you choose to include or exclude.
- Make sure the labels you provide can legally be shared under ODbL.
- fAIr will redistribute labels only under ODbL, making them available for others working with the same training area.