Skip to content

How it works

fAIr has two connected cycles: a model loop, where models are shared, run, and improved, and a data pipeline, where imagery and labels become predictions that people validate and reuse.

The model loop

A model starts in the shared collection. A mapper runs it over their own area and reviews the output. When someone retrains a model for a specific area, the local version is published back to the collection, so the next mapper can start from it.

The loop has four steps:

  1. A collection of models is shared with everyone.
  2. A mapper runs a model on their own area.
  3. They validate the results: approve, fix, and feed back.
  4. Improved or locally adapted models flow back to the collection.

Developers build models, mappers use them, and advanced users retrain and republish local versions.

The three roles that meet fAIr at this loop are described in User bases.

The data pipeline

A model is only as good as the examples it learns from. fAIr takes labels and imagery, produces predictions, lets people validate them, and returns clean results.

The fAIr data pipeline: gather imagery and labels, train and predict, validate, then use and share the data.

1. Inputs

fAIr needs two things:

  • Imagery. Openly licensed, very high-resolution RGB imagery from OpenAerialMap (which uses CC BY 4.0) or an open TMS source (a tiled imagery server).
  • Labels. Examples that teach the model what to find: existing OpenStreetMap features in the area, finding where objects are with MapSwipe as a starting point, for those new to mapping, tracing in HOT's Tasking Manager for experienced mappers, or features drawn by hand for advanced users.

2. Inside fAIr

Labels and imagery are used to train a model. A trained model then predicts matching features across a chosen area and returns them as GeoJSON, a common open format for map features.

3. Validate

Predictions are reviewed before they are used. A mapper can check them on fAIr's map or in QGIS, verify them on the ground with FieldTM (with QField), or confirm them at scale through MapSwipe Validate Footprints projects; see Validation.

4. Use and share

Predictions can be downloaded as points, lines, or polygons. HOT publishes validated outputs as open data layers on the Humanitarian Data Exchange and Hugging Face. See Datasets for the published outputs.

Bringing AI-generated features into OpenStreetMap is outside fAIr's aim; see fAIr and OpenStreetMap.