Production release announcement for fAIr, HOT’s AI assisted mapping community service
07-20, 12:30–12:50 (Poland), Room CA7

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 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 for resilience and our responsibility for our communities and the role we play within humanitarian mapping.

The talk will announce the production release of fAIr tool that is available for wider OSM communities, illustrate the academic background and showcase the research developed in recent months to assess how the accuracy of the training process performs, evaluating the accuracy metric currently used and comparing against different metrics scenarios.

Why fAIr? The Humanitarian OpenStreetMap Team (HOT) sees that mappers can, on average, map between 1000-1500 buildings per working day without AI assistance. During an AI-assisted mapping pilot (2019-2020) supported by Microsoft, 18 million building footprints were extracted from satellite imagery for all of Tanzania and Uganda. HOT discovered during this pilot that this average mapping nearly doubled to 2500-3000 buildings being added to OpenStreetMap (OSM) per day with the assistance of high-quality AI open-source datasets.

fAIr seeks to solve three foreseen problems:

  1. AI models openness: AI-assisted mapping for humanitarian purposes feels like a black box. Useful open-source results coming from AI exists (e.g. META's global roads dataset available in RapiD, Microsoft’s global buildings dataset and Google’s open building in Africa). However, the models (code/training data) are currently not open-source.
  2. Model bias: Having model biases means predicting over satellite imagery would be biased toward the training dataset used to teach the AI model and the nature and quality of imagery is very different across the globe.
  3. Lack of feedback: There is no enhancement applied easily on the intelligence and accuracy of the AI models and humans are out of the loop when building the AI models, this due to AI models being either closed source or were built once and made available to end users so enhancements would require repeating the process from scratch.

What is behind fAIr?

The deep learning model behind fAIr is based on RAMP (Replicable AI for MicroPlanning), whose architecture and approach originates from an Eff-UNet model.

The research presented in this talk, falls within the broader spectrum of research on understanding the fine tuning process for geographic domain adaptation in image analysis validation, in particular for building footprints detection.

The talk will describe the procedure put in place to perform a reproducible analysis of the performance of the ML model training, from the training datasets selection process, the more technical details of the backbone model underneath, the choice of the metrics used to measure accuracy, and finally the analysis of the results obtained testing for different metrics.

Anna has a background in Physics of the Atmosphere and Remote sensing, with applications to environmental studies (water management and desertification), spatial inequalities and sustainable transport. Currently based in the UK, working on ML for buildings footprints detection in collaboration with HOT (Humanitarian Openstreetmap Team). She is interested in anything maps and open source.

Omran NAJJAR is the AI Product Owner in Humanitarian OpenStreetMap Team with 15 year experience in software engineering, information systems, advanced data management and artificial intelligence. Omran holds a MSc in EEE, Computer Science, Turkey (2020) with research thesis and experimental application on Extreme Learning Machine (ELM) algorithms and a MSc in Data Analytics and Information Systems Management, Germany (2023) with research focus on spatial and temporal analysis on OSM - Nepal. Omran has been working in humanitarian and development context since 2014, specialized in monitoring and evaluation, Information and communication technology and AI for social good. Currently, Omran uses that experience to pursue justice, ethical and open source tech to amplify the connection between human[itarian] needs and open map data.

Kshitij Raj Sharma is a product owner for map data access services at HOTOSM from Nepal, He is a passionate spatial developer with a love for open-source software and open data. His expertise in spatial data and deep interest in mapping led him to explore the potential of AI. Kshitij has been experimenting and advocating for FOSS and contributing to open data initiatives for the past seven years. He is also a maintainer of several open-source tools including free and open-source AI tool: fAIr, OSM Export Tool, Raw Data API, geojson2osm, OSMSG etc and an founding member of OSGEO Nepal