Machine learning

From OpenStreetMap Wiki
Jump to navigation Jump to search

 Machine Learning is a subdomain of  Artificial Intelligence centered around the idea that machines can be trained to recognize patterns based on annotated input data. In the context of maps and OSM, a common application is to annotate a sample set of imagery (aerial or street-level) in order to automatically derive information and/or map features from a larger dataset. By the types of features it looks at, specific tasks can be:

  • Road extraction (sometimes takes GPX too!)
  • Building segmentation (footprints)
  • land cover classification, for landuse tagging
  • Other types of object detection.

As with many areas of computer vision, competitions are sometimes held to assess and encourage better models to be made. One example is DeepGlobe 2018.


This is a list of existing applications of Machine Learning designed to support OSM.

Open Source

Name Description Year Organisation External Links
Looking glass Tool to identify buildings within satellite imagery. It uses a segmentation algorithm to label each individual pixels in an image as to whether it's part of a building or not. 2018 Development Seed Source code
Mapping the Electric Grid Using machine learning to augment human tracing of high-voltage infrastructure 2018 Development Seed Documentation and Source code
MapSwipe ML Tools for user input from MapSwipe as input for ML applications. More specifically, judging whether an area is build up. 2017 Missing Maps Source code
ml-enabler A registry and facilitator for machine learning models for the OpenStreetMap ecosystem. 2019 Humanitarian OpenStreetMap Team and Development Seed Blog post and source code
OSM Analytics Gap Analysis Compares OSM data with machine learning derived prediction of built up areas to show where gaps in OSM data may exist. 2018 Humanitarian OpenStreetMap Team Source code and data set
RapiD An extension of the iD editor for mapping with AI-generated features. 2019 Facebook Website, source code, documentation
RoadTracer Automatic extraction of road networks from aerial images. 2018 MIT CSAIL Documentation and source code
Robosat Generic ecosystem for feature extraction from aerial and satellite imagery. 2018 Mapbox Source code
Telenav.AI Sign detection from street level imagery 2018 OpenStreetCam (Telenav) Blog post Source code

The following project uses other fields of artificial intelligence (finite state machines) rather than machine learning techniques:

Name Description Year Organisation External Links
Trip Simulator A tool for generating simulated raw GPS location telemetry. 2019 SharedStreets Source code

Propietary Software

Name Description Year Organisation External Links
Recognize and Label Objects in the Wild Object and feature detection from street level imagery (usable from iD) 2016 Mapillary Blog post
AI-Assisted Road Tracing Road detection from aerial imagery and making it available to the mapper through the RapiD editor. 2016 Facebook Wiki page
US building footprints Building footprints detected from ML 2018 Microsoft Blog post and data
Urchn Urchn is a tool for urban change detection. It shows you how a city is changing, and helps to keep the map up-to-date. 2018 Development Seed and Radiant Solutions Blog post

Relevant Training Data

  • offers "45 000 manually annotated images containing more than 55 000 signs divided in 23 different classes". A sample dataset with 1000 images is also available.
  • Mapbox allows all users to use Mapbox Satellite images for machine learning processes[1]

Conference talks and presentations

See also

  1. Pratik Yadav (18 May 2018). “Mapbox Satellite for Machine Learning”. “All Mapbox users, using an access token from their own account, are allowed to create derivatives from Mapbox Satellite for contribution to OpenStreetMap via manual or automated processes for free.”