One Challenge, Many Challenges: Machine Learning for Mapping

OpenDRI is in Dar es Salaam, Tanzania this week for the annual conference of Free and Open Source Software for Geospatial (FOSS4G 2018). With over 1000 expected attendees, this large gathering of geospatial enthusiasts is a prime opportunity to learn and share about the latest technology in mapping.

Part of the pre-conference program, Tuesday morning OpenDRI is organizing a workshop to explore how machine learning technology – a subfield of artificial intelligence (AI) – can be used to support mapping workflows in humanitarian and disaster risk management (DRM) projects. The concept of training computers to automatically recognize patterns in aerial imagery is not a novel idea, but it’s only in the last few years that such algorithms can be run at scale…and without advanced training in remote sensing or computer science.

The workshop introduces participants to open source machine learning tools applied to mapping. Developers from ETH, Mapbox, Development Seed, and the SpaceNet program demonstrate how their algorithms work, explain the software and hardware requirements to run them, and show examples of how they were successfully applied in mapping workflows.

Sample output of building extraction with Robosat by Daniel Hofmann

Following the demo session, a user-centered design exercise guides participants in exploring ideas on how such technology can aid in humanitarian and DRM mapping tasks. The output is organized into thematic areas for discussion in the final roundtable, together with presentations from local stakeholders and mapping teams to capture the view on this technology from implementers in the field.

Topics of the roundtable discussion cover challenges around computing requirements, training data quality, scaling on large projects, working in disconnected environments, and ethical implications of automated mapping. Outcomes and lessons from the workshop will be used to inform a guidance note on machine learning that GFDRR will publish in fall 2018.

The FOSS4G workshop also introduces the new Tanzania Open AI Challenge, recently announced by our colleagues at WeRobotics, with the objective of automatically extracting building footprints and their structural conditions. Training and validation data for this challenge comes from the incredible work done by the Zanzibar Mapping Initiative (ZMI), a team of local innovators that, with support from the World Bank, mapped the entire island of Zanzibar with drones. Each building was then manually digitized, and their condition (e.g. complete structure, partial structure, foundations only) recorded as attributes. Such detailed and “fresh” geospatial data provides the local government with invaluable information for a variety of uses from urban planning, to calculating vulnerability of population and assets to natural disasters.

The Zanzibar Mapping Initiative Website

Realizing how the value of this unique dataset isn’t limited to traditional uses, we decided to select it for the Open AI challenge competition. While all drone imagery collected by ZMI is already published in OpenAerialMap, a portion of the 30,000 digitized buildings will be made available for challenge participants, and the remaining set used for testing and validating results. We’re excited to see how machine learning tools will be developed, or existing ones adapted, to solve this challenge. As more and more imagery is being collected, the need for automated interpretation becomes compelling, as well as designing careful integrations in traditional mapping workflows.

In this direction, while teams of mappers and local government stakeholders are mapping their cities through the Open Cities Africa initiative, we look for innovative ways to support their efforts with new technology. Working with developers from Geoinformation Technology (HeiGIT) at Heidelberg University we extended the OSMA platform to include visual analytics of potential gaps in OpenStreetMap data. This application use GHSL settlement data, automatically derived from satellite imagery, to show estimated completeness of OSM data. As some of the Open Cities Africa teams are also collecting their own drone imagery and training data, we design automated ways of processing and extracting information to augment what already compiled in OSM. To explore how machine learning and automation can be integrated in OSM mapping tools, we will also be participating in a design workshop with our colleagues at Humanitarian OpenStreetMap Team, at the National Museum, 11:00 am on Friday. 

Drone imagery over flooded areas of Monrovia collected by Open Cities Africa team

If you are in Dar es Salaam this week, join us to explore machine learning for mapping, and stay in touch with OpenDRI to be notified when the guidance note will be published.