Kicking Off 2019, Strengthening Collaborations, Advancing Machine Learning
By Hamed Alemohammad, Lead Geospatial Data Scientist, Radiant Earth Foundation
As a neutral entity, Radiant Earth Foundation enjoys the unique opportunity to work with diverse partners worldwide and recognizes the important role it plays in engaging and facilitating people, as well as organizations, to work together for the benefit of the development community.
As the community around machine learning for Earth observations continues to grow, we are honored by the leadership role Radiant Earth Foundation is able to play, leveraging our neutral standing and drawing on our convening capability.
That’s why in December we convened machine learning and remote sensing experts from around the world for an evening of networking and sharing of ideas in Washington. D.C., on the heels of the 2018 American Geophysical Union’s Fall Meeting. Sponsored by Omidyar Network, the community-building event included more than 80 academic, government, business, and non-profit representatives.
The reception also provided the opportunity to introduce Radiant Earth Foundation’s new initiative, ML Hub Earth, an open source, machine learning commons for Earth observation. ML Hub Earth will lead to a sustained, community-wide central library of labeled training data, models and standards. While this library supports the global good through the development of advanced, open source analytical tools, the priority efforts include different applications, including crop type and land cover classification.
Peter Rabley, venture partner at Omidyar Network, was with us at the event and underscored the need to build capacity for the global development community through collaborative initiatives, specifically those that result in labeled images and algorithms to advance machine learning for the Global South.
We also discussed how the Foundation is actively assembling the first of these open library datasets to identify major crops in Africa. The team is pulling together the best ground truth data on crops grown in Africa and creating a training dataset using Sentinel-2 imagery. Furthermore, a second training dataset for land cover classes on a global scale is being developed. We look forward to further engaging the community through providing an expert crowdsourcing tool to label satellite imagery on Radiant Earth Foundation’s platform.
“ML Hub Earth intends to be the catalyst for democratizing machine learning applications.”
Learning more about ML Hub Earth
The training datasets discussed above will be hosted on ML Hub Earth with a Creative Commons license, leading to a living, open image library for machine learning and Earth observations.
ML Hub Earth intends to be the catalyst for democratizing machine learning applications. As such, we call for participation from all interested groups and organizations that would like to work on open source training datasets or models using Earth observations.
You can join the conversation via the slack channel or contact the team via email@example.com
Capitalizing on surging interest
Interest in machine learning and Earth observation has surged in recent years, partially because of the potential it holds to create value from big data and to generate solutions for persistent global challenges.
Dr. Keely Roth, senior remote sensing scientist with The Climate Corporation, summed up the views of many describing the future of machine learning as, “The opportunity to implement scientific solutions at new scales and worldwide.”
Ph.D. candidate Cascade Tuholske expressed his interest in using machine learning to map poverty in cities more quickly and effectively.
Attendees also discussed the opportunity to examine more of the world’s data, improving its usability across the development sector and geographic areas. Interests identified at the event included the ability for machine learning to achieve global coverage for building footprints and roads, to strengthen the operational applications of Earth observation for decision-support, and to create global land cover and land use change maps.
“Machine learning models, and more open source and benchmark training data matched with the source imagery are required for different applications.”
What’s next for machine learning in Earth observation
Many of the reception participants stated that a key priority for 2019 should be engagement, ensuring that a diverse group of stakeholders actively contribute to open data advancements, specifically machine learning. Facilitating engagement is particularly critical to addressing challenges of the Global South, where a community-wide effort is needed to ensure new models are built and targeted given geographical challenges and opportunities.
It was noted that machine learning models, and more open source and benchmark training data matched with the source imagery are required for different applications. Achieving this goal requires different stakeholders to contribute to open source training data, as well as standards for cataloging these datasets.
Additional participant ideas to further advance machine learning in Earth observation
- Developing and publishing open training datasets matched with imagery at different spatio-temporal resolutions;
- Creating a central registry and repository of labeled training datasets;
- Identifying benchmark training datasets to advance machine learning models;
- Compiling best practices in machine learning modeling for remote sensing data;
- Documenting best practices for data management and cataloging;
- Surveying on-going projects to identify existing opportunities and eliminate duplicate efforts; and
- Developing and organizing tutorials, webinars and training workshops on machine learning techniques for Earth observations.
Thanks to all who attended the event and have already contributed to ML Hub Earth. We look forward to strengthening and expanding the machine learning for Earth observations community in the coming year.
Cheers to all for a collaborative 2019!