ii2030 Challenge: AI for Earth Observation
Identifying sustainable business models for open machine learning ready Earth observation repositories
When Gedeon Jean* first realized the power of Earth observation (EO) data to detect environmental changes, he was mesmerized. As a Machine Learning Research Engineer, he saw the potential of combining machine learning (ML) and EO to develop diverse predictive applications for Rwanda, his native country. Rwanda is increasingly experiencing natural disasters due to climate change, including landslides, floods, and earthquakes, which take a socio-economic toll on an already vulnerable population.
Gedeon decided to write a ML algorithm that could find patterns in remote sensing data to forewarn when flooding could occur. Flood prediction models map flood-prone areas and can improve the accuracy of early warning systems to minimize the destruction of natural habitats, food sources, infrastructure, and loss of human lives, which is associated with floods. But building the application proved to be trickier as he struggled to find high-quality labeled training data that were ML-ready, on which he could train his model. In addition, the lack of training data relevant to his area of interest is known to produce biased or incorrect results. Training data, the building block for ML algorithms, needs to capture the geographical diversity of real-world data to help the model identify patterns more accurately.
Things changed when Gedeon discovered Radiant MLHub, an open library aggregating training data, models, and standards for ML applications on Earth observation data. Radiant MLHub hosts various training datasets and models for diverse applications relevant to advancing the Sustainable Development Goals (SDGs). The open access to Radiant MLHub training data made it easier for Gedeon to focus his efforts on training a model based on the SEN12 flood detection dataset. It gave him a starting point to test his model’s validity, the knowledge thereof, which he then transferred to his area of interest. Using the same approach with datasets found on Radiant MLHub, Gedeon is currently writing an application that various stakeholders in Rwanda can use to estimate crop yield, crop diseases, fertilizer maps, and crop life cycle analysis. This application can give farmers information on crop growth, land productivity, and support market pricing and insurance decisions. For policymakers in countries that depend on agriculture, this information is essential to support market intelligence and import and export decisions.
The wide variety of high-quality training datasets that Gedeon has access to allows him to explore various solutions and innovate on a local scale. Imagine always having access to high-quality data for use with ML models to build applications that can predict relevant information on soil, water, plant growth, pests, disease, weather-related damages, and climate impacts, to list a few examples? How would innovative solutions scale on a local level?
Radiant Earth Foundation, which manages Radiant MLHub, facilitates the publication and uptake of geospatial training datasets to help individuals like Gedeon work more efficiently on applications related to the SDGs. However, building geographically diverse quality datasets and maintaining the infrastructure of these freely and openly accessible ML-ready repositories requires resources.
Building, deploying, and maintaining ML-ready geospatial training data and models comes at a price tag out of reach, especially for non-commercial organizations working to achieve a more sustainable future. If everyone across the community has access to high-quality benchmark training data, they can focus their resources on advancing models and applications to serve their target community. As a result, we can address more considerable challenges and serve broader communities.
Open repositories such as Radiant MLHub fill that gap for all users of ML on EO working to tackle global challenges such as food (in)security, conservation, climate modeling, and many other relevant challenges. It empowers a data ecosystem that curates and publishes training data, models, and knowledge for practitioners and organizations globally.
To maintain Radiant MLHub, Radiant Earth has successfully generated funds through grants and other philanthropic support for the last couple of years. Furthermore, there is a growing discussion and call to action across the community to invest in the generation and publication of benchmark geospatial training datasets. This leads us to a critical question:
How might we ensure open access to high-quality ML-ready Earth observation data on a sustainable basis?
To help answer this question, we’ve designed a new inclusive innovation 2030 (ii2030) challenge with support from the German Federal Ministry for Economic Cooperation and Development (BMZ), digilab, Endeva, and FAIR Forward. ii2030 is a proven methodology to catalyze systems change and leverage technology to achieve the SDGs. Rather than creating another tech solution, the ii2030 challenge will focus on a strategic collaboration method between a diverse group of innovators to co-create ongoing open access to EO ML data on a sustainable basis for Radiant Earth and other players working on public data repository ecosystems. The process consists of a consultation phase, bringing key stakeholders together to analyze the current system and identify leverage points. The co-creation phase will follow to ideate new and sustainable solutions for this problem.
Read the concept note to learn more about being part of this journey, and mark your calendars to join the two online consultation sessions on September 29 and October 4 at 9:00 AM EST / 3:00 PM CEST. If you have any questions, contact us at p.golbach@endeva.org.
*Name has been changed to protect the identity of the individual.
