FDL Europe is a research collaboration between the ESA’s ɸ-Lab, ESA Esrin and Trillium Technologies and the University of Oxford, in partnership with private industry players such as Google Cloud, Nvidia – Scan, D-Orbit and Planet, in which the latest tools and techniques in Artificial Intelligence (AI) and Machine Learning (ML) are applied to research priorities in support of ESA science and exploration.
The primary vision of FDL is to advance the application of machine learning, data science, and high performance computing to problems of material concern to humankind. As such, the work of FDL is conducted in the spirit of collaboration for mutual and universal benefit, with the full intention of being published and open sourced according to academic best practice
FDL Europe has consistently developed state-of-the-art in AI4EO applications with a close-eye on the needs of beneficiaries, such as UNICEF, the World Food Program and UNOSAT. FDL’s work has informed the development of Digital Twin for Earth (DTE) initiatives as well as extending the capabilities of mission operations. FDL’s work has also supported climate science research and disaster response.
Illustrative successes include ‘WORLDFLOODS’ – a Machine Learning pipeline and globally labelled dataset able to provide a vector segmentation of flooding, land and clouds from orbit and rapidly return the results to Earth. The dataset has been downloaded by researchers around the world. In mid-2021, WorldFloods flew as a ML payload on D-Orbit’s Wildride mission, running Unibap’s SpaceCloud hardware. WorldFloods was able to provide a vector segmentation of a Sentinel 2 flood image for the first time. The pipeline was also retained for an RGB camera on D-Orbit’s Ion spacecraft and the model successfully re-uploaded, demonstrating for the first time the potential for both image vector compression and the porting of ML between instruments in orbit.
FDL’s work Earth Digital Twinning and Cloud Classification have also shown the potential of ML to efficiently emulate high-performance-computing (HPC) to better understand, predict and improve the resolution of complex planetary processes – the latter winning ‘best paper’ at the prestigious ClimatechangeAI workshop at NeurIPS, where both projects were showcased. Both research projects produced benchmark datasets (‘RAINBENCH’ and ‘CUMULO’) available for the community.
FDL is a year-long research initiative that produces published research in peer-reviewed journals, however the ML pipelines are developed during an intensive nine week sprint hosted at ESA and Oxford University during the summer – the best time to broker interdisciplinary teams of Phd-level researchers from around Europe.
What makes FDL different is the emphasis on high-risk / high-reward applied AI research that takes bold concepts from low TRL (Tech Readiness Levels) through to published research and, in some cases, deployed ML pipelines, while in the process developing enhanced data products and shareable tools for the research community to build and improve on, such as ‘SatExtractor’ an open source, cloud native public satellite data extractor, allowing the rapid development of diverse analysis ready data.