Φ-week 2021: artificial intelligence and quantum computing
In three key sessions today, the Φ-week event dealt with the real nitty-gritty of advanced computational methods in Earth observation. Contributors discussed Artificial Intelligence for Earth Observation and Science (AI4EO and AI4SCIENCE), along with the emerging potential of quantum computing.
This morning’s session was entitled ‘AI4EO – Learning from Earth Observation Data to Understand Our Planet’. In summing up the role of Artificial Intelligence (AI), session co-chair Pierre Philippe Mathieu (Head of the Φ-lab Explore Office) gave a simple definition: “When AI meets Big Data from space – that’s where the magic happens.” He went on to describe just why AI is becoming such an essential tool: “Classical methods for information retrieval are now reaching their limits as a result of the volume and variety of data generated. AI offers a completely new view of acquired data, but also automates and accelerates data insight. One of the main aims of the session is to illustrate how to deliver the maximum value from the marriage of AI and EO data, while highlighting the limitations of the technology.”
Adopting AI methods is not without its challenges, and Φ-lab visiting professor Begüm Demir kicked off the session with a talk on the issues surrounding label noise in the training of Deep Learning algorithms. The presentation included proposals for identifying and reducing the impact of noisy image labels in land-cover annotations. The question of uncertainty was also taken up in talks on Machine Learning (ML) in climate change prediction and spatial mapping, while human development featured strongly in presentations on slum improvements, urban settlement and food security.
A feature of AI4EO which is central to the aims of the Φ-sat-1 experiment is filtering data directly on the satellite. A talk from Frontier Development Lab presented research on ML-driven unsupervised novelty detection which has, as Postdoctoral Research Fellow Valentina Zantedeschi from INRIA and University College London explained, already produced some promising results: “We’ve shown that unsupervised ML techniques improve the detection of relevant change while being less susceptible to noise than standard differencing methods. They also promise to reduce compute and memory requirements in terms of data processing and storage.”
“The broad range of talks has given an invaluable insight into the possibilities and hurdles associated with AI in EO,” commented the session’s other chair, ESA Research Fellow Rochelle Schneider. “I was particularly impressed by the diversity of our speakers and their shared passion for both innovation and pushing the boundaries of EO information processing.”
First up after lunch was the related topic of AI4SCIENCE. The session was chaired by ESA Open Science Platform Engineer Anca Anghelea: “AI is now showing enormous potential in helping us answer some fundamental geoscience questions. Machine Learning in particular can play a major role in rapid knowledge discovery by learning patterns and models from data.” With contributions from NASA and Microsoft as well as academia, the talks covered explainable AI and causal discovery, big Earth systems science, sustainability and seasonal forecasting.
Recent advances in quantum computing are expected to unleash unprecedented computing power in the near future for processing and analysing EO data, for example from Synthetic-Aperture Radar (SAR), multispectral and hyperspectral imaging. Following the launch of ESA QC4EO at last year’s Φ-week, the final session of the group showcased some of the work carried out in this field in the last twelve months. CERN presented both the status of its research in quantum computing and the associated shared experiences for EO and particle physics. Other institutions taking part included ECMWF and Oxford University, with the latter’s Timothy Palmer presenting the imaginatively titled ‘Quantum Computing for Earth Observations: the Good, the Bad and the Noisy’.
“Quantum computing and artificial intelligence are new digital technologies that can be combined to get better EO products, and get them faster,” said ESA Digital Technologies Engineer and session facilitator Bertrand Le Saux. “We need to develop and harness the power of Quantum Machine Learning for EO Big Data analytics and, as today’s speakers have shown, we’re now establishing the groundwork for exploring how the unique potency of quantum computing can be used for extracting information from EO data.”
Photo Copyright: IBM Research