The inaugural Space@Oxford conference
On 22nd September, around 200 delegates from 30+ organisations/companies and 80+ academic researchers from across the University of Oxford gathered at St Catherine’s College for the first-ever Space@Oxford conference.
The James Webb Space Telescope (JWST)
The chair of Space@Oxford, Prof. Neil Bowles opened the event, followed by the first presentation on early science using the first images from the James Webb Space Telescope (JWST) by Dr Gareth Jones. The University of Oxford and other researchers are already finding new ways to study galaxies thanks to JWST. The data (some of) are open for anyone of us to study (though, apparently, not so easy to use 🤔 🙃).
Panel Discussion ‘Beyond Earth: New Worlds, New Markets.’
I remember, ten years ago, when I was participating in the Space Studies Summer Program 2012 (SSP12), we talked about space as the final frontier and I got the book “Farthest Shore” by some of the organisers to read… 👩🚀 Today, space is becoming the commercial frontier, 🛰️ not such as far away shore anymore... Panellists including Laura Gonzales Llamazares and William Birch from Satelite Applications Catapult, as well as Portia Bowman from D-Orbit and Charlie Clark from Magdrive, discussed this new reality of the space field. In relation to the quite crowded Lower Earth Orbit (LEO) and space debris, the ideas of insurance and premiums on satellites were discussed (when someone hits “your” satellite, they have to pay). The fact that space is still underregulated was another point — from space mining to the fact that Elon Musk simply put Tesla into space and nobody could do anything about it. 🥲
AI Onboard
Next, we had parallel sessions, out of which I joined the one called ‘AI Onboard’. This was really interesting for someone like me (working on remote sensing applications for social computing) as a lot of the discussion revolved around AI for Earth Observation (EO). For example, delegates from Space Alliance discussed how adding AI onboard satellites will enable ‘passing around less data’ (as we can then send results instead of raw data from satellites) and that is one of the big hurdles nowadays. This AI must be Trusted, was another call from them. The delegate from Telespazio focused mainly on remote sensing. One of the first examples was how ML and AI helped to analyse a small lunar crater in the dark that helped to plan a rover entry.
In general, in EO, they argued that at the moment, there is too much data and too little labelled data. Even when we have labels, there is often class imbalance (e.g., flood mapping, or forest fires, which are rare events and cover a small portion of the area). Another data topic discussed was data fusion — sometimes there are useful things to be learned from lower-resolution imagery but how do we fuse it with the high-resolution one that has become available?
ESA representatives accented the need for explainable and trusted AI to be used by operators and ground segment engineers — as otherwise, at the moment, they are not open to using AI. Moreover, as agencies and companies collaborate, they at times require privacy-preserving AI.
Next was the presentation by Atılım Güneş Baydin, a Machine Learning Prof. from the University of Oxford who works with Space@Oxford and FDL (see my previous post about FDL 2018 in London). He presented four cases of, in my opinion, the most exciting and neat applications of ML for space. Case 1: ML for collision avoidance (e.g., in LEO). Case 2: Space weather and thermospheric density estimation. Case 3: ML onboard cube sat. A U-Net semantic segmentation model was fit to perform flood detection onboard. This one was launched into space. 🚀 Case 4: Onboard detection of molecular biosignatures (for Dragonfly mission to Titan) for predicting molecular complexity of life on Saturn’s largest moon.
The last was the presentation by the OXCAV (Oxford Control & Verification) group. Prof. Alessandro Abate and Licio Romao collaborate with ESA on building reinforcement learning (RL)-based control systems for space, such that they avoid ‘unsafe actions,’ that is, that RL takes only safe actions even during training.
‘Why connecting the unconnected matters’
In this workshop, we heard how Viasat’s communication satellites support construction sites (e.g., workers often have poor to no WIFI/5G connection on the sites due to walls or secluded areas) — internet signal availability from satellites can considerably ease and make them safer at work.
‘Earth Observation for the new Leverhulme Centre for Nature Recovery’
Researchers showed that EO data can help understand e.g., reindeer pressure on the environment (as it was not sure before how the areas that were less resilient to the climate nearby emerged), as well as that it can be used even for root depth estimation! 🌱
‘Machine Learning for near-real-time Twinning of Earth Systems
Finally, the cherry on top for me. James Parr (CEO of Trillion Technologies and Director of FDL Europe) introduced this year’s FDL. We did a set of exercises in teams, where we brainstormed challenges and issues for three main themes: (i) twinning and simulation (EO data + AI), (ii) integrating knowledge (interdisciplinary research and closing the knowledge-action gap), and (iii) decision intelligence (‘for the first time in history we have the ability to make decisions based on accurate models of the future’).
