07 Jul 2020
As more satellites go up, there's a boom coming in processing their data - and that needs image processing, artificial intelligence (AI) and high performance computing (HPC).
For the last year, Elon Musk’s StarLink project has been flinging shedloads of satellites into space. The StarLink program will have more than 1,500 devices orbiting by the end of 2021, aiming to deliver improved Internet connection.
Like anything else the technology mogul does, StarLink has been controversial - and there is billionaire-on-billionaire action here: Jeff Bezos of Amazon plans to send up more than 3,000 of his own Project Kuiper satellites, but is waiting on approvals.
Alongside Musk and Bezos’ orbiting status symbols, there are many other satellite initiatives going up right now. Outfits like Myriota are putting up tiny nano-satellites for Internet of Things applications, while earth-observation organizations like Capella Space have orbiting cameras which can resolve details less than 0,5 m on the ground.
The most interesting players are the earth observers. Using AI and HPC, some are looking for shoals of fish for conservation, or tracking weather to optimisee agriculture. Others are watching for climate change, and of course others are for government surveillance, following political signals such as troop movements.
All these satellites are capturing vast quantities of data. They need to get it down to earth from space, capture it and process it to find the information they want.
It turns out the real 'space race' might be on the ground, not just up in orbit. Satellite communications were once the preserve of NASA and its European counterpart ESA, but now competitors are emerging. And one of their competitive features will be their processing ability.
Amazon is in there with Ground Station, a service that installs downlinks that connect to the cloud capacity of Amazon Web Services (AWS), and offer “ground station as a service”. Myriota and Capella Space have already opted into AWS Ground Station.
AWS is aiming for ten physical stations by the end of this year, which will cover the world. It already has facilities in place in Ohio, Oregon and Bahrain. The plans is that no AWS ground station will be more than 9.5 milliseconds from a data center.
That actually only narrows it down to a distance of a few hundred to a thousand miles away, so there are plenty of other cloud and hyperscale-standard facilities that could do the job, besides AWS. Some are close to centres of scientific research (such as my blog host - Kao Data).
Others have access to satellite-specific expertise - and are aiming to provide a more specialist service.
At the remote end of Cornwall in the UK, Goonhilly began life as a BT satellite link station for Telstar in the 1960s. It’s now run by an independent company which is building a business on the 60-plus antennas on the site, along with a community springing from 50 years of experience in satellites.
Satellite signals need processing to extract their data, and there’s a long tradition of maths wizardry aimed at doing just that. For instance, signals are often obscured by interference between the source and the receiver; astronomers have developed ways to get those signals back.
“We understand the maths used by satellite communications engineers and radio astronomers,” says Ian Jones, CEO of Goonhilly Earth Station, in an interview for DCD. “And that maths happens to be the same as that for machine learning,”
Goonhilly has an Nvidia-based supercomputer on site; a team from Hertfordshire University is running algorithms on it which were created by radioastronomers to detect distant galaxies. Now they are being used to take images from earth observation satellites, and spot features which were previously obscured by clouds.
Having the supercomputer on site makes for a quicker response - and also cuts the cost of network transport
Moving the processing to the expertise at the ground station might turn out to be such a good fit that other AI applications move there, which have nothing to do with space. A lot of other machine learning tasks can use those algorithms from space, and training AI systems is normally done in isolation.
“We believe that, regardless of whether it's satellite or not. we can run that training model at Goonhilly,” said Jones. “And it will be more cost-effective than running it on Amazon and GCP.”