Artificial intelligence is becoming more and more present in our lives, and timidly is gaining its place in the sun even outside our atmosphere : 2 September an experimental satellite the size of a box of cereal was ejected by a rocket along with 36 other small satellites. The satellite in question, PhiSat-1, is now traveling further 27. 500 kilometers in a synchronous orbit with the Sun at approximately 530 km from our heads.

PhiSat-1 contains a new hyperspectral camera- thermal and hardware on board that performs artificial intelligence calculations , more precisely a VPU (Vision Processing Unit) Intel Movidius Myriad 2 , which we already find in several smart cameras or even on drones. PhiSat-1 is a part of a pair of satellites on a mission to monitor polar ice and soil moisture , as well as testing communication systems between satellites in order to create a future network of federated satellites.

Myriad 2’s first task is to manage a large amount of data generated by high fidelity video camera like the one present on PhiSat-1. “The ability of sensors to produce data increases by a factor 39 every generation, while our data download capabilities are increasing, but only by a factor of three, four or five per generation “, said Gianluca Furano, data systems and onboard computing lead of the European Space Agency (ESA), which led the collaborative effort that led to create PhiSat-1.

At the same time, at all times two thirds of the planet’s surface is covered by clouds . This means they are usually captured, rescued, sent to Earth and reviewed by a scientist or algorithm many useless images, which are then deleted hours or days later. An example of low efficiency that does not fit well with the scientific field. “ The artificial intelligence on the satellite can give us a hand, like the cavalry in western movies,” he explained Furano. The idea was therefore to insert some computing capacity on board to identify and discard the cloudy images, thus saving the 30% of available bandwidth.

The Irish startup Ubotica built and tested PhiSat-1’s artificial intelligence technology, working closely with cosine , camera manufacturer, together at the University of Pisa and Sinergise to develop a complete solution. “The Myriad VPU was designed from the ground up to have impressive processing capacity but with very low power consumption and that is suitable for space applications,” said Ubotica.

Myriad 2, however, was not meant for the hostile environment of space , so Ubotica has verified how it responded to radiation , subjecting the chip to a series of tests to understand how to handle any resulting errors or wear. ESA “had never tested a chip of this complexity for radiation resistance,” Furano explained. “We were doubtful if we could test it properly … we had to write a manual from scratch on how to do a complete test and characterization for this chip.”

The first test, 36 consecutive hours of radiation exposure at CERN at the end of 2018 , represented “a very high pressure situation”, but that test and two subsequent tests luckily went well. The Myriad 2 VPU has passed the tests without requiring changes . But there was another challenge to be won : generally AI algorithms are trained using large quantities of data to learn how to do something, in this case identify a cloud. The problem, as Furano explains, is that because the camera was new, “we didn’t have some type of data. We had to train our application with synthetic data extracted from existing missions “.

All this process took four months to complete . “Intel provided us with support with the Myriad device when we needed it, so we could enable PhiSat-1 AI using our CVAI technology,” Ubotica explained. Unfortunately, a series of unrelated events – delays in rocket setup, the pandemic and hostile summer winds at launch – meant that we had to wait more than a year for find out if PhiSat-1 would work as expected.

The 2 September launch from French Guiana (involved launcher Italian Vega) went very well. For initial verification, the satellite saved all images and recorded its cloud detection choices for each shot, so that on the ground they could verify that the VPU was behaving as expected. After three weeks, the team detected the first hardware-accelerated AI inference directly from an orbiting satellite .

By sending only the useful pixels, the satellite makes better use of the band, significantly reducing downlink costs aggregates, as well as saving time for scientists on the ground . Looking to the future, the operation of PhiSat opens the door to new scenarios, with satellites capable of performing multiple tasks. “Instead of having dedicated hardware in a satellite that does one thing, you can switch networks in and out,” said Jonathan Byrne, head of technology at Intel Movidius.

A satellite so to speak “programmable” , when it flies over fire-prone areas it can locate them and alert local rescuers in minutes instead of hours. In oceans, which are generally ignored, a satellite can detect pirate ships or environmental incidents. By hovering over forests and farms, a satellite can monitor soil moisture and crop growth. On ice, it can check