On September 3, 2020, the European Space Agency launched Phi-sat-1, a shoebox-sized CubeSat that ESA describes as Europe’s first artificial intelligence in space and the first demonstration of AI hardware for Earth observation. It rode a Vega rocket from Europe’s spaceport in French Guiana as an enhancement to the FSSCat mission. The point was to run a neural network onboard the satellite itself rather than sending all the raw data home to be processed.
The problem Phi-sat-1 tackled is mundane but expensive: a large fraction of Earth-observation images are spoiled by clouds, and transmitting useless cloudy pictures wastes precious downlink bandwidth. Phi-sat-1 carried an Intel Movidius Myriad 2 vision processing unit running a convolutional neural network called CloudScout, built to detect cloud cover and discard the bad images before they were ever sent to the ground. The AI payload was developed by Ubotica Technologies, and the cloud-detection network was trained with the University of Pisa.
By doing the screening in orbit, the satellite cut the volume of data it needed to downlink by roughly a third. Phi-sat-1 was a proof of concept for “edge AI” in space - putting inference on the spacecraft - and ESA followed it with Phi-sat-2 to push onboard processing further.