Amazon Web Services (AWS) claims to have delivered a first in operating a suite of machine learning software running on a satellite in orbit, as a testbed for others to collect and analyze data directly on orbiting satellites using its cloud.
The experiment, conducted over the past 10 months in low Earth orbit, involved using AWS machine learning (ML) models to study satellite imagery in real time, with AWS IoT Greengrass providing cloud management and analytics services.
Announced at the cloud giant's AWS re:Invent 2022 conference, the satellite system could enable customers to examine large volumes of raw satellite data in orbit, and only downlink the most useful images for storage and further analysis, AWS said. This would drive down costs and enable more timely decision making, it claimed.
"Using AWS software to perform real-time data analysis onboard an orbiting satellite, and delivering that analysis directly to decision makers via the cloud, is a definite shift in existing approaches to space data management. It also helps push the boundaries of what we believe is possible for satellite operations," said AWS VP Max Peterson.
For the experiment, AWS said it worked with D-Orbit and Unibap.
D-Orbit operates in the "space logistics and transportation service industry," and the project used one of its ION satellites to carry the hardware for the AWS in-orbit experiment. The hardware itself comprised a space-qualified processing payload built by Unibap, a Swedish technology outfit specializing in space-borne computer systems.
According to AWS, the teams worked to build a software prototype that would include the tools they had jointly identified as essential for the Earth observation mission, and this included the ML models and AWS IoT Greengrass, the company's cloud-managed edge runtime.
The compute hardware is Unibap's iX5-100 SpaceCloud "infrastructure computer." This features a processing core based on AMD's G-series embedded processors with up to 4 CPU cores and integrated Radeon GPU, 2.5GB of DDR3 memory, and an FPGA from the Microsemi SmartFusion2 family.
The FPGA is apparently used to implement an ARM Cortex-M3 processor running FreeRTOS which is used to monitor the system, while the AMD chip typically runs a version of Linux such as Lubuntu.
According to Unibap, the iX5-100 has interfaces for sensor readout and payload telemetry downlink, plus local SSD storage. The system has also been validated with S- and X-band radios.
The compute here is hardly cutting edge: the G-series chips are based on AMD's Jaguar cores, which were introduced about a decade ago, but stability is a key consideration with embedded applications.
Throughout the experiment the team applied various ML models to satellite sensor data to identify specific objects in the sky, such as clouds and smoke from fires, as well as surface objects including buildings and ships, AWS said.
The team overseeing the project came up with a couple of technical fixes to help operate the orbital payload more effectively, according to the company. They developed a way to break down the large satellite images into smaller data files, using AWS AI and ML services to reduce the size by up to 42 percent, which it claimed enabled real-time inferencing in orbit.
They also enabled the bidirectional movement of data over multiple ground station contacts by setting up a reliable TCP/IP proxy between the satellite and the AWS Cloud. This made it simpler for ground crews to manage file transfers, removing the need to manually process downlinks over multiple contacts, AWS said.
It should be noted that AWS already has its own AWS Ground Station service, created to provide satellite operators with the ability to control satellites and download data without having to build and maintain their own infrastructure.
AWS said that the satellite remains in space, but along with Unibap and D-Orbit, it has been using the space-borne experiment hardware to test new capabilities beyond the original set of test objectives. These include other techniques for processing raw data in orbit and better methods of data delivery. ®
Remote Desktop, domain controllers giving you gyp? Gamble on these fixes
As per employee model replaces old subscription, user bills could soar
Opinion Nothing good - the recent layoffs hit its best and brightest leaders hard
Even with instructions staring them in the face, this genius couldn't get it right
Welcome to the real world, kids. And for the rest of us, a future at which Meta is gulp! - better at large-scale analytics
The irony, it burns
Calls to avoid C/C++ and embrace Rust grow louder