Training AI in space for autonomous detection
Researchers in Germany have demonstrated an AI framework being trained in low earth orbit for the autonomous detection of images in space.
The SONATE-2 satellite developed at the Julius-Maximilians-Universität Würzburg (JMU) has been monitoring the Sahara desert for the last year. However the satellite has been designed to support training of its AI computer vision framework, and has been able to autonomously identify new features.
This is a key step forward for embedded AI systems with constrained resources to update the AI frameworks.
The team led by Hakan Kayal, head of the Professorship for Space Technology, spent three years developing and building the CubeSat. The aim was for SONATE-2 to learn to independently recognize and photograph anomalies on the Earth’s surface.
- Sakura-1 accelerator ready for AI in space
- Eight core RISC-V chip for AI in space
- Imsys develops RISC-V core, looks to AI in space
The Sahara was selected as the target region for detecting anomalies. The satellite first took 270 photos and stored them in two data sets of 90 and 180 images. The neural networks stored on board were trained using data from Sentinel-2 and Landsat-8 for Slovenia with the images assigned as agricultural area, forest, meadow, water, clouds, snow and artificial area in the EfficientDetD1 and Faster R-CNN models.
Running on an Nvidia Jetson NX GPU, SONATE-2 reliably recognized objects that do not normally occur in a desert, such as the river Nile and the adjacent green region. The paper is here: AI in orbit
Regardless of the complexity of the landscape captured, all images were processed within the specified 5s. The GPU did not work at the maximum possible clock frequency of 1 GHz, but only at 300 MHz and in 10W mode. This demonstrated not only the execution of the pre-trained networks, but also the ability to train independently in orbit with autoencoder models.
“In the future, these methods can be used to autonomously detect interesting features during interplanetary missions, for example on unknown planets or celestial bodies,” says Professor Kayal.
In addition to the primary AI payload, there are three other payloads on SONATE-2 that have also been successfully tested: the MultiView multi-head miniature star sensor, a pulsed plasma thruster and an amateur radio payload that offers two amateur radio services to radio amateurs worldwide.
SONATE-2 is still fully functional and now available, for example, for the practical training of aerospace computer science students.
Further experiments are theoretically possible on board via uploads, but this will only be possible for a limited time as the orbit decays. “We estimate that our satellite will function for another year and a half and then burn up completely when it enters the Earth’s atmosphere,” said Kayal.
www.informatik.uni-wuerzburg.de/en/space-technology/
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
