AI On-board of Satellites, Towards Autonomous Scientific Instruments
summary
Deployed a tiny foundational model RaVAEn on a D‑Orbit satellite in 2023 and runs the world’s first PyTorch on-board training for a few‑shot cloud classifier using only six labeled images; on‑sat inference runs on a VPU/CPU and a 5×5 km change detection finishes in <0.11 s, a 10‑epoch retrain completes in ~1.55 s. New work targets hyperspectral EMIT (285 bands) and replaces slow radiance calibration + matched‑filter steps with end‑to‑end models tested on a Jetson Orin Nano, cutting preprocessing from >2 minutes to <0.3 s (>99% time reduction).
AI-written summary · about
chapters (14)
- 0:00AI on satellites introduction
- 1:02Why deploy AI on satellites
- 4:14Space-based AI challenges
- 4:35Unsupervised learning for images
- 7:05Onboard training with limited data
- 9:33Hardware constraints in space
- 10:37Hyperspectral data challenges
- 12:01Methane leak detection
- 14:33Hyperspectral processing bottlenecks
- 15:57Faster processing solutions
- 17:24Hyperspectral model design
- 18:27Methane detection improvements
- 19:51Future work and applications
- 21:39Q&A
related talks
06mMETASAT Demonstrator: Mixed Criticality, Accelerated AI Computing for Future Space Systems
28mMachine Learning on the Edge
49mMonitoring global events with AI-driven satellite swarms
22mAI On The Farthest Edge
48mThe Power of Edge AI: Revolutionizing Future of Real-Time, Autonomous Processing
15mKeynote: Incremental Deep Learning For Satellite with KubeEdge...
19mMission-Critical AI in Space and Sky: SWaP-Constrained Intelligence Wi...
01h 02mAI on the Edge