Constellations - Explore Space Network Technologies with Industry Leaders 218 - AI for EO, Neural Network Supervisors and Overcoming the Clouds
Oct 29, 2025
Aubrey Dunn, co-founder and CTO of Ubotica, focuses on onboard AI solutions for Earth observation satellites. He discusses how cloud cover complicates imagery collection, introducing innovative strategies for cloud detection and removal. Aubrey highlights real-time insights gained from AI-driven techniques, like dynamic targeting and lossless compression. He shares exciting stories, including detecting ships with minimal latency, and addresses privacy concerns with neural network supervisors. Overall, a deep dive into improving satellite efficiency and overcoming technological challenges.
AI Snips
Chapters
Transcript
Episode notes
Clouds Severely Limit EO Value
- Cloud cover makes optical Earth observation frequently unusable and reduces the effective value of downlinked imagery.
- Removing cloudy pixels on-orbit raises usable data efficiency and lowers per-image cost for end users.
Do Cloud Removal Or Dynamic Targeting On-Orbit
- Process images on-satellite to detect and remove clouds and only downlink useful pixels.
- Or use onboard look-ahead imaging to autonomously target the least-cloudy areas before capture.
Three-Step Onboard Cloud-Processing Flow
- On-orbit cloud removal uses a three-step flow: detection, removal, and lossless compression.
- A segmentation UNET produces per-pixel masks so satellites can discard cloudy tiles and compress the rest.