A satellite photograph has two possible customers: the person with a question today, and the person who will notice something unexpected tomorrow. Letting an AI choose which parts reach Earth gives the first customer a faster route to useful information. It can also decide what the second customer will never see.
An October preprint from Thomas Goudemant and colleagues at IRT Saint Exupéry explores that choice using 68 harbour, coastal and canal scenes. The researchers tested a YOLOX-S vessel detector on Versal embedded hardware, then estimated the image volume that different selection rules would send home.[1]
Two ways to preserve vessel images
| Selection | Raw scene volume sent | Vessels retained |
|---|---|---|
| Crops around detections | 2.8% | 83%Source 1 |
| Tiles containing detections | 28.7% | 98%Source 1 |
Tiles preserve a missed vessel when another vessel in the same tile triggers a detection. A close crop provides less insurance. On sparsely occupied open sea, that neighbour advantage can disappear.[1]
A smaller picture is a different product
Compression offers another way to reduce the transmission burden. ESA’s Φsat-2 compression application encodes images aboard the spacecraft and reconstructs them with a decoder on the ground. Its developers assessed reconstructed imagery through a downstream task: predicting building masks and comparing those with masks obtained from the original images.[2]
That is a useful distinction for anyone buying satellite information. One question concerns how faithfully an image survives being squeezed into fewer bits. Another concerns whether the relevant part of the image was selected at all. A beautiful reconstruction cannot restore a patch of coastline that the sender chose to omit.
The maritime figures are extrapolations from measured detector errors, using corrected ground imagery. They are not an orbital demonstration; some vessel-centred training tiles can overlap other data splits. The reported retention rates need testing on independent scenes in a complete onboard chain.[1]
The value of keeping the picture
ESA encountered the broader value of imagery during Φsat-2’s commissioning. The original plan emphasised transmitting results generated by its onboard applications. By the start of its science phase in July 2025, the team had also decided to provide Level 1 image products to help developers train and improve applications. The mission offers an actual example of useful data extending beyond an AI’s immediate answer.[3]
For a harbour operator requesting a quick vessel inventory, a tightly selected image may be enough. For a researcher comparing activity over time, the surrounding scene could become part of the evidence. The right transmission rule follows from what the recipient needs to recover later—not simply from how small the file can become.
Sources & further reading
- AI-Based On-Board Maritime Object Detection for Earth Observation Payload Data Reduction on Versal Embedded Hardware — version 2 Thomas Goudemant and colleagues / arXiv · 9 Oct 2026 · Accessed 18 Oct 2026
- AI for image compression European Space Agency · Accessed 18 Oct 2026
- Φsat-2 begins science phase for AI Earth images European Space Agency · 15 Jul 2025 · Accessed 18 Oct 2026