A European satellite is now hunting for ocean anomalies with AI running onboard, skipping the usual ground station round trip.
The European Space Agency's Phi-sat-2 mission carries a lightweight pipeline that segments sea from land, encodes marine regions with self-supervised features, and flags anything that deviates from a normal sea state, with an option to characterize specific anomaly types. Researchers first trained and validated the system on simulated Phi-sat-2 imagery to confirm it could run on the satellite's resource-constrained hardware. Once in orbit, real acquisitions turned out to differ significantly from the simulated data used for training. The team retrained the pipeline on actual Level-1 imagery with a reworked annotation strategy for handling ambiguous marine regions, which substantially improved performance.
That gap between simulation and reality is the real story here. Pre-launch simulation is useful for catching hardware and engineering problems before a satellite is irretrievably in space, but it is not a stand-in for scientific validation, which needs real in-orbit data. For a field racing to put more AI directly on spacecraft to cut bandwidth and speed up responses to things like oil spills or ships in distress, that distinction matters more than the demo itself.
It is the same lesson every machine learning team eventually learns on the ground: a model that aces its training set can still stumble the first time it meets the real world. In orbit there is no quick patch. Engineers needed a full retrain using actual satellite imagery before the system was reliable.