Satellites can now decide what's dangerous before sending data home — using only compressed patterns, not raw images.

Researchers tested whether satellites can identify hazards and filter imagery onboard using only compressed data patterns (embeddings) sent from Earth, rather than storing full images. It turns out the system works differently depending on the task: spotting clouds works best with nearest-neighbor matching, while detecting changes over time works best with statistical averages — but both use the same compressed data, and the whole system transmits less than 1 kilobyte per query.

Satellites have always faced a hard constraint: they can't send everything back to Earth because the radio link is too narrow and expensive. This paper shows you can solve that constraint with a two-step system that works at scale. Ground stations send up compact mathematical patterns instead of algorithms, and the satellite picks which task-solving method to use depending on what it's looking at. The real implication is bandwidth. If this holds up in practice, satellites stop being just cameras and become decision-making systems, which opens up distributed Earth monitoring for disaster response, infrastructure inspection, and environmental tracking where speed matters more than perfect data.