An Earth model runs above the Earth.

The human problem

Earth-observation satellites collect more imagery than they can always send down quickly, while many decisions benefit from identifying useful scenes earlier.

The prior constraint

Raw imagery must often wait for bandwidth and ground processing before screening.

AI’s actual role

The model performed inference onboard to identify flood and cloud features before full data return.

The documented result

NASA reports demonstrations on two orbital platforms. Prithvi was trained on 13 years of Harmonized Landsat and Sentinel-2 data; the flight result is currently described in a preprint.

Why it may matter

Onboard screening could help prioritize limited bandwidth, but a demonstration is not a deployed warning system.

Limitations

The result is a preprint.

Two demonstrations do not establish broad satellite reliability.

Flood detection onboard does not equal an official flood alert.

Unresolved questions

What accuracy and compute tradeoffs followed compression?

How did radiation and hardware constraints affect inference?

Which outputs were verified against ground truth?