SentinelBrief

A background agent that turns Sentinel-1 radar into a flood situation brief, and refuses to state anything it cannot measure. Gemini plans and writes; every number passes a deterministic confidence gate first.

Run an analysis

Window origins snap to the model's 32-pixel tile grid; an unaligned origin would change every reported number.

Result

No analysis run yet.

Ask the agent

Measured performance

domainIoUprecrecall
feni_2024
trained on this event		 0.72--
sylhet_2022
zero-shot transfer		 0.37 0.93 0.38
Recall 0.38 out of domain. When this model flags water it is usually right, but it misses roughly two thirds of actual flooding. Absence of a flag is not evidence of safety.

What it cannot do

  • No forecasting - it describes imagery already captured
  • No water depth - output is binary extent
  • No optical imagery - Sentinel-1 SAR only
  • Never evaluated outside Bangladesh

These are enforced, not advisory: the gate rejects text implying any of them.

Recent runs