Radar-to-rescue intelligence for flood disasters.
See who the flood cut off, and what to fix first.
Multimodal SAR and optical change detection, probabilistic cut-off routing and a bilingual AI copilot. Lifeline turns raw Sentinel-1, Sentinel-2, elevation and pre-flood OpenStreetMap data into the villages that lost their road to care, and the repair that reconnects the most, in about two minutes.
Free satellite dataRuns from raw radarEnglish and नेपाली
One run, from raw radar to a ranked repair list.
Hover or focus a number to see how we know it. Numbers we have not measured yet are left out.
What a rescuer needs to know on day two.
Each answer comes from a different source, and each carries its uncertainty with it.
Open water, debris, and the ground radar could not see.
Sentinel-1 radar sees through monsoon cloud. Lifeline compares same-orbit passes from before and after the flood, checks open water with a trained model, and marks steep slopes the radar could not see as not observed. Where the sky cleared, Sentinel-2 shows the same ground in colour: the grey river of debris is the flood.
Every building, road and bridge, checked against the flood map.
Lifeline overlays the flood layers on OpenStreetMap as it stood before the flood. Each feature gets a probability that it was hit, so a washed-out bridge and a muddy road are not treated the same.
Which villages can no longer drive to a hospital.
Lifeline routes every village to its nearest hospital or clinic on the road network before and after the flood. It repeats the routing many times over uncertain damage, so each village gets a probability of being cut off.
A ranked list for the people deciding where to go first.
Ranked by the chance a village has no road route to care. Walking time is the fallback when the road is gone; it follows only roads and paths mapped in OpenStreetMap, so where bridges are down it is an upper bound.
Repairs that reconnect the most villages
| Village | Chance cut off | Drive before | Drive now | Walk now | Nearest care | Fix that reconnects it |
|---|
Draft field message
Every number comes from the facts ledger and passes the same number guard as the copilot.
For field teams
Open in Google Earth, on a phone map app, or in QGIS.
Ask in English or Nepali. Every number comes from the map.
Every number the copilot may say lives in the map's facts ledger. A guard checks each answer and rejects any number that is not there, so the reply either cites its facts or refuses.
Invented numbers caught
Red-team prompts that tried to make the copilot state a number outside the ledger. The guard caught .
The same ledger writes the one-page situation report: English · नेपाली
Run any valley.
Pick a past flood or draw a box, choose the date, and Lifeline finds the images and runs the whole pipeline.
~2 minfor a new 10 × 10 km valley on the hosted engine, from raw Sentinel-1 and Sentinel-2 to a cut-off list and a Nepali report (125 s measured from this page; most of it waits on OpenStreetMap). The first run after a quiet spell adds about 25 s while the engine wakes.
Scored against Copernicus EMS, after the fact.
The official rapid-mapping products are read only by the scoring step, after the run has finished; the pipeline never reads them while it runs. We tuned the settings on one official area (Bidur) and score the other three as held out.
| Area | Flood F1 | Buildings found | Bridges found | Roads agree | Note |
|---|
Four free data sources go in, and one ledger of facts comes out.
Each step writes files into one run folder. The copilot and the report can only read the last one.
-
Inputs
Sentinel-1 radarBefore and after, same orbitSentinel-2 opticalCloud-free scenes when they existCopernicus DEMElevation, slope, height above drainageOpenStreetMapSnapshot from before the flood -
Where
Radar changeOpen water, debris and scourModel checkSegmentation model checks the water mapNot observedRadar shadow and layover masked -
What
DamageHit probability per building, road and bridge -
Who
AccessRouting to care, many times over uncertain damageChokepointsRepairs ranked by villages reconnected -
Say
Facts ledgerEvery number, with its source fileCopilot and reportA guard rejects numbers not in the ledger
Data rules
- The pipeline never reads the official damage maps while it runs; a test fails the build if any reference file is read outside the scoring step. Settings were tuned on one official area and tested on the other three.
- OpenStreetMap is a snapshot from before the flood (never later than 27 July 2026, the newest date the ohsome service holds), so mappers' post-event edits cannot leak in.
- Every number shown to a user exists in the run's facts ledger, with the file it came from.
- Pixels radar could not see stay unknown. They are never counted as dry.
Trained water model
What Lifeline cannot do.
It waits for the satellite
Each Sentinel-1 orbit track repeats every 12 days; with three tracks over this valley a new image arrives every few days. The first after image of this flood came two days after it.
It does not warn ahead
Lifeline maps what already happened. It cannot predict a glacier collapse or a lake outburst.
Steep slopes hide ground
In narrow gorges radar shadow and layover hide parts of the valley floor. Those areas are marked not observed.
Maps have gaps
Remote trails, footbridges and small settlements are often missing from OpenStreetMap. Lifeline cannot route what is not mapped.
Debris can look like bare ground
Fresh debris on a dry gravel bar can scatter radar much like it did before. Some of it will be missed.
Nepali needs a native reviewer
The Nepali answers and report are machine-written. Until a native speaker reviews them, treat the English as the reference.