Flood Damage from Space

Where did the flood hit ┬╖ What was damaged ┬╖ Who is cut off

No area open
Answersopen or run an area
Flooded / debrisтАУ
Buildings hitтАУ
Road under floodтАУ
Bridges reachedтАУ
Settlements cut offтАУ
Long detourтАУ
Map key click to hide / show
See the evidenceradar before & after ┬╖ accuracy ┬╖ tests

The evidence

How we know the map is right

Every answer on the dashboard comes from free satellite data and the map as it was before the flood. Here is the proof, step by step, for the Trishuli flood of 26 August 2026.

01

Radar sees the flood through the clouds

Sentinel-1 sends its own radar waves, so it works through cloud, day and night. We compare the same valley before and after the flood. Drag the line to compare. This is the 3.8 km around Rasuwa Gadhi, where the radar changed most.

Radar image before the flood
Radar image after the flood
BeforeAfter
Sentinel-1 radar, before тЖФ after
Radar change on the valley floor
What changed, on the valley floor only: blue = smoother after (mud, water) ┬╖ red = rougher after (debris)
Our flood map of the same place
Our flood map: orange = mud and debris ┬╖ blue = new water

02

Why radar: the photo after the flood is full of clouds

Sentinel-2 takes normal colour photos, but monsoon clouds hide the valley. Only 67.9% of the valley was cloud-free after the flood, so photos only confirm the radar where the sky is clear.

Photo before the flood
Photo before ┬╖
Photo after the flood, mostly clouds
Photo after ┬╖
The answer: flood map, cut roads and cut-off villages
The answer: flood, cut roads and cut-off villages (rescue priority)
03

How it works, from satellite to answer

Draw a box, give the flood date, press Run. Everything below happens by itself in 2тАУ6 minutes, for any area and any date.

  1. Sentinel-1 radar3 passes before and 2 after, from both directions; change detection finds new mud, debris and water.
  2. Copernicus height modelOnly the valley floor is checked (low above the river, not too steep), so mountain slopes don't give false alarms.
  3. Sentinel-2 photosWhere the sky is clear, the photos confirm the radar (water and bare mud look different from forest).
  4. Possible damageA second, lighter map: land next to a real damage patch that is no higher above the river than the debris we confirmed there. The confirmed map stays separate.
  5. Map from before the floodOpenStreetMap as it was before the flood: buildings, roads, bridges, villages, hospitals.
  6. Road networkBroken roads are removed; a village with no road left to a town or hospital is cut off, then ranked by rescue priority.
  7. AI copilotWrites the situation report in English or Nepali. It writes only blanks; our code fills in every number.
04

Checked against the official EU emergency map

Copernicus EMS mapped this flood by hand from very sharp images (0.3тАУ0.7 m). We used them only to check our result, never as an input.

Our flood map compared with the official EMS map
Green = our confirmed map and EMS agree ┬╖ light green = our possible-damage zone and EMS agree ┬╖ orange / light orange = only ours ┬╖ purple = only EMS ┬╖ black dots = buildings EMS marked as destroyed. Credit: European Union, Copernicus Emergency Management Service (EMSR927).

Honest reading: what we map as confirmed is almost always real (high precision), but alone it misses part of the outer edge of the debris, where radar and photos show little change. The lighter possible damage zone (land next to real damage that is no higher above the river) closes most of that gap, and it adds nothing on the no-flood test below. Its four settings were picked after checking against these maps, so we show both numbers.

05

No flood, no alarm

The same box on a dry-season date (15 April 2026), run with the same settings. A system that "sees" floods everywhere would fail this test.

Real flood result
Flood date ┬╖ 26 Aug 2026
Dry season result: almost nothing found
Dry season ┬╖ 15 Apr 2026
06

Not built for one valley: the same system on other floods

No settings were changed between these runs. Each one is a saved result you can open.

07

An AI that cannot invent numbers

In a disaster, a wrong number can send a rescue team to the wrong place. So the language model is only allowed to write blanks; a checker rejects any digit it writes, and our code fills each blank from the maps.

08

Who needs help first

Every cut-off village gets a rescue priority out of 100, computed by code (not by the AI) from four parts:

Size 40%Damage 30%Isolation 20%Far 10%
09

For people in the flood, and what we cannot do

"I need help" side

People inside a mapped flood area can check their danger level, call emergency numbers, and send their GPS location to the rescue team's inbox (PIN-protected). Requests from outside the flood area, or far from the chosen place, are refused, which stops casual fake requests.

Open the "I need help" side тЖТ

Honest limits

  • No early warning: satellites pass every few days, so we map damage after the event.
  • 10 m pixels: small houses, narrow roads and footbridges can be missed.
  • Steep gorges: radar can't see some slopes, so we under-map debris edges.
  • "Hit" is not "destroyed": we don't grade damage.
  • Map gaps: missing buildings or roads in OpenStreetMap mean missing answers.
  • An educational prototype, not an operational tool.
тЖС Map