WaterGuard: Flood Impact from Space

Inspiration

When a glacier lake bursts in the Himalaya, the monsoon clouds don't part politely for your satellite. Roads vanish, bridges go downstream, and every rescue team asks the same question: "Which villages can we still reach?"

Normally the answer takes days. We wanted it in minutes, using nothing but free satellite data. 🛰️

What it does

Point at a map, pick a date, and WaterGuard tells rescue teams where the flood hit, what it broke, and who is now cut off.

🌊 Where did the flood hit? Sentinel-1 radar (which sees straight through monsoon cloud) is fused with Sentinel-2 optical change, kept to the valley floor where water can actually reach.
🏚️ What was damaged? Buildings, roads, bridges and health facilities from pre-event OpenStreetMap, laid over the impact map.
🚧 Who is cut off? Settlements that had a road to a town or health post before the flood and have none now.
💬 "So what do I tell my team?" A bilingual copilot (English + नेपाली) that won't say a number the maps can't back up.
📄 "I need it on paper." A one-page situation report, one click.
🧭 "Where will it go next?" A flood-path tracer from any upstream point.

How we built it

Terrain: Copernicus DEM → HAND (height above river), slope and D8 flow paths using WhiteboxTools.
Radar: Sentinel-1 RTC. The first image after the event is compared with up to 5 earlier images from the same orbit track, giving per-pixel z-score change. Layover and shadow are masked.
Optical: Sentinel-2 L2A cloud-free before/after composites → NDVI loss and NDWI gain.
Fusion (OpenCV): valley floor AND (optical change where clear, radar elsewhere), then morphology. Patches must touch the river.
Damage and access: a pre-event OSM snapshot becomes a road graph. We remove impacted segments and damaged bridges, then test whether settlements can still reach towns, health posts and roads out of the area.
Copilot: it has no data of its own. It can call only six read-only tools over the run's results. Every number in its draft must appear in a tool result. If one doesn't, the model gets one retry, and if it's still wrong the number becomes [unverified]. Every answer ends with "✓ N numbers checked". It runs on Amazon Bedrock (Nova Pro), with an offline fallback.
Stack: Python, FastAPI, OpenCV, MapLibre, Docker, and EC2 behind CloudFront.

Challenges we ran into

The flood wasn't water. The 26 Aug 2026 glacier-outburst flood on the Bhote Koshi–Trishuli corridor was mostly debris and landslides. A "find the water" model misses most of it, so we look for change instead.
Honesty is hard to enforce. EMS damage maps and post-event OSM edits are never inputs. Each run uses an OSM snapshot from 24 h before the event and fails if any feature was edited after it. EMS data is read by exactly one validation module that the pipeline never imports.
Crying wolf. We tried to make it cry wolf. A placebo date a month before the flood flagged 0 ha, and the untouched Langtang valley flagged 0 ha.
Clouds. Optical imagery in monsoon season is mostly a rumour, so radar picks up the slack where Sentinel-2 can't see.

Accomplishments that we're proud of

We checked our homework against the official Copernicus EMS maps (EMSR927), tuning on two areas and keeping two held out:

EMS area	Split	F1	Valley-only baseline
Syapru Besi	tuning	0.94	0.86
Timure	tuning	0.90	0.89
Bidur	held-out	0.92	0.56
Phosretar	held-out	0.88	0.53

🏚️ At Bidur it caught 2475 of 2479 EMS-destroyed buildings.
🌉 It flagged 20 of 20 EMS-damaged bridges at Bidur and 5 of 5 at Syapru Besi.
🐺 No wolf-crying on the placebo and untouched-valley tests.
🌍 It ran unchanged on the 2021 Melamchi flood and the Sept 2024 Kathmandu Valley floods, and a flat city is about as far from a Himalayan gorge as you can get.

What we learned

Radar sees through clouds, optical sees what radar can't, and neither helps if you forget to mask layover and shadow.
A baseline matters. "Mark the whole valley" scoring 0.89 at Timure kept us humble about how much the satellites were really adding.
An AI that says [unverified] is more useful than one that sounds sure. In disaster response, a confident wrong number can send a rescue team the wrong way.

What's next for WaterGuard: Flood Impact from Space

Next up:

More flood events and regions
Feedback from people who actually do flood response
Better estimates of stranded people (right now it's buildings × 4.37, the Nepal 2021 census household average)
More languages for the copilot