01 · Sense — For engineers, researchers, and the agencies that own gauges
Stormwatch
It keeps reading when the grid goes down.
River gauges are scarce, and the ones that exist go quiet with the power. Stormwatch is a solar camera that finds the floodwater, reads its depth and grades its severity on the device — day, night, and through the storm.
- from the grid
- 0 W
- day and NoIR night
- 24 h
- inference, no cloud
- On-device
- to service in the field
- Hand tools
The problem
The measurement dies exactly when it starts to matter.
A telemetered river gauge costs more than most municipalities can justify for a hazard that arrives eight or nine times a year. So the network is thin to begin with — and thinnest in the small catchments where flash flooding actually kills people.
Then the storm arrives and takes the rest. Power fails, backhaul fails, and the instrument goes quiet in the exact window its reading was worth something. What comes out the other side is a reconstruction: a high-water mark on a wall, a phone video, someone’s memory of how far up the door it reached.
Stormwatch replaces the gauge with a camera and a model. Two lenses — one of them NoIR, so darkness is not an outage — watch the water. A segmentation model running on the station itself finds the floodwater in the frame and estimates its depth, which means the reading is produced where the water is, not in a datacenter the storm can cut it off from. Solar power and a chassis we print ourselves handle the rest.
How it works
Four moves, end to end.
- 01
Watch
Two cameras on a solar mast, one of them NoIR, framing the river and its banks.
- 02
Segment
A YOLOv11 model finds the floodwater in every frame, on the station itself.
- 03
Measure
Surface-normal estimation turns extent into a water level and a severity class.
- 04
Report
Readings go out when the link is up, and queue on the station when it isn’t.
Real-time flood segmentation
The camera finds the floodwater.
A YOLOv11 segmentation model isolates floodwater in every frame and tracks how far it has spread, turning ordinary video into a measurement. No staff gauge in shot, no calibration target, no person standing in the water with a ruler — the water itself is the instrument.
Water level
4.12 m
▲ 0.4 m in the last hour
- Severity
- Alert
- Critical at
- 5.4 m
- Confidence
- ± 0.15 m
Severity estimation
Then it reads the depth.
Segmentation alone tells you where the water is, not how bad it is. Pairing it with surface-normal estimation recovers the scene geometry, which turns extent into a water level and a severity class calibrated against the site. That is the part the paper is about, and the part that makes the reading comparable across stations.
Edge AI, off the grid
It runs on sunlight.
Inference happens on a Hailo accelerator at the station, powered by a panel. Connectivity is for reporting, not for computing — so when the network drops, the station keeps measuring and backfills when it can. Designing for the outage rather than around it is the whole point.
17:40 · RGB
02:14 · NoIR
All-weather, day and night
Built for the worst night of the year.
Dual cameras, one tuned for low light, in a weather-sealed PETG chassis we print in-house and can service in the field with hand tools. Anything that requires fair weather and a technician to keep working is not a flood instrument.
Where this stands
Field prototype. The method is peer-reviewed; the deployment is not yet at scale, and we would rather say so.
Built on
Push on Stormwatch. We answer the hard ones ourselves.
Questions go to the engineer who built the part you’re asking about — not a sales sequence.