Inspiration
On 26 August 2026, a glacier collapse in northern Nepal sent ice, rock and water down the Bhote Koshi–Trishuli corridor. When roads, bridges and phone lines fail, rescuers still need three answers: where the flood hit, what was damaged, and which villages are cut off. Satellites cover every Himalayan valley for free, but raw imagery is not an answer. We built FloodWatch to turn it into one.
What it does
Draw an area and a flood date. FloodWatch answers three questions:
- Where did the flood hit? Flooded and debris-affected areas from before/after Sentinel-1 radar change detection (same orbit track, about 12 days apart), with Sentinel-2 optical imagery used only where the sky is clear.
- What was damaged? Pre-event OpenStreetMap buildings, roads and bridges are overlaid on the flood mask and marked hit, not hit, or uncertain.
- Who is cut off? A road-network analysis finds settlements with no remaining route to the nearest town or hospital, ranked by estimated population and distance to a hospital.
Extras: a confidence layer showing how many independent sources agree on each area, a flood-path tracer that follows the elevation data downstream from any upstream point and lists the settlements along it, and a one-page situation report in English and Nepali generated by the system. Every number in the report comes from the computed maps, never from a language model.
AI component
We trained a flood segmentation model on the Kuro Siwo dataset and evaluated it on scenes held out from training, comparing it against a plain threshold baseline. Results are shown on the app's model page.
How we built it
- Data (allowed inputs only): Sentinel-1 GRD, Sentinel-2 L2A, Copernicus DEM GLO-30, and the pre-event OpenStreetMap snapshot (27 July 2026) through the ohsome API.
- Pipeline: radar change detection, DEM slope masking, multi-source fusion, OSM intersection, road-graph connectivity analysis and D8 flow routing.
- App: FastAPI backend, React and MapLibre dashboard, cached jobs, and a provenance log recording every data source used in each run.
- Data rules: Copernicus EMS (EMSR927) and other published damage maps were never used as inputs. They are used only in a separate validation module to check our results.
Case study: Trishuli flood vs EMSR927
We mapped the August 2026 event and compared our flood extent with the official EMSR927 reference in a side-by-side swipe view, with overlap metrics. The reference was used for checking only.
Challenges
- Radar shadow and layover in steep valleys cause false positives, so we mask them using DEM slope.
- Floodwater, debris and landslide scars are hard to tell apart, so we show confidence instead of a single hard mask.
- Sentinel scenes from different orbit tracks cannot be compared pixel by pixel, so the system pairs only same-track scenes.
- Full satellite products are too large for a live demo, so we fetch only the clipped area requested.
Limitations (what it cannot do)
- It cannot warn of a glacier collapse in advance. Sentinel satellites revisit an area only every few days.
- Optical imagery is blocked by monsoon cloud. Radar sees through cloud but is noisier.
- A resolution of 10–20 m misses small buildings and narrow footpaths.
- OpenStreetMap is incomplete in remote valleys, so some damage and some cut-off villages will be missed.
- Road passability is inferred from the flood mask, not observed.
- This is an educational prototype, not an operational tool.
What we learned
Trustworthy disaster mapping is mostly about honesty: separating what the data shows from what it only suggests, tracking where every input came from, and showing uncertainty instead of hiding it.
What's next
Coherence-based debris detection, Sentinel-1 SLC processing, and field feedback to verify cut-off settlements.
Responsible use
No images of victims are used anywhere in the demo, report or app.
Attribution
Contains modified Copernicus Sentinel data 2026.
Produced using Copernicus WorldDEM-30 © DLR e.V. 2010–2014 and © Airbus Defence and Space GmbH 2014–2018 provided under COPERNICUS by the European Union and ESA; all rights reserved.
© OpenStreetMap contributors.
Comparison data: European Union, Copernicus Emergency Management Service data (EMSR927), used for checking only.
Training data: Bountos et al., 2024 (Kuro Siwo, NeurIPS 2024).
Log in or sign up for Devpost to join the conversation.