The flood outran the warning.

The collapse began at 5,200 to 5,400 m on the north side of the Langtang Lirung massif at 08:37 Nepal time (02:52 UTC). The US Geological Survey recorded its seismic signal as event us7000tbwb; ICIMOD put its volume at 100 to 200 million m³.

The debris came down the Lhende Khola and reached Gyirong Port, on the border at Rasuwagadhi, about seven minutes later. It followed the Bhote Koshi past Timure to Syabrubesi, where the river becomes the Trishuli, and on past Betrawati to Bidur (Trishuli Bazar) in Nuwakot. It is Nepal’s deadliest natural disaster since the 2015 Gorkha earthquake.

Drag the pen on the trace at the top of the page, or press Play: the map follows the modelled front down the river, minute by minute, and marks each place as it is reached.

Very-high-resolution imagery from before the event and from the days after it, and PlanetScope from the morning itself.

Rapid-mapping services such as Copernicus EMS map damage in selected areas within days, but not the whole corridor. Open satellite archives, open building and road data and public timing records now make a corridor-wide, validated reconstruction possible without commercial data.

We compared satellite pictures of the valley from the same time of year, before and after the flood, and marked every spot that had changed.

Against the Copernicus EMS rapid mapping of the event, the footprint reaches F1 scores of 0.83 to 0.89, with precision of 0.93 to 0.98. The misses lie at the margins of the mapped extent.

The thresholds were not set by hand. A sweep of 1,008 combinations was scored against the EMSR927 mapped extent inside each area’s analysed region, on a 20 m grid. Leave-one-area-out tuning chose settings on two areas and scored them on the third. The Langtang Khola valley, untouched in 2026, served as a control: settings that flagged more than 2 % of its valley floor were rejected.

The footprint contains 92 % of the buildings Copernicus graded as destroyed (2,936 of 3,174), 87 % of the 56.8 km of road graded destroyed, and all 30 destroyed bridges within 40 m.

Buildings count as inside when their 20 m pixel is flagged. Road length inside is sampled every 5 m along each line. Bridges count within 40 m of the footprint, hydropower sites within 100 m, since both sit at the river’s edge.

The front reached Syabrubesi about 12 minutes after the collapse, 26 minutes before the SMS alert, and Betrawati about 43 minutes after it.

The river profile is the MERIT Hydro main stem, ordered downstream by upstream area, with bed slope S smoothed over 500 m either side. From Rasuwagadhi to Syabrubesi the front is a debris flow at a steady speed of at most 52 m/s, the speed observed between the collapse and Gyirong Port. Below Syabrubesi it is a kinematic flood wave whose celerity follows Manning’s equation:

The single unknown k is fitted to the observed timing: Gyirong Port 7 minutes after the collapse, Syabrubesi by 13 minutes (its gauge stopped transmitting at 08:50), and Betrawati at 38 to 48 minutes, a ±5 minute band around 09:20, when its gauge stopped transmitting. A Monte Carlo of 10,000 draws across these windows gives the arrival times and their 5 to 95 % ranges.

The fitted model implies a flood depth of 5.5 to 8.5 m (median 6.7 m, Manning’s n = 0.05), of the same order as the 9 m rise in 30 minutes that ICIMOD reported on the Trishuli at Galchhi, further downstream, although depth was not used in the fit.

The Flood Forecasting Division learned of the flood at 09:05, and SMS alerts went out at 09:15, after the front had passed Timure and Syabrubesi and as it reached the Upper Trishuli plants and Betrawati. Near the border, even an instant automatic alert gives only 5 to 10 minutes. Further downstream the time existed but was not used: an alert triggered automatically, from the collapse’s seismic signal or from upstream sensors with data shared across the border, would have given 34 to 41 minutes at the midstream hydropower sites and Betrawati.

Open data reproduced the official damage mapping within nine weeks of the event and without commercial imagery. Three choices made the difference: change vector analysis for debris on already-bare ground, an event corridor that keeps other rivers and the late-September rains out, and a control valley that stopped tuning from rewarding false alarms outside the reference areas.

Two attempts failed and are worth reporting. Ending the Sentinel-2 window before the 24 to 28 September rains left too few cloud-free monsoon scenes, and F1 fell to 0.1 to 0.3. A snow mask widened by 100 m erased much of the corridor, because water and wet debris pass the snow index test as well as snow does.

Limitations

The flood reached the border settlements within minutes, before any warning could be issued, but the midstream hydropower sites and towns had 34 to 41 minutes that an automatic alert could have used. For Nepal’s transboundary valleys, investment in automatic detection and data sharing is likely to save more lives than faster mapping after the fact.

Every number on this page is read from the rasuwa package’s outputs when the site is built. The package has one module per analysis and a command-line step for each. The footprint is released as 20 m GeoTIFF tiles (UTM 45N) with the scores and counts as JSON. Code, footprint and results are archived on Zenodo at doi.org/10.5281/zenodo.23887627.