Case study · Nepal & Tibet Autonomous Region, China
Reconstructing a transboundary flood
On 26 August 2026 a flood came down the Bhote Koshi from across the Chinese border and destroyed communities in Rasuwa. We reconstructed the causal chain from terrain and orbit — and published the correction when our own central claim failed a back-test.
The geography reporting could not resolve
Early coverage placed the source vaguely "upstream in Tibet". The border crossing at Rasuwagadhi is in fact a confluence, and the two branches above it are not equal. Flow-routed delineation on 30 m elevation data gives the north-eastern branch — the Lhende Khola — 899 km², or 73% of the 1,233 km² draining to the border, against 328 km² for the north-north-western branch. Forty-nine per cent of the Lhende basin lies above 5,000 m.
That single distinction determines which valley matters, which glaciers are upstream of the disaster, and which lakes are even candidates.
The largest lakes near the channel sit north of the drainage divide and drain away from Nepal entirely. Ranking by proximity includes them. Only flow routing excludes them.
Why a small release stayed lethal
A landslide dam of the observed scale impounds somewhere between 0.5 and 25 million cubic metres — computed by flood-filling the digital elevation model, restricted to that dam cell's own upstream area so the lake cannot leak down-valley and inflate the answer. That is a small release for the damage that followed, which means the magnitude has to come from somewhere else: entrainment on the way down, and geometry.
The corridor below the border stays 180 to 480 m wide for 60 km, measured 50 m above the channel bed. A surge that would have spread and attenuated across a floodplain instead stayed deep for the entire distance to the settlements. Confinement, not volume, is what made this event what it was — and it is a property of the valley, computable in advance, anywhere.
Five attempts to see the source, all of which failed
We tried to image the detachment scar itself. Every route failed, and the accumulated failure turned out to be the finding.
- Optical imagery two days after the event: 98.5% cloud over the detachment envelope, measured from the scene classification band rather than estimated.
- A commercial crisis-response archive with sub-metre sensors: those sensors were tasked over the damage corridor in Nepal and never over the source in Tibet.
- The coarser scenes in that archive that do reach the source: 96.6% cloud.
- Radar change detection on a matched ascending Sentinel-1 pair: no scar detected. Flagged change inside the detachment envelope ran 0.31% against 0.28% across the whole basin — statistically indistinguishable.
- A further optical pass three days after the event: 99.8% blocked.
Published parameters for the detachment do not uniquely locate it either: 26.45 km² across 306 separate clusters match them equally well. So we shipped a detachment envelope and five ranked candidates rather than a fabricated pin on a map. We state radar's result as not detected, never as did not happen — steep-terrain layover can hide a scar, and an ice-on-ice detachment gives little backscatter contrast.
Where the sky did clear, over the Nepal reach, the measurement is unambiguous. Comparing cloud-free pixels on both dates, bare sediment increased 2.21× — up 121%, with roughly 1.1 km² of channel margin newly stripped.
Our original analysis argued that no lake in the conduit basin was large enough to have caused the flood — 50 mapped lakes totalling 2.66 km², the largest 0.44 km² — and concluded that the trigger was therefore a mass movement rather than a lake burst.
We then ran the identical pipeline, with no per-site tuning, against three floods with established causes. It scored 2 of 3. The miss was Thame, 2024: a genuine lake outburst from a lake of 0.389 km² — smaller than this basin's largest. Any threshold that classifies all three known events correctly must lie between 0.025 and 0.389 km², and 0.442 km² sits above that entire band. On our own evidence, the basin classified as lake-plausible.
The claim was withdrawn — in the client document, in the deck, and in the briefing that followed — and the corrected version states plainly that a small lake is not a safe lake. The catchment delineation, the confinement measurement and the impoundment volumes were never dependent on lake size and stand unchanged. The mass-movement account now rests where it belongs: on the field agencies' reported observation, not on our arithmetic.
Why this case matters beyond Nepal
Roughly 15 million people worldwide live within 50 km of a glacial lake, and 9.3 million of them are in High Mountain Asia — more than half in India, Pakistan, Peru and China. The properties that made this flood lethal are not exotic: a confined corridor, a cross-border source, and an archive that could not see the mountain on the day it mattered. All three are measurable in advance, and all three are what our Pakistan early-warning system was rebuilt around.
Method: Copernicus GLO-30 elevation with a hand-rolled priority-flood fill, D8 flow routing, accumulation and reverse-graph upstream search; impoundment by DEM flood-fill masked to the dam cell's upstream area; corridor width measured 50 m above thalweg; optical change from Copernicus Sentinel-2 with SCL-derived cloud fractions; radar change from a matched-geometry Sentinel-1 RTC pair; hydrography and place-names from OpenStreetMap via Overpass. Back-test run on three events with documented causes using the identical pipeline and no per-site tuning. Contains modified Copernicus Sentinel data 2026.