← Back to the demos Research note, 27 August 2026

Was the Nepal barrier lake visible from orbit before it burst?

On 26 August 2026 a rock and ice avalanche dammed the Lhende Khola on the Nepal–Tibet border. The lake behind the debris burst and the surge ran 60 km down the Bhote Koshi through Timure and Syabrubesi. We went looking for that lake in the last satellite images taken before it broke.

48 hfrom the last orbital look to the collapse

3independent instruments, one answer

0water bodies found, at any threshold

4 daysthe observation gap it fell into

The questionCould anyone have seen it coming?

A barrier lake is the one part of this cascade that sits still long enough to be photographed. If it was there on the last pass, free radar saw it and nobody looked. If it was not, the warning case is for a different instrument entirely.

Timeline of Sentinel-1 passes over the Lhende catchment in August 2026, showing the last pass on 24 August and a four-day gap containing the 26 August collapse — Three Sentinel-1 tracks interleave to give a look every three to five days. The last one before the collapse was Sentinel-1D at 00:18Z on 24 August. The next was the 28th. The event fell in the gap.

The answerNothing was there

Log scale chart of water body area showing the radar detection floor at 2000 square metres, the 10,000 to 100,000 square metre band a lake would need to occupy to cause the flood, and no detections anywhere in the detectable range — A lake capable of driving this flood would cover something like 10⁴ to 10⁵ m², which is hundreds to thousands of pixels and far above the noise floor. The detectable range came back empty.

Three instruments, measuring three unrelated things, all say the same. That matters because they fail in unrelated ways: cloud shadow and radar shadow have nothing to do with each other.

Instrument	What it measures	What it found on 24 Aug	Verdict
Sentinel-1 C-band radar, 10 m	Backscatter. Smooth water reflects away from the sensor and goes dark	0 clusters. Excess over control never clears the −5 dB speckle floor	No lake
Sentinel-2 optical, 10–20 m	Reflected sunlight. Liquid water is dark in near infrared and shortwave	0 pixels below 0.10 in both bands. The 6.9% flagged as water is shadow at 0.116 / 0.117	No lake
Open-Meteo precipitation	Whether enough water was arriving to build a lake at all	0.3–2.3 mm/day across 20–25 Aug	No water to fill one

Verdict: the lake was not visible, because it was not yet there. The blockage formed and failed inside a four-day observation gap, and probably inside a few hours. Nothing running at Sentinel-1 cadence could have warned about it. The window that existed ran from the seismic detection of the collapse, which is t=0 and is an instrument that already fired, to the surge reaching Timure. That is a seismometer and a siren, not a satellite.

The evidenceThree ways of finding nothing

Overlaid histograms of radar backscatter change for the 12 August control and 24 August test scenes, showing a small uniform offset and no excess in the dark tail — **Radar.** Valley floor, 309,370 pixels. The 24 August scene is 0.8 dB darker everywhere, which is rain on the ground mid-monsoon. A filling impoundment puts a lump in the dark tail in one place. There is no lump. Broad and uniform, not local and deep.

Time series of radar dark area in the glaciated source zone from 25 June to 24 August, non-monotonic, including a fall on 31 July — **The glacier hypothesis, tested.** Independent analysis placed an active water area 35 to 38 km upstream, and ICIMOD attributed the 2025 flood in this corridor to a supraglacial lake. Our first run excluded that by capping detections at 5,000 m, so we took the cap off. The dark area wanders: it falls 1.87 km² on 31 July, four times the move between the last two passes, and a control zone in the same elevation band swings just as widely. A lake fills monotonically. This tracks the weather.

Daily catchment-mean rainfall from 17 to 31 August 2026, showing a dry spell from the 20th to the 25th, the collapse on the 26th, and heavy forecast rain from the 29th — **Rainfall.** The week before the collapse was close to dry, which is what finding nothing on the 24th should look like, and it points away from rainfall as the trigger. The bars from 27 August are forecast, and they land on a catchment that now holds a second impoundment.

Why it mattersThe null is the useful part

It rules out a class of warning system. Nobody needs to build orbital monitoring for this hazard at this cadence, because the failure window is shorter than the revisit. That is worth knowing before money is spent on it.
It points at the instrument that would have worked. The collapse registered 5.2 on local seismometers. Seismic detection gives an exact origin time and minutes of downstream warning. It is free, it already fired, and it is where the effort belongs.
It constrains the timeline. The blockage formed inside 48 hours, which narrows what anyone reconstructing this event needs to explain.
It is checkable. Every threshold, date and figure here can be reproduced from free archives with no login.

LiveWhat we are doing now

A second lake formed upstream near the Chhochen Khola and Purepu Tsangpo confluence. China's Ministry of Water Resources put it at about 2 million cubic metres on 27 August and forecast 3 million more over three days, with a high risk of breach. That figure rests on a single source, so we are measuring it independently.

Map of the Bhote Koshi and Trishuli corridor showing the flood path in red running 60 km from Rasuwagadhi past Timure and Syabrubesi, and a dashed zone in the upper catchment where the second barrier lake is reported — The corridor in red is the flood path, traced by steepest descent down the Copernicus DEM rather than drawn by hand, so the route and the distances along it are measured. It runs 70 km from the border to the Trishuli, of which roughly the first 60 carried the damage. The dashed zone is where the second lake is reported. It is drawn as a reach and not a point because no source has published a coordinate for it, which is one of the things the 28 August pass should settle.

2 Mm³reported volume, single source, 27 Aug

3–4%volume spread across three elevation models

3passes booked: 28 Aug, 31 Aug, 5 Sep

nonebreach time we can honestly give

A lake surface is flat, so its shoreline is an equipotential: intersect a radar outline with an elevation model and the shoreline elevation is the water level. Integrate below it for volume. Two passes give a filling rate, which makes the lake its own flow gauge, and the river gauges that would normally do that job were destroyed in the flood.

Two checks before trusting any of it. Running one fixed footprint through three independent elevation models from three different epochs returns volumes within 3 to 4 percent, so the elevation model is not the weak link, which is not what we expected. The weak link is the water outline and the surface estimate. And for 3 million cubic metres to arrive on rainfall alone needs a catchment of roughly 270 to 480 km², which is plausible for these two catchments with melt on top. The Chinese figure is consistent with the weather.

What none of it produces is a breach time. The debris dam is in no elevation model we have, so there is no freeboard, and freeboard sets the clock. We can say how much water is behind it and how fast that is changing. We cannot say when it goes, and nor can anyone else working from orbit.

CorrectionsThree things we got wrong first time

An elevation cap that excluded the leading hypothesis. Capping detections at 5,000 m ruled out a supraglacial source by construction, before testing it. Removed, then tested.
A doubling that was an artifact of counting. We reported the largest source-zone cluster going from 411,000 to 919,000 m². That was separate patches merging across a threshold into one label. Total area, the number that means something, moved 12.37 to 12.67 km². Two percent.
Assuming optical was blind. We justified a radar-only study on monsoon cloud without checking. Sentinel-2 was 75.6% clear over the source zone on 24 August. We reached for the instrument that had worked before instead of asking what the question needed, and it took someone else asking why we were using only one satellite to catch it. For a company whose whole offer is independent verification, that should not have needed pointing out.

Why we publish negative results

Because the alternative is selling things that do not work. We ran this the day after the event expecting either a quiet confirmation or a striking image of a lake nobody had noticed, and got neither. A method you can check is worth more than a result you cannot.

Data, all free and none of it needing a login: Copernicus Sentinel-1 RTC and Sentinel-2 L2A via Microsoft Planetary Computer; Copernicus DEM GLO-30, NASADEM and ALOS World 3D-30m for the elevation spread; Open-Meteo for precipitation. Half-hourly GPM IMERG would have been the better rainfall source, but the Planetary Computer copy ends in May 2021. ICESat-2 has no usable recent track over this catchment. SWOT does cover it, on a pass pattern that should put a look somewhere around 30 August, and it measures water surface elevation directly, which is the number our shoreline method can only approximate.

Method: descending orbit 19, Sentinel-1D, VV gamma0 at 10 m. Baseline is the median of 25 Jun, 7 Jul, 19 Jul and 31 Jul; control 12 Aug; test 24 Aug. Detection AOI 85.30 to 85.70 E, 28.20 to 28.60 N. New water is VV at or below −15 dB and at least 3 dB under baseline, on slopes under 8 degrees, in clusters of at least 2,000 m². Same-track baselines throughout, because comparing an ascending scene against a descending one in this terrain compares two shadow maps rather than two days. Wind roughening raises the backscatter of open water and can hide a small lake. In a confined gorge an early impoundment may be 20 to 50 m wide, which is two to five pixels, below what we would claim to see. Piping failure gives no surface warning at all. Detecting water is not assessing stability: seeing a lake tells you an impoundment exists, never when it will go. The pipeline is a few hundred lines of Python and we will send it to anyone who wants to check the numbers.