A three-decade satellite record finds 2,377 glacial lakes across the China–Nepal Himalaya, with their combined area up by more than a quarter. A bigger lake is not always a more dangerous one, and one of the deadliest Himalayan floods of 2026 began nowhere near a lake.
As Himalayan glaciers thin and retreat, meltwater collects in the hollows they leave behind, often penned in by ridges of rubble called moraines. When such a dam fails, or when an avalanche or landslide throws the water out of its basin, the result is a glacial lake outburst flood, known by its acronym GLOF. Nepal’s valleys have seen them before, and counting the lakes that could produce the next one has become a core task of mountain hazard science.
What the inventory found
A study by Wu Yu-Hong and colleagues, published in Advances in Climate Change Research in 2026, mapped glacial lakes across the China–Nepal Himalaya from satellite imagery between 1992 and 2022. It identified 2,377 lakes. Their combined area grew by 53.79 square kilometres, a rise of 27 per cent over the 30 years, or an average of roughly 1.8 square kilometres of new lake surface a year. Of the lakes, 76 were classed as potentially dangerous, including 4 at very high risk and 14 at high risk.
The authors also estimated what lies in the path of a flood from those lakes: about 260 kilometres of road, 0.39 square kilometres of buildings and nine hydropower facilities. The paper is separate from ICIMOD’s own regional inventory, which ranked 47 potentially dangerous lakes in the Koshi, Gandaki and Karnali basins.
How a lake becomes potentially dangerous
Classifying a lake as potentially dangerous is a judgement built on several criteria rather than a single measurement. Studies of this kind typically weigh the type of dam, whether moraine or ice, the size of the lake and how fast it is growing, the steepness of surrounding slopes, the proximity of glaciers or unstable ground that could send ice or rock into the water, and what lies downstream. The criteria often differ between studies, so counts of dangerous lakes are not directly comparable, which is why ICIMOD’s 47 and the 2026 study’s 76 refer to different regions, periods and methods.
Growth is not the same as danger
The 27 per cent figure is a total. It says the region’s lakes have collectively spread, not that each is more threatening. A lake’s danger depends on the nature of its dam, its volume, the slopes and glaciers above it that might send a wave into the water, and the settlements and infrastructure below.
The variation is large. A separate preprint, not yet peer reviewed, tracked four potentially dangerous lakes in central-eastern Nepal from 1992 to 2024. Lower Barun grew by 258 per cent, from 0.77 to 2.76 square kilometres. Hongu 2 grew by 137 per cent and Lumding Tsho by 119 per cent, while Thulagi expanded by 29 per cent. Growth in area is a warning that volume may be rising, but it is not a forecast of failure.
Area is also easier to measure from orbit than volume. A lake that spreads across a flat basin can hold less water than a smaller, deeper one, and only field surveys or modelled bathymetry can tell them apart. Cloud cover, seasonal ice on the lake surface and the resolution of the imagery add further uncertainty, particularly for small lakes. The 27 per cent figure is best read as a well-founded indicator of direction and scale, not as a measure of stored water or of hazard.
Not every flood begins in a lake
On 26 August 2026 a rock-ice avalanche fell from the north face of Langtang Lirung into the Lhende Khola valley on the Nepal–Tibet border. ICIMOD states that the trigger was bedrock failure that took the glacier with it, and that no glacial lake was involved at the origin. It contrasts the event with a flood in the same valley in July 2025, which came from the Purepu glacier lake.
The two events in the same valley system were little more than a year apart, one from a lake and one from the rock above it.
The distinction is important because the hazard maps built around lakes could not have flagged the collapsing rock wall. Chamoli, in India’s Uttarakhand, offers a similar precedent: the February 2021 disaster was caused by a rock and ice avalanche, according to a study in Science. These are cascading events in which ice, rock and water interact, and a lake inventory captures only part of that chain.
Lakes can also appear after an event. Reports in the days following the August collapse described two barrier lakes forming behind debris, creating a secondary flood risk. Such lakes are landslide dams and belong to no glacial-lake inventory that ends in 2022.
Monitoring options range from repeat satellite mapping, which is cheap and regional, to sensors and cameras on individual lakes, which are costly and reserved for the few judged most dangerous.
What exposure counts are for
The counts of roads, buildings and hydropower plants downstream are not predictions of loss. They identify where monitoring and warning systems would protect the most, and they show that energy infrastructure sits within the reach of lake hazards as well as of avalanche ones.
The exposure counts also carry a caveat. They measure what lies within modelled flood paths, which depend on assumed lake volumes and dam-failure scenarios, so they estimate potential loss rather than realised loss.
An inventory is a map of one class of hazard, drawn as carefully as satellites allow. It answers where the water is and how much there is. What it cannot yet say is when a given lake will fail, or whether the mountain above it will move first. The harder question raised by 2026 is whether comparable maps exist for the rock walls that hang above the valleys.
– Srinivas VR Yadavalli
