Late on the night of 3 October 2023 a mass of frozen moraine slid into South Lhonak Lake in northwestern Sikkim, India, and the wave it pushed across the lake cut through the natural dam holding the water back. The flood ran down the Lhonak Chu into the Teesta, destroyed the 1,200 megawatt Teesta-III dam at Chungthang in the early hours of 4 October, and carried on into West Bengal. Around a hundred people died or were never found. The hazard was known, the lake was on published watch lists, and the warning chain still did not reach the people in the valley.

The night of 3 October 2023

South Lhonak Lake sits in far northwestern Sikkim at an elevation commonly reported at about 5,200 meters, in a basin left behind as the South Lhonak Glacier retreated. On the night of 3 October a section of the lateral moraine slope above the lake failed and slid in. A large multi-author study published in the journal Science in 2024 reconstructed the sequence and put the volume of frozen moraine material that entered the lake at roughly 15 million cubic meters and the water released by the resulting breach at roughly 50 million cubic meters. Other published reconstructions use different methods and give different numbers, so anyone who needs a figure for design work should read the primary literature rather than take mine.

The flood traveled down the Lhonak Chu into the Teesta system, over a channel distance to Chungthang generally given as on the order of 60 kilometers. Reconstructions place the arrival at Chungthang in the hours around midnight, and the times at each downstream point come from those reconstructions rather than from anyone standing there with a watch. It was raining hard in the region that night, which complicated the early explanation of what had happened and also meant the river was already up before the surge arrived.

Below Chungthang the flow ran through Sikkim’s populated Teesta corridor, past Dikchu, Singtam and Rangpo, and across the state boundary into the Jalpaiguri district of West Bengal. It arrived in the middle of the night, in the rain, in a narrow valley where the road, the powerhouses, the bridges and most of the buildings sit close to the river because there is nowhere else level to put them.

What a glacial lake outburst flood actually is

A glacial lake outburst flood, usually shortened to GLOF, is the sudden release of water impounded by ice or by glacial debris. The impoundment is not engineered and was never designed to hold anything, and it exists only because a retreating glacier left a ridge of loose rock and buried ice across a valley, or because ice itself is blocking a drainage path. Three types matter operationally: moraine-dammed lakes, ice-dammed lakes, and water stored inside or beneath a glacier, the last of which produces the Icelandic jokulhlaup.

Moraine dams fail in a small number of well-documented ways. A wave generated by rock, ice or debris falling into the lake overtops the crest and cuts it down, which is what the Science reconstruction describes at South Lhonak. Water seeping through the loose fill can pipe out fines and enlarge a channel until the crest collapses, and buried ice inside the moraine can melt out and let the crest settle. Earthquakes, extreme rainfall and inflow from an upstream lake failing can trigger any of these. The point for planners is that failure is generally driven by something falling into the lake or something changing inside the dam, not by the lake slowly filling and spilling over.

What comes down the valley is not clean water. A GLOF entrains moraine, channel sediment and boulders, and in steep upper reaches it can behave as a debris flow with a bulked volume several times the water released. That is why a discharge estimate alone understates the destructive effect, and why a structure rated for a large rainfall flood can still be demolished by a smaller GLOF carrying rock.

South Lhonak, a lake that was already on the list

South Lhonak was not an unknown feature. The lake grew substantially over recent decades as the glacier retreated, from a fraction of a square kilometer in mid-twentieth-century imagery to roughly one and a half square kilometers by the early 2020s, with published figures varying according to the imagery date and the method used to delineate the shoreline. It appeared repeatedly in Indian and international assessments of potentially dangerous glacial lakes, and Sikkim authorities and researchers had flagged it specifically.

Action followed the assessment, up to a point. Around 2016 a state-supported expedition installed high-density polyethylene siphon pipes to draw water off the lake and lower its level, a genuinely difficult piece of work at that altitude, and it is widely reported to have reduced the level by a few meters. That is real mitigation and it deserves credit, and it also treats one failure pathway, the hydrostatic one, while doing very little about a large slope failure dropping into the lake and generating a wave.

An instrumented early warning installation had also been placed at the lake. Press reporting and state statements after the event indicated that monitoring equipment installed shortly before October 2023 was not delivering usable data at the moment it was needed, with accounts describing damage or a loss of transmission. I want to be careful here, because the reporting on exactly what was installed, what was operating and what transmitted has not always agreed, and the authoritative account belongs to the Sikkim authorities and the agencies that funded the installation rather than to me.

Chungthang, and a dam that became part of the flood

The first major structure in the path was the Chungthang dam of the Teesta Stage III project, a 1,200 megawatt scheme that was the largest power project in Sikkim. The dam was breached in the early hours of 4 October 2023. Once it failed, the water and sediment stored behind it joined the flood, and the downstream reach received a surge larger than the one that arrived from the lake.

Two questions about that failure remain matters for the official record rather than for me. Whether the spillway gates were opened, and when, was disputed in reporting immediately afterward, with the operator and various officials offering accounts that did not fully agree. Whether the design basis for the structure incorporated a GLOF from South Lhonak, and on what assumed magnitude, is a question for the project’s approval documents and for the inquiries that followed. Both have been examined by Indian authorities and by expert committees, and readers who need the answer should go to those proceedings. What is not in dispute is that the dam was destroyed and that its destruction increased the flood downstream.

The broader pattern is one I would ask any planner with a dam upstream to sit with. A dam in the path of a debris-laden surge is not automatically a mitigation feature. If it holds, it attenuates, and if it fails, it converts stored water into an additional pulse arriving at the worst possible moment, without any of the warning that a slow reservoir rise would give. Which outcome you get depends on freeboard, on how fast gates can be opened, on whether anyone is present at eleven at night to open them, and on whether the incoming load was ever contemplated when the structure was sized.

The toll, the ordnance, and where accounts differ

The casualty figures published for this event differ, and the difference is mostly about the split between confirmed dead and missing. Sikkim state disaster management bulletins issued through October and November 2023 reported dozens of bodies recovered and a larger number of people still missing, and compiled totals of dead plus missing were widely reported in the range of roughly 90 to 100. Some accounts published a figure closer to 40 in the first days, which was a body count rather than a total. If you need a defensible statement, say that around a hundred people died or remain missing according to figures published by Sikkim state authorities in late 2023, and name the source and the date alongside the number.

The Indian Army reported on 4 October 2023 that 23 soldiers were missing after a camp in the Teesta valley was inundated, with one later reported rescued. Vehicles and stored ammunition were swept from that site into the river. In the weeks that followed there were reports of ordnance carried tens of kilometers downstream and turning up on sandbars in West Bengal, and of civilians, including children, injured after handling items they found. A flood that passes a military installation, a chemical plant, a fuel depot or a hospital does not stop being a hazard when the water goes down.

Physical damage was extensive along the Sikkim corridor. Reporting in October 2023 described more than a dozen bridges destroyed, with counts in the mid-teens commonly given, along with long sections of National Highway 10, the main road link into the state, and thousands of people displaced into relief camps. The Teesta channel itself was left heavily aggraded with sediment, which raised flood levels for subsequent monsoon seasons downstream and is a slow consequence that outlasts the emergency phase by years.

Why the inventory of dangerous lakes keeps growing

Glacial lakes form where a retreating glacier uncovers a basin it had previously filled with ice. As glaciers lose mass, more of these basins are exposed, existing lakes deepen and lengthen, and new ones appear in places that held ice within living memory. A global assessment published in Nature Climate Change in 2020 found that glacial lake volume worldwide grew by roughly half between 1990 and 2018, with similar growth in lake count and area. The trend is not subtle and it is not confined to the Himalaya.

Exposure grew alongside it. A study published in Nature Communications in 2023 estimated that on the order of 15 million people live within 50 kilometers of a glacial lake, with the largest share in High Mountain Asia and significant populations in the Andes. Those numbers depend heavily on the distance threshold chosen, so treat them as a description of scale rather than as a count of people at risk.

Two other trends compound the first. Lakes that expand up-valley move closer to the steep ice and rock walls above them, which shortens the distance a failure has to travel to reach the water and increases the size of wave it can generate. Warming also degrades permafrost in high mountain rock and in the moraines themselves, weakening the material in the slopes above lakes and in the dams holding them. A lake assessed as low risk in 2005 can be a different feature by 2025, which makes an inventory a maintenance obligation rather than a project you finish.

What GLOF early warning actually requires

An early warning system for a GLOF is a chain, and it delivers nothing unless every link works on the night. It starts with sensing at the lake, which typically means water level, a seismic or geophone element to catch a slope failure or the breach itself, a camera where power allows, and weather instrumentation. It requires a communications path out of a place with no grid power, no cellular coverage and no road, which in practice means solar and battery with satellite backhaul, sized for winter and for equipment that ices over. It requires a decision rule that fires without a human interpreting a graph, because the whole event may last less than the time it takes to convene a call.

Then it has to reach people. Cell broadcast reaches phones that are switched on and in coverage, sirens reach people within earshot who know what the tone means, and both fail against a population asleep in the rain at one in the morning unless they have been drilled. The downstream community also has to know, in advance and without thinking, which direction is uphill and where the assembly point is. Vertical evacuation of a few tens of meters, done in the first minutes, is the entire survival strategy in a narrow valley, and it cannot be improvised.

The link most often missed is sustainment. Instruments at 5,000 meters need someone to visit them, and the visit is a mountaineering expedition. Batteries degrade, lightning kills electronics, snow buries sensors, and a station that stops reporting looks identical to a station reporting nothing wrong unless the system alarms on loss of heartbeat. Whoever installs a high-altitude station should be funded and tasked for its maintenance for its whole life, and the downstream siren network needs a named owner who tests it on a published schedule. Both of those are budget lines rather than engineering problems.

Infrastructure built without the hazard in the design basis

Most river infrastructure is sized against a flood derived from rainfall and snowmelt hydrology, expressed as a return period or as a probable maximum flood. A GLOF does not come from that distribution. It is a single-source release of a measurable volume, it arrives with a steep front rather than a rising limb, it carries a sediment load that a rainfall flood of the same discharge would not, and it can happen on a clear day. If the design basis was built from a gauge record and no GLOF occurred during the period of record, the hazard is simply absent from the calculation.

The consequences show up in specific, boring ways. Bridge piers are sized for water force and not for boulder impact, and their approach embankments sit at a level chosen for a rainfall flood. Powerhouse intakes and desilting works are designed for a sediment concentration that a debris-laden surge exceeds by orders of magnitude. Roads follow the valley floor because that is where the grade is, and camps, labor colonies and staff quarters cluster at the river because that is where the work is. Gates require power, hydraulics and an operator, and none of those are guaranteed at midnight on the night the surge arrives.

The remedy is unglamorous, and it begins with adding a GLOF scenario to the design basis for anything new in a glaciated catchment, using a volume derived from the actual upstream lakes rather than from a statistical flood, then running the resulting inundation to identify what sits inside the footprint. For existing assets it means honest identification of what cannot be defended and would have to be evacuated instead, siting new construction above the modeled surge even when that costs grade and money, and treating the zero-lead-time band near the lake as a place where nothing occupied should be built. That is mitigation by siting, and it is the only control that works when travel time is measured in tens of minutes.

Takeaways

• On the night of 3 October 2023 a slope failure dropped frozen moraine into South Lhonak Lake in Sikkim, and a study published in Science in 2024 put the entering material at roughly 15 million cubic meters and the released water at roughly 50 million cubic meters, while other reconstructions differ.
• The 1,200 megawatt Teesta-III dam at Chungthang was breached in the early hours of 4 October 2023, and its failure added to the flood downstream rather than attenuating it.
• Casualty figures differ by source and by date, with Sikkim state bulletins through October and November 2023 supporting a total of roughly 90 to 100 dead and missing, while early figures near 40 were body counts rather than totals.
• The Indian Army reported 23 soldiers missing from a Teesta valley camp on 4 October 2023, and ammunition swept from that site was later reported washing up far downstream, injuring civilians who handled it.
• South Lhonak was already on published lists of potentially dangerous lakes and had been partially mitigated by siphoning around 2016, which addressed the overfilling pathway and not the wave-from-slope-failure pathway.
• Glacial lake volume worldwide grew by roughly half between 1990 and 2018 according to a study published in Nature Climate Change in 2020, so a hazard inventory in a glaciated basin has to be re-surveyed rather than filed.
• A GLOF early warning system is a chain of sensing, satellite backhaul, an automatic decision rule, nighttime dissemination and a drilled evacuation, and it fails completely if any single link is down, including through unfunded maintenance at altitude.
• Infrastructure sized against rainfall-derived floods has no GLOF in its design basis, and the practical remedy is a lake-derived scenario, honest inundation mapping, and siting people and structures out of the surge footprint where warning time is too short to matter.