A Landslide Dam Forms in Seconds and Fails on Its Own Schedule

A Flood With No Storm Attached

Strong ground shaking is over in well under a minute. The flood it sets up can arrive a week later, or two months later, in clear weather, with nothing in the rainfall record to account for it. The object standing between the two events is a pile of rock and soil that fell into a river channel during the shaking and stayed there.

Geologists call it a landslide dam, or a barrier lake when the emphasis falls on the water rather than the debris. The U.S. Geological Survey studied the general case in The Formation and Failure of Natural Dams, Open-File Report 87-392, which compiled landslide dams worldwide and tabulated how long they lasted. Its central finding is uncomfortable: these structures do not fail on any schedule that can be pinned down, and among those that do fail, most fail well after the response has wound down.

The failure mode is not the landslide. The landslide is the visible event, the one that gets counted. The failure mode is the interval between blockage and breach, in which a reservoir fills behind a structure no engineer designed, no agency inspected, and no instrument is watching.

The Chain

The sequence runs in a fixed order, and each step operates on a different clock. That difference in timescales is the problem.

1. Shaking. Seismic waves load steep hillslopes past the strength of the material holding them. The USGS ground failure product works from the premise that "landsliding is more likely where shaking is strong and slopes are steep." Seconds.

2. Detachment and runout. Material leaves the slope. Report 87-392 names the responsible movement types plainly: "The most common types of mass movements that form landslide dams are rock and debris avalanches, rock and soil slumps and slides, and mud, debris, and earth flows." Seconds to a few minutes.

3. Blockage. The debris reaches the valley floor and spans the channel. Whether it spans it completely is the pivot. Of the six geometries in the report's classified inventory of 81 dams, Type I, at 11 percent, is the only type it defines as not spanning the valley. Immediate.

4. Impoundment. The river keeps delivering water to a channel that no longer drains. A lake grows. The filling rate is not a property of the landslide at all: "In many cases the amount of flow in a stream and thus the rate at which a natural lake will fill is directly proportional to the size of the upstream drainage area." Hours to months.

5. Breach. Water reaches the low point of the debris crest, spills over, and erodes a channel down through loose fill. Outflow accelerates as the notch deepens. Minutes to hours, at the end of a wait that may have lasted months.

Six Ways a Valley Gets Plugged

A landslide dam is not a small version of an engineered dam. It is the opposite of one in nearly every property that governs whether a dam holds. An engineered embankment is built in compacted lifts of graded material, keyed into foundation rock, fitted with a spillway sized to pass a design flood, and instrumented for pore pressure and seepage. A landslide dam is dumped: whatever the hillside happened to contain, at whatever angle it came to rest, over whatever alluvium was in the channel, with no outlet works of any kind. The only spillway it will ever have is the one the water carves.

The geometry still varies, and Report 87-392 sorted it into six types by how the debris sits in the valley. Type II, at 44 percent of the classified inventory, "span the entire valley flood, occasionally depositing material high up on opposite valley sides." Type III, at 41 percent, "move considerable distances both upstream and downstream from the landslide failure." Those two together account for 85 percent of the classified dams, a sum computed here from the report's breakdown. Types IV, V, and VI are rare, at 1, 1, and 3 percent. Type V is worth naming because it defeats the intuition that one landslide produces one problem: it occurs "when a single landslide sends multiple tongues of debris into a valley and forms two or more landslide dams in the same reach of river."

What none of the six types determine is survival time. That is set by a ratio the landslide knows nothing about: the volume and coarseness of the debris against the water arriving from upstream. Report 87-392 states the consequence directly. "The duration of a landslide dam is apt to be short whenever a small landslide blocks a stream with a large drainage area." A modest blockage on a major river may be gone in a day. A very large blockage on a small headwater stream may hold for a season while the lake behind it grows to a size that makes the eventual breach far worse than an early one.

The Failure Clock

Report 87-392 compiled the one dataset that matters for the delay question. "Seventy-three examples of landslide-dam failures in which the time to failure is known are compiled in figure 6." The resulting curve is the most complete timing record that exists for these structures, and it reads as follows: "Twenty-seven percent of the dams failed within 1 day of formation; 41 percent failed within 1 week; about 50 percent failed within 10 days; 80 percent failed within 6 months, and 85 percent failed within 1 year of formation."

Five published points. Everything that follows in this section is arithmetic performed on those five points, and each derived figure is marked as such.

First derived figure. Between one week and six months the published curve climbs from 41 percent to 80 percent, a gain of 39 percentage points of the sample. But 100 is the wrong denominator for a sample member that had not yet failed at one week. The right denominator is the 59 percent of the sample that had not yet failed. Dividing 39 by 59 gives 0.66. Computed here from the report's own figures: roughly two out of every three of the failures that had not yet occurred by the end of week one arrive before the six-month mark. The report publishes the cumulative curve, not this conditional value.

Second derived figure. Twenty-seven percent fail on day one and 85 percent within the year, leaving 58 percentage points in between. As a share of all first-year failures, 58 divided by 85 is 0.68. Computed here: about 68 percent of the first year's failures occur after day one is over. The single day that gets all the attention accounts for under a third of the annual total.

Third derived figure. The two intervals compare as rates. From day 1 to day 10 the curve rises 23 points across 9 days, about 2.6 points per day. From day 10 to the six-month mark, treated here as day 180, it rises 30 points across roughly 170 days, about 0.18 points per day. Computed here: the daily share of the failure sample landing between day one and day ten runs roughly fourteen times the daily share over the following 170 days. That thinning is why the risk feels finished long before it is. The report's sample cannot say whether the underlying hazard falls with it; it contains no dams that never failed.

The report attaches a limit to all five figures: "these percentages pertain only to landslide dams that have failed." The curve describes the timing of failures, not the probability of one. The report notes separately that "numerous other landslide dams have formed and not failed," and USGS Bulletin 2130 records that "some landslide dams become long-lasting geologic features."

Overtopping Does the Cutting

The mechanism of failure is narrower than the range of timing suggests. Report 87-392 is unambiguous: "Overtopping is by far the most common cause of failure." The associated process is fluvial erosion by headcutting, in which the eroding notch works progressively upstream until it reaches the impounded lake and the reservoir is connected to the falling channel below.

The alternatives barely register. Piping, meaning internal erosion along a seepage path through the fill, is the mode engineers watch for in built embankments and is nearly absent here: as of that report, "only two cases where a landslide dam actually has failed in this manner are known." Slope failure of the dam faces is rarer still, and Report 87-392 explains why through the deposit's own geometry: "nearly all upstream and downstream faces of landslide dams are at the angle of repose of the material or shallower, and the dams are much wider than tall. In such cases, slope failures are rare."

That geometry explains why the timing is so unpredictable while the mechanism is so consistent. A broad, low, gently faced pile does not fall over and does not slump. It waits, entirely stable in the structural sense, until the reservoir behind it reaches the lowest point on its crest. Everything governing when that happens is upstream hydrology, not geotechnics. Stable does not mean safe, and the two get conflated because in most engineered structures they coincide.

Even the rare piping failures make the same point about time. USGS Bulletin 2130 records that in 1966 a rock- and debris-fall dam on the Tegermach River "failed by piping after having been stable for 131 years."

The Observation Gap

Three observation systems touch this hazard, and the gap between them is structural rather than a matter of funding or effort.

The first is modeling. The USGS ground failure product delivers "near-real-time spatial estimates of earthquake-triggered landslide and liquefaction hazard following significant earthquakes worldwide," built from ShakeMap ground motions and susceptibility proxies, with a preferred landslide output resolving to about 250 meters. Its documentation is explicit about the limits: the models "are not able to account for local characteristics of topography or geology nor to predict specific occurrences." It names this exact hazard as a blind spot, noting that population exposure "does not account well for indirect remote effects (e.g., landslide dam hazards, blocked or damaged roads)."

The second is imagery. After a significant earthquake, USGS analysts review satellite imagery to map what actually failed, and the constraints are physical. In one such assessment the agency reported it "was unable to assess some of the priority areas due to cloud coverage or lack of available imagery," and that the result was "not a complete representation of landslides triggered by the two earthquakes." On blocked valleys specifically: "The USGS did not identify any landslide dams in the areas reviewed, but landslide dams could exist in areas that were not reviewed or that were obscured by cloud cover." That is not hedging. It is an accurate description of what an imagery review can establish.

The third is the river network itself, and here the division of labor matters. The USGS operates a network of roughly 7,000 streamgages and "has no responsibility or authority for making forecasts." Forecasting sits with the NOAA National Weather Service, distributed across 13 River Forecast Centers, whose inputs are radar, gage networks, soil zones, temperature, and coordinated reservoir releases. A gauge measures water that has already reached it. A forecast model built on precipitation and soil moisture is not looking for a reservoir that appeared without precipitation, upstream of every instrument, behind an obstruction that is not in any dam inventory.

The domestic mapping baseline is thinner than the streamgage count suggests. The USGS compilation of landslide inventories across the United States exists precisely because "spatial data on landslide occurrence across the U.S. varies greatly in quality, accessibility, and extent," assembled from inventories "collected and maintained by different agencies and institutions, usually within specific jurisdictional boundaries, and often with varied objectives." The associated national inventory contains 613,724 mapped landslides, and the susceptibility model's documentation notes that it "does not explicitly account for long-runout and does miss a handful of landslides," a share the same documentation places at under 1 percent. Long runout is the property that carries debris all the way into a channel.

One further timing note extends the window past the breach of any single blockage. USGS assessments of earthquake-triggered landslide hazard state that "landslides can still happen for the next few days to weeks," that "aftershocks and heavy rainfall can trigger additional landslide activity," and that "areas downstream, or downslope, from earthquake-triggered landslides may be at an increased risk from landslides and debris flows for several years." A second blockage can form in the same reach long after the first has emptied.

Where This Doesn't Apply

• Low-relief terrain. The chain requires a slope steep enough to deliver material into a confined channel. Across flat or gently rolling country, strong shaking produces liquefaction and lateral spreading rather than valley-blocking runout, and no reservoir forms.
• Incomplete blockages. Type I dams do not reach from one valley side to the other. Flow continues past them. They may aggrade or deflect the channel, but they impound no lake, and the failure clock never starts.
• Large drainage areas behind small blockages. When the upstream basin delivers far more water than the debris pile can hold back, the structure is overtopped almost immediately, resolving inside the earthquake response itself. Those cases sit in the sample's day-one 27 percent, not in the delayed window.
• Blockages that reach equilibrium. Some never breach. Where the deposit is coarse and permeable enough that seepage matches inflow, the lake level stabilizes and the structure can persist indefinitely. USGS documentation of one such case describes inflow and seepage outflow that "appear to be in equilibrium."
• Reaches under active observation. The observation gap is a default, not a law. Where an overflight, a clear satellite pass, a temporary gauge, or a road crossing establishes that a valley is blocked and how high the lake stands, the hazard becomes a monitored engineering problem rather than an unmeasured one.
• Non-seismic blockages. Landslide dams also form from rainfall, snowmelt, volcanic activity, and undercutting with no earthquake involved, while glacial ice and moraines dam valleys through different processes entirely. The failure statistics above cover landslide dams as a class; the seismic trigger is one path into it.

This failure mode does not persist because the physics are obscure. Overtopping followed by headcut erosion has been the documented dominant mechanism since at least 1987, and the survival curve has been in print just as long. It persists because the hazard's clock and the response's clock run in different units. Shaking is measured in seconds and mapped in minutes. The blockage is measured in cubic meters and mapped when the clouds clear. The lake is measured in days of inflow. The failure, when it comes, is measured against an attention span that expired somewhere near the middle of the curve.