It is difficult to reproduce the natural conditions of the real landslide dam with the experimental model established in the laboratory. Therefore, this study placed the experiment site in a river channel with natural water flow to simulate the failure process of natural landslide dams. The experiment site is located in Huashi Gully, which is the branch of Mianyuan river, Tianchi Township, Mianzhu City, Sichuan Province, China, as illustrated in Figure 1. The 2008 Wenchuan earthquake induced many barrier lakes in Mianyuan River, such as Xiaogangjian barrier lake, Xujiaba barrier lake and Wenjiagou barrier lake. Hence, the in-site model tests of barrier lakes in Mianyuan river basin are very representative. The Huashi gully is sparsely populated and convenient for transportation, there are rural roads along the gully. In Huashi Gully, the measured average flow during the experiment period was 0.25 m³/s.

According to the historical statistics of grain size distribution of landslide dam materials, clay-sand-rock mixture and sand-rock mixture are the most common landslide dam materials (Jiang et al., 2018). Based on the field investigation, there are large number of natural graded granular materials (sand and gravel) in Huashi Gully near the experiment site, which are directly used in the test to construct the landslide dam. The material of test dam is non-cohesive and wide grading unconsolidated material, the particle size distribution range is 0.005–100 mm. Granular materials with particle size greater than 100 mm are removed to reduce the effect of large size particles on the dam failure process. Although there are boulders which are not in the statistical range of particle size distribution curve in natural landslide dams, the boulders are not considered in the establishment of experimental materials due to the lack of measurement data to quantify the specific particle size and quantity of boulders. The mean grain diameter (d ₅₀) of dam material is 10.0 mm. As shown in Figure 2, the uniformity coefficient (Cu) and the curvature coefficient (Cc) are 50 and 2.64, respectively, which represents the wide and continuous distribution of the particle size.

The scale relationship derived from the similarity analysis is often contradictory in large-scale model test. The model test of landslide lake, which is strictly similar to the prototype, is very rare. In order to study the special phenomenon, it needs to relax the similarity criterion, abandon the generality of similarity, and only focus on the macro effect similarity when designing the large-scale mode test. Therefore, the test model can be designed according to the dominant similarity condition, so that the model can predict the prototype. In order to simulate the erosion of an erodible landslide dam, the material of the model should be stable in physical and mechanical properties, and have the following three characteristics: 1) permeability, 2) sudden collapse and, 3) multiphase flow characteristics of the burst flood. It is found that the dammed lake breaching flood is a kind of structural fluid (mostly diluted debris flow), which meets the similar conditions of flow and sediment transport (Zhang et al., 2015). However, due to the complexity of calculation of bed load transport rate and sediment carrying capacity of debris flow, and the limitation of research level, the accuracy is still difficult to meet the requirements of model design. Meanwhile, the time distortion is not allowed in the landslide dam breaching test, that is, the model time scale should be determined by the flow movement time scale, so the similar conditions of flow sediment transport and riverbed deformation can’t be considered in model design.

Besides geometric similarity and water quantity similarity, the following similarity conditions should also be considered in the design of dammed lake failure model (Zhang et al., 2015).

The similarity conditions of flow gravity: λ v = λ H (1)

The similarity conditions of flow resistance (downstream channel of barrier lake): λ n = 1 λ v λ H 2 / 3 (2)

The similarity conditions of flow process: λ t 1 = λ L λ v (3)

The similarity conditions of sand and gravel initiation velocity: λ v c = λ v (4)

The similarity condition of debris flow velocity: λ v n = λ v (5) Where λ _L is the horizontal scale; λ _V is the flow velocity scale; λ _h is the vertical scale; λ _n is the roughness scale; λ _t1 is the flow movement time scale; λ _VC is the sediment initiation velocity scale; λ _VN is the debris flow velocity scale.

In addition, in order to ensure that the flow pattern of the model is similar to that of the prototype, two limiting conditions of flow pattern and surface tension should be satisfied. Because the dam breaching model is prone to form turbulence and the water depth is large, these two conditions cannot be considered. Since this test does not aimed at any specific landslide dam, therefore, the parameters such as the size of the dam, the material of the dam and the flow pattern of the burst flood are not designed specifically during the test. The similarity analysis shows that the results of this study can reflect the main characteristics of the real landslide dam breach, such as the flood routing process, breach development, downstream sediment accumulation and so on. According to the similarity criterion, the test results can be deduced to the relevant characteristics of the actual dam failure process, and provide reference for risk assessment and emergency disposal of real landslide dam.

Three model tests were carried out in this study, these dams after construction and before test are shown in Figure 3. The cross-section shape of the model landslide dam is trapezoidal, the shape and size of test no.3 are shown in Figure 4. The test dam body adopted the method of layered filling, with the thickness of each layer of 50 cm, mechanical throwing filling and no rolling. To simulate unconsolidated landslide dam material, the granular material was fully mixed and poured into the river using an excavator to form the model dam. The void ratio of each dam material layer was between 0.75 and 0.82, which is consistent with the results of field investigations of nature landslide dams that range between 0.59 and 1.11 (Chang and Zhang, 2010; Zhang et al., 2019). After the construction of dam, the density of dam material was measured by Pycnometer method, and the specific gravity is 2.6, the mechanical parameters of the dam soil as summarized in Table 1.

In the Test No. 1, except for the width of dam crest is 1.0 m and the volume of dam body is 105 m³, other geometric dimensions, dams filling method and test conditions are exactly the same as the other two tests. In the Tests No. 2 and No. 3, the dimension of the model landslide dams is the same, the width of landslide dam crest is 3.0 m, the width of landslide dam bottom is 10.5 m, the length of landslide dam is 9.0 m, the height of landslide dam is 2.5 m, the slope of upstream and downstream dam body is 1:1.5, and the volume of dam is 152 m³, as summarized in Table 2. The difference between these two tests (No. 2 and No. 3) is the discharge channel position on the model landslide dam. In the Test No. 2, discharge channel was located one third away from the left bank of the dam, i.e., 3 m from the left bank, while the discharge channel of test No. 1 and test No. 3 were located at the middle of the dam. The test conditions were set to reveal the influence of dam width and location of discharge channel on the failure mechanism of the landslide dams.

After the completion of the dam filling, the discharge channel is manually excavated at the pre-designed position of the dam crest. The notch is inverted trapezoidal, with a height of 20 cm, a bottom width of 15 cm and a top width of 40 cm. Before the test, a rectangular grid was painted on the downstream slope of the dam to quantitatively observe the erosion failure of the downstream slope of the landslide dam. The grid dimension was 30 cm × 30 cm. In order to prevent the dam body from being washed away during the dam construction, a diversion pipe with a diameter of 400 mm and a length of 10.7 m was embedded in the river channel before the dam body was constructed. Digital cameras are installed upstream and downstream of the dam to record the entire dam breaching process. Furthermore, an unmanned aerial vehicle (UAV) hovered over the dam to record the continuous video. When the construction of the model landslide dam is completed and the impounding begins, the diversion pipe was blocked with a waterproof retaining plate.

According to the water balance equation, the discharge during landslide dam breaching can be calculated as follows: Q o u t = − d V d t + Q i n (6) dV dt = dV dH dH dt (7) where Q _out is the outflow (m³/s), Q _in is the inflow (m³/s), V is the dammed lake storage capacity (m³) and H is the water level (m).

However, the outflow is usually composed of seepage and breach discharge. Field observation shows that seepage exists in the process of impoundment and dam failure. In addition to surface runoff, seepage into the ground, which brings difficulties to the measurement. Considering that the seepage is very tiny and will not cause significant impact on the test result, hence, the seepage flow is not considered in the calculation of dam breach flow.

Different from laboratory test, the topography of the dammed lake area in the field test is irregular. In this study, terrestrial laser scanner (TLS) technology was used to obtain the point cloud of the barrier lake and establish a three-dimensional surface model. The TLS used was a RIGEL VZ-2000i, a pulse-based scanner produced by RIGEL MTS Austria, with a measurement accuracy of 5 mm and a repeatability of 3 mm. The curve between the water level and the storage capacity of the reservoir can be obtained by the three-dimensional surface model, as shown in Figure 5.The relationship between the lake volume and the water level (d _V /d _H ) was determined by calculating the water surface area of the lake under different water levels. According to the principle of hydrostatic pressure, the relationship between water level and time (d _H /d _t ) is obtained by measuring the water level with a piezometer installed in the dammed lake, with an accuracy of 0.01 mm. In this study, the inflow is the average flow during the experiment period, which was 0.25 m³/s. Base on the water balance equation, combined with the water level-storage curve and real-time water level change curve of the reservoir, the discharge process curve can be calculated. Before the test, the accuracy of discharge process is consistent with the actual situation through field calibration.

Combined with the change process of water level and discharge, and video-recordings data, the development process of dam failure can be divided into three stages: initiation stage, development stage and failure stage, as shown in Figure 7.

Stage I: it is the initiation stage and mainly characterized by downstream slope erosion, formed obvious erosion ditch in the downstream surface. The corresponding time was 0–270 s (Figure 7A and Figure 7B), the reservoir water level and flow increased slowly. t = 0 s is the moment that water travel along the initial notch and finally reaches the Point A The dam crest breach is widened to 0.5 m and the alluvial width at the toe of the slope finally extended to 2 m. Under the action of gravity, the water flows along the downstream slope to find the optimal path to the downstream channel, and the optimal path is the path with the smallest material particles on the slope. At first, the flow velocity is slow, which drives small particles. With the increase of flow velocity, it drives slightly larger particles and gradually forms a fixed channel. In the range of near the dam crest, the flow is small strands. Because the flow drives the fine particles on the slope, the velocity is not enough, and the particles move slowly, so they accumulate near the downstream slope toe and flow radially near the slope toe. The sediment carried by the slope water deposited in fan shape around the downstream slope toe gully, forming an alluvial fan.

Stage II: it is the development stage, which mainly shows that the upstream and downstream of the dam crest are connected, and the dam breach develops from the downstream of the dam crest to the upstream of the dam crest, connecting the whole dam crest. The corresponding time is 270–900 s (Figure 7C and Figure 7D), and the reservoir water level and flow continue to increase slowly. The velocity and discharge of overtopping flow increased gradually, the erosion capacity increased gradually, and the gully widened and deepened gradually. In the loose and concentrated area of fine particles, slope materials are easy to be carried away by water flow and eroded to form gullies. Due to the inconsistency of the density of the model landslide dam and the uneven distribution of the material particles, the anti-erosion ability of the dam body is quite different. The water flow forms a multi-stage stepped small drop sill on the slope surface and a large alluvial fan at the foot of the slope. At this stage, the dam crest breach is widened to 2.4 m, and the alluvial width at the foot of the slope extend to 8 m.

Stage III: it is the failure stage, which is mainly characterized by the undercutting erosion and lateral erosion of the breach, the rapid decline of the bottom elevation of the breach, and the collapse of the slopes on both sides, with a corresponding time of 900–1800 s (Figures 7E–H). The velocity and discharge of the breach increase rapidly, then decrease rapidly after reaching the peak value, and the reservoir water level decreases rapidly. The width of the breach develops to 6.5 m and remains stable. The alluvial width at the foot of the slope reaches 9 m, that is to say, it develops to the banks on both sides of the river. About two thirds of the dam body is washed away by high-speed flow. This stage is the key of the whole breach process, and the development of breach, the decline of reservoir water level and the change of discharge mainly occur in this stage. When t = 1,800 s, the breach discharge is equal to the upstream inflow, and the flow is clear, which indicates that the breach tends to be stable and the dam breaching ends. The change of water flow from turbid to clear can be used as a sign to judge the end of outburst.

For the same dam size and test conditions, the shape and size of discharge channel are the same, the position of discharge channel is different, the failure process of landslide dam is similar, and the residual volume of dam after breaching is different. As shown in Table 3, the residual volume of the dam in Test No. 2 is larger than that in Test No. 3. The residual volume of the dam in Test No. 2 and Test No. 3 are 78 m³ and 51 m³, respectively, and the proportion to the original dam volume is 51 and 34%, respectively. As shown in Figure 8, the final breach width of Test No. 2 is 4.5 m, which is less than that of Test No. 3 of 6.5 m. As shown in Figure 9, the total breaching time of Test No. 2 is approximately equal to the Test 3, 1,760 and 1,800 s, respectively. From the stage time point of view, each stage time of Test No. 2 is consistent with that of Test No. 3, which indicates that the different position of discharge channel has no effect on the dam failure process. The development stage time and failure stage time are very close, which indicates that the peak discharge process is similar. The dam body has high erosion resistibility, and there will be no sudden dam failure during overflow (Liu, 2014). As the discharge channel is located near the left bank in the Test No. 2, the water flow will destroy the dam body on the left side of the discharge channel and wash out all the dam body near the left bank easily, and the rock mass on the left bank is exposed, which has strong anti-scouring ability, so as to reduce the scouring of the dam body. The larger volume of residual dam, the smaller erosion of dam caused by the burst flood, that is, the less sediment carried by the burst flood. Under the same flow condition and topographic features, the less sediment carried by flood, the less impact on downstream channel sedimentation and narrowing. The tests results show that the discharge channel should be set near the bank to reduce the dam erosion, which can provide a scientific reference for the emergency disposal and risk assessment of landslide dam.

Under the same discharge channel and test conditions, the dam failure process is different with different dam crest width. As shown in Figure 10, the peak discharge Q _P of Test No. 1 is 7.55 m³/s, much larger than that of Test No. 3, which is close to 2 times of Test No. 3. As shown in Figure 9, the total failure time of Test No. 1 is 690 s, which is far less than that of Test No. 2 and Test No. 3, 1,760 and 1,800 s, respectively. From the stage time point of view, each stage time of Test No. 1 is significantly less than that of Test No. 2 and No. 3. Compared with each stage time of Test No. 3, each stage time of Test No. 1 was 47, 39, 36%, respectively. As shown in Figure 8, the final breach width of Test No. 1 is 7.5 m, and the final breach width of Test No. 3 is 6.5 m. Combined with the present observation, the collapse process of Test No. 1 is a typical instant collapse, which is very dangerous for the infrastructures and lives in the downstream area. The crest width of Test No. 1 is 1 m, compared with Test No. 3, the heights of Test No. 1 and No. 3 are the same, the dam volume and crest width of the Test No. 1 is small and the stability is poor. In addition, due to the increase of the peak discharge, the sediment carrying capacity of burst flood is also improved, which aggravates the erosion of dam materials, accelerates the damage rate, and shortens the time, accordingly. Under the same hydraulic conditions, the dam is easier to be destroyed, the breaching time is shorter, the flood peak flow is larger, and the harm to the downstream is greater.

The development of breach is an extremely complex process, which is a process with soil-water coupling, involving hydraulics, soil mechanics, sediment movement and other disciplines. Head-cut erosion, slope instability and “two-helix flow” erosion are the three main breach mechanisms of dam failure (Zhang et al., 2009; Zhu et al., 2011; Jiang et al., 2018), which were observed in this experiment. During the field tests, it can be observed that the development of the breach is a trumpet shaped radial development with the outlet of the discharge channel on the dam crest as the origin. However, many existing breach models predict discharge by assuming that the breach linear development along the longitudinal and lateral directions, therefor, the development process of breach should be in-deep research. The channel is formed by the erosion of the downstream slope, then the width of the breach remains unchanged, and the channel is scoured upward and downward at the same time. When the breach develops to the upstream of the dam crest, the development of the breach accelerates, and the width and undercutting depth of the breach increase rapidly, accompanied by slope instability (includes shear failure and dumping failure) and “two-helix flow” erosion (Figure 7E and Figure 7F). Helix flow exists on both sides of the breach, and the flow pattern of the flow channel is composed of two parts: bend flow on both sides and direct flow in the middle. The water flow on both sides is rolling and turbid, and the water flow in the middle is calm and clear. The existence of two-helix flow increases the carrying capacity of flow and promotes the lateral expansion of dam breach. The non-viscous material dam breaching develops rapidly, the flood peak process time is short, and the discharge peak value is large. The results of the three tests show that the discharge of the breach increases sharply and the water level in the upstream decreases rapidly.

The initiation stage of breach has an important influence on the development of later breach, which should be paid enough attention to. At present, it is generally recognized that with the continuous rise of reservoir water level, the water will naturally overflow at the lowest part of the dam crest, and gradually erode on the downstream slope, forming a flow channel, so that the dam breaching begins. The initiation mechanism of debris flow plays a dominant role in the occurrence, development, and damage of debris flow. Moreover, migration, accumulation, and blockage of fine particles in the soil will lead to soil failure, and then develop into debris flow (Cui et al., 2017). As shown in Figure 7B and Figure 7C, after the water overflows, the fine particles are first taken away on the slope. Due to the low velocity and gravity, the fine particles move slowly in the flow and deposit at the foot of the slope temporarily. With the accumulation of fine particles at the foot of the slope, the outflow channel is blocked and the water flows to both sides, resulting in the lateral deposition of fine particles and the formation of an obvious alluvial fan. With the widening of the breach and the increase of the velocity, the ability of erosion and transportation of water flow is enhanced, and the fine particles of the alluvial fan at the foot of the slope are taken away by the flow, and the water is concentrated to form a flow channel. Later, with the increasing of the flow velocity, the flow channel is deepened and widened. In the initiation stage, the factors that affect the erosion ability of materials, such as the material characteristics, particle composition and compaction degree, have a greater impact on slope erosion and flow channel development, thus affecting the change of breaching velocity and the breaching development. Therefore, an in-depth understanding of the initiation stage of landslide dam failure process is imperative for providing guidance for the prediction, mitigation, and disposal of landslide dams.