The 2008 Wenchuan earthquake triggered rapid local geomorphic changes, shifting abundant material through exogenic processes and generating vast amounts of loose material. The substantial material movement increased the geohazard (flash floods, landslides and debris flows) risks induced by extreme precipitation in the area. Intervention measures such as check dams, levees and vegetated slopes have been constructed in specific locations to reduce sediment transport and thereby mitigate the impact of ensuing geohazards.
This study assessed the short–medium-term effects of interventions, including multiple control measures, in a post-earthquake mountainous region. Taking the Xingping valley as an example, we used CAESAR-Lisflood, a two-dimensional landscape evolution model, to simulate three scenarios, unprotected landscape, present protected landscape and enhanced protected landscape, between 2011 and 2013. We defined two indices to assess the intervention effects of the three scenarios by comparing the geomorphic changes and sediment yields.
The results show that the mitigation measures are effective, especially the geotechnical engineering efforts in combination with ecological engineering in the upstream area. The spatial patterns of erosion and deposition change considerably due to the intervention measures. Additionally, the effectiveness of each intervention scenario shows a gradual decline over time, mainly due to the reduction in the reservoir storage capacity. The enhanced scenario performs better than the present one, with a more gradual downward trend of effectiveness. The simulation results evaluated the ability and effectiveness of comprehensive control measures and will support optimal mitigation strategies.
Strong earthquakes can trigger co-seismic landslides and discontinuous rock
masses in mountainous areas, which can increase erosion (Huang,
2009). Consequently, the movement of material through co-seismic landslides
and attendant mass failures modifies mountain landscapes through various
surface processes for days, years and millennia
(Fan et al., 2020). The 2008 Wenchuan earthquake
with a surface-wave magnitude (
To mitigate the above-mentioned hazards and protect the landscape, including downstream settlements, structural mitigation measures have been developed in the affected area, depending on the different site-specific conditions, in addition to technical and economic feasibilities. For example, slope protection with vegetation was conducted to stabilise source material on hillslopes (Cui and Lin, 2013; Forbes and Broadhead, 2013; Stokes et al., 2014). Check dams were also widely used to intercept upriver sediment (Yang et al., 2021; Marchi et al., 2019). Lateral walls and levees, which are longitudinal structures (Marchi et al., 2019), are used to protect settlements near main channels with relatively high levels of sediment discharge.
Although comprehensive control measures have been taken in potentially dangerous sites, improvement of mitigation performance in the Wenchuan earthquake-stricken area is still ongoing. The seasonal and periodic occurrence of massive sediment transport often particularly affects the mountainous area. This might be caused by intense precipitation and the failure of mitigation measures due to rough terrain, vague information about source storage and sometimes relatively low-cost mitigation measures (Yu et al., 2010; Cui et al., 2013). Therefore, understanding and quantifying the effectiveness of intervention measures is crucial for mitigation strategies. Many studies have focused on establishing post-evaluation effectiveness index systems that are not supported by sufficient practices (Zhang and Liang, 2005; Wang et al., 2015). Some researchers compared the changes before and after intervention measures by recording long-term on-site measurements, which require a great deal of time, energy and financing (Zhou et al., 2012; Chen et al., 2013). More recently, studies have compared disaster characteristics before and after mitigation actions through quick calculations using numerical simulations (Cong et al., 2019; He et al., 2022). Nevertheless, these studies ignore the lasting effects of earthquakes on geomorphic changes (longer than the duration of a single event). Therefore, the short–mediu- term (from the duration of a single event to decades after) geomorphic changes obtained from simulations provide more details to interpret engineering measures in notable locations, even in locations inaccessible to humans.
CAESAR-Lisflood (C-L), a two-dimensional hydrodynamic surface landscape evolution model based on the cellular automata (CA) framework, has powerful spatial modelling and computing capabilities (Coulthard et al., 2002, 2013; Van De Wiel et al., 2007; Bates et al., 2010). C-L is used widely in rehabilitation planning and soil erosion predictions in post-mining landscapes (Saynor et al., 2019; Hancock et al., 2017; Lowry et al., 2019; Thomson and Chandler, 2019; Slingerland et al., 2019) as well as studies in channel evolution and sedimentary budget planning for dam settings (Poeppl et al., 2019; Gioia and Schiattarella, 2020; Ramirez et al., 2020, 2022). The applications presented demonstrate the efficiency of C-L model to simulate the surface material migration and landscape evolution after anthropogenic and natural disturbances, which indicates the potential to simulate the complexity of surface processes integrated with different interventions. In addition, many studies applied C-L to investigate the landscape evolution after the Wenchuan earthquake (Li et al., 2020; Xie et al., 2022a, b, 2018). The configuration of the model can be referenced to the study of intervention scenarios in the same post-earthquake region.
In this study, we investigated the impact of different interventions on sediment dynamics and geomorphic changes in an earthquake-stricken valley. Hourly rainfall data over 3 years were generated by daily downscaling to capture extreme events. We then simulated and compared the geomorphic changes and sediment yield in three scenarios that varied in their mitigation compositions and intensities in the catchment. The objectives were (1) to assess the effectiveness of a set of mitigation measures to reduce sediment transport, (2) to analyse the role of each measure on geomorphic changes and (3) to determine the influence of vegetation on catchment erosion.
The study area was the Xingping valley, the left branch of the Shikan River
(a tributary of the Fu River) in north-eastern Sichuan Province (Fig. 1).
Nearly 200 settlements are scattered in the study catchment. The
catchment has a total drainage area of approximately 14
An overview of the study area.
The basement rocks in the study area are mainly metamorphic sandstone, sandy
slate, crystalline limestone and phyllite of the Triassic Xikang Group
(T
Six debris flow flash-flood disaster chain groups have been found in the Xingping valley over the decade after the earthquake. Based on the published work of SKLGP (State Key Laboratory of Geohazard Prevention and Geoenvironment Protection), the geological survey of local government and our biannual field surveys since 2012, we catalogued the time of occurrence, total rainfall and corresponding disaster details of each event (Table S1 in the Supplement). A massive amount of sediment was transported soon after the devastating earthquake in 2008 and 2009. Extensive loose materials were then delivered and deposited in the channel triggered by the extreme rainfall events in 2013 and 2018. Considering the transport processes of landslide material, we divided the study area into three subregions: the source area, the transitional area and the deposition area (Fig. 1). The dashed white lines in Fig. 1c indicate that the loose material can be easily transported from the source area to the deposition area through the transitional zone.
An engineering control project was constructed in the study valley to
intercept the upriver material in October 2010. The project included two
check dams, with one located in the upper source area and the other located
in the transitional zone (Feng et al., 2017) (Fig. 1c). The
upper dam has a storage capacity of
In this study, we examined the intervention effectiveness through the morphological response and sediment yield in the Xingping valley using the C-L simulations. The research entailed four main steps: (1) setting three scenarios with different intervention measures, (2) pre-processing the model input data, (3) calibrating the hydrological component and (4) simulating geomorphic changes and analysing the intervention effectiveness during 2011–2013.
The abundant material mobilised by landslides should be controlled in order to reduce the sediment transport. Therefore, we designed three scenarios by integrating geotechnical engineering with ecological engineering to assess the effectiveness of intervention measures. In Scenario UP, unprotected landscape means the sediment is transported without anthropogenic intervention. In Scenario PP, present protected landscape means that only the present two check dams trapped sediment during 2011–2013 without dredging work over this period (see Sect. 2.2). In Scenario EP, enhanced protected landscape represents the addition of slope protection with vegetation in the source area and levees in the deposition area, in addition to the two check dams in Scenario PP.
Figure 1c shows the locations of the existing two check dams in both Scenario PP and Scenario EP. We determined the placements of additional measures in Scenario EP according to a field survey, which demonstrated that the continuous supply of sediment is mainly from the source area. Therefore, vegetated slopes were designed in the upstream area to prevent erosion by stabilising the topsoil and enhancing the soil's infiltration capacity via roots (Lan et al., 2020).
Considering the damage caused by flash floods to the residential area downstream, the levees (see Fig. S1 in the Supplement and Sect. 3.2.2), i.e. artificial barriers, were placed to protect agricultural land and buildings by preventing water and sediment from overflowing and flooding the surrounding areas. Table 1 shows the scenario descriptions, initial model conditions and input rainfall. The details about the model and input data are introduced in Sect. 3.2.
Scenario settings.
The C-L integrated the Lisflood-FP 2D hydrodynamic flow model
(Bates et al., 2010) with the CAESAR landscape evolution model (LEM)
(Coulthard et al., 2002; Van De Wiel et al., 2007), which is described in detail by Coulthard et al. (2013).
The catchment mode of C-L was applied in this study, in which the surface
digital elevation model (DEM), the bedrock DEM (bedDEM), the grain size
distribution and a rainfall time series are required to simulate the
geomorphic changes and sediment transport. There are four primary modules
within C-L that are implemented as follows:
a hydrological module generates surface runoff from rainfall input using
an adaptation of TOPMODEL (topography-based hydrological model)
(Beven and Kirkby, 1979), a hydrodynamic flow routing module based on the Lisflood-FP method
(Bates et al., 2010) calculates the
flow depths and velocities, an erosion and deposition module uses hydrodynamic results to drive
fluvial erosion by either the Einstein (1950) or the
Wilcock and Crowe (2003) equations, which are applied to
each sediment fraction over nine different grain sizes, and a slope module of material movement from the hillslope into the
fluvial system, taking into account both mass movement when a critical slope
threshold is exceeded and soil creep processes, where sediment flux is
linearly proportional to surface slope.
The C-L model updates variable values stored in square grid cells at
intervals such as DEM, grain size and proportion data, water depth, and
velocity. For the three scenarios, the initial conditions, such as the DEM,
bedDEM, rainfall data and the
To clearly describe the control process, especially the two dams and levees
in the catchment, we unified grid cell scales to 10 m for all input data of
the C-L. The GlobalDEM product with a
The spatial heterogeneity in the source material (Fig. 1c) results in differences in the erodible thickness, which equals the difference between the surface DEM and the bedDEM. We divided the study area into five regions according to the erodible thickness (Fig. S1) by checking the relative elevation of the foundations of buildings, the exposed bedrock and the deposition depth of landslides with respect to ground level. The average thicknesses in upstream low- and high-elevation areas were set to 10 and 3 m, respectively, and the thickness of the erodible layer in the downstream area was set to 3 m. For the river channel and outlet, where there would be a large amount of deposition, the thickness of erodible sediment was set to 5 and 4 m, respectively. As the dams in Scenario PP and the levees in Scenario EP were non-erodible concrete, we set the erodible thickness of these features to 0 m. Eventually, the DEM data were formatted to ASCII raster data as required by C-L. The additional levees and vegetated slopes in Scenario EP and the pre-processes of the DEMs and bedDEMs are shown in Fig. S1.
Another parameter required in each scenario simulation was the
In this research, we compared three scenarios by matching precipitation data
between 2011 and 2013, as mentioned in Sect. 3.1. The source data of
precipitation in 2011–2013 (Fig. 2a) were obtained from the Resource and Environment Science and Data Center ( extracting the hourly rainfall of specific days in 2016 closest to the daily rainfall in 2011–2013 through the threshold setting and producing the genetic operators using the extracted hourly rainfall dataset; mixing the genetic operators by an algorithm (Goldberg, 1989) composed of reproduction, crossover and mutation and repeating these processes until the distance between the sum of hourly rainfall and the actual daily rainfall was less than the set threshold; normalising the hourly precipitation to keep the daily rainfall value unchanged.
Figure 2c shows the downscaled rainfall series between 2011 and 2013. The
downscaled hourly rainfall better captured the hydrological events at an
hourly scale compared to the hourly mean rain (5.27 mm) on the day with
extreme rainfall (126.5 mm), which was far from the actual situation.
Corresponding to the
As introduced by Skinner et al. (2018), the C-L
model is sensitive to a set of input data for a catchment with a grid cell
size of 10 m, such as the sediment transport formula, slope failure
threshold and grain size set. The grain size distribution of sediment was
derived from sampling at 14 representative locations in the same study basin
by Xie et al. (2018). Given the grain
size distribution in this study, the Wilcock and Crowe formula was selected
as the sediment transport rule, which was developed from flume experiments
using five different sand–gravel mixtures with grain sizes ranging between
0.5 and 64 mm (Wilcock and Crowe, 2003). Considering the
steep slopes on either side of deep gullies, a higher slope failure
threshold was determined to replicate the geomorphic changes between 2011
and 2013. Additionally, we found that the probability of shallow landslides
increased with increasing slope gradient from 20 to 50
Because the basin was ungauged before 2015, we replicated the flash flood event in July 2018 using C-L simulations to calibrate the hydrological components. Based on Scenario PP (with two check dams), we used the 2-week hourly precipitation of July 2018 as the input (Fig. S2a), which was recorded by a rain gauge located 2.5 km from the catchment (Fig. S2b). The simulation results (Fig. S2c and d) yielded an erosion map and a maximum water depth map in Scenario PP on 15 July 2018. We selected three locations to compare the deposition and inundation in the simulation results with satellite images and photos (Fig. S3). The simulated sediment thickness and water depth were close to those measured from the images, which indicated that the flash flood event was well replicated by the C-L using the input data.
The C-L model outputs of each scenario include hourly water and sediment
discharge at the basin outlet and EleDiffs (the difference between modelled
DEM at a specified time and initial DEM). We validated the model outputs by
comparing the hourly discharge and EleDiffs reflecting the depth of sediment
deposition or erosion (
Based on the visual analysis and quantitative results, we defined two
formulae to assess the effectiveness of the intervention. The conservation
ability (Ca, Eq. 3) was calculated based on variables in the sediment
balance system (Fig. 3). The sediment volume of deposited sediment
(
The sediment balance system in the study area (region
Additionally, we designed the relative efficiency (Re, Eq. 4) to depict the
efficiency of intervention measures in Scenario PP and EP in sediment loss,
with the comparison to Scenario UP.
Figure 4 shows the input rainfall data and modelled discharge hydrograph
between 2011 and 2013 (Fig. 4a). The comparisons of simulated mean discharge
in April through July and the whole year with field survey materials in the
two locations are also presented (Fig. 4b and c). Concerning the discharge
hydrograph, the peak discharges (63.7, 54.9 and 50.3
The input and output of the hydrograph.
Typical cross-sections are generated (Fig. 5) based on the replicated landform changes in Scenario PP. The first site is located on the upriver road, which is eroded to a depth of 5.7 m according to the simulation results, while the photo shows a depth of no less than 4.0 m without an apparent eroded base. Cross-section #2 and the site photo of the gully show that the eroded depth is approximately 1.0 m. Meanwhile, a clear sediment boundary is found in the building located in the deposition area (# 3), indicating a slightly lower deposition depth than the model predicted.
The comparison of cross-sections from the simulation results to the field measurements after 2013 in Scenario PP.
Figure 6a compares the three annual landform changes in each scenario, which
are classified into nine categories according to natural breaks for
EleDiffs: extreme erosion (
The total area of erosion and deposition in the three scenarios is
calculated to compare the impact of sediment transport (Fig. 6b). The
affected area in Scenario UP is approximately 0.76
Figure 6c compares the extent of geomorphic changes in three situations using the ranges that varied in depth. The areas of light and moderate erosion are greater than the areas of extreme and heavy erosion in all three scenarios. The zone of each erosion degree in UP is more extensive than that in PP, followed by that in EP. In addition, the greater the deposition depth is, the smaller the area of deposition. In particular, the extreme deposition area is greater than the area of heavy deposition in the UP scenario. Further analysis shows that the extreme, moderate and light deposition areas decrease in the order of UP, PP and EP. The heavy deposition area shows the opposite trend, mainly attributed to the check dams and slope protection with vegetation.
As shown in Fig. 7, the control measures and surroundings for the three scenarios are further investigated. Behind the two dams upriver in Scenarios PP and EP, the evident orange clusters indicate deposition. In contrast, these locations are dominated by erosion, shown in green, in scenario UP. Further analysis of the sediment depth shown in Fig. 8 shows that the deposited depth behind the dams in Scenario EP is lower than that in Scenario PP. Additionally, in Scenario PP, sediment trapped by dam 1 is less than that of dam 2, but both have deposition thicknesses of more than 10 m, which exceed the dams' heights (dam 1's height is 10 m, dam 2's height is 9 m). For the simulation results in Scenario EP, the values of deposition depth behind the two dams are nearly 8 m, which is lower than the dams' heights.
Geomorphic changes at key locations of the simulation results for
the UP, PP and EP scenarios.
The depth of deposited sediment in the dams' placements.
The additional ecological protection measure alters the material produced
from the upriver tributary gullies. A sediment volume of
Figure 9 shows the erosion and deposition volumes in the source, transitional and deposition areas and compares the conservation ability (Ca) in each scenario. For all three scenarios, the deposition volume in the source area is less than that in the transitional area, and the largest amount of sediment accumulates in the deposition area. Regarding the eroded sediment, the largest volume is in the transitional area, followed by the transitional area, and the source area presents the lowest volume. Moreover, sediment transport is best controlled in the deposition area and worst contained in the source area under any intervention conditions.
The volumes of sediment and the conservation ability (Ca) in the three areas for each scenario (S: source area, T: transitional area, D: deposition area).
Compared with the Ca of the source area in Scenario UP, the value increases by 138.1 % in Scenario PP, which is attributed to dam 1. Likewise, dam 2 in the transitional area effectively reduces sediment loss, which is reflected by a 52.5 % increase in Ca. Furthermore, the mitigation measures in Scenario PP with vegetated slopes and levees in Scenario EP act best. The conservation ability in the source area increased by 161.9 % due to the dam retainment and slope protection with vegetation, and the levees helped increase the Ca by 3.49 % in the deposition area.
The cumulative sediment yield time series for each scenario and the relative
efficiency of scenarios UP and EP are presented in Fig. 10b and a,
respectively. The steep curve of the output cumulative sediment indicates a
significant increase in deposition. Three increasing stages are consistent
with the rainfall intensity in the three monsoons (May–September). The total
sediment output in UP is the largest at
The relative efficiency over the period of controlling measures by human intervention in PP and EP (Fig. 10a) indicates three distinct stages. Stage I shows that the intervention measures in both scenarios completely prevent sediment transport. Later, stage II shows a peculiar period when the effect of enhanced protective measures in EP is less than that in PP through repeated experiments. In stage III, the relative efficiency of the intervention measures in EP is greater than that in UP, which achieves the long-term effect and stable conservation of solid material.
Calibration and uncertainty analysis are important issues in the CAESAR-Lisflood (C-L) simulation of the geomorphic response to intervention measures based on the CA framework (Yeh and Li, 2006). A preliminary calibration was carried out in our study by reproducing the geomorphic changes and water depth driven by an extreme rainfall event that occurred in 2018. The results (Fig. S3) demonstrated that the C-L model can well replicate the flash flood event using the initial conditions and model parameters. The calibration of the geomorphic response to the intervention measures was derived from a direct comparison between the model results and observed measurements (Figs. 4 and 5). As a result, the simulated water discharge was greater than the measured discharge but on the same order of magnitude. Moreover, the errors of erosion and deposition depth between the simulation in Scenario PP and photographic evidence at three locations were less than 20 %. These results suggest the robustness of the model settings and parameterisation.
The source of uncertainty is mainly from the model parameters and driving factors. Skinner et al. (2018) provided a detailed sensitivity analysis of C-L, indicating that the sediment transport formula significantly influences a smaller catchment modelled by 10 m grid cells. The sediment transport law and the Wilcock and Crowe equations (Wilcock and Crowe, 2003) have been proved suitable in the Xingping valley (Xie et al., 2018, 2022a, b; Li et al., 2020). Nevertheless, the empirical models of sediment transport overpredict bedload transport rates in steep streams (gradients greater than 3 %) (D'Agostino and Lenzi, 1999; Yager et al., 2012). Additionally, the input hourly rainfall data downscaled from the daily sequence is an unrealistic situation. Various sediment transport equations and downscaled hourly rainfall data need to be tested in the C-L model to further decrease uncertainty.
In this study, various measures were taken to represent three intervention scenarios with the goal of controlling sediment transport. The C-L model simulated the geomorphic responses to intervention measures and suggested the considerable influence of intervention measures on spatial modifications and sediment yield. The intervention measures lead to reductions in the total affected area (7.9 %–19.7 %) and lower sediment yields (16.7 %–36.7 %), as demonstrated by the overall evidence (see Figs. 6 and 10). The model's prediction of the overall catchment-scale dynamics in response to extreme events is in line with the viewpoints of other authors (Chen et al., 2023, 2015; Lan et al., 2020).
The mitigation measures change the soil conservation ability considerably in the subregions including source area, transitional area and deposition zone, especially in the source area. Compared to the other two subregions, we postulated that the decreased erosion in the source area was caused by the interactions of loose material and topographic constraints. First, most of the loose solid material triggered by the strong earthquake has stabilised since the 2008 debris flow (details in Table S1). Second, the long and deep gullies are mainly located in the transitional area (Yaogouli, Shicouzi, Yangjiashan) and deposition area (Qinggangping). These gullies provide a greater sediment supply than the source area. As shown in Fig. S4, the movement of the material occurs mainly in the branch valleys in the transitional and deposition zones.
Moreover, morphological changes and the ability of soil conservation in the three scenarios show the unique role played by different intervention measures. For example, check dams are most effective in blocking sediment, and vegetated slopes can further strengthen the conservation ability. The synergetic effect of the combination of check dams and vegetation coverage increases the soil conservation ability by more than twofold. The levees can pose a discernible impact on sediment conservation with specific object-oriented protection.
The effectiveness of mitigation measures decreases over time. We performed
an additional 10-year experiment to reveal the declining trend over an
extended period. We randomly selected one of the 50 repeated rainfall
datasets (year 2016–year 2025) downscaled by
Li et al. (2020), which were generated from
the NEX-GDDP product (spatial resolution:
Rainfall input over 10 years and relative efficiency of sediment
intervention measures.
The storage capacity of the check dams decreases with sediment accumulation, and this decrease necessarily leads to a gradual reduction in intervention effectiveness. However, slope protection with vegetation remains operationally effective in reducing sediment transport by stabilising topsoil over the period when the role of dam reservoirs gradually fails due to the lack of dredging work. Therefore, the vegetation protection strategy is vital for “green development” to reduce sediment loss but requires further efforts.
We built the check dams and levees in our simulations by increasing the
elevation in specific locations where they could not be eroded (see
The methods applied in the study further demonstrate that the C-L is an effective tool for understanding short–medium-term or long-term geomorphic changes (Ramirez et al., 2022; Li et al., 2020; Coulthard et al., 2012a) and testing the effectiveness of intervention measures under different scenarios. Our simulations indicate that the mitigation measures in this study are effective, especially the combination of check dam and vegetated slopes in the upstream area, which could help decision makers optimise the management strategies to control mountain disasters. Though geotechnical engineering is a mature technology that can effectively prevent geohazard occurrence (Cui and Lin, 2013), it has disadvantages such as extensive cost and the difficulty of maintenance. In green development, the planting and maintenance of vegetation cover can effectively prevent erosion by strengthening topsoil and absorbing excess rainwater via roots (Reichenbach et al., 2014; Stokes et al., 2014; Forbes and Broadhead, 2013; Mickovski et al., 2007). Alternatively, these methods can be used to study the impact of tree planting patterns on sediment dynamics.
In this study, scenarios involving check dams, vegetated slopes and artificial barriers were simulated using the C-L model to outline the erosion and deposition areas, measure the impacts of sediment blocking and retention, and thus examine how vegetated slope help stabilise slopes. Four key findings were obtained. First, the geotechnical engineering used for controlling sediment transport are efficient, and their performance in protecting the fragile environment can be improved by integrating with other intervention measures, such as ecological engineering and artificial barriers. Second, the effectiveness of mitigation measures decreases over time. Third, the characteristics of the sediment transport patterns are considerably altered due to the intervention measures. The stabilising sediment ability in the source area increased by 161.9 % with the additional effect of slope protection with vegetation. To sum up, the present intervention measures need to be refined with regular dredging works to maintain the effectiveness of reducing sediment transport.
The version of CAESAR-Lisflood used in this work is available at
The daily precipitation data are available at
The supplement related to this article is available online at:
DW: conceptualisation, methodology, software, writing – original draft preparation. MW, KL and JX: supervision, methodology, writing – reviewing and editing, validation.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the special issue “Hydro-meteorological extremes and hazards: vulnerability, risk, impacts, and mitigation”. It is a result of the European Geosciences Union General Assembly 2022, Vienna, Austria, 23–27 May 2022.
This research was supported by the National Key Research and Development
Plan (2017YFC1502902). The financial support is highly appreciated. The
authors would also like to thank Tom Coulthard and his team for
their excellent work on the freely available C-L model (
This research has been supported by the National Key Research and Development Program of China (grant no. 2017YFC1502902).
This paper was edited by Nadav Peleg and reviewed by Jorge Ramirez, Christopher Skinner and one anonymous referee.