The activities of debris flow (DF) in the Wenchuan earthquake-affected area significantly increased after the earthquake on 12 May 2008. The safety of the lives and property of local people is threatened by DFs. A physics-based early warning system (EWS) for DF forecasting was developed and applied in this earthquake area. This paper introduces an application of the system in the Wenchuan earthquake-affected area and analyzes the prediction results via a comparison to the DF events triggered by the strong rainfall events reported by the local government. The prediction accuracy and efficiency was first compared with a contribution-factor-based system currently used by the weather bureau of Sichuan province. The storm on 17 August 2012 was used as a case study for this comparison. The comparison shows that the false negative rate and false positive rate of the new system is, respectively, 19 and 21 % lower than the system based on the contribution factors. Consequently, the prediction accuracy is obviously higher than the system based on the contribution factors with a higher operational efficiency. On the invitation of the weather bureau of Sichuan province, the authors upgraded their prediction system of DF by using this new system before the monsoon of Wenchuan earthquake-affected area in 2013. Two prediction cases on 9 July 2013 and 10 July 2014 were chosen to further demonstrate that the new EWS has high stability, efficiency, and prediction accuracy.
After the Wenchuan earthquake on 12 May 2008, large areas of slopes were determined to be potentially unstable (Zhang et al., 2016). These can be readily transformed into landslides by intensive rainfall and thereby supply abundant loose solid material for debris flow (DF) formation. The increasing availability of solid material enhanced the activity of DFs, especially in the earthquake-affected area. The enhanced DFs will threaten the lives and property of local residents for a substantial period of time (Tang et al., 2009; Xie et al., 2009). The construction of DF control projects is an effective mode of DF defense. However, there are so many DF valleys in need of disaster mitigation that the construction of DF control works will cost a great deal of money and take a long time (Papa et al., 2013). Therefore, building an early warning system (EWS) to predict DFs within the Wenchuan earthquake-affected area is essential for disaster mitigation.
Disasters such as floods, landslides, collapses, and DFs can be predicted using EWSs. However, the EWSs in this study will only focus on the function of DF prediction. The EWS has become an essential feature of civil protections and is commonly recognized as a vital risk prevention tool (European Commission D. G. Environment, 2008). In recent years, there has been a tendency towards increasing the number of operational landslide and DF EWSs (Berenguer et al., 2015). For example, the EWS used for predicting DF and landslides in Japan (Osanai et al., 2010), the landslide EWS in Hong Kong (Chen and Lee, 2004), as well as some regional systems for DF prediction in China (Wei et al., 2007) and in Italy (Aletti, 2004). All the mentioned EWSs are based on the empirical rainfall thresholds derived from the analysis of past rainfall events that induced DF. Since the pioneering work by Caine (1980), empirical rainfall thresholds for shallow landslides and DFs have been developed at local, regional, and global scales (Giannecchini et al., 2012; Borga et al., 2014; Vennari et al., 2014). However, rainfall thresholds can only be obtained from the DF valleys where certain amounts of rainfall information and related DF data are available (Papa et al., 2013; Zhang et al., 2014b). Additionally, Guzzetti et al. (2007, 2008) reported that a majority of empirical rainfall thresholds available in the literature were defined using non-objective and poorly reproducible methods (Gariano et al., 2015). The above-mentioned problems in identifying empirical rainfall thresholds are mainly attributable to the lack of a physical framework of DF formation, and hence the accuracy of an EWS based on this empirical mode is relatively low (Wei et al., 2004). For example, a false positive rate of approximately 40 % is reported for threshold relationships in the Alpine region (Badoux et al., 2012). A physics-based EWS that accounts for the formation mechanism of DF is a potential path to solving these issues (Zhang et al., 2014a; Papa et al., 2013).
The evolution of shallow landslides into DF is one type of formation mechanism. It is believed that this evolution is derived from a single slope (Iverson et al., 1997). This evolution process is now generally accepted as “the slope debris flow” (Li et al., 2010) or “debris flow on slope” (Zeng et al., 2004; Berti and Simoni, 2005; Kim and Lee, 2015). This type of debris flow formation has been well studied, and its corresponding prediction models have been established based on the common view that pore water pressure is the key factor triggering this type of DF formation (Cui, 1991; Iverson et al., 1997). However, these prediction models do not allow conclusions regarding whether there will be DF formation at a catchment level, as they tend to focus on a single slope. The debris flow formation at the catchment scale is defined as the gully-type debris flow (Ni, 2015). For example, debris flows that occur in the Jiangjia gully are a typical gully-type debris flow, and landslides are the dominate mechanism of supplying the material for debris flow in this gully (Kang, 1987). Coe et al. (2008) identified the Chalk Cliffs in the USA as producing runoff-induced debris flow, as they found that there was no landslide source when debris flow occurred. Kean et al. (2013) agreed with this identification result during field observation of Chalk Cliffs. These observation results lead us to conclude that rainfall-induced landslides are another kind of supplementary mode to provide material for debris flow formation at a watershed level, in contrast to the entrainment of river bed sediment material for debris flow formation. The volume of expected point source landslides is key in understanding the timing and volume forecasting for DFs (Borga et al., 2014). Papa et al. (2013) argued that the total number of rainfall-induced landslides is the key factor for DF formation in a watershed, and he used a variable to relate failure percentages of landslides to intensity–duration (I–D) curves. Using the curve and forecasted rainfall data, DF in a watershed can then be predicted. This pioneering work can derive an I–D curve from a certain number of physical numerical simulations regardless of whether adequate observation data are available in a watershed. As a consequence, this method overcomes the drawback of empirical mode that depends on a large amount of rainfall and DF data. However, the Papa model is only a quasi-physically based method, because it attempted to predict DF in a watershed without accounting for the effect of rainfall-induced runoff.
In fact, rainfall-induced landslides and rainfall (or rainfall-induced runoff) are the two necessary factors for DF formation, and the interactions between them can induce DF formation at a watershed scale (Cui et al., 2013). Hence one can see that the descriptions of DF formation at a watershed scale can be divided into two coupling stages. The first stage involves the interaction between rainfall and slopes leading to soil-mass failure, and the second stage includes rainfall-induced runoff and solid material from landslides leading to DF formation (Zhang et al., 2014b).
According to descriptions of the two main coupling processes, the authors developed a prediction model based on water–soil coupling mechanism. In this model, the total volume of rainfall-induced landslide and runoff is simulated at each time step in the first stage, and based on which, the density of the soil–water mixture is calculated. In the second simulating stage, using the density of the soil–water mixture as the key parameter, the formation probability of DF at a watershed scale is identified. This model has been applied at different spatial scales (Jiangjia valley in Yunnan province, China; Sichuan province, China) to verify its precision and applicability, and the forecasting results are excellent with low false-negative and false-positive rates (Zhang et al., 2014a, b).
Based on this forecasting model, the main objective of this work is to develop an EWS for DF forecasting that accounts for the formation mechanism of DF at a watershed scale. We also hope to introduce the operational effectiveness of this EWS in the weather bureau of Sichuan province.
Map of the 2008 Wenchuan earthquake-affected area.
Analysis of soil mass stability influenced by rainfall infiltration.
The 2008 Wenchuan earthquake-affected area (IIX–XI intensity), with a total
area 40 162.7 km
The eastern part of the earthquake-affected area is adjacent to the Sichuan Basin, and the other borders (i.e., the northwest, southwest, and northeast) are adjacent to steep mountains. The terrain variations within this region are large, with elevations ranging from 500 to 5000 m. In general, the western terrain within the region is higher than the eastern. In addition, numerous deep canyons created by large rivers (e.g., the Jialing River, the Fujiang River, and the Minjiang River) are common in this region, and their huge elevation differences and steep slope gradients can provide the energy conditions necessary for DF formation. The elevations and slope gradients of the earthquake-affected region are shown in Fig. 2.
As the main geological structure of the earthquake-affected region, the Longmen Shan fault zone consists of three fault zones, including the Wenchuan–Maoxian fault, the Yingxiu–Beichuan fault, and the Peng–Guan fault. Due to its large scale, long geological development, and frequent activity, the Longmen Shan fault zone has provided abundant loose solid material (Xie et al., 2009). The dramatically increased quantity of loose solid material from the Wenchuan earthquake-induced geological disasters (e.g., collapses, rockfalls, and landslides) further enhanced the development conditions of DFs (Wei et al., 2012).
Due to the blocking action of the Longmen Shan Mountains, the warm and wet air from the southeast makes the piedmont region one of the rainstorm centers of northwestern Sichuan Province, whereas the leeward slopes to the west are dry. Although the mean annual precipitation in the north of Wenchuan, Maoxian, and Lixian counties is 500–800 mm, the maximum daily precipitation can reach 35–75 mm. The mean annual precipitation in the south of Wenchuan, Qingchuan, and Pingwu counties is 800–1200 mm. The mean annual precipitation of the other regions (e.g., Beichuan, Anxian, Shifang, Mianzhu, and Dujiangyan) is greater than 1200 mm and sometimes even reaches 2500 mm (Xie et al., 2009). The runoff caused by rainfall is not only the water component of DFs but also provides the power for DF formation.
The two soil–water coupling processes described above are the main
simulation modules in our prediction system. Detailed descriptions are
listed in the literature (Zhang et al., 2014a, b). Therefore, we will
just provide a brief introduction for the basic principle of this system.
Stability variation simulation induced by rainfall – Influenced by
rainfall infiltration, the increasing soil water content will decrease the
matrix suction of the soil mass. This is the main cause of failure in soil
masses on slopes. Therefore, based on the failure criterion of the
unsaturated soil, the limit equilibrium analysis method with the safety
factor ( where Calculation of the density of the water–soil mixture under the influence
of rainfall – DF is a water–soil mixture with a high density, and the density
represents the material quantity in the DFs. A certain density of mixture is
necessary for the formation of DF at the watershed scale. Therefore, the
density of the water–soil mixture, calculated by Eq. (2), can be employed to
estimate the DF formation probability at the watershed scale. where Estimation of the probability of DF formation based on the mixture
density – High density is a key characteristic that distinguishes pure fluid
or hyper-concentration debris flow. If there is inadequate soil material
from landslides DFs cannot develop even during extreme rainfall, because
inadequate landslides cannot guarantee a water–soil mixture reaching to the
density standard of DF. The volume of soil mass from landslides is the most
important influencing factor on the mixture density. When the volume of
unstable soil material from landslides increases, there is a larger density
of water–soil mixture, and this creates a more favorable condition for
debris flow formation. Therefore, we conclude that when the mixture is
denser, the volume of unstable soil material induced by rainfall is greater
and debris flow formation is more likely. However, there is no function that
can be used to describe this qualitative relationship. According to the
research of Kang, the density of the DFs in nature varies in the range of
1.1–2.3 g cm
Formation probability of DFs according to the water–soil mixture density.
Prediction process of the debris-flow prediction system based on water–soil coupling mechanism.
The workflow of the DF prediction system based on the water–soil coupling
mechanism consists of six steps (Fig. 4): basic data preparation, dynamic
data inputs, potential DF watershed identification, hydrological process
simulation, water–soil mixture density calculation, and estimation of DF
formation probability. The core parts of this system are the hydrological
process simulation and the calculation of the mixture density, and their
objects are potential DF watersheds. In every prediction process, the soil
water content value is first updated and saved from 1 January of the
prediction year to the prediction end-point. The value of the soil water
content just before the prediction end-point is used as the initial value
for the following DF prediction. However, this prediction process takes a
long time and cannot satisfy the service requirements of DF prediction. To
improve the prediction efficiency, the system should be run once to save the
value of the soil water content prior to the rainy season.
Acquiring dynamic data – The dynamic data (meteorological data and
predicted precipitation data) are provided by the local bureau of
meteorology. The meteorological data (e.g., measured rainfall, temperature,
relative humidity, wind velocity, and sunshine duration) for the system
should be in a raster format with the same range and resolution as the digital elevation model (DEM).
This data can be obtained by using the spatial interpolation method to
accommodate the data supplied by weather stations. The predicted
precipitation data shares the same format, range, and resolution with the
meteorological data. This data can be obtained by using data extraction and
resampling techniques to handle the data from Doppler radar. Identification of potential DF watersheds – The system uses the DF
watersheds as the prediction units and estimates the formation probability
of DFs at the watershed scale. Therefore, it is the principal premise to
determine that watersheds in the prediction region have the possibility of
DF occurrences. According to previous research, the area of DF watersheds
varies from 1 to 100 km Xiong (2013) selected the relative height and area of small watersheds as
the identification indices to determine the watersheds with the possibility
of DF occurrence, and the study established an identification model of
potential DF watersheds based on energy conditions. ,
where Based on the energy conditions, an identification subsystem of the potential
DF watersheds was established using GIS and database techniques. This
subsystem was embedded in our prediction system as an independent module.
The identification steps of this subsystem are as follows. (1) Hydrologic
modeling is used to extract the small watershed data from the DEM of the
prediction region. (2) The raster-to-polygon tool is used to convert the
watershed data into vector format data. (3) The calculate-geometry tool is
used to compute the area of each watershed. (4) The extract-by-mask tool is
used to obtain the DEM of each watershed, as well as the relative height
(the difference between the maximum and minimum elevation values), and area
of each watershed. (5) The identification model is used to judge whether
each watershed is a potential DF watershed. Simulation of hydrological process in watersheds experiencing rainfall
–
According to Eq. (2), to use the mixture density For a given watershed, the slope parameters, i.e., the soil cohesive force,
internal friction angle, and slope gradient, are provided. According to the
limit equilibrium Eq. (1), the safety factor ( Calculation of the soil–water mixture density in watersheds experiencing
rainfall – Taking the soil water content and matrix suction as dynamic
inputs, the safety factor ( where where Estimation of the probability of DF formation in the watersheds – By
comparing the density value
The land use information in the earthquake-affected area.
The characteristic soil hydraulic parameters in the earthquake-affected area.
The soil depth information in earthquake-affected area.
Data for the hydrological process simulation – The hydrological process
simulation is performed using grid cells as the basic units. A greater
number of grid cells results in lower computational efficiency but higher
prediction precision, and vice versa. To improve the prediction efficiency
and ensure that the prediction precision meets the demand, a compromise
should be found for this problem. Because the commonly used DEM resolutions
are 30, 90, and 250 m and the area of each DF watershed is 0.1–300 km The original underlying surface data, such as land use type, soil type, and
soil depth, in this region can be obtained from the FAO database
( Soil mechanical parameters in the earthquake-affected area – The
assessment approach for the soil mechanical parameters in the
earthquake-affected area consists of three steps. (1) Based on the
geological map of earthquake-affected area, the lithology of the region is
obtained. (2) Based on the handbook of rock mechanics, the mechanical
parameters (such as the soil cohesion
Map of the DF watersheds and potential DF watersheds.
With the DEM of the earthquake-affected area as the input, the identification subsystem can extract the potential DF watersheds in this region. The identification results show that there are 631 potential DF watersheds in this region. As shown in Fig. 7, we compared the identification result (polygon data) and the 669 identified DF watersheds (point data) based on field observation. In total, 98.7 % of the identified DF watersheds are located in the potential DF watersheds, and certain identified DF watersheds may be located in the same potential DF watershed. Some missing points are near the edge of the potential DF watersheds, and the other missing points are almost all located in the Sichuan Basin.
Input of meteorological data – The original meteorological data, including
the 65 weather stations in the earthquake-affected area and nearby regions
(Fig. 8), are provided by the weather bureau of Sichuan province. The
meteorological data from each weather station consist of measured rainfall,
temperature (maximum, mean, and minimum temperature), wind velocity,
relative humidity, and duration of sunshine. According to the coordinate
information of each weather station, these data are converted into raster
data with a resolution corresponding to the DEM via the Kriging spatial
interpolation method. Input of radar-predicted precipitation data – The original predicted
precipitation data in the earthquake-affected area on 17 August 2012 are
provided by the weather bureau of Sichuan province, but they do not directly
meet the demands of this prediction system. The predicted precipitation data
for this prediction system should have the same resolution and coverage area
as the DEM of this region. Therefore, according to the original predicted
precipitation data and this regional DEM, the predicted precipitation data
for this prediction system can be generated using resampling techniques,
etc. (Fig. 9).
Distribution of weather stations in the earthquake-affected area and nearby regions.
Predicted distribution of 24 h of cumulative precipitation.
The formation probability and the warning level of each potential DF watershed in the earthquake-affected area were predicted for 17 August 2012 via the DF prediction system based on the water–soil coupling mechanism. The prediction results at Beijing time zone 19:00 p.m. are shown in Fig. 10. According to the DF events reported by the Land and Resources Department of Sichuan Province, 156 watersheds experienced DFs in the Wenchuan earthquake-affected area. The results of this prediction system show that there are 161 watersheds that are at risk for DFs. The false negative rate (in which DFs occurred but were not predicted) is 19 %, and the false positive rate (in which DFs are predicted, but no DFs occurred) is 21 %. Further analysis of the watersheds with false negative shows that there are 4 watersheds (labeled as “Unsuccessful watersheds: 1” in Fig. 10) that are not classified as potential DF watersheds, and that the 24 hours of cumulative precipitation is less than 20 mm in 6 watersheds (labeled as “Unsuccessful watersheds: 2” in Fig. 10). If these 10 watersheds with false negatives are excluded for the two reasons mentioned above, there are 19 watersheds with false negatives (labeled as “Unsuccessful watersheds: 3” in Fig. 10). We therefore conclude that the actual false negative rate for the system is only 12 %. As for the 34 watersheds with false positives, a portion of them were actually caused by the precision of the radar prediction rainfall. The radar prediction precipitation in the 16 watersheds with false positives nearly exceeded 100 mm on the prediction day. As a consequence, the warning level generated by the system ranged from green to orange with the radar prediction rainfall as input. However, the observed rainfall (Fig. 11) in these 16 watersheds was less than 10 mm (labeled as “Watersheds: false-positive” in Fig. 11). This was far smaller than the radar prediction precipitation and represents a condition in which DFs are difficult to trigger. Thus, if the 16 watersheds with false positives are excluded for this reason, the false positive rate for the system is only 11 %.
Prediction results of the debris-flow prediction system based on the water–soil coupling mechanism.
Distribution of observed rainfall for the prediction day in the earthquake-affected area.
The formation probability of DFs in this region on the prediction day is also forecasted by the system based on the DF contribution factors from the weather bureau of Sichuan Province. The so called DF contribution factors mean the variables that can contribute to DF formation, which include rainfall, fault activity, lithology, and slope, etc. (Wei et al., 2006, 2007). This system divides the critical rainfall into a series of grade intervals according to the different conditions of DF formation, and it then utilizes fuzzy mathematics to determine the formation probability of DFs based on the predicted precipitation interval range and the underlying surface conditions (Wei et al., 2006, 2007). The prediction results for 17 August 2012 are shown in Fig. 12.
Comparison of the prediction results of the two systems.
According to the Fig. 12, the expression mode of the warning level in the prediction system based on the DF contribution factors is entirely grid-unit based. But DF generally occurred in a relatively small watershed (Wei et al., 2008), and the grid unit is difficult to express the morphology of a DF watershed. So the expression format using the grid unit should be replaced by the watershed unit. There are two rules that govern the above-mentioned transformation process: the rule of highest warning level (the warning level of a watershed is assigned the highest warning level within the watershed) and the rule of most quantity warning level (the warning level of a watershed corresponds to the predominant warning level within the watershed). Figures 13 and 14 indicate that some watersheds can even adopt two different warning levels according to the above-mentioned two rules; therefore the transformation results are somewhat uncertain. Additionally, this transformation process requires a significant amount of manual labor, and it is therefore not beneficial to the real-world DF prediction.
Prediction results of the debris-flow prediction system based on contribution factors.
Transformation of results via the rule of highest warning level.
Transformation of results via the rule of most common warning level.
The results obtained by the above-mentioned two prediction systems are compared to test the capacity of our proposed system. The comparison results listed in Table 5 indicate that the prediction results of our system are quite good, with false negative and false positive rates of 19 and 21 %, respectively. These rates were lower than those based on the DF contribution factors (32 and 45 %, respectively). However, it should be noted that the calculation of initial soil water content for DF forecasting is time consuming. This situation is not beneficial to the operation and represents a disadvantage relative to the current early warning system based in contributing factors.
Due to advantages of the new prediction system compared to the system currently used in operational forecasting projects, the new system was installed at the Sichuan weather bureau in the Wenchuan earth quake-affected area before the rainy season in 2013. The prediction results for 9 July 2013 and 10 July 2014 as generated by the new prediction system are respectively shown in Fig. 15. The false negative and false positive rates of the prediction result for 9 July 2013 are 16 and 22 %, respectively (Table 6).
The DF events in 2014 were not very remarkable, and the exact information of
DF events (e.g., the time and place) is not available for this research.
However, there is still some helpful information on DF events that can be
collected from the Internet. For example, on 10 July 2014, it was
reported that there were DF events in six townships in Lixian County, one
town in Xiaojin County, and three towns in Wenchuan County. The radar
rainfall on 10 July 2014 was used in the new prediction system. The
DF prediction results shown in Fig. 15b show a good agreement with the
reported information from the Internet
(
Analysis of forecasting results on 9 July 2013.
A novel physics-based EWS for DF forecasting was developed and installed in the weather bureau of Sichuan province. The three typical cases studies demonstrated that this EWS has a high operational stability and high prediction accuracy. However, there are several problems with this EWS that need to be studied further. First, the hydrological simulation in this system is so complicated that it will result in a time-consuming prediction process. The calculation of the initial soil water content value for DF prediction is especially time intensive which creates difficulty in meeting the demand of an operational prediction as this calculation process is from 1 January to the prediction end-point with a calculation step of 1 h. To improve the prediction efficiency, antecedent rainfall could be solved by continuously running the model, in this way, the antecedent conditions are ready every day for the computation of possible DF triggering. Another issue is that the total volume of unstable soil mass is used in this system to calculate the density of the water–soil mixture within a watershed. It is assumed that once a soil mass fails, its entire extent will participate in the water–soil coupling process. However, portions of the unstable soil mass may remain on slopes instead of moving into the channel. This will result in over-estimation of density values calculated by this system. Additionally, another issue that can influence the soil mass volume participation in the water–soil coupling process is the soil porosity. This variable can also influence the water volume due to pore water. To obtain more accurate density values, the movement process of unstable soil masses and the role of soil porosity require further study. An additional problem with the system is that the density of the water–soil mixture is used to quantitatively characterize the complex dynamic formation process of DFs. Therefore, this method cannot completely describe this complex dynamic process and represents a grey-box model. To reflect the formation of DFs at the watershed scale more completely, the dynamic DF formation process in watersheds needs to be studied further. Lastly, there are two supplement patterns of material source for DF formation, rainfall-induced landslides from slopes in watersheds and deposits in channel beds eroded by the runoff. The underlying theory of this prediction system is based on the coupling of runoff and solid material from rainfall-induced landslides. Obviously, this prediction system will not work well for predicting DFs with the other sources of material.
This research was supported by the Science and Technology Service Network Initiative (No. KFJ-SW-STS-180), the foundation of the Research Fund for Commonweal Trades (Meteorology) (No. GYHY201006039), the Science and Technology Support Project of Sichuan Province (No. 2015SZ0214), the National Natural Science Foundation of China (No. 61501064, No. 41501114). We appreciate the weather bureau of Sichuan Province for the data services. Edited by: H. Mitasova Reviewed by: two anonymous referees