Flash flooding prediction in regions of northern Vietnam using the KINEROS2 model

Recently, there has been a growing interest in the damage caused by natural hazards (Creutin et al. 2013; Kousky & Walls 2014) and flash floods (FFs). The most devastating floods often cause heavy loss of life (Gupta 2006; Ashley & Ashley 2008; Brauer et al. 2011). In northern Vietnam, the development of FFs has induced an imperative need to mitigate their impact (NCHMF 2011). Despite this need, few attempts have been made to mitigate floods in the region. This lack of flood mitigation might be explained by the complexity of FFs themselves and by the remaining prevailing uncertainties (Montz & Gruntfest 2002; Estupina-Borrell et al. 2006; Ntelekos et al. 2006). An extensive amount of research has been accumulated worldwide on many aspects of FFs (Montz & Gruntfest 2002; Morin et al. 2009); many of these works suggested that a feasible approach to FF mitigation is to identify their occurrences early (Khavich & Benzvi 1995; Alfieri et al. 2012; Looper & Vieux 2012; Quintero et al. 2012; Versini 2012). We used a modelling method employing the Kinematic Run-off and Erosion Model (KINEROS2) to assess a past FF on 23rd June 2011 (R23^rd), the Bridging Event and Continuous Hydrological (BEACH) model to calculate θ as input to the KINEROS2, and the results from some meteorological models, namely the Global Spectral Model (GSM) (Krishnamurti et al. 2006) and the High Resolution Model (HRM) (Majewski 2009), for the forecast stage.

In 11 years, from 1995 to 2005, Vietnam experienced up to 300 FF events that resulted in the following losses: 968 people died, 628 people were injured, and material losses worth €71 million occurred. Most of the FFs were in the north of Vietnam (NCHMF 2011). The north of Vietnam, in general, and Yen Bai province, in particular, are identified as being very prone to FFs, and the people are highly exposed to FF problems. In Yen Bai province (case study area), the R23^rd event took four lives in the Nam Khat catchment area. FFs have not only resulted in loss of life and property but also negatively impacted society or had severe social consequences (Ruin et al. 2008). It is believed that for an advanced understanding of FFs, multidisciplinary approaches must be integrated (Villarini et al. 2010). This study focuses on the physical techniques of modelling FFs and their forecast prospects.

FF modelling is complex and linked to the problem of uncertainties (Beven 2001). Although there are large numbers of studies on FFs, their behaviours are not fully understood (Sahoo et al. 2006). Many questions have been addressed, such as questions on the accuracy of model inputs and outputs (Li et al. 2010, 2013; Li & Xu 2014), model structure (Bloschl et al. 2008; Garcia-Pintado et al. 2009; Looper & Vieux 2012), initial conditions and boundary conditions (Abderrezzak et al. 2009; Vincendon et al. 2009; Seo et al. 2012), temporal and spatial scales (Amengual et al. 2007; Reed et al. 2007; Younis et al. 2008), and threshold-runoff uncertainties (Gupta 2006; Ntelekos et al. 2006). El-Hames & Richards (1998) suggested that, for a successful application of flood prediction, complex and comprehensive techniques are often required. Furthermore, as FFs occur shortly after the onset of rainfall events, hydrological models used for FF forecast must have the ability to evaluate the level of risk in a short time (Janal & Stary 2012).

We used the robust model of KINEROS2 for estimating and predicting Q_C (a complete study framework is presented in the Materials and methods section). The model has already proven to be reliable for such tasks in semi-arid regions (Volkmann et al. 2010). The goal of our study is to evaluate its merits for a more humid sub-tropical environment, to identify the key variables that determine its output, and to assess its suitability to predict FF events based on different precipitation input data.

This study focuses on regions in northern Vietnam featuring similar climatic and morphological conditions of typical tropical, steep and dense drainage-network systems. Some representative watersheds in Yen Bai province were chosen for the actual modelling implementations.

Because the event-based KINEROS2 model does not compute inter-storm θ conditions, the antecedent θ (θ_ant) must be provided as the initial condition at the beginning of the model run. The daily soil moisture calculated by the BEACH model could be a good solution.

BEACH was developed by Sheikh et al. (2009) and is a spatially distributed daily basic hydrological model. The main aim of the model is to estimate the daily θ. More details about the model are referenced in Sheikh et al. (2009).

SWAT is a well-known hydrological distributed model. Details about the model and its applications can be found in Neitsch et al. (2009) and others. As the ETa is extremely difficult to validate, we used the SWAT ETa for calibrating the BEACH model.

The land use/land cover (LULC) of Yen Bai province was mapped using Landsat 5 TM scenes (30 × 30-meter resolution) acquired in 2009. The map consists of six classes, including forest, shrub, agricultural land, grassland, water bodies and bare land. The soils of the regions are categorized into six major soil groupings: fluvisols, calcisols, ferralsols, alisols, acrisols and gleysols. The soils were derived from a custom soil map produced by the Environment and Resource Centre-Agricultural Institute of Plan and Design, Vietnam. The Digital Elevation Model (DEM) used for geomorphological inputs was provided by the Vietnam Resources and Environment Corporation and produced in 2009 using the Aerial Photogrammetric technology of the Intergraph Corporation, USA.

We used several rainfall sources for this research including satellite-based (Sat-P), radar rainfalls (Rad-P), gauged and forecasted rainfall (FR). All the data were provided by the Vietnamese National Centre for Hydrological Forecasting.

The Sat-P of the R23^rd event and the rainfall event on 30th June 2011 (R30^th) was derived from 70 images. We developed a model (called SATAS) using the Model Builder application running within the ArcMap environment. The model extracts the rainfall from the original images.

Daily gauged-3-hour rainfall (Gau-P) was recorded by the rain gauge in Nam Kim (Figure 1) and extracted for the KINEROS2 validation of the rainfall on 8th (R8^th) and 31st (R31^st) July 2011. These data were accumulated to daily values and used as input for the SWAT model.

As the forecasted rainfall is a prerequisite for the FF prediction, we used the FR from the numerical weather prediction (NWP) of GSM (3.5 day forecast) and of HRM (5.5 day forecast). The GSM and HRM are operated in local watersheds at times of 00:00, 06:00, 12:00 and 18:00 and at times of 00:00 and 12:00, respectively. The Thiessen approach (Thiessen 1911) can be applied for generating distributed forecasting precipitation.

The in situ measured Q (daily and hourly) recorded by the stream gauges (Figure 1) was available for Nam Kim and Ngoi Hut. The daily data were used for the SWAT, and the hourly data were used for the KINEROS2 model.

The model was calibrated over the five years from 2001 to 2005 for Nam Kim and Ngoi Hut. Some sensitive parameters were adjusted, as summarized in Table 1. Note that some final values were averaged from the values of hydrologic response units (HRUs). Model validation was applied after every calibration from 2006 to 2012. The goodness of agreement between the simulated data and the observed data were evaluated using the coefficient of determination (R²), the Nash–Sutcliffe simulation efficiency coefficient (NSE) and graphical methods. These methods were also applied for BEACH and KINEROS2.

The BEACH calibrated parameters are represented in Table 2, based on the results of the calibration and validation of SWAT. As the BEACH estimates ETa for different types of crops, the SWAT ETa was calculated for each HRU representing the corresponding crop class for these comparisons.

The KINEROS2 model was calibrated for the R23^rd event using the hourly gauged Q and Sat-P. Some significant changes of parameters were implemented in the calibration stage to make the model an acceptable simulator of the real hydrological system. These parameters are summarized in Table 3. Afterwards, the model validation was performed for the R30^th (using Sat-P), R8^th and R31^st (employing Gau-P) events.

While other inputs, such as θ and N, were kept unchanged, in general, using Rad-P produced lower peaks than using the Sat-P (≈observed) in both of the watersheds (Table 4). Although the total Sat-P and Rad-P rainfalls were not much different, the estimated time to peak of the Rad-P was 0.73 hours later than that using Sat-P in Nam Kim and 0.26 hours later than that in Nam Khat.

In performing FF prediction, we face great challenges and uncertainties concerning the input data and model structure (Montz & Gruntfest 2002). Although it is difficult to overcome these obstacles (Beven 2001), it is certainly not impossible. Many previous studies have failed to predict the stream runoff due to insufficient θ condition information (Ntelekos et al. 2006; Javelle et al. 2010; Marchi et al. 2010). Hence, the result of the certified BEACH model might play an important role to address this issue. In addition, all other inputs used for this study, such as the DEM, soil map, or LULC, were produced following the national norms or evaluated using a thorough accuracy assessment.

Concerning the model structure and parameter evaluations, all models were calibrated and validated. In particular, KINEROS2 was tested with several rainfall inputs; such testing had not been performed in many other works. This difference gives us opportunities to compare and contrast the outcomes. Some similarities of using the Gau-P and Sat-P outputs were found, whereas the supervised output of Rad-P underestimating the observed data was found. However, we do not judge that the Rad-P is less accurate than the Sat-P. It might be a question of the old radar generation used in Vietnam. As FFs often occur a short time after rains (4–6 hours) (NWS 2002), a good FF warning system should provide timely information (Lin et al. 2002). This requirement is for the computational operation of models as well. KINEROS2 was validated using an Intel core™ i3, 2GB RAM computer, which only required approximately 20 seconds to run (at Nam Kim). This short run was one of the reasons for choosing this robust model. Only new updated FR is required for an up-coming forecast operation. Other parameters, such as the LULC, are recommended to be updated seasonally, and the soils and DEM are renewed in decades if significant changes are found.

Manual parameter calibration is a time-consuming implementation. Thus, the results of model parameter sensitivity tests might be significant for this work and for future studies as well. These study findings of substantial sensitivities of the Ksat, S and N to the model outputs are consistent with those of Memarian et al. (2013) and Martínez-Carreras et al. (2007). An important point is that this result of model calibration could be applied to ungauged ‘neighbour’ watersheds (Nam Khat for example) using a so-called ‘regionalization approach’ or similar methods already applied by Boughton & Chiew (2007), Makungo et al. (2010) and Servat & Dezetter (1993). Therefore, previously unresolved questions of flooding threshold determination for ungauged catchments might be answered. In addition, Carpenter et al. (1999) considered that the FF threshold is an essential part of FF warning systems. Unfortunately, the FF threshold varies naturally from watershed to watershed; as a result, traditional methods of evaluating the threshold using statistical reports were often not convincing (considering impacts of climate change) and inapplicable for ungauged catchments.

The uncertainties of a hydrological model's outputs are directly linked to rainfall estimations (Villarini et al. 2010). Currently, Adjei et al. (2015), Castro et al. (2014) and Xu et al. (2015) have introduced the newest rainfall estimation for tropical regions using satellite techniques. In addition, Finsen et al. (2014), Kang & Merwade (2014) and Wu et al. (2015) have made efforts to improve radar rainfall and to provide hydrological models with near or real-time data as well. Altogether, the above-described work gives more accurate rainfall availability for hydrological modellers. However, for operational FF prediction, very precise precipitation forecasts are needed at any given time. The GSM and HRM rainfalls were not expected to be the most suitable, but were the only ones available at the time of this study in Vietnam. Because most parameters of the NWP models are highly nonlinear and naturally inconsistent (Majewski 2009; JMA 2013), uncertainties grow when the forecasted period expands further into the future. This might limit the average goal of the USA FF warning systems of 65% of correct alarms (Smith et al. 2005). As such, much work is required to improve the forecast of rainfall particularly with regard to extremes and the consequence of the prediction of FFs.

The θ_ant conditions were found to be extremely important for predicting the flooding magnitude and are therefore very helpful for providing FF guidance. However, unlike discharge and precipitation data, θ data are not routinely observed over a long period. Hence, the use of BEACH for estimates is a promising approach. The Ksat was found to be the most sensitive parameter to determine the simulated Q_C. Although the variance of S and the Ksat had no impact on the time of Q's peak, changes in both parameters had significant effects on Q's volume. KINEROS2 represented a dynamic, robust hydrological model with the capability to simulate discharges (hydrographs) that fitted well to measured data using different rainfall sources. Therefore, we conclude KINEROS2 could be an appropriate model for the purpose of predicting the Q_C and thus for use in forecasts. The applications of the KINEROS2 model with the FR (from GSM and HRM) revealed the possibility to predict the time, magnitude and location of approaching FFs.

The Vietnamese National Centre for Hydro-Meteorological Forecasting and the Vietnamese Environment and Resources Corporation are greatly acknowledged for providing meteorological and topographical data. The authors acknowledge the anonymous reviewers for their valuable comments.