Application of deterministic distributed hydrological model for large catchment: a case study at Vu Gia Thu Bon catchment, Vietnam

Water is seen to be the premier constitutive factor of human beings. Thus, it always plays an important role in the development of human society (Gleick 1993; Watkins 2006). However, water distribution is not equal across time and space. These imbalances result in many negative impacts to human societies and activities. Annually, the natural disasters related to water issues, such as floods, droughts, and storms, produce significant severe damage, not only to property but also to human life (Johnson et al. 1997; Guha-sapir et al. 2013). Furthermore, in recent years, under the impact of climate change, the consequences of catastrophic flood and drought events have been more frequent and serious (Vázquez et al. 2002). Gaining a deep knowledge of the hydrological factors and processes in a catchment is an essential objective for the hydrologist community in order to be able to transform disadvantages to advantages or to mitigate catastrophic damage. Nowadays, with the development of mathematics and computer science, the simulation of the hydrological cycle for a catchment has become easier and more accurate (Blöschl & Sivapalan 1995). These improvements linked to modeling methods help hydrologists to gain more concrete and truthful insights about what happens in hydrological processes and support the decision-making process for reducing natural hazard impacts over the catchment.

To date, many hydrological models have been developed with different theories to simulate catchments' hydrological phenomena. They may be classified according to the description of the physical process as conceptual or physically based, and according to the spatial description of catchment processes as lumped or distributed (Refsgaard 1996). Each of these models has advantages and inconveniences for simulating the hydrological process. Lumped models are simple hydrological models which assume that all characteristics are constant across the catchment (Chow 1972), such as HBV, MIKE11/NAM, TANK, TOPMODEL, XINANJIANG. Lumped parameter models are considered much simpler in their treatment of spatial variation. In this kind of model, each parameter is described by a value that is uniform over the whole catchment. The parameters of lumped models cannot be determined directly from the physical characteristics of the catchment and they are generally determined via calibration (Chow 1972; Madsen 2003). Consequently, this kind of model cannot represent all hydrological processes precisely. Alternatively, a distributed model is constructed with sub-units that represent the divided catchment. Each unit integrates and represents all physical characteristics of the real area. This kind of model maintains the physical details at a given grid size and considers the distributed nature of hydrological properties such as soil type, slope, and land use (Refsgaard 1997; Vansteenkiste et al. 2013). In principle, the parameters of a distributed model could be obtained from the catchment data (Refshaard et al. 1995). Due to this characteristic, a distributed model is supposed to be able to translate hydrological processes in a catchment more accurately and concretely. One additional advantage of a distributed model is that the outputs, such as water levels, discharges, and other hydrological variables, could be perfectly extracted at any location in the catchment (Graham & Butts 2005). These efficiencies of the distributed model help to overcome difficulties in the lack of observed data, especially in developing countries, and which have a great significance for simulating the hydrological processes in a large catchment. As a result, they can provide more accurate predictions. From these pre-eminences, many distributed models have been developed and applied in recent years, such as MIKE SHE, SWAT, SWIM, LISFLOOD, and WETSPA (Cunderlik 2003). Nevertheless, which model is the best to simulate hydrological processes according to the objectives of the water-related project, such as resource optimization or long-term perspective? Until now, this question has not been clearly answered. In fact, there exist certain arguments related to the quality of different models that the lumped models, in many cases, perform just as efficiently as distributed ones with regard to rainfall–runoff simulation when sufficient calibration data exist (Refsgaard & Knudsen 1996). Moreover, even if the distributed model has drawbacks on computation time, initial parameter definition, or lack of spatial data for setting up the model as well as validation (Beven 2011), most of the modelers greatly estimate the capacity of the distributed model over the lumped one (Beven 1996; Refsgaard & Knudsen 1996; Golmohammadi et al. 2014).

The Vu Gia Thu Bon is one of the largest catchments located in central Vietnam. It is annually confronted with severe damage due to natural disasters such as catastrophic flood and drought events. Furthermore, according to the prediction of the IPCC's scenario (Pachauri & Reisinger 2007), under the impact of global warming, sea level increase, changes in the hydrological cycle, abnormal phenomena, e.g., El Nino and La Nina in the Vu Gia Thu Bon basin, flood and drought disasters are forecast to happen more frequently and be more extreme. This situation will generate more severe consequences for people, livelihoods, and socio-economic development. Hence, in order to mitigate the impact of these catastrophes on the region, an efficient tool is required to help hydrologists and authorities to have a good understanding of what is happening in the stream flow regime and its potential variations in the future. For this purpose, and as outlined above, a fully deterministic distributed hydrological model – based on MIKE SHE modeling system from DHI – has been chosen to simulate the hydrological processes in the Vu Gia Thu Bon catchment. In other respects, the study also aims: (1) to confirm the performance of distributed model types to simulate the hydrological processes and the capacity to translate climate scenarios for large catchments with very limited data sets (almost an ungauged catchment); (2) to evaluate the sensitivity of model parameters; (3) to compare the calibrated efficiency of multi-site versus single-site models; and (4) to analyze the uncertainty of input data and the model structure.

Due to its performance, MIKE SHE has been used in a broad range of applications. It is being used operationally in many countries around the world by organizations ranging from universities and research centers to consulting engineering companies (Refshaard et al. 1995). MIKE SHE has been used for the analysis, planning, and management of a wide range of water resources and environmental and ecological problems related to surface water and groundwater, such as river basin management and planning, water supply design, management and optimization, irrigation and drainage, soil and water management, groundwater management, interaction between the water surface and ground water, ecological evaluation of flood plain studies, impact of land use and climate change. Graham & Butts (2005) applied the MIKE SHE model to simulate the hydrological process of the Senegal River Basin. This model was developed on an area of 375,000 km². The result was relatively representative of the catchment's characteristics with good statistical coefficients. Thompson et al. (2004) used this model to simulate the hydrological system in lowland wet grassland in southeast England. These authors used MIKE SHE coupled with MIKE 11 to present the hydrologic factors in Elmley Marshes catchment during 36 months. This research gave remarkable results in simulating surface flooding, groundwater, and flow in the channel. The application of the coupled MIKE SHE/MIKE 11 modeling system to the Elmley Marshes has demonstrated its potential to represent complex hydrological systems found within many wetland environments. By simulating the stream flow process at catchments (<100 km²) in China and in Hawaii, USA, the works of Zhang et al. (2008) and Sahoo et al. (2006) have already proved the capacity of MIKE SHE to describe the flow in mountainous regions. Besides that, this model has a proven considerable competence in simulating the hydrological process in semi-arid areas. The study of McMichael et al. (2006) is an example demonstrating the above capacity of MIKE SHE. From the previous applications and regarding the specific issues in the Vu Gia Thu Bon catchment, MIKE SHE represents a reasonable alternative for hydrological modeling in a data scarcity situation. MIKE SHE is a deterministic modeling system based on physical laws. Most of the used parameters are physical variables that evolve in a range of values that can be defined according to physical processes. In such a context, realistic assumptions could be made on such variables and then allow the development of an efficient hydrological model over the large Vu Gia Thu Bon catchment. This approach has the objective of overcoming the lack of data that is a recurrent situation in many large catchments located in developing countries. The approach has already been implemented with success by Hundecha et al. (2002) and Ma et al. (2013). The developed deterministic model is supposed to provide a reasonably accurate estimation of the hydrological processes under the current climate conditions, and then to be used for assessing the impacts of the climate dynamic under various scenarios that define future conditions.

The MIKE SHE model is built with all of the available model components, e.g., overland flow, river and lake, unsaturated flow, evapotranspiration, and saturated flow (Figure 2). As a result, the model is expected to describe accurately hydrological processes in the Vu Gia Thu Bon catchment as well as to reduce the uncertainty when simulating future climate scenarios. The following data sets have been used.

The elevation data used in the model are taken from SRTM DEM (digital elevation model) with a horizontal resolution of 90 m from NASA (Figure 1) (http://www.cgiar-csi.org).

By analyzing the effect of spatial rainfall distribution on the stream flow, the study of Vo & Gourbesville (2014) showed that Kriging is the most suitable method to interpolate the rainfall distribution in this catchment. Hence, the simulations use rainfall data that are re-distributed spatially based on daily rainfall data from 15 rain gauge stations with the Kriging method. Unfortunately, no hourly data are available over the catchment. However, due to the size of the catchment – 10,350 km² – and of the different tributaries, the concentration times are always above 24 hours and up to 48 hours. The scale of the catchment contributes to reducing the underestimation of the high intensities that is the main difficulty with the daily data. At the same time, the planned application of the model is dedicated to addressing flooding processes and changes at the catchment scale. The quick hydrological processes are not the priority of the approach. The available data can match this objective knowing that the discharges and water level are recorded with more accurate time steps.

Data are inherited from the study of Vu et al. (2008) .These authors calculated the potential evapotranspiration in the Nong Son basin by using the Penman–Monteith equation. A monthly mean potential evapotranspiration for each vegetation type and average over the catchment were constructed.

The harvest schedule is set up for main plant types such as forest, homestead, rice, sugarcane, and grass. Each kind of crop is specified by vegetation property. The vegetation property in this simulation is from DHI results (DHI 2012a).

In order to simulate the river flows better, the MIKE SHE model is coupled with a hydrodynamic MIKE 11 model (1D model). The model is developed over 44 major branches with a length varying from 20 to 202 km (Figure 3(c)). The geometry of each river branch is specified via cross sections. The cross sections applied in this model are from two sources: a few of them downstream are taken from the measurements, and the remaining ones are extracted from the DEM. The initial bed resistance is set up with a Stickler roughness coefficient (M) varying from 15 to 25 m^1/3/s for upstream tributaries, and a value changing from 30 to 50 m^1/3/s for downstream branches.

The overland flow appears after the net rainfall rate exceeds the infiltration capacity of the soil; water then ponds on the ground surface. The main parameter to calculate this flow is the Stickler roughness coefficient (M). For Vu Gia Thu Bon, this parameter is determined depending on the land use map and at 2–90 m^1/3/s (Vieux 2001; Nguyen 2005a, 2005b; DHI 2012b).

DHI suggested three methods for describing the flow in this zone: Richards' equation, gravity flow and two-layer UZ. However, the application demonstrates that the various approaches do not provide very different results. For the current application, the simple two-layer water balance method is chosen to reduce the computational time. The physical property of each soil type is presented via the water content at saturation, water content at field capacity, water content at wilting point, and saturated hydraulic conductivity.

The groundwater is supplied by the Central Vietnam Division of Water Resources Planning and Investigation (http://www.ceviwrpi.gov.vn). The characteristic of the aquifer is mainly presented by horizontal hydraulic and vertical hydraulic conductivities.

In addition, the sensitivity analysis quantifies the dependent rate of runoff on the change of these parameters. As a result, these rates make the calibration easier and allow acceptable values to be more quickly obtained. It is seen to be a prior step to the calibration process. For the model applied to the Vu Gia Thu Bon catchment, the sensitivity of each parameter is analyzed based on the response of discharge in the Nong Son and Thanh My stations. In such a complex system as hydrology, the implication of the change in magnitude of several model parameters due to other parameters is undeniable (Sivapalan et al. 1987; Muleta & Nicklow 2005; Wang et al. 2007). It is clear that the sensitivity analysis will be better when accounting for the interactions between model parameters (Mishra 2009). However, with a model containing many parameters such as the MIKE SHE model applied to a large catchment, the evaluation of the interaction between model parameters during the sensitivity analysis is complicated and requires many simulations. In this analysis, the sensitivity analysis was done manually by varying the values of the parameters individually, one by one. This method has been applied by many authors (Refsgaard 1997; Andersen et al. 2001; Wang et al. 2012). The model is set up for a long period, even so the sensitivity analysis process is only based on the runoff variation during the period of two typical years, 2003 for the ‘dry year’ and 2004 for the ‘wet/flood year’. The parameter sensitivity is demonstrated via the variation tendency of the base flow and peak flow at the Vu Gia Thu Bon river system. Subsequently, the result of this process is expected to provide valuable information for the calibration process.

The available data over the Vu Gia Thu Bon cover the 1990–2010 period. The chosen strategy for calibration and validation has to be implemented within this period. The first period of 11 years from 1990 to 2000 is used for calibration. Over these 11 years, the recorded climate situations include two major flood events – 1996 and 1999 – with a return period of 100 years (ICEM 2008) and severe drought events in 1992–1993 and 1998 (Vu et al. 2014), as well as a moderate drought event in 1998 (Nguyen 2005a, 2005b). For the validation phase, the 11 year period from 2000 to 2010 also includes two major flood events – 2007, with a return period of 100 years; 2009, with a return period of 20 years (Chau et al. 2015); and a moderate drought event in 2002 (Nguyen 2005a, 2005b). The two periods allow analysis of the hydrological dynamic at the catchment scale and ensure representative results. Obviously, the availability of data – lack of measurements – is a serious constraint, but the proposed approach tries to overcome the difficulty by maximizing information production from the available data. In order to stabilize the model and establish proper initial conditions, the first year of each period is used for warming up the model. Hence, in this analysis, only 10 years of daily data are taken to calibrate and validate the model. The calibration is done manually. This process, of course, is based on the results of the sensitivity analysis. Only a few sensible model parameters are used for the calibration process.

The Vu Gia Thu Bon catchment has a complicated river system. Although the length of the two main rivers reaches 200 km, there is only one flow measuring station in the middle of each main river: Nong Son on the Thu Bon branch and Thanh My on the Vu Gia branch. This situation creates several difficulties for comparison of the results between simulations and observations. Especially, this is not only an inconvenience for predicting flood risk in the downstream region, but it is also a factor that produces uncertainty when assessing the impact of climate change on runoff. The lack of observation data for comparing with simulation results degrades the performance of the distributed model. Many hydrologists have suggested realizing calibration by multi-site, with not only the discharges but also the water levels (Wang et al. 2012). Accordingly, the water level at Ai Nghia station, Cam Le Station on the Vu Gia branch and Hiep Duc, Giao Thuy, Cau Lau stations on the Thu Bon branch are also compared with the MIKE SHE model outputs in order to increase the confidence and simultaneously to reduce the uncertainty in the projected climate scenario. The model assessment is performed with statistical measures of the root mean squared error (RMSE) (Equation (2)), the correlation coefficient (R) (Equation (3)), and Nash–Sutcliffe coefficient (E) (Equation (4)). The smaller the RMSE value is, the higher the model performance will reach. The perfect value of RMSE is 0 (Moriasi et al. 2007). The performance levels of R and E are classified in Table 1. The optimal values of these two factors are all 1 (Safari et al. 2012; Wang et al. 2012).

Most of the parameters in Vu Gia Thu Bon's MIKE SHE model are analyzed in order to estimate the runoff response. The results demonstrate the effects of the parameters on the simulated stream flows. These effects are not similar for all parameters. The results show that the difference is not only about the quantity but also about the timing of the peaks. The responses of river flow versus the variation of main parameters are shown in Table 2.

Parameters for the precipitation-dependent time step control are put forward to reduce the numerical instabilities (DHI 2012b). These parameters define the maximum rainfall value per time step and they are expected to have a great impact on river flows, at least on peak flows. In fact, the sensitivity analysis results in Table 2 demonstrate the role of these factors to peak flow. Accordingly, if the Max precipitation depth per time step (P_{Max depth}) increases, the peak flow will reduce. For the Vu Gia Thu Bon catchment, the reduced quantity is around 70–120 m³/s for an increase from 10 to 200 mm for P_{Max depth}. This tendency happens similarly with Input precipitation rate requiring its own time step (P_{Input rate}). However, the impact of P_{Input rate} on runoff is not as high as the P_{Max depth}. The increase of this factor from 0.1 to 10 mm/h leads to a reduction of around 20–50 m³/s of peak flow at the catchment. An important aspect for this factor is related to simulation time. The smaller P_{Input rate} is, the longer the model is running. Regarding the correlation between the Nong Son and Thanh My stations in connection with precipitation parameters, the results show that change at the Nong Son station is generally two times higher compared to the Thanh My station. This difference might be related to the characteristics of the catchment.

Overland flow simulates the movement of ponded surface water across the topography. It can be used for calculating flow on a flood plain or runoff to streams (DHI 2012b). In this case, the finite difference method is selected to solve the overland flow for the Vu Gia Thu Bon catchment. The Manning number (M), which is equivalent to the Stickler roughness coefficient, is estimated as the basic factor of the Overland flow module. Therefore, this part considers mostly the effects of the Manning number on runoff. In the Vu Gia Thu Bon catchment, there are many kinds of land use and soil. These lead to there being many corresponding Manning values. The effect of different Manning parameters corresponding to different land use types regarding river runoff is expected not to be equal. The disparity is due to the diversity in Manning values and distribution of land use. The difference is shown in Table 2. Nevertheless, there is one common point, that the change in Manning value greatly affects the peak flow, but mostly does not affect the base one.

The Manning number (M) is used to represent bed resistance in MIKE 11 as well. The river flow seems quite sensitive to the change in the bed resistance. It is expected to be the key factor in the calibration process. Nevertheless, in the Vu Gia Thu Bon catchment, it seems that the bed resistance only affects flood flow, specifically the peak flow is significantly impacted in the interval of Manning values from 5 to 10 m^1/3/s. If the M value changes in this interval, the peak discharge could rise to 200 m³/s at Thanh My and 400 m³/s at Nong Son. The flow apparently does not change if the Manning value in MIKE 11 is higher than 10 m^1/3/s.

The unsaturated zone is usually heterogeneous and characterized by cyclic fluctuations in the soil moisture as water is replenished by rainfall and removed by evapotranspiration and recharge to the groundwater table. Hence, this process plays a significant role in river runoff. Correspondingly, the variation of parameters in the unsaturated zone will deeply impact the runoff factor, on both base flow and peak flow. For simulating the unsaturated zone of the Vu Gia Thu Bon catchment, the two-layer UZ soil method is used. In this model, the physical characteristics of each soil are supplied, such as water content at saturation, at field capacity, at wilting point, and saturated hydraulic conductivity, yet only saturated hydraulic conductivity (K_uz) has a huge impact on the flow. The response of flow versus the K_uz variation is expressed in Table 2. According to this table, the peak flow goes down quite quickly when increasing the typical infiltration parameter. This change can reach more than 1,000 m³/s. Furthermore, the reduction of K_uz also makes the base flow decrease. The response of the base flow is small, around 10–20 m³/s, but it has a big value in comparison with observed base flows. With variation in soil distribution, the change in the amount of runoff against soil property variation in Thanh My and Nong Son is very different, as presented in Table 2.

The groundwater plays a crucial role in the behavior of hydrological processes within the catchment. Thus, the variation of this component will significantly influence the river flow, especially the base flow when the discharge from groundwater is seen as its principal source. The groundwater is present in the saturated zone. The flow in the saturated zone is characterized by aquifer properties, of which horizontal saturated hydraulic conductivity (K_h) proves the most influential on saturated flow. For this reason, in the model only this factor is considered. For Vu Gia Thu Bon, the results in Table 2 demonstrate the primary impact of this factor on peak flow. The peak flow tends to reduce quickly when increasing K_h. The reduction is very clear, so when K_h rises from 1e⁻⁵ to 8e⁻⁵ m/s, the runoff goes down around 500 m³/s at both the Thanh My and Nong Son stations. In contrast, K_h increases in the interval of 1e⁻⁵ to 8e⁻⁵ m/s, making the base flow rise by approximately 30 m³/s at Thanh My and 50 m³/s at Nong Son. In spite of the insignificant variation, it is quite important for adjusting the base flow.

MIKE SHE mode parameter values are varied based on the above sensitivity analysis. The optimal values reached from the calibrated process are shown in Table 3.

The efficiency of the MIKE SHE model is also shown through the statistical coefficients in Table 4. Observed and simulated daily and monthly discharges are compared. These numbers demonstrate the accuracy of the model and its efficiency for describing the hydrological processes of the Vu Gia Thu Bon catchment. The R and E coefficients at Nong Son and Thanh My for the calibration period are 0.92, 0.89 and 0.82, 0.78, respectively. In the validation period, these factors slightly decrease. R and E coefficients at Nong Son station are 0.92 and 0.82 and at Thanh My, 0.90 and 0.69. The RMSE coefficients at Nong Son and Thanh My in both periods are relatively small. Over the 10 year period, the RMSE of simulation is only 132.3 m³/s at Thanh My versus maximum observation 4,440 m³/s–minimum observation 12.2 m³/s. The value at Nong Son is 288.7 m³/s versus maximum observation 8,920 m³/s–minimum observation 21.7 m³/s. The RMSE of validation at Thanh My is 123.2 m³/s versus maximum observation 4,540 m³/s–minimum observation 20.2 m³/s and at Nong Son is 250.5 m³/s versus maximum observation 8,410 m³/s–minimum observation 21.9 m³/s. Due to these big differences between maximum and minimum values of observed data, the values of the normalized root mean square error at these stations are quite small, with a value below 0.05. For the monthly analysis, the good performance of the model is obviously demonstrated by the figures in Table 4.

In order to verify the efficiency of the MIKE SHE model for the Vu Gia Thu Bon catchment, the water levels recorded at several stations are compared with the results. However, the accuracy of simulated water levels is not as good as for discharges. Due to these differences, the statistical coefficients for water level comparison between simulation and measurements are not as high as the ones obtained for the discharges (Table 4). The relation coefficient in all stations is around 0.8–0.9, but Nash–Sutcliffe coefficients are low and vary from 0.6 to 0.7 at the upstream station and get smaller in stations under tidal effect. The source of this inaccuracy can be explained by the coarse resolution of topographic data, the lack of measured cross sections, and the large distance between two computed sections (dx) in MIKE 11 (Vázquez et al. 2002). Despite the fact that statistical coefficients for water levels are still low, calibrating the model relying on these factors probably adds a certain value to show the correlative level of model results with real data.

By mean of the analysis above, we can judge that the flow in the Vu Gia Thu Bon catchment is reproduced rather impressively by MIKE SHE. Therefore, this model is able to be applied to a flood event or the variability of stream flow under the impact of climate change. Moreover, these results also demonstrate the performance of deterministic distributed models in simulating the hydrological process, especially with a large catchment.

Although trying to reflect most truthfully the hydrological dynamic in the catchment, the model has not yet gained the optimal results when inaccuracies still remain. The statistical coefficients, such as Nash–Sutcliffe and correlation coefficient (RMSE) are still weak. These coefficients could not get maximum values for many reasons. The model has many potential uncertainties for simulating hydrological processes. Uncertainty in hydrologic modeling may arise from several sources: model structure, parameters, initial conditions, and observed data used to drive and evaluate the model (Liu & Gupta 2007).

One of the most important factors regarding the model structure, and which may have a significant influence on the model accuracy as well as on the uncertainty of simulation, is the cell size issue (Egüen et al. 2012). The advantage of the distributed hydrological model is that it represents hydrological characteristics of the catchment cell by cell. However, the resolution used in the model is still coarse due to the limitation of topographic data and computation time. The 90 m topographic grid data used may not describe precisely enough the surface of the catchment. Thus, it derives some differences in surface flows between reality and the model. The land uses, the soil properties, or the roughness coefficients, which are simplified in order to optimize the calibration are the major causes of underestimation or overestimation for the model.

Another issue significantly influencing the model uncertainty is the rainfall, which is a key factor in the hydrological dynamic. Rainfall spatial variation heavily affects both runoff generation and the hydrologic process in a catchment (Moon et al. 2004). The spatial variability in rainfall may introduce a significant uncertainty in model parameters during the calibration process (Chaubey et al. 1999). The quality of spatial rainfall distribution usually depends on the characteristics of the study area and other factors, in particular, the rain gauge density. The network of rain gauge stations in Vu Gia Thu Bon is sparse, with, on average, one station for an area of 700 km². In the constructed model, the rainfall inputs are re-interpolated and could be considered as a great source of uncertainty. As well as the spatial distribution, the time factor is also a great potential source of uncertainty (Dendy 1987), especially within the Vu Gia Thu Bon catchment where the concentration time is short due to the steep topography. Using daily rainfall in this simulation probably affects the rising limb and the peak flow appearance. However, these data are the unique complete rainfall data set available for long-term simulations for the Vu Gia Thu Bon catchment. The analysis regarding the rainfall distribution in space and in time demonstrates again the impact of lack of data on simulation uncertainty.

The groundwater is an unignorable component when simulating hydrological processes (Winter 1999). In terms of input data, the insufficiency of ground water data is seen as a major source of uncertainty for simulating hydrological processes. The quality of the groundwater data of the Vu Gia Thu Bon catchment is not very good, and the collected data do not present the groundwater properties concretely. For the whole catchment, the model integrates a unique geological layer.

Regarding the modeling methods implemented in MIKE SHE, the selection of one or another method can potentially generate several sources of uncertainty. For example, there are three functions to select for unsaturated flow: Richards' equation, gravity flow, or the two-layer UZ. The Richards' equation is supposed to be the best method for simulating unsaturated flow. However, in the current application, the two-layer UZ was chosen due to the limited available data sets and the short processing time.

The coupling between MIKE SHE and MIKE 11 contains additional potential uncertainties. It could notably affect water exchanging between the floodplain and riverbed. The number of simulating branches and intervals in MIKE 11 is considered as a main source of uncertainty. In the model, the coupling between MIKE SHE and MIKE 11 is not set well due to the limited number of cross sections. Additionally, there is a difference between overland flow and river flow, but in the analysis, the river network presented in MIKE 11 is merely over 44 large branches. Understandably, the set-up manner of the coupling between MIKE SHE/MIKE 11 is estimated, producing a large uncertainty.

Based on Table 4, it is easy to recognize that statistical indices counting with monthly data are better than counting with daily data. Hence, an additional element affecting the quality of the hydrologic model concerns the time factors. The time step applied in this MIKE SHE model is fixed at 1 day. This time step is still high, so it may not present thoroughly what the hydrological cycle is in the catchment. As well, the timescale of the model is limited to 10 years. The simulation time is thought not sufficient to convey adequately extreme events from natural phenomena. Unavoidably, simulating the stream flow in the Vu Gia Thu Bon catchment will contain particular uncertainties.

The aim of the approach was to assess the capacity of a deterministic distributed hydrological model based on the MIKE SHE modeling system to simulate the hydrological process and to estimate climate change impacts on the runoff in the Vu Gia Thu Bon river system. The developed model integrates most of the hydrological processes – from surface flow to groundwater flow, and evapotranspiration – and is expected to reproduce the hydrological cycle within the catchment. The results demonstrate clearly the performance of the approach with reasonable physical hypotheses that contribute to simplifying the calibration procedure. At the same time, due to its physical foundation, the model is hoped to produce a good assessment, with an estimated accuracy, of the climate change impacts over the catchment. One of the advantages of fully deterministic distributed models is the possibility of overcoming the weakness and the lack of systematic data that is a frequent situation in many developing countries. The difference between actual and future runoff regimes could be compared anywhere over the catchment. Hence, an overview of changes in runoff regime across the whole catchment could be generated without difficulty. These points confirm the capacity and the efficiency of distributed models for assessing climate change impacts over hydrological processes.

The developed model was calibrated and validated against daily data and monthly data for the period 1991–2000 and 2001–2010, respectively. The performance of the model is demonstrated via the hydrograph shapes and statistical indices, comparing simulation results with data from seven gauging stations. The model efficiency is likewise confirmed by the capacity to predict the extreme peak flow and base flow. However, the model still contains uncertainties that originate from different sources, such as the model structure, the cell size, and the input data. Reducing these uncertainty sources is likely impossible. At the same time, it is not completely necessary, regarding the models whose assigned objectives are dedicated to operational purposes. Obviously, improvements may address priority cell size, data accuracy, and solving algorithms.

In the traditional vision of hydrological analysis, the availability of limited data resources, and the necessary computation performance to some extent, reduce de facto the use and the application of deterministic distributed models. However, they represent a relevant alternative and may provide useful results for operational management of catchments. Obviously, the calibration process is a key and compulsory task in the model development. The sensitivity analysis appears as an essential step in order to efficiently support the calibration procedure. The performed analysis suggests that the model comparisons should be carried out not only on the discharges, but also on the water levels in all locations where measured data are available in order to maximize the use of available information. The results of calibration confirm that it is necessary to compare the measurement and simulation at multi-sites. The multi-site calibration helps to increase the accuracy of translating what happens in the nature of the model and in minimizing the global uncertainty of the model as well. The analysis demonstrates likewise that the model validation versus extreme high and low values is also quite important. The validation on extreme runoff values is helpful for reducing the uncertainty in simulations.

The sensitivity analysis has to be focused on the factors and variables that are directly affecting the runoff process. The analysis has demonstrated that the runoff variation is strongly driven by parameter changes. In the Vu Gia Thu Bon catchment, the peak flow is significantly affected by most of the model parameters while the base flow, as expected, is merely influenced only by the horizontal saturated hydraulic conductivity of the saturated zone and the saturated hydraulic conductivity of the unsaturated zone. The analysis has demonstrated the interest of the sensitivity analysis in the calibration procedure, which requires a distributed deterministic hydrological model. At the same time, the sensitivity analysis approach helps to determine a useful range for each physical variable used for the stream flows and contributes to simplifying the calibration process.

Thanks to the Hydro Meteorological Center in mid central Vietnam, the Central Vietnam Division of Water Resources Planning and Investigation, the LUCCI project, the P1-08 VIE project that provided the data used for this analysis. The MIKE SHE modeling system was provided by DHI.