Long-Term Simulation of Daily Streamflow Using Radar Rainfall and the SWAT Model : A Case Study of the Gamcheon Basin of the Nakdong River , Korea

In recent years, with the increasing need for improving the accuracy of hydrometeorological data, interests in rain-radar are also increasing. Accordingly, with high spatiotemporal resolution of rain-radar rainfall data and increasing accumulated data, the application scope of rain-radar rainfall data into hydrological fields is expanding. To evaluate the hydrological applicability of rainradar rainfall data depending on the characteristics of hydrological model, this study applied Rgauge and Rradar to a SWAT model in the Gamcheon stream basin of the Nakdong River and analyzed the effect of rainfall data on daily streamflow simulation. The daily rainfall data for Rgauge, RZ, and KDP were utilized as input data for the SWATmodel. As a result of the daily runoff simulation for analysis periods using R Z(P) and RKDP(P), the simulation which utilized Rgauge reflected the rainfall-runoff characteristics better than the simulations which applied R Z(P) or RKDP(P). However, in the rainy or wet season, the simulations which utilized RZ(P) or R KDP(P) were similar to or better than the simulation that applied Rgauge. This study reveals that analysis results and degree of accuracy depend significantly on rainfall characteristics (rainy season and dry season) and QPE algorithms when conducting a runoff simulation with radar.


Introduction
With the influence of climate change and climate variability, the magnitude and frequency of extreme hydrological events increase, which makes the world suffer from natural disasters such as floods, droughts, landslides, avalanches, and so forth.Particularly in Korea, people are suffering from localized heavy rainfall and flash flood in some years and from severe drought in other years.Therefore, a study on a more accurate and comprehensive analysis and prediction is required in order to adapt to climate change and climate variability and reduce human life and property damages from floods or droughts caused by extreme hydrological events.To this end, the necessity of high-resolution hydrometeorological data with high accuracy is emphasized.In particular, rainfall is used as basic data for all interpretations related to hydrologic cycle and the water resources plan and management and also has a nonlinear relationship with other hydrological factors and environmental ones (i.e., runoff, soil moisture, erosion, water quality, etc.) [1].For this reason, acquiring exact rainfall data is very important.
In a runoff simulation for the analysis of floods and droughts, rainfall is used as input data for rainfall-runoff model and many studies have been conducted for measuring exact rainfall in terms of time and space.Particularly for improving the accuracy of the spatiotemporal rainfall, hydrological and meteorological fields have recently paid more interest in estimating the precipitation data with a new Remote Sensing Technology including rain or weather radar (in the following, we will frequently omit the term "rain" or "weather") and artificial satellite.
In particular, radar data has advantages in that it can continue to provide rainfall information with high spatial and temporal resolutions.In early times of rainfall observation using radar, the rainfall intensity can be determined from the horizontal reflectivity ( H ) observed using singlepolarization radar.However, there is difficulty in measuring rainfall because its changing characteristics can vary depending on the type of clouds or developmental conditions, temporal and spatial location, and type of hydrometer.It has since become possible to measure various dual-polarization variables ( H ,  V ,  DR ,  DP ,  HV , etc.) using dual-polarization radar.In this process, precipitation estimation techniques can be developed.
Due to the advantages of radar, many studies in hydrological and meteorological fields that utilize radar data are being actively carried out.Utilization of radar data can be largely divided into two aspects.The first is general radar image data which analyzes the current status of rainfall and the second is grid-type rainfall distribution data for calculating flood flow [2].In hydrological fields, two major research areas include hydrologic phenomena in various basins (including natural basins and urban basins) [2][3][4][5][6][7][8][9][10][11][12][13][14][15] and the spatial and temporal variability of rainfall [16,17].To evaluate previous studies which applied radar rainfall data to the Soil & Water Assessment Tool (SWAT) rainfall-runoff model used in this study, Di Luzio and Arnold [18] applied NEXRAD Stage III data to the SWAT model as part of the Distributed Model Intercomparison Project (DMIP) to carry out the first simulation of daily runoff and Jayakrishnan et al. [19] used WSR-88D data obtained from the Sondu River in Kenya to carry out a simulation of runoff and water quality.Recently, Jeong et al. [20] used CAPPI data of single-polarization radar to measure the optimal grid size for radar reflectivity and used the SWAT model and as a result of the runoff simulation proposed the optimal grid size of radar rainfall data in the applicable basin to be 4-8 km.Furthermore, Sexton et al. [21] and Price et al. [22] used the SWAT model to compare the runoff discharge from NEXRAD data with that from rain gauge data and suggested that radar rainfall data could be utilized in a useful way in basins where rain gauge data is insufficient.
As shown above, previous studies have presented various evaluation results of the hydrological applicability of radar rainfall data depending on the radar data used, target area, and hydrological model applied.However, most of the studies have largely focused on short duration rainfall events or longterm runoff simulation for about a month or a season using radar data.In other words, there are few studies on the hydrological applicability in terms of long-term runoff for water resource management.However, the accumulated rain radar rainfall data are considered to be used sufficiently for a long-term hydrologic analysis.
Therefore, this study aims to review the hydrological applicability of radar rainfall data in a long-term runoff analysis for more than one year with the characteristics of single-and dual-polarization radar rainfall data.
To this end, the procedure employed in this study consists of the following steps: ( In this study, the radar data used was the Mt.Bisl rain radar data from MOLIT (Figure 1).MOLIT constructs a radar rainfall estimation system by carrying out a series of processes including quality control of radar data and application of Quantitative Precipitation Estimation (QPE) algorithms and provides rain radar-derived rainfall data.
The major procedures includes (1) import of observed radar data, (2) data quality control (removal of nonmeteorologic echoes), (3) creation of radar rainfall field (3 types of spatial fields: LEMAP (Lowest Elevation MAP), PPI (Plan Position Indicator), and CAPPI (Constant Altitude Plan Position Indicator)), ( 4) radar rainfall estimation (3 types of algorithms:   ,  DR , and   DP ), ( 5) radar rainfall adjustment using ground rainfall, (6) calculation of point rainfall and areal rainfall of subwatershed using adjusted radar rainfall, and ( 7) storage of rainfall data in a DB system.In particular, LEMAP (Lowest Elevation MAP) shows a radar Rrinfall field based on the radar reflectivity data observed at a very close altitude from the ground.Figure 2 represents the procedure of the radar rainfall estimation system currently employed by MOLIT.
In this study, we applied the following rainfall estimation algorithms to compare the hydrological applicability of  radar .First,   and   DP are calculated using ( 1) and (2).Equation (1) has the same shape as the existing single-polarization with the so-called - relationship ( =   ): The inverse of this equation is  = 300 1.4 .Herein,  is the reflectance of radar in the horizontal direction observed with single-polarization radar,  H . Equation ( 1) is an empirical equation appropriate for rainfall types between straight form rainfall and convective rainfall, which are at intermediate level or higher in terms of rainfall intensity according to the characteristics of rainfall in the summer [23].
Next, (2) and (3) were proposed by Ryzhkov et al. [24] as a prototype for dual-polarization radar of WSR-88D in the United States.Generally, the radar reflectivity (), differential reflectivity ( DR ), and specific differential phase shift ( DP ) are used in dual-polarization radar rainfall estimation.Also, the drop size distribution (DSD) depends on the rain intensity.Therefore, to adjust for any errors in the process of obtaining   DP , each formula used depends on the rain intensity calculated by   , as shown in (3) [24].One has 1 ( DR ) = 0.4 + 5.0      dr − 1      Here,  1 and  2 , the functions of   and   DP , are determined by utilizing (1), (2), and (4) and the reflectivity,  1 ( DR ) and  2 ( DR ), where  dr is an adjustment factor depending on the shape of the drop size.
In this study, single-polarization rain radar-derived rainfall which considers the reflectivity () and can be described by (1) alone and dual-polarization rain radar-derived rainfall which involves all of the dual-polarization variables shown in (1) to (4) expressed as   DP ,   , and   DP were utilized.The rainfall intensity (mm/hr) had an observational radius of 150 km, a temporal resolution of 2.5 min, and a spatial resolution of 125 m × 125 m while applying the rainfall estimation algorithms shown above.To utilize as input data in the SWAT model,   and   DP data were converted to daily rainfall by multiplying the rate of the observational cycle (2.5 min) and time.In this case, radar-derived point rainfall data ( (P) ,   DP (P) ) that belongs to the subbasin were recreated (Figure 4).[26,27].In particular, the SWAT model has advantages in that it can allow a hydrologic analysis of ungauged basins by conducting a predictive simulation of long-term rainfallrunoff and sediment movement within the basin.It also has the ability to quantify relative effects of water quality depending on forms of cultivation and climate/vegetation changes.To make a temporal/spatial analysis of hydrology and water quality using the SWAT model, it is necessary to obtain meteorological data that changes over time (daily amount of precipitation, temperature, wind speed, amount of sunshine, and relative humidity), the current status of land use spatially, soil attributes, and the Digital Elevation Model (DEM).The SWAT model is widely used because it is easy to generate major input values and it is possible to analyze the runoff of rainfall in basins, the occurrence of nonpoint pollution, and temporal/spatial changes.

SWAT Model
In this study, the DEM was set at 30 m × 30 m so that runoff of rainfall in the basins and actual stream within the basins can be well reproduced.As a land use map, the 1 : 25,000 classification land use map provided by WAMIS was used.As a soil map, the 1 : 50,000 reconnaissance soil map provided by WAMIS was used.Meteorological data including the mean daily wind speed (m/sec), daily average relative humidity (%), daily maximum/minimum temperature ( ∘ C), and daily quantity of horizontal solar radiation (MJ/m 2 ) were obtained from western meteorological observing stations.Table 1 summarizes the input and output data of the SWAT model.

Model Parameter Calibration and Validation.
To correct the parameters, we applied a trial-and-error method and calibration tool to increase the predictive accuracy of runoff discharge in the SWAT model.If the calibration procedure is properly planned, daily data collected over the course of one year is sufficient for the model calibration to obtain conceptually realistic estimates.In addition, the use of older data does not greatly influence the adjustment of parameters [28].The periods of correction and calibration were 2010 and 2011, respectively, and the ground rainfall data and daily discharge data within basins used for the periods of correction, calibration, and simulation were provided by WAMIS and the Korea Hydrological Survey Center (KHSC).
As CANMX, CN2, ESCO, GW REVAP, SOL AWC, SOL K, REVAPMN, and GWQMN among the parameters related to runoff discharge in the SWAT model react sensitively, CN2 was adjusted to correct observational values of the runoff discharge.In addition, to correct the base runoff, the parameters related to underground water (GW REVAP, REVAPMN, and GWQMN) were calibrated.In other words, if the base runoff is simulated to be higher, GW REVAP and GWQMN are increased and REVAPMN is reduced.If base runoff is simulated to be lower, the coefficients are calibrated reversely.The range of parameters and the shape of the input data are summarized in Table 2.

Model Applicability Evaluation Index.
In this study, to evaluate the applicability of the SWAT model for the calibration, validation, and simulation periods, the Nash-Sutcliffe efficiency (NSE), percent bias (PBIAS (%)), and RMSEobservations standard deviation ratio (RSR) were used.To determine the optimal value of each index, NSE = 1, PBIAS = 0, and RSR = 0, as shown in (3)-( 5).The NSE is a normalized statistic that determines the relative magnitude of the residual variance ("noise") compared to the measured data variance ("information") [29].PBIAS measures the average tendency of the simulated data to be larger or smaller than their observed counterparts [30].RSR was calculated as the ratio of the RMSE and standard deviation of measured data [25]: Herein,  obs  is the th observed streamflow,  sim  is the th simulated streamflow,  mean is the mean of the observed streamflow, and  is the total number of observations.Ramanarayanan et al. [31] suggested that if  2 is 0.5 or higher and NES is 0.4 or higher, the model simulates natural phenomenon well.Moriasi et al. [25] claimed that, based on examples of existing various models and research data, index values of the model simulation of NSE > 0.50, RSR < 0.70, and PBIAS ± 25% are satisfactory.In particular, Moriasi et al. [25] proposed the criteria for setting the general performance rating in the model of runoff discharge, as shown in Table 3.This criteria is based on monthly unit runoff discharge, but the model simulation is poorer with a shorter time step than a longer time step (e.g., daily versus monthly or yearly) [32].Therefore, these criteria can be used to evaluate the results of calibration, validation, and simulation obtained in this study.Radar-derived rainfall data can be used in a useful way in basins where ground observation data (rainfall station) is not sufficiently guaranteed.Therefore, this study creates radarderived point rainfall data ( (P) ;   DP (P) ) at central points of 42 subbasins divided when building the SWAT model to demonstrate the advantages of radar-derived rainfall data.
In other words, the SWAT model uses the closest rain gauge station (Buhang 1 station) to interpret #7 subbasin (#7 basin is at the utmost bottom of the figure), as shown in Figure 3(b).Therefore, it is difficult to simulate appropriate runoff discharge if there is difficulty in taking into account the temporal and spatial characteristics of rainfall because the size of basin is large or there are not many rain gauge stations in the basin.

Results and Discussion
In this study, we utilized 2012 as the year for simulation taking into account the observation period and accuracy of each data set, established the SWAT model in the basins prior to runoff simulation, and created  (P) and   DP (P) , as shown in Figure 3.In addition, we made a simple comparative analysis of  (P) ,   DP (P) , and  gauge .As a result of comparing the average basin rainfall accumulated during the period of simulation (2012),  gauge is 1,281.4mm and  (P) and   DP (P) are 1,272.6mm and 1,450.6 mm, respectively.Compared to the ground observation rainfall,  (P) is underestimated by about 0.7% (8.8 mm) and   DP (P) is overestimated about 13.2% (169.2 mm).In the scatter plot which shows the accumulated rainfall and correlation of  gauge and  (P) or   DP (P) (inside Figure 4), the correlation coefficients and root-mean-square deviation error values between  (P) or   DP (P) and  gauge are 0.968 and 0.976 and 2.926 and 2.848, respectively.This suggests that the values obtained for   DP (P) are better.However, as a result of comparing the rainfall of the Nile depending on the period,  (P) and   DP (P) represent the characteristics of rainfall in the rainy or wet season (heavy rainfall and in the summer when typhoons occur frequently: Jun. to Sept.) relatively well but are overestimated in the dry season (periods other than the rainy or wet season: spring, fall, and winter).In particular, in the dry season, the PBIAS results utilizing  (P) and   DP (P) are in the range of about 11% to 37% (overestimated).Given that the annual PBIAS is about 0.7% to 13.2%,  (P) and   DP (P) are very low in terms of accuracy for rainfall estimation in dry seasons (Table 4).Tables 5 and 6 show data for the ten (10) days in which daily rainfall data of  (P) and   DP (P) are overestimated and underestimated the most compared to  gauge .As shown in the overestimation list,   DP (P) has higher variations except for the abnormal observation at  (P) (Sept.12, 2012) and in the underestimation list,  (P) has higher variations.

Simulated Streamflow Results.
As mentioned earlier, the entire period of analysis (2010 to 2013) was into periods of calibration (2010), validation (2011), and simulation (2012).The parameters were corrected through runoff simulation for the period of correction using ground rainfall data.Using these parameters, the runoff analysis was carried out based on the SWAT model for the periods of calibration and simulation.The results from the runoff analysis were analyzed based on the general applicability evaluation criteria of the model presented by Moriasi et al. [25] in Section 3.3.
Figure 5 compares the observed and simulated streamflows for the periods of correction and calibration.The simulated streamflows obtained using the observed streamflow and  gauge during the period of correction are 20.7 m 3 /sec and 20.5 m 3 /sec, respectively, and the means of the simulation streamflows using the observed streamflow and  gauge during the period of calibration are 22.7 m 3 /sec and 22.4 m 3 /sec, respectively.As shown in Figure 5, the results of the runoff  charge using  gauge for the periods of correction and calibration describe the entire characteristics of the daily unit runoff discharge relatively well.However, 2010/8/10 and 2010/8/15∼ 16 during the period of calibration and 2011/8/9 during the period of validation are underestimated because the run-off discharge is relatively high.Moreover, as a result of evaluating the applicability of the SWAT model during the periods of correction and calibration, as presented in Table 7, the NSE values are 0.97 and 0.78, respectively, the PBIAS (%) values are 1.44 and −24.13, respectively, and the RSR values are nearly 0.47.The results from the runoff analysis during the periods of correction and calibration are a natural result of the application of the optimized model parameters of the correction period to the calibration period.
Figure 6 shows the results of daily streamflow hydrographs which applied  (P) and   DP (P) to the model that completed its validation using the parameters calibrated earlier.
The mean of the observation streamflow during the period of simulation (2012) is 19.8 m 3 /sec.The mean of the comparison of both seasons, the NSE, RSR, and PBIAS (%) values in the rainy or wet season showed similar or more significant values when  gauge was used compared to when  (P) and   DP (P) were used.But, in the dry season, the analysis did not match the ground observation rainfall data well ( gauge ).Based on the results obtained to date, the QPE algorithms used in this study are highly applicable in runoff simulation from Jun. to Sept. (summer; rainy season) and less applicable in other periods (dry season; winter).
The results from this study suggest that it is necessary to select radar observation strategies and algorithms appropriately depending on the intended purpose of radar rainfall data.Therefore, further studies are needed to improve the bias correction and algorithms (in real time) to increase the usability of radar data in analyzing long-term runoff for more than one year (both daily and monthly time steps).Still, there is a limit to the accuracy of Quantitative Precipitation Estimation.But if the accuracy of Quantitative Precipitation Estimation can be improved sufficiently, the hydrological application scope of rain radar rainfall will be expanded sufficiently and more exact hydrologic analysis will become possible.

Figure 2 :
Figure 2: Radar data quality control and rainfall estimation procedures employed by MOLIT.

Figure 4
represents the results of the comparison of the basin average rainfall accumulated for the period of simulation (2012).

Figure 6 :
Figure 6: The results of the daily streamflow simulation (2012).

2. Study Area and Rainfall Data
Analyze the characteristics and applicability of  gauge ,  (P) , and   DP (P) in long-term runoff analysis.
3.1.SWAT Model and Input Data Buildup.The SWAT (Soil and Water Assessment Tool) model is a unit model of a basin developed by the USDA Agricultural Research Service (ARS)

Table 1 :
Input and output data of the SWAT model.Application of Radar Rainfall Data in the SWAT Model.A rainfall station should be installed to represent the local distribution of rainfall in a basin.In this case, five rainfall stations (Seonsan, Gimcheon, Jirye, Buhang 1, and Buhang 2) are located in the Gamcheon stream basin of the Nakdong River and the density of rainfall station is about 201.1 km 2 /station (basin area is 1,005.3km 2 ).This density of the rainfall station is above the minimum criteria recommended by World

Table 3 :
General performance ratings for a monthly time step. [25]].

Table 4 :
Summary of the statistics of basin average rainfall for the different rainfall data types.

Table 5 :
List of the ten days where  (P) and   DP (P) results overestimated the rain gauge rainfall data the most.

Table 6 :
List of the ten days where  (P) and   DP (P) results underestimated the rain gauge rainfall data the most.