Recent climate change and abnormal weather phenomena have resulted in increased occurrences of localized torrential rainfall in South Korea, including urban areas. The urban hydrological environment has changed in relation to precipitation, in terms of reduced concentration time, decreased storage rate, and increased peak discharge. These changes have altered and increased the severity of damage to urban areas. The notable recent Seoul floods in 2010 and 2011 are examples of such disasters [1
]. Accordingly, extensive research has been conducted on urban flood forecasting using only point rain gauges; however, this approach is inadequate for developing efficient disaster prevention measures, as such gauges, can not detect the spatial variations in rainfall [2
The mean areal precipitation (MAP) is usually used as input to hydrologic and hydraulic models. When MAP is measured with a low-density gauge network, noticeable errors are found in the simulated peak discharge [3
]. Previous studies have confirmed that the spatial resolution of rainfall data substantially affects rainfall runoff simulations [5
]. In particular, previous rainfall and hydrologic studies in urban basins, using high-density rain-gauge networks, have mentioned the necessity of accurately measuring intense heavy rainfall that is occurring in a specific area over a short period of time, and using spatially distributed rainfall for urban flood analysis [6
]. However, it is not always realistic to install numerous rain gauges to measure spatially representative rainfall estimates over a metropolitan area for geographic and economic reasons [8
]. Therefore, various researchers have attempted to use radar rainfall data. Nevertheless, the results of studies using radar rainfall data to prevent urban flood disasters in Korea were found to be inefficient because of the low accuracy and low spatial resolution of these radar rainfall data.
For the design and analysis of urban hydrological systems, rainfall data should have minimum temporal and spatial resolutions of 5 min and 1 km2
, respectively. Furthermore, the uncertainty of rainfall intensity should be less than 10%, and within a range of 10–50 mm/h [9
]. In particular, recent research has suggested a minimum resolution of 1 min/100 m for smaller catchments with an area of 0.01 km2
or less [10
]. Urban areas are usually divided into small catchments for urban hydrological applications; therefore, the use of high-resolution radar data is recommended [11
Thus far, research on urban flood estimation and forecast has mostly employed radar data as inputs to a hydrological model. James et al. [12
] studied flood forecasting using radar and gauge data on the Yockanookany watershed of the Mississippi River in the United States (U.S.) The results showed that the accuracy of the hydrographs was improved after the radar data were calibrated, using gauge data based on a Kriging coefficient method. Pessoa et al. [13
] used a Distributed Basin Simulation model (DBS) to test the sensitivity of rainfall estimations to radar data input. They reported that flood estimation was improved because the radar provided more accurate average rainfall accumulations and improved the determined spatial distribution of rainfall. Improved flood forecasting results were obtained by Mimikou and Baltas [14
] by using both radar and gauge rainfall data as inputs to the HEC-1 (Hydrologic Engineering Center) rainfall-runoff model. Sun et al. [15
] compared flood forecast results with different rainfall estimates, including those from rain-gauge data alone, Kriging of rain-gauge data, radar data alone, and co-Kriging of both radar and rain-gauge data. They concluded that rainfall that was estimated by the co-Kriging method considerably improved flood estimation. Therefore, rainfall data have also been obtained from weather radars for hydrological applications. Kim et al. [16
] used various approaches combining radar and rain gauges, including Kriging of rain-gauge data alone, using only radar data, using the mean-field bias of both radar and rain gauges, and the conditional merging of both radar and rain gauges. Subsequently, they evaluated the performance of flood estimates with the product of these approaches using a physics-based distributed hydrologic model. Their results showed that the conditional Kriging method provided the most accurate results for flood estimation.
As shown in the literature review, ground-observed rainfall is necessary for the improved accuracy of radar-estimated rainfall in order to reflect the true quantitative value. For this purpose, various techniques have been suggested to improve the accuracy of radar rainfall estimates by using rain gauges [17
]. One approach is to merge radar estimates with gauge measurements operationally in order to obtain quantitatively accurate and spatially continuous radar-derived rainfall fields [19
]. Past studies have shown the necessity of denser rain-gauge networks for improving the accuracy of radar rainfall field data. In the past, rain gauges were quite scarce in urban areas; however, high-density networks have been recently established over wide areas, such as in metropolitan cities, states, and countries, for research and commercial purposes. In the U.S., Japan, and other countries, telecommunication companies have built local observation networks to provide meteorological information. Using the base stations of these companies saves additional costs related to the transmission of data, because the networks are already established and the stations provide optimal locations for gauge system installation without any surrounding canopy (vegetation and trees). For example, WeatherBug operates the large surface weather network in the world with over 8000 surface weather stations across the U.S.; the observed data are used for severe-weather prediction [21
]. Minnesota’s high-density rain gauge network combines many observer organizations with different aims and different instruments for estimating extreme rainfall frequencies, using 45 years of observed daily rainfall [22
]. NTT Docomo, which is a major telecommunications company in Japan, established an environmental sensor network for its 4000 mobile network base stations around the country in 2011 and 2012. The stations are approximately 10 km from each other and the installed equipment observes temperature, humidity, direction and speed of wind, and precipitation for weather forecasting and disaster mitigation [23
]. Currently, SK Planet (SKP), a subsidiary of SK Telecommunication Company in Korea, has established 1089 weather observatory facilities in the Seoul region. Compared to other high-density observation networks, SKP created the world’s densest weather observatory network with similar or newer equipment.
The aims of this study were to produce quantitative precipitation estimation (QPE) products by using data from the densest rain gauge network and weather radar data, and to analyze the effect of high rain-gauge density and radar data in terms of the amount and spatial distribution of rainfall. The availability of QPE products for urban runoff simulation was also analyzed. The objectives were (I) to produce and assess high-resolution QPE products for Seoul, using high-density observed rainfall and radar rainfall data and (II) to assess the applicability of the simulated runoff with QPE products using an urban runoff model. The study area, data sources, and methodologies are explained in Section 2
, and the storm water management model is presented in Section 3
. The results of the QPE products according to high gauge density and radar data in terms of urban runoff analysis are presented for a rainfall event in Section 4
and the conclusions are presented in Section 5
2. Study Area and Data
2.1. Study Area
The study area is located in the Gangnam District of Seoul, the capital of South Korea. It covers an area of 605 km2
(126°46′15″ to 127°11′15″ E longitude, 37°25′50″ to 37°41′45″ N latitude; Figure 1
a). Seoul is one of the most highly urbanized cities in Korea; impervious areas reached 48.64% of the total urban area in 2014. In addition, several residential areas and industrial facilities are located near flood plains, such that the damage that is caused by urban flooding is exacerbated. For instance, Seoul experienced heavy rainfall events in 2010 and 2011. One of these (26–29 July 2011) resulted in 67 casualties and 37.7 billion KRW (Korea Republic Won) worth of property damage, which amounted to 50% of the total damage that was caused by natural disasters in Korea in 2011. Moreover, the greatest volume of rain among all of the rainfall events recorded since the beginning of 1908 was deposited in Korea on 21 September 2010. These torrential rains flooded 4727 residential areas, 1164 shopping areas, and 126 factories.
The drainage system of Seoul is managed by dividing it into 16 drainage zones. These are separated by the tributaries and the main outfalls of the Han River that flow directly into the river. These 16 drainage zones are further divided into 239 drainage districts within the region.
Gangnam District lies at the intersection of Gangnam-dong and Teheran-ro; it is representative of Seoul’s downtown area where residential and commercial areas are concentrated. Geomorphically, it is a relatively low-lying location with a complicated sewer network as compared to the surrounding areas. Thus, this area is vulnerable to flooding during heavy rainfall and the events in 2010 and 2011 caused flood damage to it. Gangnam has five drainage districts (Nonhyeon, Yeoksam, Seocho 1, Seocho 2, and Seocho 3), which collectively cover an area of 7.4 km2
. The drainage system comprises 4170 manholes, pipelines with a total length of 200,698 km, and two drainage pump systems at the Sapyeong and Seocho stations (indicated by black boxes in Figure 1
2.2. High-Density Rain-Gauge Network
In Korea, weather observations are traditionally carried out through the observation network of the Korea Meteorological Administration (KMA). This network has 95 automated surface observing systems (ASOS) and 493 automatic weather stations (AWSs) nationwide. Of these stations, 34 are used for meteorological observations in Seoul.
In recent years, SKP has been operating an integrated meteorological sensor network, using the base station infrastructure of the SK Telecom company. This network provides meteorological information, thereby offering weather services that can be used in disaster prevention. By May 2013, SKP had installed 264 operational meteorological sensors over the city of Seoul (Figure 1
a), and another 131 around Incheon. In 2014, it also established 694 sensors in Gyeonggi Province. A total of 1089 meteorological sensors have been installed at intervals of 1–3 km, forming the world’s densest weather observatory network. The meteorological information that was collected by SKP can be used to enhance the accuracy of weather observations by the existing disaster prevention system. The SKP sensors are set to the KMA standard to collect meteorological data, such as from the AWS of the KMA, every minute in real time [24
]. In addition, SKP regularly maintains the observed data by using automatic and semi-automatic observation information quality-management systems (www.weatherplanet.co.kr
The data used in this study were collected through the AWSs of SKP and KMA. The data contained errors that had occurred at the AWSs, such as system malfunctions, calibration deviations, and bias errors, in addition to errors that were associated with electronic and communication malfunctions [25
]. Using data with such errors decreases the accuracy of (or increases the uncertainty of) the hydrological runoff products. Therefore, data quality tests were conducted, and gauge stations with high missing rates were eliminated following a quality control technique in order to use the observed data from the AWSs of SKP and the KMA, effectively. The criteria of quality control were determined for missing values or outliers, as shown in Table 1
. Here, R10min
is the 10-min accumulated rainfall, Pi
is the observed rainfall at station “i
”, and σ
is the spatial standard deviation of rain gauges within 7 km. The Madsen-Allerup method [28
] is a non-parameterized spatial checking method that is based on analyses of median values and upper and lower quartiles of data from stations in the surrounding area. If Tit
exceeds 2.0, then it is judged to be “suspicious”.
The accuracy of SKP data from 262 stations was evaluated for three months from July to September 2013, targeting the KMA data using the abovementioned criteria to ensure data stability. The time interval of observations was 1 min. The average missing ratio was 18.71% for the three months, and the missing ratio of the 262 stations ranged from 2.00% to 77.11%. Missing values were caused by system errors during wire transmission and equipment defects. In order to ensure data stability, 156 SKP stations were selected, which had less than 20% missing ratio and less than three times the average standard deviation of cumulative rainfall for three months. The cumulative rainfall average of these stations was 878.2 mm and the standard deviation (σ) was 136.6 mm for three months.
The average missing ratio, which was recorded as “null”, was 7.09% for SKP and 0.46% for KMA for three months. The density of available rain gauges was approximately 3 km2
per gauge and the average distance between rain gauges was 977 m in the Seoul area. SKP rain gauge stations nearest to the KMA stations were selected and the observed rainfall estimates were compared one-to-one. The average relative error was within 0.25% for all of the evaluated stations. The error was determined according to the difference between observation equipment and location; therefore, the accuracy of the SKP data matched that of the KMA data [30
]. The 156 SKP rain gauge stations were selected among the 262 stations through the quality control and the 34 gauge stations operated by the KMA were used additionally (indicated by black stars in Figure 1
a). In total, 190 rain gauge stations were used in this study.
2.3. Weather Radar Data
For the purpose of rainfall estimation in terms of spatial variability in urban areas, radar-derived rainfall was estimated using data from the Gwangdeok Weather Radar Station (GDK). The radar site is located on Gwangdeok Mountain in Gangwon Province, approximately 100 km northeast of Seoul, and it is used to investigate the three-dimensional structure of clouds. The GDK radar is an S-band type, with a beam width of 1° and a gate size of 250 m. The GDK radar data are in the Universal Format (UF), and most of the noise, such as ground, sea, sun strobe, and AP clutter, was removed using the Worldwide Integrated Sensors for Hydrometeorology group (WISH) algorithms from the Next Generation Weather Radar (NEXRAD), U.S. As there are many mountains near the radar site and study area, volume data for the radar reflectivity data were extracted to obtain constant altitude plan position indicator (CAPPI) data at a height of 1.5 km in order to eliminate ground clutter using a bilinear interpolation program, which was based on the algorithm, as suggested by Mohr and Vaughan [31
]. The spherical coordinates of the volume data were transformed to Cartesian coordinates. The spatial and temporal resolutions of the extracted CAPPI data are 250 × 250 m2
and 10 min, respectively.
2.4. Drainage Network and Topographic Data
To obtain input data for urban runoff analysis, the data from pipes and manholes were collected from the Seoul drainage network map and then simplified. The pipe input data were simplified into an urban hydrological model using diameters of 600 mm and lengths of 100 m. In total, input data from 773 manholes, 1059 pipes, and 772 sub-basins were used. The Gangnam area is mostly occupied by buildings and paved streets that extend up to the foot of Umyeon Mountain. The areas of the sub-basins range between 0.0002 km2 and 0.435 km2, with an average area of 0.01 km2. There are no manholes at the sub-basins of the southwestern part of the area, including Umyeon Mountain. The slope was calculated using a digital elevation model (DEM) at a resolution of 5 m, with slopes ranging from 0.001% to 10.092%. The average slope of the northeast region is gentle and the Gangnam area is a relatively low-lying area. Furthermore, the Seoul Biotope Map was used to determine the distribution of runoff curve numbers (CN) and impermeability rates, which were 47–95 and 10.6–100%, respectively.
In this study, quality-controlled rainfall data from high-density ground-based rain-gauge networks covering Seoul, South Korea, were used with radar data to examine the estimations made by the QPE products at a resolution of 250 m and for urban runoff simulation. To examine the accuracy of the QPE products and for hydrological analysis, the data from the integrated meteorological sensors from the KMA and SKP were analyzed. A real-time quality control technique that can eliminate missing values and outliers was applied. The QPE products were used to make estimations based on four rainfall events that occurred in 2013. In addition, the three QPE products, which were classified according to the data used and the QPE method, were cross-validated. The quantitative accuracy of QPE2, which was only using weather radar data, was found to be underestimated when compared with the values that were derived from the ground gauge data. The accuracies of QPE1 and QPE3 were found to be good considering the rainfall amounts. Although the quantitative accuracy of QPE3 was lower than that of QPE1, QPE3 adequately simulated the irregularity of the spatial rainfall distribution. Each estimated QPE product was used for urban flood analysis and was tested for its applicability to urban flood analysis. In the evaluation of urban hydraulic and hydrologic impacts, according to three rainfall fields, QPE3, using both available rain gauges and radar data, simulated the peak runoff and overflow phenomena most accurately. This is because rainfall is generally non-homogeneous and QPE3 best reproduced spatial fluctuations of rainfall for small sub-basins. Therefore, urban runoff was reasonably simulated at the surface and in pipes. Hence, this study confirmed that spatial variations in rainfall and runoff can exist in a small urban area and that the use of high-resolution rainfall data is desirable for urban runoff analysis.
This study’s results show that radar data has few advantages over a sufficiently dense rain gauge network when estimating spatial rainfall distribution. However, the heterogeneity of the spatial distribution of rainfall can be better realized by radar; our results show that the accuracy of runoff analyses in urban watersheds is improved by this method. A radar will generally provide better quantitative precipitation estimates with a higher resolution, especially at the considered basin (30~50 km) in this study.
The limitation of this study lies in its generalization, as only a particular watershed was assessed for a few storm events. However, it is meaningful that the spatial variability of rainfall and the hydrological analysis of the urban watershed performed well using high density observation network and radar data for a wide urban area. In order to build on these results, future work should be conducted with more various storm events, and the errors should be corrected at each analysis step.