1. Introduction
Rainfall data are the primary inputs for most hydrological systems. Historically, rainfall has been monitored with a network of rain gauge stations, and those measured data have been used in hydrological modeling to generate runoff data at various points on a river. This method of estimating river runoff data is conventional and accepted worldwide. However, there have been several discussions on the advantages and disadvantages of using rain gauge data for hydrological modeling [
1,
2,
3,
4,
5,
6]. Rainfall measurements from rain gauge data provide some of the best ground truth rainfall information at single point stations. During the hydrological simulation of a river basin, available rain gauge networks within a basin are considered and interpolated to cover the spatial distribution of rainfall using various interpolation techniques [
2,
6,
7,
8]. With a very dense network of rain gauge stations, it could be possible to determine a reliable spatial distribution of rainfall [
4]; however, this is not possible practically [
9]. Therefore, a limited rain gauge station network has been used to generate information about the spatial distribution of rainfall for a given time. However, interpolation does not always provide a sufficiently precise picture of rainfall distribution [
10]. Furthermore, limited and interpolated rain gauge data can introduce large uncertainties into predictions made by hydrological models [
11].
Remote sensing technologies are the most common methods used to address the low density of rain gauge stations. Satellite observations can be used to monitor the spatial and temporal distribution of rainfall across a large domain. A comparison of satellite-derived maps with observational (station) data indicates large errors in rainfall derived from ground-based stations, especially for heavy and orographic rainfall [
12,
13,
14]. In recent years, radar rainfall data of different spatiotemporal resolutions have been made available for research and operational purposes globally. For example, Japan has used C- and X-band radars for operational purposes, and their spatial and temporal resolutions are 500-m and 5-min and 250-m and 1-min, respectively. In the United States, S-band radars have mostly been used operationally, with resolutions from 1000-m and >5-min, and most European countries have used C-band radars with spatial resolutions of about 1000-m and 5-min temporal resolutions. The quality of radar rainfall data has improved substantially recently, allowing the application of weather radar for hydrometeorological research, including the use of radar rainfall data in hydrological modeling [
9,
15,
16,
17,
18,
19,
20,
21,
22,
23].
Traditionally, hydrological models had to be suitable for use with rainfall point data. However, despite improvements in hydrological models to allow high spatiotemporal resolution radar rainfall data to be input to simulate runoff data on a high temporal basis, many hydrological models are still not suitable for use with high temporal resolution inputs, as the output of the model is always for more than an hour—i.e., for days, months, and so on. Nevertheless, several types of hydrological models have been developed for specific purposes, some of which are freely available for research purposes. One of these is the Hydrologic Engineering Center–Hydrologic Modeling System (HEC-HMS), which has been widely used worldwide. The accuracy and performance of this model using radar rainfall has also been widely recognized [
4,
5,
17,
19,
22,
23]. Since many types of hydrological models have been developed for research and operational purposes, no standard exists for selecting the type of hydrological model used. In many cases, semi-distributed or lumped models have been used [
5,
17,
19,
22,
23,
24,
25,
26,
27,
28], whereas physically distributed models have been used in certain other cases [
20,
21,
29]. Each type of hydrological model has its merits and demerits depending on performance and the objectives for which it is used. However, regardless of the model used, rapid hydrological simulation of river basins is always expected to mitigate water-related disasters.
Several researchers agree that the use of accurate spatiotemporal rainfall variation data in hydrological models is essential for monitoring river discharge and may help to improve our understanding of water balances [
8,
22,
24,
25,
30,
31]. Therefore, the use of high-resolution weather radar data in hydrological applications has grown significantly as an alternative to traditional rainfall observations with rain gauges [
5,
8,
17,
19,
20,
21,
22,
31], and several studies suggest that the use of high-resolution rainfall data in hydrological models may offer a more realistic output [
4,
8,
16,
21,
22,
24,
25,
30,
31]. However, there is no clear guideline on the optimum spatiotemporal radar rainfall resolution for this purpose. This is partially due to the differing spatiotemporal resolutions of rainfall data from different countries.
It is important to be careful when modeling hydrological issues using rainfall data. For example, rapid hydrological simulation of an extreme event is very important for mitigating water-related disasters. In some cases, forecasts of floods and probable inundated areas along a river should be issued as early as possible, especially in urban areas. From a practical approach, use of high-resolution radar rainfall data may increase the model computation time such that it might not be easy to run on a real-time basis. Moreover, considering the cost–benefit analysis, it is preferable to determine the most suitable spatial resolution of radar rainfall data for hydrological modeling.
There are several reviews on the effect of the spatial resolution of rainfall on hydrological applications, most of which were theoretically based or focused on the analysis of rain events over different parts of the world while using rain gauge or remote sensing based data. For example, Schilling [
29] suggested that at least 1–5-min and 1-km resolutions of gridded rainfall should be used for urban hydrological modeling. His study was based on rain gauge observations and review studies. In most cases, the effects of the spatial resolution of rainfall on the results of hydrological modeling for a river basin were evaluated based on available rain gauge networks and using interpolation techniques [
1,
2,
6,
7,
18]. However, such interpolations of rainfall data from networks where data availability is low can produce misleading results for the real spatial distribution of rainfall [
1,
5,
6,
7,
11]; notably, such hydrological applications have been carried out for a long time. Therefore, many studies have commented on the possible errors associated with the spatial resolution of rainfall data based on such work, and have recommended that the optimal spatial resolution of rainfall be derived for hydrological modeling of any river basin [
7,
29]. However, comparisons of simulated hydrological outputs using different spatial resolutions of rainfall will be limited, especially over medium to large river basins, in part due to the limited number of rain gauge networks.
Recently, the availability of radar rainfall data that can be used for operational and research purposes has increased in most countries. Fabry et al. [
32] analyzed radar rainfall data at different spatial resolutions and found that the error could increase under increasing the spatial resolution; hence, they recommended the use of high-resolution radar rainfall data, especially in urban hydrology applications. However, there was no clear opinion given on the optimal use of radar rainfall at different spatial scales. Einfalt et al. [
16] reviewed radar-estimated rainfall data by focusing on its application to urban drainage analyses. They first considered studies on the accuracy of the radar rainfall data and then considered its potential use in urban hydrology. They concluded that the analysis of extreme events in specific catchments should be meaningful considering the different types of rainfall. Based on these studies, it is recommended that finer resolution data (i.e., 1–5-min and 100–500-m) be used for urban hydrological applications [
16,
32]. However, early research did not compare hydrological simulation results for radar rainfall data at different spatial resolutions. Berne et al. [
33] considered high-resolution radar rainfall for different small-to-medium catchments, examining a few heavy rainfall events over different catchments. While they provided convincing evidence that the use of high spatial resolution radar data is important, their concerns pertained mainly to the location of the watershed given the radar observation point. However, they revealed that the temporal structure of radar rainfall data is also important for urban hydrology research. All of the above-mentioned studies paid more attention to the quality of radar data and evaluated uncertainties in analyzing it rather than comparing simulated hydrological results with observed hydrological data. Similar kinds of opinions have been presented in several studies [
1,
3,
4,
6,
7,
24]. A recent study by Ochoa-Rodriguez et al. [
25] agreed that the spatial resolution of rainfall input is strongly dependent on the drainage area of interest. Their study focused on different very small river basins (3–8 km
2) in Northwest Europe, and they determined that radar rainfall input data with a spatial resolution lower than 500 m should be sufficient for the simulated runoff of very small drainage basins. For larger river basins, however, input spatial resolutions of 1–3-km can be sufficient [
15,
29].
Very few rain events have been considered in the previous studies, which did not evaluate the types of different rainfall events separately in hydrological simulations. Moreover, comparisons of simulated hydrologic data that considered different resolutions of spatial rainfall data have been limited to very few river basins. Most of these studies focused on small river basins; however, it is still uncertain how hydrological modeling would differ in small-to-medium river basins with various resolutions for different types of rainfall (e.g., stratiform vs. convective or orographic vs. typhoon-associated rainfall). It should be noted that the use of appropriate rainfall data resolutions may save on computation time and make the approach more economically advantageous. Herein, we therefore attempt to estimate river basin runoff using various degrees of very high spatial resolution radar rainfall data to determine its effect on hydrological simulation. While the pattern of rainfall characteristics is different in each region, this study based on Japanese river basins can be a good starting reference on the optimal use of spatial radar resolutions in hydrological modeling. We also attempt to categorize the types of rainfall events and make recommendations on appropriate spatial resolutions to use during specific time intervals.
This study considers the use of high spatial resolution radar rainfall data for different rainfall events during summer to determine its effect on simulating runoff in a Japanese river basin. The results will allow us to provide recommendations on the most suitable radar spatial resolutions for hydrological modeling purposes. The HEC-HMS model is used to simulate discharge from the basin. In Japan, the Ministry of Land, Infrastructure, Transport and Tourism established an eXtended RAdar Information Network (XRAIN) that uses an operational data processing system developed by the National Research Institute for Earth Science and Disaster Resilience (NIED; [
33]). XRAIN consists of X-band multi-parameter radars and has spatial and temporal resolutions of 250-m and 1-min, respectively. This product is one of the best high-resolution radar rainfall systems in the world, and it is available to the public and private sectors in real time.
4. Discussion and Conclusions
Quantitative precipitation estimations from radar observations have improved recently, and composite maps of radar rainfall based on several observation points have been generated at very high spatial resolutions over a specified region with an accuracy that is quite close to ground data [
33,
34,
35]. This represents an improvement in the quality of available radar estimated rainfall data, which is being used in an increasing number of hydrological applications. To study the optimal spatial resolution of radar rainfall data for use in hydrological simulations, we selected the Tsurumi River Basin, which is one of the most important urbanized river basins in Japan; it is located in the heart of Tokyo. Altogether, 20 independent rain events were selected based on higher river discharge at the outlet of the basin. Among those selected rain events, 10 rain events represented isolated rainfall or CR types of patterns, while the remaining events represented the widespread type of rainfall within the Tsurumi River Basin. Of course, the selection of appropriate hydrological models is a key issue in the hydrological community that needs deeper investigation. In this study, we first adopted the HEC-HMS to utilize the very high spatial resolution of XRAIN data as the main input. Simulated results of each event were validated and calibrated separately. To obtain a good correlation between the observed and simulated discharge at the outlet of the sub-basins, some of the model parameters were optimized within the reliable range.
We used the gridded SCS CN, ModClark, and exponential recession as the loss, transform, and base flow methods, respectively, during the model set up. We selected the entire model for HEC-HMS given the suitability of the gridded rainfall data. Previous research showed that the SCS CN and ModClark methods are especially good for radar rainfall data [
25], and although there are several other models available in the HEC-HMS, we did not check them because this would have been beyond the scope of this study. However, there have been only a very limited number of studies carried out in Japanese river basins using HEC-HMS; therefore, this study used global reference parameters, particularly for the SCS CN method. The model parameters of each method were optimized to achieve close relationships between these and the observed data. Some of the model parameters were found to be sensitive for this river basin. The most sensitive model parameters were found to be the SF of the loss-gridded SCS CN and the TC and ST of the ModClark method. Previous studies have shown that SF can vary greatly during event analysis [
25]. Therefore, variations in SF are common and this parameter should be calibrated before it is applied in any river basin. TC and ST were optimized similarly and were very sensitive in the peak discharge cases [
25,
40]. The model parameters of the base flow were found to be less sensitive. One important point is that we selected individual events from different periods and based the time and optimization of model parameters on the peak-weighted root mean square error method for each event. This could be one of the reasons why the baseflow parameters were less sensitive than the others.
Several previous researchers studied the use of the spatial resolution of rainfall data in hydrological modeling in two main ways: with rain gauge networks [
1,
2,
6,
7,
18,
29,
40,
46] and via remote sensing data—e.g., satellite or radar rainfall data [
10,
14,
17,
19,
20,
30,
31,
32,
42]. Some studies used both approaches [
3,
4,
5,
12,
21,
22,
23,
26]. However, uncertainty can arise during the interpolation of rainfall data from low availability rain gauge networks [
1,
5,
6,
7,
11]. Additionally, satellite-based rainfall data may have high uncertainties, especially in small to medium river basins [
12,
13,
14]. High spatial resolution rainfall data can be obtained from radar observations; however, previous studies focused on the qualities of radar rainfall data rather than comparing them with simulated hydrological results. We believe that XRAIN has maintained its quality because it has been used for operational purposes [
33] and shown that is quite close to ground data [
33,
34,
35]. In this study, detailed research was conducted on the effects of spatial resolutions of radar rainfall on hydrological modeling. We focused our research mainly on simulated hydrological outputs rather than on comparative analyses of rainfall. Moreover, we included several events and separated them by type for the simulations, which had not been done effectively in previous research. Statistical assessment tools showed that the simulated results were better for the WR event cases. This may be due to the wide and smooth coverage of the rainfall distribution. In this study, the temporal resolution of XRAIN data were averaged every 5 min, and the default temporal resolution of the XRAIN data is 1-min. Changing the temporal resolution of the rain events could be another good topic for hydrological applications; however, consideration of the fixed temporal resolution of each rain event is one of the limitations of this study.
Next, we studied the effect of the spatial distribution of rainfall on the simulated discharge at the different sub-basins. The default spatial distribution of XRAIN is 250 m, and the simulated discharge using this dataset was considered as the base reference data for inter-comparison analysis. Rescaling of the spatial resolution of the rainfall data from coarse to high resolution could be very difficult, but the opposite approach is very simple. Therefore, rescaling of the spatial resolution was done based on the nearest neighbor approach. The spatial resolution of the rescaled XRAIN data was set as 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, and 5 km; it can be extended to coarser resolutions, but we divided the Tsurumi River Basin into several sub-basins, with their basin areas varying from about 8 to 50 km2. Hence, considering the basin area, we limited the rescaled rainfall data up to 5 km. We compared the accumulated average basin rainfall of each rescaled XRAIN data event separately. The coarse and fine average rainfall values were similar.
Finally, each rescaled rainfall event was separately considered as the main input for the model. The same optimized mode parameters for each event were used for all the data for those events. All the simulated discharges at different sub-basins were compared with reference discharges obtained from the default XRAIN data (i.e., 250-m spatial resolution). In general, the simulated discharge appears less sensitive using rescaled rainfall data with up to 2-km spatial resolutions for all of the event cases at the outlet of the river basin, which is a range similar to that used in previous studies [
25,
29,
33]. More specifically, we can suggest that the hydrological response using a spatial resolution of ≤5 km may somehow have a similar trend during the case of WR for a river whose basin area is greater than 25 km
2. This scenario is slightly different than that reported in previous studies because the type of rain was not classified in the latter.
However, the difference among the hydrological outputs is more noticeable for a small sub-basin (<8 km
2). In the case of the isolated rainfall or CR cases, the simulated data error started to increase rapidly using greater than 1-km spatial resolution radar rainfall data over almost all the sub-basins, except its outlet. Overall, the degrees of fluctuation in the statistical results were also higher for CR events than for WR events (
Figure 12,
Figure 13 and
Figure 14) for all sub-basins. In the case of small sub-basins (≤8 km
2), the optimal spatial resolution of WR data should be ≤2 km to obtain reliable simulated discharge data, but the optimal spatial scale of the radar rainfall data should be less than 1 km for isolated rainfall or CR cases. This was almost in agreement with previous studies [
16,
25,
32]. When we studied the small sub-basins (8–25 km
2), we also found that radar rainfall data with a spatial resolution greater than 1-km may provide high uncertainties in certain cases of CR. This is a finding unique to this study.
We also found that WR was less sensitive in hydrological simulations of medium-scale river basins in Japan and are very curious if a similar approach could be defined for such river basins in other regions. However, radar rainfall data is not yet commonly available in very high-resolution formats. In previous research, all types of rain events were generalized before an appropriate spatial resolution was selected for use in the hydrological modeling. This study clearly shows that the optimum scale of the spatial distribution of rainfall should depend upon the type of rainfall distribution, especially in small urban river basins. Hence, the type of rainfall must be distinguished before an optimum scale for radar rainfall data is selected for hydrological modeling in any river basin.
The importance of the spatial resolution of rainfall in hydrological modeling has been well emphasized [
15,
16,
30]. Since the rainfall distribution varies by season and location, it is not easy to generalize an appropriate spatial resolution of rainfall that can be used for hydrological modeling over any given river basin. However, some events have been analyzed in various countries, and some guidelines have been presented for very small river basins [
25]. This study provided a statistical summary based on 20 events over the different sub-basins of the Tsurumi River Basin in Japan. These findings may provide an important basis for the selection of appropriate spatial resolutions of rainfall data for hydrological modeling over similar river basins. We expect to carry out similar analyses for different river basins in different environments so that the optimal scale for spatial rainfall data can be determined on a global basis.