## Abstract

This study investigated the effects of activating convective parameterization (CP) in higher-resolution domains of a regional climate model, the Weather and Research Forecast (WRF) model, on watershed-scale precipitation by means of sensitivity analysis. The sensitivity analysis was conducted over three watersheds in the Hokkaido region, Japan. Three nested domains of 27, 9, and 3 km of horizontal resolution were set over the three study watersheds. Then, two types of sensitivity analysis were conducted. First, 36 of the single-year simulations were run with 12 combinations of two cloud microphysics (MP) schemes and six CP schemes and with three types of CP activation; activating a CP scheme only in the outermost domain, in the outermost and middle domain, and in all three domains. After the single-year simulations, long-term (32-year) simulations were conducted using an MP × CP combination with the three types of CP activation. The two-sensitivity analysis shows that activating CP in the higher-resolution domains does not always improve watershed-scale precipitation, but it has a high possibility of improving the results. Moreover, the improvement via activating CP in higher-resolution domains may continue for a long-term period.

## INTRODUCTION

Accuracy of dynamical downscaling is currently an important issue in water management because it has been more frequently utilized to approach water-related issues, particularly for climate change studies. Dynamical downscaling is a technique to obtain atmospheric variables at a higher resolution from a coarse-resolution atmospheric dataset using a regional climate model (RCM). For climate change studies in water management, dynamical downscaling is utilized to obtain precipitation at a regional or watershed scale and analyze climate change impacts on precipitation (e.g. De Paola *et al.* 2014; Harding & Snyder 2014; Berg *et al.* 2015; Ishida *et al.* 2017, 2018a; Toride *et al.* 2018). Meanwhile, because dynamical downscaling can produce not only precipitation but also other variables, hydrological modeling with dynamical downscaling can be used for analyses of climate change impacts on hydrological variables such as snow (e.g., Trinh *et al.* 2015; Ishida *et al.* 2018b) and flow discharge (e.g., Ali *et al.* 2015; Gu *et al.* 2015; Amin *et al.* 2017; Trinh *et al.* 2017; Gorguner *et al.* 2019).

The reliability of these climate change studies may largely depend on the accuracy of the dynamical downscaling. Climate change studies generally require year-round data to calculate summations or peaks for one year, each season, or each month. Furthermore, these studies require year-round data for at least one decade to reduce the interannual variabilities, and then calculate a trend and/or differences between two time slices. Therefore, dynamical downscaling accuracy during a year or decades would be of great interest for climate change studies in water management.

An important factor related to the accuracy of dynamical downscaling is the configuration of an RCM. An RCM partially relies on parameterization that describes physical processes such as precipitation, radiation, and land surface processes. The configuration of parameterizations in an RCM largely affects the dynamical downscaling accuracy. Among these parameterizations, then, this study mainly focused on convective parameterization (CP), which accounts for convective precipitation at a subgrid-scale. There are many schemes of CP for an RCM. Selection of a CP scheme is important to improve dynamical downscaling accuracy with respect to precipitation. Meanwhile, there is another issue regarding the use of CP. Because CP describes subgrid-scale precipitation processes, CP is not adequate in model domains whose resolution is sufficiently high to be able to reasonably resolve small-scale precipitation processes. However, it is still unknown whether CP should or should not be activated within a certain range of grid resolution, termed a ‘gray zone.’ In general, a gray zone is considered to be from 1 km to 5 km (Hong *et al.* 2004; Wootten *et al.* 2016).

There are several studies that have investigated the effect of CP activation in the gray zone on precipitation. Lee *et al.* (2011) simulated three severe storm events over the Korean Peninsula using an RCM. They used the nested four domains of 27-, 9-, 3-, and 1-km grid resolutions. Then, they investigated the effects of activating a CP scheme, the Kain–Fritsch (KF) scheme (Kain 2004), in the 3-km resolution domain on simulated precipitation. Their study shows that activating the CP scheme even in the 3-km resolution domain improves precipitation reconstruction for severe storm events. Sun & Barros (2014) also used the KF scheme and investigated the role of activating the CP scheme in a high-resolution domain to simulate a tropical cyclone over the southeastern United States. Their results also show that activation of the CP scheme in the 3-km resolution domain obtained better simulated precipitation results. While these two studies focused on precipitation during storm events, Wootten *et al.* (2016) used an RCM over Puerto Rico over one year with three nested domains of 30-, 10-, and 2-km grid resolutions. Then, they employed two CP schemes, the KF and Tiedtke (TD) schemes, to investigate the impacts of activating a CP scheme in the high-resolution domain. In their results, activating the CP in the 2-km resolution domain improved the monthly accumulated values of the domain-averaged precipitation and distribution of annual accumulated precipitation during the year using both of the CP schemes.

These previous studies show some improvement for precipitation reconstruction by activating a CP scheme in a high-resolution domain. Previous studies have focused on only one or two CP schemes (KF and TD) to discuss how activation of CP in a high-resolution domain improves simulated precipitation by additionally investigating atmospheric conditions in detail. In water management, however, it may be more important to pursue the possibility to improve simulated precipitation. Meanwhile, climate change studies generally employ time series of year-round data or their peaks, as previously mentioned. Watershed-scale precipitation may be a more appropriate measure in water management. Although the aforementioned studies targeted time series of domain-averaged precipitation or distribution of accumulated precipitation, it would be important in water management to investigate the effects of activating a CP scheme in a high-resolution domain on time series of watershed-scale precipitation at a higher temporal resolution.

In this context, this study analyzed the sensitivity of activating CP in a high-resolution domain to time series of watershed-scale precipitation at daily temporal scale during one year and a long-term period. This study employed the Weather and Research Forecast (WRF) model as an RCM and selected three watersheds in the Hokkaido region, Japan, as the study area: the Ishikari River watershed (ISRW), the Teshio River watershed (TSRW), and the Tokachi River watershed (TKRW). Then, three nested domains were set over the Hokkaido regions. The horizontal resolutions of the outermost, middle, and innermost domains were 27 km, 9 km, and 3 km, respectively. The basin-averaged values were used as watershed-scale precipitation. For the sensitivity analysis, first, three types of CP activation were considered as follows; CP activation only in the outermost domain (D1), in the outermost and the middle domains (D1-2), and in all the domains (D1-3). With the three types of CP activation, multiple CP schemes were considered. An RCM has other parameterizations, as aforementioned. The effects of CP activation in a high-resolution domain may change via a combination with a different scheme of parameterization other than CP. Especially, precipitation is generally very sensitive to the cloud microphysics (MP). Two MP schemes were therefore utilized together with multiple CP schemes in this study. A total of 36 single-year simulations were run with 12 combinations of two MP schemes and six CP schemes. Then, daily basin-averaged precipitation was calculated for the three study watersheds compared to the corresponding observations. Next, the WRF was run for a long-term period (32 years) with one of the 12 MP × CP combinations. The three types of CP activation (D1, D1-2, and D1-3) were also conducted for this long-term simulation. After calculating daily basin-averaged values, statistics between the time series and the peaks of the simulated and observed precipitation during the 32-year period were examined.

## STUDY WATERSHEDS

The ISRW, TSRW, and TKRW are the three largest watersheds in the Hokkaido region, Japan (Figure 1). The ISRW, TSRW, and TKRW are 14,330 km^{2}, 5,590 km^{2}, and 9,010 km^{2} in area, respectively. The Ishikari, Teshio, and Tokachi rivers are 268 km, 156 km, and 256 km in length, respectively. The Ishikari River is the third longest river and the ISRW is the second largest watershed in Japan. The Teshio River is the fourth longest river and the TSRW is the tenth largest watershed in Japan. The Ishikari and Teshio rivers flow into the Japan Sea and the Tokachi River flows into the Pacific Ocean. The highest elevation peaks in the ISRW, TSRW, and TKRW are 1,967 m, 1,558 m, and 2,077 m, respectively.

## METHODOLOGY

This study employed the Weather and Research Forecast (WRF: Shamarock *et al.* 2008) model as an RCM. Three two-way nested domains were set as illustrated in Figure 1 such that the innermost domain contains the three study watersheds. The outermost domain (Domain 1) has 32 × 44 simulation grids with 27 km of horizontal resolution. The middle domain (Domain 2) has 64 × 100 with 9 km of horizontal resolution. Domain 3 is the innermost domain. The number of grids is 133 × 94 and its horizontal resolution is 3 km. All the domains have 40 vertical levels. For all the simulations in this study, the Bougeault and Lacarrere scheme was used for planetary boundary layer parameterization, the Dudhia scheme for short-wave radiation parameterization, and the rapid and accurate radiative transfer scheme for long-wave radiation parameterization.

The initial and boundary conditions for the sensitivity analysis in this study were obtained from a reanalysis dataset, ERA-Interim (Berrisford *et al.* 2009). ERA-Interim is a second-generation reanalysis dataset that digests more observational data in its data assimilation system compared to first-generation reanalysis datasets. ERA-Interim is available over the entire globe and temporally covers greater than 35 years from 1979 to the present. The spatial and temporal resolution of ERA Interim is 0.75 degree × 0.75 degree with 37 vertical pressure levels and 6-hourly data, respectively.

This study utilized the Asian Precipitation – Highly-Resolved Observational Data Integration Towards Evaluation dataset version 1207 (APHRODITE: Kamiguchi *et al.* 2010) as observational data. APHRODITE is a gridded daily precipitation dataset based on gauging observations. The resolution of APHRODITE V1207 covers the whole of Japan with 0.05 degree × 0.05 degree resolution grids. The temporal coverage of APHRODITE V1207 is from 1900 to 2011.

Two types of simulations were conducted to investigate the effects of CP scheme activation in a high-resolution domain on watershed-scale precipitation. First, a total of 36 single-year simulations using the WRF were run over the Hokkaido region for the year 2011. Two MP schemes and six CP schemes in the WRF were utilized consisting of 12 combinations of the MP and CP schemes. The two MP schemes were the Lin *et al.* (LIN) and new Eta microphysics (ETA) schemes. The six CP schemes were the KF, Betts–Miller–Janjic (BMJ), Grell–Freitasthe (GF), new Grell (G3), TD, and new Tiedtke (NTD) schemes. Three simulations with each of the 12 combinations were conducted with CP scheme activation only in the 27-km outermost domain (D1), in the 27-km outermost and the 9-km middle domain (D1-2), and in all the three (27 km, 9 km, and 3 km) domains (D1-3), respectively. In total, 36 single-year simulations were conducted. After calculating basin-averaged values from the simulated precipitation, the basin-averaged precipitation was compared to the corresponding observations over the three study watersheds to investigate the sensitivity of activating different CP schemes on watershed-scale precipitation.

Next, 32-year simulations from 1980 through 2011 were conducted using the WRF to investigate effects of CP scheme activation in a high-resolution domain on watershed-scale precipitation for a long period. For the 32-year simulations, a set of the MP × CP combination was selected from the 12 combinations that were used for the single-year simulations. Then, three 32-year simulations were run with the selected MP × CP combination and with the three types of CP activation: activating CP only in the outermost domain (D1), in the outermost and middle domains (D1-2), and in all the three domains (D1-3). Similar to the single-year simulations, daily basin-averaged values were calculated from simulated precipitation for the 32-year period. Then, the basin-averaged values of the simulated daily precipitation were compared to the corresponding observational data.

### Single-year simulation

Figures 2–4 show the times series of the simulated basin-averaged precipitation over the ISRW, TSRW, and TKRW, respectively, for the 36 cases together with those of the corresponding observations. The largest peak in the daily basin-averaged precipitation over the ISRW occurred on September 2, 2011, whose value is 106.1 mm in the observations (Figure 2). There are some other peaks over the ISRW although the values of these peaks are relatively small. The second and the third peaks occurred during the middle of July and August, respectively, 63.2 mm and 48.4 mm, respectively. The TSRW has a large single peak in the daily basin-averaged precipitation (Figure 3), whose value is 87.9 mm in the observational data. The single large peak over the TSRW occurred on the same day as the largest peak over the ISRW. Compared to this single large peak, the other precipitation peaks are much smaller over the TSRW. Over the TKRW, there are two relatively large peaks during 2011 (Figure 4). The larger one occurred on September 5, a different day from the single large peak over the ISRW and TSRW. The smaller one occurred on July 14, the same day as the second peak over the ISRW. The value of the former peak is 64.1 mm and that of the latter peak is 60.2 mm. Both the daily precipitation peaks over the TKRW are smaller than the single large peak over the ISRW and the TSRW.

In all 36 cases, the simulated daily basin-averaged precipitation over all three study watersheds generally agrees well with the corresponding observations through the year 2011, as shown in Figures 2–4. Over the ISRW, peaks in the daily basin-averaged precipitation are mostly underestimated (Figure 2). Particularly, activating a CP scheme only in the 27-km outermost domain (D1) underestimates the three largest peaks during 2011 using all 12 combinations of the MP × CP schemes. CP scheme activation in higher-resolution domains (D1-2 and D1-3) does not improve the second and the third peaks in their results. However, most cases with D1-2 and D1-3 increase the largest peak of the simulated daily basin-averaged precipitation over the ISRW during 2011. Consequently, the largest peak in the simulated daily basin-averaged precipitation matches that observed in some of the results (e.g., Lin-KF, Lin-GF, Lin-NTD, and Eta-G3) by activating CP in higher-resolution domains (D1-2 and D1-3).

The single large peak in the daily basin-averaged precipitation over the TSRW is overestimated in nearly all cases (Figure 3). In general, Lin for MP overestimates the single large peak over the TSRW. In contrast to the ISRW, the single large peak value decreased by employing a CP scheme in higher-resolution domains (D1-2 and D1-3) compared to activation of a CP scheme only in the outermost domain (D1). Consequently, the single large peak was reasonably reconstructed in one D1-3 case using Lin (Lin-G3) and most D1-3 cases using Eta (except Eta-GF).

As for peaks in the daily basin-averaged precipitation, the simulated results are worst over the TKRW among the three study watersheds, as shown in Figure 4. In all cases, the two largest peaks are underestimated over the TKRW. Unlike over the ISRW and TSRW, clear differences among the three types of CP activation (D1, D1-2, and D1-3) are not found in the two largest peaks.

Figure 5 shows the annual average daily basin-averaged precipitation over the three study watersheds. Lin for MP generates more precipitation than Eta over all three study watersheds. Lin overestimates the annual average precipitation, but Eta reasonably reconstructs it over the ISRW and TKRW. In contrast, the annual average precipitation is overestimated in all cases over the TSRW. In general, CP activation in higher-resolution domains reduces the annual average daily basin-averaged precipitation over the three study watersheds. Only in the cases using KF or G3 for the CP did activating CP in the outermost and the middle domains (D1-2) decrease, and that in all domains (D1-3) increases the annual mean values of the daily basin-averaged precipitation over the three study watersheds.

The correlation coefficients between the simulated and the observed daily-averaged precipitation over the three study watersheds during 2011 are shown in Figure 6. The simulated results show a high correlation coefficient value over all three study watersheds in all 36 cases. The correlation coefficient is greater than 0.812 over the ISRW, greater than 0.877 over the TSRW, and greater than 0.826 over the TKRW. The highest value is 0.882 over the ISRW in the case of Era-NTD-D1-3, 0.908 over the TSRW in the case of Lin-NTD-D1-3, and 0.886 over the TKRW in the case of Lin-NTD-D1-3. The highest correlation coefficient value is found in the case using NTD for CP and CP activation for all the domains (D1-3) over all the study watersheds.

CP scheme activation in all three domains (D1-3) shows a higher correlation coefficient value over the ISRW for the 12 MP × CP combinations except the Lin-BMJ and Eta-GF combinations, compared to CP activation only in the outermost domain (D1) and in the outermost and middle domains (D1-2). However, D1-3 increases the correlation coefficient only for five and six of the 12 MP × CP combinations over the TSRW and TKRW, respectively. Only NTD for CP with both MP schemes (Lin and Eta) improves the correlation coefficient value by employing CP in all three domains (D1-3) over all three study watersheds. On average, over the 12 MP × CP combinations, the correlation coefficient is 0.829 for D1, 0.838 for D1-2, and 0.855 for D1-3 over the ISRW; 0.903 for D1, 0.899 for D1-2, and 0.899 for D1-3 over the TSRW; and is 0.846 for D1, 0.841 for D1-2, and 0.860 for D1-3 over the TKRW. The average of the correlation coefficient is increased by activating a CP scheme in all three domains (D1-3) over the ISRW and TKRW, but it is decreased over the TSRW.

Figure 7 plots the Nash–Sutcliffe coefficient between the simulated and observed daily-averaged precipitation over the three study watersheds during 2011. The Nash–Sutcliffe coefficient ranges between 0.655 (Eta-KF-D1) and 0.771 (Lin-NTD-D1-3) over the ISRW. Combinations of Lin for MP show a lower value of Nash–Sutcliffe coefficient over the TSRW, compared to those of Eta. The lowest value is 0.463, obtained in the case of Lin-NTD-D1-2. Meanwhile, the Nash–Sutcliffe coefficient is greater than 0.675 for the combinations with Eta over the TSRW. The highest value of 0.809 is found over the TSRW in the case of Lin-NTD-D1-3. Over the TKRW, the Nash–Sutcliffe coefficient is relatively high compared to that of the other two watersheds. Its value is greater than 0.681 and the highest value is 0.781 over the TKRW in the case of Lin-NTD-D1-3.

Similar to the correlation coefficient, activating CP in all three domains (D1-3) improved Nash–Sutcliffe values over the ISRW for most of the 12 MP × CP combinations. Meanwhile, D1-3 shows a higher Nash–Sutcliffe coefficient value for six of the 12 combinations over the TSRW and TKRW. Moreover, only NTD for CP with both MP schemes (Lin and Eta) improved the Nash–Sutcliffe value by activating the CP in all three domains (D1-3) over all three study watersheds. However, the Nash–Sutcliffe coefficient is more sensitive to CP activation in higher-resolution domains than the correlation coefficient. The Nash activation of Sutcliffe coefficient is 0.679 for D1, 0.693 for D1-2, and 0.715 for D1-3 over the ISRW on average over the 12 MP × CP combinations. The difference in the average value of the Nash–Sutcliffe coefficient is 0.036 between D1 and D1-3 and 0.026 in the correlation coefficient. The average value of the Nash–Sutcliffe coefficient over the 12 combinations is 0.653 for D1, 0.662 for D1-2, and 0.692 for D1-3 over the TSRW. A large improvement in the Nash–Sutcliffe coefficient via activating CP in all three domains (D1-3) can be found in the MP × CP combinations consisting of both of the MP schemes and the BMJ, TD, and NTD CP schemes. Consequently, the average value of the Nash–Sutcliffe coefficient increases by employing CP in higher-resolution domains, while the average value of the correlation coefficient decreases. The average value of the Nash–Sutcliffe coefficient is 0.713 for D1, 0.704 for D1-2, and 0.735 for D1-3 over the TKRW. The difference in the average of the Nash–Sutcliffe coefficient is also larger than that in the correlation coefficient over the TKRW.

### Thirty-two-year simulation

Dynamical downscaling was conducted over the Hokkaido region during the 32-year period from 1980 through 2011 to investigate effects of activating CP in a high-resolution domain to watershed-scale precipitation during a long-term period. For the 32-year simulations, Lin was used for MP and NTD for CP because this combination with CP activation in all the three domains (Lin-NTD-D1-3) shows the highest values of Nash–Sutcliffe coefficient over the ISRW and TKRW and the fourth highest over the TSRW among the 36 cases for the single-year simulations (Figure 7). For the combination of the Lin MP scheme and the NTD CP scheme, moreover, the Nash–Sutcliffe coefficient obtained with CP activation in all the domains (D1-3) was higher than that only in the outermost domain (D1) and in the outermost and middle domains (D1-2). Therefore, the combination of the Lin MP scheme and the NTD CP scheme (Lin-NTD) was selected for the 32-year simulations. Then, the 32-year simulations were conducted using the three types of CP activation (D1, D1-2, and D1-3).

Figure 8 shows the simulated and observed daily basin-averaged precipitation over the ISRW, TSRW, and TKRW during the 32-year period from 1980 through 2011. The simulated results using all three types of CP activation agree with the corresponding observations through the 32-year period. As shown in Table 1, the correlation coefficient is greater than 0.824, and the Nash–Sutcliffe coefficient is greater than 0.523 over the three study watersheds during the 32-year period for the three types of CP activation (D1, D1-2, and D1-3).

Case | Mean (mm) | Correlation coefficient | Nash–Sutcliffe coefficient | |
---|---|---|---|---|

Ishikari | Obs | 3.43 | ||

D1 | 4.36 | 0.859 | 0.667 | |

D1-2 | 3.92 | 0.861 | 0.711 | |

D1-3 | 3.75 | 0.876 | 0.753 | |

Teshio | Obs | 3.16 | ||

D1 | 4.05 | 0.824 | 0.523 | |

D1-2 | 3.55 | 0.830 | 0.617 | |

D1-3 | 3.41 | 0.859 | 0.712 | |

Tokachi | Obs | 2.88 | ||

D1 | 3.44 | 0.876 | 0.742 | |

D1-2 | 3.29 | 0.878 | 0.758 | |

D1-3 | 3.11 | 0.890 | 0.788 |

Case | Mean (mm) | Correlation coefficient | Nash–Sutcliffe coefficient | |
---|---|---|---|---|

Ishikari | Obs | 3.43 | ||

D1 | 4.36 | 0.859 | 0.667 | |

D1-2 | 3.92 | 0.861 | 0.711 | |

D1-3 | 3.75 | 0.876 | 0.753 | |

Teshio | Obs | 3.16 | ||

D1 | 4.05 | 0.824 | 0.523 | |

D1-2 | 3.55 | 0.830 | 0.617 | |

D1-3 | 3.41 | 0.859 | 0.712 | |

Tokachi | Obs | 2.88 | ||

D1 | 3.44 | 0.876 | 0.742 | |

D1-2 | 3.29 | 0.878 | 0.758 | |

D1-3 | 3.11 | 0.890 | 0.788 |

There are no clear differences in the correlation coefficient for the 32-year period over the ISRW and TKRW among the three types of CP activation (Table 1). The differences are at most 0.003. Activating CP in higher-resolution domains slightly increases the correlation coefficient over the TSRW. The correlation coefficient increases 0.009 using D1-2 and 0.015 using D1-3 over the TSRW, compared to D1. However, the Nash–Sutcliffe coefficient for the 32-year period clearly increases by the use of CP for higher-resolution domains (Table 1). Compared to D1, D1-2 shows 0.044, 0.103, and 0.018 higher values of Nash–Sutcliffe coefficient and D1-3 shows 0.068, 0.155, and 0.027 higher values over the ISRW, TSRW, and TKRW, respectively.

The Nash–Sutcliffe coefficient was calculated for each year to investigate the transition of effects of activating CP in higher-resolution domains (see Figure 9). Mostly activation of CP in all the three domains (D1-3) obtains the highest value of the Nash–Sutcliffe coefficient over the three studied watersheds for each year through the 32-year period among the three types of CP activation. It should be noted that activating CP in all the three domains (D1-3) consistently provides a high Nash–Sutcliffe coefficient value except during 1980 and 2009 over the ISRW and during 1980 over the TSRW. For example, D1-3 provides 0.595 for the Nash–Sutcliffe coefficient during 1984 over the ISRW although D1 and D1-2 provide 0.182 and 0.379, respectively (Figure 9(a)). Furthermore, significant improvement by activating CP in the highest-resolution domain is found over the TSRW nearly throughout the 32-year period. While the yearly Nash–Sutcliffe coefficient is frequently less than 0.5 over the TSRW in the case of D1, the yearly value is mostly greater than 0.5 in the case of D1-3, as shown in Figure 9(b). Particularly, the yearly Nash–Sutcliffe coefficient is −0.157 during 1984 and 0.069 during 2007 in the case of D1. Nevertheless, the value increases to 0.668 during 1984 and 2007 by activating CP in the 3-km resolution domains (D1-3).

The average values of the daily basin-averaged precipitation over the three study watersheds during the 32-year period from 1980 through 2011 are overestimated in all the dynamically downscaled results (Table 1). Particularly, activating CP only in the outermost domain (D1) generates much more precipitation than the corresponding observations. Meanwhile, activating CP in higher (9 km and 3 km) resolution domains can reduce precipitation over the three study watersheds. Activating CP in the 3-km resolution domain (D1-3) provides the smallest values of the 32-year mean basin-averaged precipitation over all three study watersheds. Compared to the average daily precipitation in the case of D1, D1-3 decreases by 0.61 mm over the ISRW, by 0.64 mm over the TSRW, and by 0.33 mm over the TKRW.

Figure 10 shows the significant improvement in the annual peak values of the daily basin-averaged precipitation by activating CP in all three domains (D1-3) over the TSRW. Indeed, activating CP in higher-resolution domains shows a higher correlation in the annual peak values between the simulation results and the corresponding observations over the TSRW (Table 2). The differences in the correlation coefficient are 0.050 between D1 and D1-2 and 0.138 between D1 and D1-3. Moreover, the root mean square error (RMSE) and the mean absolute error (MAE) are reduced by employing CP in higher-resolution domains (Table 2). Compared to D1, D1-2 reduced the RMSE by 12% and MAE by 17%, and D1-3 reduced the RMSE by 40% and MAE by 34% over the TSRW. However, the plots in Figure 10 do not show clear differences among the three types of CP activation over the ISRW and TKRW. CP activation only in the outermost domain (D1) shows the highest correlation over these two watersheds (Table 2). D1-3 shows the best results for RMSE and MAE over the ISRW while D1 show the best results for RMSE and MAE over the TKRW.

Case | Correlation coefficient | RMSE (mm) | MAE (mm) | |
---|---|---|---|---|

Ishikari | D1 | 0.850 | 15.99 | 13.05 |

D1-2 | 0.843 | 16.74 | 12.20 | |

D1-3 | 0.840 | 15.82 | 11.54 | |

Teshio | D1 | 0.628 | 26.25 | 20.68 |

D1-2 | 0.678 | 23.21 | 17.22 | |

D1-3 | 0.766 | 15.75 | 13.74 | |

Tokachi | D1 | 0.792 | 19.87 | 14.75 |

D1-2 | 0.786 | 20.23 | 15.30 | |

D1-3 | 0.781 | 20.98 | 15.69 |

Case | Correlation coefficient | RMSE (mm) | MAE (mm) | |
---|---|---|---|---|

Ishikari | D1 | 0.850 | 15.99 | 13.05 |

D1-2 | 0.843 | 16.74 | 12.20 | |

D1-3 | 0.840 | 15.82 | 11.54 | |

Teshio | D1 | 0.628 | 26.25 | 20.68 |

D1-2 | 0.678 | 23.21 | 17.22 | |

D1-3 | 0.766 | 15.75 | 13.74 | |

Tokachi | D1 | 0.792 | 19.87 | 14.75 |

D1-2 | 0.786 | 20.23 | 15.30 | |

D1-3 | 0.781 | 20.98 | 15.69 |

## DISCUSSION

The results of the single-year simulations over the Hokkaido region show no simple relationship between the three types of CP activation (D1, D1-2, and D1-3) and the accuracy of their results. For example, activation of CP in D1-2, which uses CP in two of the three model domains, sometimes generates the smallest annual mean basin-averaged precipitation and other times the largest values (Figure 5). D1-2 frequently provides the worst results with respect to the correlation coefficient or the Nash–Sutcliffe coefficient among the three types of CP activation (Figures 6 and 7). Meanwhile, CP activation only in the outermost domain (D1) obtained the highest value of the correlation coefficient or the Nash–Sutcliffe coefficient for some of the 12 MP × CP combinations (Figures 6 and 7). Activating CP in all the three domains (D1-3) shows the highest value for some other MP × CP combinations (Figures 6 and 7).

The effects of CP activation in higher-resolution domains are not closely related to a selection of MP schemes and even a selection of CP schemes. For example, CP activation in all the domains (D1-3) using the Lin MP scheme over the TSRW produced the smallest annual mean daily basin-averaged precipitation with two of the six CP schemes and the largest annual mean with the other four CP schemes (Figure 6). CP activation only in the outermost domain (D1) using Eta over the TKRW is the best with respect to the Nash–Sutcliffe coefficient with one of the six CP schemes, but D1 is the worst with two of the six CP schemes (Figure 7). Meanwhile, D1-3 using the GF CP scheme shows the best correlation coefficient using the Lin MP scheme, but the worst using the Eta MP scheme over the ISRW (Figure 6).

Furthermore, the effects of CP activation in higher-resolution domains are mostly different among the three study watersheds. CP activation in all three domains (D1-3) obtained the highest values of Nash–Sutcliffe coefficient in ten of the 12 MP × CP combinations over the ISRW. However, D1-3 is the best with respect to the Nash–Sutcliffe coefficient only in six of the 12 combinations over the TSRW and TKRW (Figure 7).

These results indicate that activating CP in the high-resolution domain does not always improve dynamical downscaling accuracy, although previous studies by Lee *et al.* (2011), Sun & Barros (2014), and Wootten *et al.* (2016) showed improvement in reconstructing precipitation. The effects of activating CP in higher-resolution domains depends on the study watershed, the resolution of the domain where the CP is activated, and the schemes of parameterizations (not only CP but also other parameterizations).

However, the results of the single-year simulations show a high possibility that activating CP in higher resolution domains, particularly in the 3-km resolution domain, can improve dynamical downscaling accuracy. Notably, activating CP in higher resolutions shifted the simulated values of the large peaks of the daily basin-averaged precipitation nearer the corresponding observation by increasing and decreasing it over the ISRW and TSRW, respectively. Activating CP at a higher resolution does not simply reduce the mean values and the peak value of the daily basin-averaged precipitation. However, it can work for improving the simulated results of precipitation with some combinations of MP and CP.

For the long-term (32-year) period from 1980 through 2011, activating CP in higher-resolution domains also improved the reconstruction of the daily basin-averaged precipitation over all three study watersheds (see Table 1 and Figure 9). Activating CP in higher-resolution domains significantly reduced the 32-year average value of the daily basin-averaged precipitation over the three study watersheds (Table 1). Then the 32-year average precipitation via the use of CP for higher-resolution domains was nearer the observation because all the simulations overestimated the 32-year average precipitation. Meanwhile, activating CP at a higher resolution results in a higher correlation coefficient and Nash–Sutcliffe coefficient for the 32-year period (Table 1), although the annual peaks of the daily basin-averaged precipitation improved only over the TSRW by employing CP in higher-resolution domains. This result indicates that improvement by activating CP at a higher resolution can continue throughout a long-term period.

Activating CP in higher-resolution domains increases the value of the yearly Nash–Sutcliffe coefficient over all three study watersheds, which was calculated for each year during the 32-year period (Figure 9). This result indicates that improvement by activating CP in higher-resolution domains does not work only for a specific year. If activation of CP in higher-resolution domains with a certain combination of parameterizations improves the results for a year with a specific set of parameterization schemes of an RCM, it has the potential to improve the results throughout a long-term period. This finding is very valuable for climate change studies because these studies generally require long-term simulations.

## CONCLUSIONS

To improve dynamical downscaling accuracy for a long-term period, the sensitivity of CP activation in higher-resolution domains to watershed-scale precipitation was investigated over three study watersheds in the Hokkaido region, Japan. First, 36 of the single-year simulations were run with 12 combinations of two MP schemes and six CP schemes and with three types of CP activation: activating a CP scheme only in the 27-km outermost domain (D1), in the 27-km outermost and the 9-km middle domain (D1-2), and in all the three (27 km, 9 km, and 3 km) domains (D1-3). After the single-year simulations, long-term (32-year) simulations were conducted with the best MP × CP combinations over the three study watersheds.

The results in this study indicate that dynamical downscaling accuracy with respect to precipitation is not always improved by activating CP in higher-resolution domains. The improvement by activating CP in higher-resolution (9-km and 3-km) domains depends on the study watershed and selected parameterizations schemes. However, activation of CP in higher-resolution domains, particularly in the 3-km resolution domain, has a high possibility of improving dynamical downscaling accuracy with respect to watershed-scale precipitation. For instance, activating CP in higher resolutions improved the simulated values of the large peaks of the daily basin-averaged precipitation over the ISRW and TSRW in the single-year simulations. Activating CP in higher resolutions obtained higher values of the correlation coefficient and the Nash–Sutcliffe coefficient with several combinations of MP and CP. Meanwhile, the simulated results with the selected combination of a MP scheme and a CP scheme showed that activating CP in higher-resolution domains improved the long-term (32-year) reconstruction of the daily basin-averaged precipitation over all three study watersheds. Such an improvement by CP activation in a high-resolution domain works not only for a specific year but also throughout a long-term period. These findings indicate that activation of CP in a high-resolution domain may enhance the reliability of climate change studies using dynamical downscaling.

## ACKNOWLEDGEMENTS

This study was supported by the General Collaborative Research (29G-10) of Disaster Prevention Research Institute, Kyoto University, and by the Integrated Research Program for Advancing Climate Models (TOUGOU). This study used the supercomputer of ACCMS, Kyoto University.