Future changes and uncertainties in Asian precipitation simulated by multiphysics and multi–sea surface temperature ensemble experiments with high-resolution Meteorological Research Institute atmospheric general circulation models (MRI-AGCMs)

Authors

Errata

This article is corrected by:

  1. Errata: Correction to “Future changes and uncertainties in Asian precipitation simulated by multiphysics and multi-sea surface temperature ensemble experiments with high-resolution Meteorological Research Institute atmospheric general circulation models (MRI-AGCMs)” Volume 118, Issue 5, 2303, Article first published online: 8 March 2013

Abstract

[1] This study focuses on projecting future changes in mean and extreme precipitation in Asia, and discusses their uncertainties. Time-slice experiments using a 20-km-mesh atmospheric general circulation (AGCM) were performed both in the present-day (1979–2003) and the future (2075–2099). To assess the uncertainty of the projections, 12 ensemble projections (i.e., combination of 3 different cumulus schemes and 4 different sea surface temperature (SST) change patterns) were conducted using 60-km-mesh AGCMs. For the present-day simulations, the models successfully reproduced the pattern and amount of mean and extreme precipitation, although the model with the Arakawa–Schubert (AS) cumulus scheme underestimated the amount of extreme precipitation. For the future climate simulations, in South Asia and Southeast Asia, mean and extreme precipitation generally increase, but their changes show marked differences among the projections, suggesting some uncertainty in their changes over these regions. In East Asia, northwestern China and Bangladesh, in contrast, mean and extreme precipitation show consistent increases among the projections, suggesting their increases are reliable for this model framework. Further investigation by analysis of variance (ANOVA) revealed that the uncertainty in the precipitation changes in South Asia and Southeast Asia are derived mainly from differences in the cumulus schemes, with an exception in the Maritime Continent where the uncertainty originates mainly from the differences in the SST pattern.

1. Introduction

[2] Asia is one of the most water-rich regions of the world where many people depend on water for agriculture. However, the region is sometimes inundated with too much water, which results in disastrous floods particularly in the more vulnerable countries. Therefore, reliable projections of future precipitation changes are required, for both mean precipitation and the degree of variability of precipitation. Such projections are especially necessary for impact, vulnerability, and adaptation studies in relation to climate change in Asia.

[3] The World Climate Research Programme's (WCRP's) Coupled Model Intercomparison Project phase 3 (CMIP3) [Meehl et al., 2007] multimodels has projected increased summertime rainfall over East Asia, South Asia, and most of Southeast Asia due to enhanced moisture convergence under a warmer climate, despite a tendency toward the weakening of monsoonal airflows [Intergovernmental Panel on Climate Change (IPCC), 2007b]. An increase in the frequency of intense precipitation events is very likely in East Asia and in parts of South Asia [IPCC, 2007b]. Kusunoki and Arakawa [2012] projected an increase in precipitation intensity in June–July over almost all regions of East Asia, based on the coupled atmosphere–ocean general circulation model (AOGCM) projections of the CMIP3 data set. However, large uncertainty exists in projections of Asian monsoon precipitation [e.g., Kripalani et al., 2007; Li et al., 2011; Turner and Slingo, 2009; Turner and Annamalai, 2012].

[4] Higher horizontal resolution models may be needed to better reproduce extreme rainfall associated with tropical cyclones and precipitation systems such as the Meiyu–Changma–Baiu rainband. To this end, dynamical downscaling approaches have been developed, including regional climate models (RCM), variable-resolution atmospheric general circulation models (AGCM), and global high-resolution AGCMs.Zou et al. [2010]showed that the variable–resolution AGCM of Laboratoire de Météorologie Dynamique-zoom (LMDZ4) can be a useful tool for the dynamical downscaling of rainfall variability over eastern China.Feng et al. [2011]reported that a global 40-km-mesh AGCM projected a larger change in precipitation intensity than did CMIP3 models.Kamiguchi et al. [2006], using an earlier version of the 20-km AGCM in the Meteorological Research Institute [Mizuta et al., 2006], projected an increase in heavy precipitation particularly in Bangladesh and in the Yangtze River basin at the end of the 21st century, due to the intensified convergence of water vapor flux in boreal summer.

[5] Quantifying uncertainty in projections of future climate changes is a critical issue. Kusunoki et al. [2011]assessed future changes in East Asian summer rainfall based on a combination of a single global warming projection experiment with a 20-km-mesh AGCM and ensemble projections with a 60-km-mesh AGCM, using four different SSTs and three atmospheric initial conditions. In their future climate simulation by the 20-km model, precipitation shows an increase over the Yangtze River valley in May–July (Meiyu), over Korea in May (Changma), and over Japan in July (Baiu). Simulations with the 20-km and 60-km models consistently show that the termination of the rainy season over Japan tends to be delayed until August in the future climate. Differences in physical parameterization employed in the models as well as differences in SST changes projected by AOGCMs could lead to large uncertainty in future precipitation projections. Based on the CMIP3 models,Turner and Slingo [2009] suggested that differences in convective parameterization may explain the wide range of extreme precipitation changes projected in the Indian monsoon region.

[6] We performed climate projections for the end of the 21st century using the 60-km-mesh AGCM under a multiphysics multiSST framework (twelve ensemble experiments) as well as using the 20-km-mesh AGCM. This approach allows us to address differences of impacts due to model physics and prescribed future SST changes on projected future changes in Asian climate [Murakami et al., 2012a].

[7] In this paper, we focus on projecting future changes in mean and extreme precipitation in Asia, and discuss their uncertainties. The remainder of this paper is organized as follows. Section 2 describes the models used, section 3 explains the experiment design, and section 4 describes the validation data for precipitation. Section 5verifies the mean and extreme precipitation in the present-day climate simulations, andsection 6 presents the future changes projected by the models. Section 7 discusses the uncertainty arising from the use of different future SST patterns and cumulus parameterization schemes. Finally, the conclusion and discussions are shown in section 8.

2. Models

[8] The model used in this study is the latest Meteorological Research Institute AGCM (MRI-AGCM3.2) [Mizuta et al., 2012]. The model simulations were run at horizontal resolutions of TL959 (MRI-AGCM3.2S, the 20-km model) and TL319 (MRI-AGCM3.2H, the 60-km model). The model is equipped with multiple cumulus convection schemes that can be easily switched. Three cumulus convection schemes were used for the multiphysics ensemble simulations: the prognostic Arakawa–Schubert (AS) cumulus convection scheme [Arakawa and Schubert, 1974; Randall and Pan, 1993]; a new cumulus convection scheme named as “Yoshimura scheme (YS)” [Yukimoto et al., 2011]; and the Kain–Fritsch (KF) convection scheme [Kain and Fritsch, 1990, 1993]. The YS scheme is based on the Tiedtke [1989]scheme, but modified as to represent all top-level cumulus plumes by interpolating two convective updrafts with maximum and minimum rates of turbulent entrainment and detrainment [Yukimoto et al., 2011]. The use of the YS scheme in the MRI-AGCM yields a more realistic simulation of tropical precipitation than the use of the AS scheme [Mizuta et al., 2012].

3. Experiment design

[9] A time-slice experiment [Bengtsson et al., 1996] was conducted, in which the high-resolution AGCM was forced by prescribed sea surface temperatures (SSTs). The experiment design was identical to that reported byMurakami et al. [2012a].

3.1. Present-Day Climate Simulations

[10] For the present-day climate simulation, the 20-km model with the YS scheme was integrated for 25 years (1979–2003) with the observed historical SST and sea ice data of HadISST1 [Rayner et al., 2003]. This simulation is equivalent to an Atmospheric Model Intercomparison Project (AMIP)-type experiment for an atmospheric model, which is widely adopted in numerous modeling studies. Considering that the performance of the 20-km model depends on the convection scheme, the uncertainty in the simulations should be evaluated by ensemble simulations using the 20-km model with different convection schemes. However, we used the 60-km model for ensemble simulations owing to the limited availability of computer resources (the computation time required for the 20-km model is about 15 times greater than that for the 60-km model). We conducted ensemble simulations with the 60-km model for three different convection schemes (YS, AS, and KF), using experimental settings identical to those for the 20-km model.Table 1summarizes the specifications of the present-day climate simulations. It is noted that there is no physical parameter adjustment between the 20-km model and the 60-km model using the YS scheme.

Table 1. Experiment Design
Run NameaCumulus Convection SchemeSea Surface TemperatureGrid Size (km)
  • a

    First character of the run name denotes horizontal grid size: S (Super high, 20 km), H (High, 60 km). Second character denotes target period: P (Present-day), F (Future).

  • b

    Observational data by the Hadley Centre of Met Office, United Kingdom [Rayner et al., 2003].

  • c

    Coupled Model Intercomparison Project 3.

  • d

    Multi-Model Ensemble.

Present-Day Climate Simulations for 1979–2003 (25 Years)
SP_YSYoshimura (YS)Observation HadISSTb20
HP_YSYoshimura (YS)Observation HadISST60
HP_ASArakawa-Schubert (AS)Observation HadISST60
HP_KFKain-Fritsch (KF)Observation HadISST60
 
Future Climate Simulations for 2075–2099 (25 Years)
SF_YSYoshimura (YS)CMIPc MMEd (MME)20
HF_YSYoshimura (YS)CMIP MME (MME)60
HF_YSc1Yoshimura (YS)Cluster1 (C1)60
HF_YSc2Yoshimura (YS)Cluster2 (C2)60
HF_YSc3Yoshimura (YS)Cluster3 (C3)60
HF_ASArakawa-Schubert (AS)CMIP MME (MME)60
HF_ASc1Arakawa-Schubert (AS)Cluster1 (C1)60
HF_ASc2Arakawa-Schubert (AS)Cluster2 (C2)60
HF_ASc3Arakawa-Schubert (AS)Cluster3 (C3)60
HF_KFKain-Fritsch (KF)CMIP MME (MME)60
HF_KFc1Kain-Fritsch (KF)Cluster1 (C1)60
HF_KFc2Kain-Fritsch (KF)Cluster2 (C2)60
HF_KFc3Kain-Fritsch (KF)Cluster3 (C3)60

3.2. Future Climate Simulations

[11] For the future climate simulations, the 20-km model with the YS scheme was integrated for 25 years during 2075–2099 with future SSTs. The future SST change was evaluated by the difference between the 20th Century experiment (20C3M) and future simulation under the Special Report on Emission Scenario (SRES) A1B experiments in the CMIP3 data set [Meehl et al., 2007]. The boundary SST data for the future were prepared by superposing: (1) future change (between 2075–2099 and 1979–2003) in SST projected by CMIP3 multimodel ensemble (MME) mean; (2) the linear trend of SST projected by CMIP3 MME during 2075–2099; and (3) the detrended observed SST for the period 1979–2003. Future sea-ice distribution was obtained in a similar fashion. Details of the method can be found inMizuta et al. [2008]. The prescribed future SST retains the observed year-to-year variability and El Niño–Southern Oscillation (ENSO) events in the present-day (1979–2003), but with a higher global mean and a clear increasing trend in the future climate. This is an acceptable assumption, becauseIPCC [2007a]drew no definitive conclusion regarding the future change in ENSO, based on future projections by AOGCMs. The change in global annual mean SST at the end of the 21st century (2075–2099) relative to the present-day (1979–2003) is +2.17°C.

[12] To evaluate the uncertainty of future projections due to the choice of convection scheme, simulations using the 60-km model for the YS, AS, and KF convection schemes were conducted similarly for the future as well as the present-day.

[13] Although the MME mean of future SST changes projected by CMIP3 models can be considered as one of the best estimates of future SST change, uncertainty of geographical SST distribution especially in the tropics should also be considered. Therefore, three other SST patterns were prepared using a cluster analysis, in which tropical SST anomalies derived from the 18 CMIP3 models were grouped. The detailed procedure for the cluster analysis was as follows: (1) for each CMIP3 model experiment, a mean future change in annual mean SST was computed by subtracting the 1979–2003 mean SST from the 2075–2099 mean SST; (2) the computed mean future change in SST was normalized by the tropical (30°S–30°N) mean future change in SST; (3) the MME mean of the normalized value was subtracted from the normalized value for each model experiment; (4) then, the inter-model pattern correlationr was computed between each pair of models; and (5) norms (or distances) were defined as 2 × (1 − r) for every pair, and the cluster analysis was performed using these norms. A small distance between two models indicates they share similar spatial patterns in future changes of tropical SST. Clustering is based on the single-linkage (or minimum-distance) method [Wilks, 2011], in which the smallest distance between two models (or groups) is joined step-by-step. When the final three groups are bounded, the clustering procedure is terminated. Each cluster consists of five or seven models. Only three cluster groups were identified, because too many groups would increase the risk that the results would be dominated by a single outlier model.

[14] Next, based on the result of the cluster analysis above, we composed three kinds of the future SST change as follows: (6) for each model, the monthly future changes in SST were normalized by annual-mean tropical mean change in SST (30°S–30°N); (7) the normalized data were averaged for each cluster; (8) then, the monthly changes in SST for each cluster were multiplied by a factor in order that the annual-mean tropical SST change becomes the same value as that in the CMIP3 MME mean. Note that the resultant three kinds of the SST change pattern have almost the same values of SST for not only the tropical mean but also the global mean [Murakami et al., 2012a].

[15] Figure 1shows all four prescribed future changes in the annual-mean SST. It is found that SST warming even in the CMIP3 ensemble mean is spatially inhomogeneous (Figure 1a): warming is larger in the Northern Hemisphere than in the Southern Hemisphere, and the largest SST warming occurs in the tropical and subtropical central to eastern Pacific and tropical western Indian Ocean. The three clustered SSTs are shown in Figures 1b–1d. Cluster 1 (Figure 1b) shows relatively small warming over the central Pacific. Cluster 2 (Figure 1c) is similar to the CMIP3 ensemble mean, but intensified warming occurs in the Indian Ocean and the tropical central to eastern Pacific. Cluster 3 (Figure 1d) has the largest spatial variation in the tropics among the prescribed SSTs, and shows the largest warming of SSTs in the tropical western to central Pacific and the subtropical central to eastern Pacific, and the smallest warming of SSTs in the subtropical Atlantic. After all, we conducted 12 ensemble simulations with the 60-km model (i.e., combination of 3 cumulus convection schemes by 4 future SSTs).Table 1 summarizes the specifications of the future climate simulations.

Figure 1.

Four prescribed future changes in the annual-mean SST (°C). (a) CMIP3 ensemble mean. (b) Cluster 1, (c) Cluster 2, and (d) Cluster 3, based on a cluster analysis.

4. Validation Data

[16] To verify the mean and extreme precipitation simulated in the present-day climate, we used the following five observational data sets: (1) Climate Prediction Center Merged Analysis of Precipitation (CMAP) Version 1103_standard [Xie and Arkin, 1997] with a 2.5 degree grid spacing; (2) Global Precipitation Climatology Project (GPCP) Version 2.1 [Huffman et al., 2009] with a 2.5 degree grid spacing; (3) GPCP One-Degree Daily data (GPCP 1DD) Version 1.1 [Huffman et al., 2001] with a 1.0 degree grid spacing; (4) Tropical Rainfall Measuring Mission (TRMM) 3B42 and 3B43 products in version 6 [Huffman et al., 2007] with a 0.25 degree grid spacing; and (5) Asian Precipitation Highly Resolved Observational Data Integration Toward the Evaluation of Water Resources (APHRODITE, hereafter APHRO) Version 1103 [Yatagai et al., 2009] with a TL959 grid spacing, which is prepared by K. Kamiguchi of MRI/JMA using the same algorithm as the original grid spacing (0.25 degree grid).

[17] For verification of the seasonal mean climatology, we employed CMAP, GPCP, APHRO, and TRMM 3B43. For verification of the extreme indices, we used GPCP 1DD, APHRO, and TRMM 3B42 because of their fine temporal resolution. CMAP, GPCP, and APHRO cover the whole period of the present-day climate simulation (1979–2003), while GPCP 1DD covers the 12 years from 1997 to 2008, and TRMM 3B42 and TRMM 3B43 cover the 11 years from 1998 to 2008.

5. Present-Day Climate Simulation

5.1. Mean Precipitation

[18] First, climatological seasonal mean precipitation, as reproduced in the present-day simulations, are compared with observed data.Figure 2shows the simulated June–July–August (JJA) mean precipitation climatology for the 20-km model and for the three 60-km models with the different cumulus schemes. Also shown are two observed data sets. All models successfully reproduced the observed major convection centers over the eastern Arabian Sea, the Bay of Bengal, Bangladesh, the South China Sea, the western Pacific including the Philippine Sea, and southwest Japan. In addition, the models successfully simulated regionally detailed precipitation patterns such as orographic precipitation bands which are observed on the windward side (i.e., the western side) of mountains of the Western Ghats, the Indochina Peninsula, and the Philippines [Xie et al., 2006]. However, the models except the KF scheme, i.e., SP_YS, HP_YS, and HP_AS simulate less precipitation over the northern Bengal Sea and Bangladesh.

Figure 2.

June–July–August (JJA) mean precipitation climatology (mm/day): (a) CMAP, (b) TRMM 3B43, (c) 20-km model with YS cumulus scheme (SP_YS), (d) 60-km model with YS cumulus scheme (HP_YS), (e) 60-km model with AS cumulus scheme (HP_AS), and (f) 60-km model with KF cumulus scheme (HP_KF). The average period for the climatology is 1979–2003, except for TRMM 3B43 (1998–2008).

[19] The model with the AS scheme (HP_AS) produces less precipitation over the western Pacific, and excessive precipitation over the southwestern Bay of Bengal and the eastern Arabian Sea, especially over west of the Western Ghats. Similar biases are commonly found in our previous versions of the 20-km model (i.e., MRI-AGCM3.0S and MRI-AGCM3.1S) which incorporate the AS scheme, indicated byKitoh and Kusunoki [2008] and Mizuta et al. [2012]. In association with the precipitation biases in HP_AS, the monsoon westerlies in the lower troposphere over South Asia flow more southward than those in reanalysis data (not shown).

[20] For the December–January–February (DJF) mean precipitation climatology, all four models reproduce both broad-scale precipitation and topographically concentrated precipitation associated with prevailing northwesterly winds over East Asia and northeasterly winds over Southeast Asia quite well (not shown).

[21] To assess the performances of the model precipitation quantitatively, we employed the “Taylor diagram” [Taylor, 2001], which has been widely used in the evaluation of climate model performance. Using this diagram, the geographical distribution of the seasonal mean precipitation climatology over 60°–160°E, 10°S–50°N was evaluated. Root-mean square difference (RMSD) and area-averaged biases of the precipitation climatology were also calculated because the Taylor diagram is derived from bias-corrected RMSD. The performances of our AGCMs were compared with those of CMIP 3 CGCMs to facilitate multimodel inter-comparison, although the experiment framework is different between AGCMs and CGCMs. Before calculating the statistics, the model data were re-gridded to common grid boxes with 2.5° × 2.5° horizontal resolution in order to compare them with the observed data. To estimate observational uncertainty, we also calculated the statistics between the two observed data sets, i.e., CMAP and GPCP.

[22] For the JJA mean (Figures 3a and 3b), the biases against the reference (i.e., CMAP) are within 0.5 mm/day for all AGCMs, and the RMSDs are approximately 2.0 mm/day for SP_YS, HP_YS, and HP_KF, whose values are much smaller than those of CGCMs. In the Taylor diagram, the geometric distances from the reference to SP_YS, HP_YS, and HP_KF are almost the same distance to GPCP (note that the SP_YS and the HP_YS are nearly overlapped), and they are much smaller than those of CGCMs, indicating their outstanding skill in simulating the JJA mean precipitation. On the other hand, it is found that the performances for HP_AS are lower than those for other AGCMs. For the DJF mean (Figures 3c and 3d), all four AGCMs generally outperform the CGCMs, and differences of the AGCMs' performances are fairly small, in contrast to the case in JJA.

Figure 3.

Skill of precipitation climatology in the present-day climate simulations for 4 AGCMs and 24 CMIP3 CGCMs over 60°E–160°E, 10°S–50°N in (a, b) JJA and (c, d) DJF. The CMAP data is selected for verification data. Figures 3a and 3c show root-mean square difference (RMSD) and bias. Figures 3b and 3d show the Taylor diagram for displaying pattern statistics [Taylor, 2001]. For the 20-km AGCM, the skill is shown by open blue circles. For the 60-km AGCMs, the skill is shown by closed circles colored blue, red, and green for the YS, AS, and KF models, respectively. For the CMIP3 CGCMs, the skill is shown by open black circles. Also shown is the skill of GPCP against CMAP data. The average period for the climatology is 1979–2003, except for the CGCMs (1979–1999). In the Taylor diagram, the radial distance from the origin is proportional to the spatial standard deviation of a simulated pattern normalized by the observed standard deviation. The correlation coefficient between the observed and simulated fields is given by the azimuthal position. The contours show the measure of skill, defined by the equation 4 inTaylor [2001].

5.2. Precipitation Extremes

[23] It is well recognized that changes in the frequency and intensity of extreme events are more likely to affect society rather than are changes in the mean climate. In order to assess the extreme precipitation in the models, three extreme precipitation indices are introduced. The extreme indices employed in this study are based on the core set of 27 extreme indices recommended by the World Meteorological Organization (WMO) [2009]. First one is the simple daily precipitation intensity index (SDII), defined as the annual total precipitation divided by the number of wet days, where a wet day is defined as a day with precipitation greater than or equal to 1 mm. Second one is the maximum 5-day precipitation total (R5d), defined as the annual maximum precipitation total in 5 consecutive days. Third one is the fraction of the annual total precipitation due to very wet days exceeding the 95th percentile in the present climate (R95T). Note that the 95th percentile is fixed at the value identified in the present-day climate. In addition, mean precipitation indices for annual averaged precipitation (Pav), JJA averaged precipitation (Pav_JJA), and DJF averaged precipitation (Pav_DJF) are defined for use in the following analysis. The indices used are listed inTable 2. All of these indices are calculated for each calendar year and then averaged for the entire period of data available. Note that WMO [2009] referred to R5d and R95T as RX5day and P95pTOT/PRCPTOT, respectively.

Table 2. Precipitation Indices Employed in This Study
IndexDefinitionUnit
PavAnnual averaged precipitationmm/day
Pav_JJAJune-August averaged precipitationmm/day
Pav_DJFDecember-February averaged precipitationmm/day
SDIISimple daily precipitation intensity index: total annual precipitation divided by number of wet days (≥1 mm/day)mm/day
R5dMaximum 5-day precipitation total: the annual maximum consecutive 5-day precipitation totalmm
R95TFraction of annual precipitation due to very wet days exceeding the 95th percentile in the present climate%

[24] As an example of the extreme precipitation indices, we show the distribution of R5d in the present climate simulations and two observations (Figure 4). For the observation data, TRMM 3B42 tends to show larger R5d values, although their spatial pattern is generally similar to each other. Here it should be kept in mind that the horizontal resolutions of these data sets are different: GPCP 1DD employs a 1° × 1° grid and TRMM 3B42 a 0.25° × 0.25° grid. Larger R5d values are generally distributed over the tropics, including over the Bay of Bengal, the South China Sea, the Philippine Sea, and the region from Taiwan to southern Japan, as well as over coastal regions where precipitation is enhanced by orography. It is likely that larger R5d values over the South China Sea, the Philippine Sea, and the region from Taiwan to southern Japan are due in part to rainfall associated with tropical cyclones.

Figure 4.

Maximum 5-day precipitation total (R5d; mm): (a) GPCP 1DD, (b) TRMM 3B42, (c) SP_YS, (d) HP_YS, (e) HP_AS, and (f) HP_KF. The average period for the climatology is 1979–2003, except for GPCP 1DD (1997–2008) and TRMM 3B42 (1998–2008).

[25] For the simulations of R5d, the SP_YS, HP_YS, and HP_KF models show a general resemblance to observations, and they are similar to each other, for the large-scale pattern and the orography-induced pattern. The SP_YS model tends to simulate larger R5d values over the tropical western Pacific and its surroundings than the HP_YS, as noted in the next section. On the other hand, the HP_AS model clearly underestimates R5d over tropical and subtropical Asia. It must be noted that all the models underestimate R5d over the equatorial Indian Ocean. This deficiency might be due to an insufficient representation of the eastward propagation of intraseasonal disturbances or Madden–Julian oscillation (MJO) in the models [Mizuta et al., 2012], because Jones et al. [2004] showed higher frequency of precipitation extremes in the Indian Ocean, Indonesia, and the western Pacific during periods of active MJO than quiescent episodes of the oscillation, based on the pentad GPCP data.

5.3. Regional Average Precipitation

[26] To further assess the present-day realizations, we selected nine domains over land-only grids (Figure 5) and then averaged the precipitation indices in the domains. Figure 6shows the regional averages of the precipitation indices in the present-day simulations and observations. Observed mean precipitation data show a large seasonal contrast between JJA and DJF. All the regions (except MTC) have more mean precipitation in the boreal summer than in the boreal winter. PHP, SCH, and JPK receive moderate mean rainfall in the boreal winter, brought by prevailing northeasterly or northwesterly winds. For the mean precipitation indices (i.e., Pav, Pav_JJA, and Pav_DJF), there is little difference among the observed data, and all the models successfully reproduce their spatial distribution. However, there are some exceptions, such as larger bias of Pav_JJA for HP_KF in ICH and PHP, and smaller bias of Pav_JJA for all the models except HP_KF in BGD. APHRO tend to show smaller precipitation throughout the year in MTC than the other observed data.

Figure 5.

Definition of target domains for regional analyses of land-only precipitation. India (IND, 70°E–87°E, 5°N–27°N); Bangladesh (BGD, 87°E–93°E, 20°N–27°N); Indochina (ICH, 93°E–112°E, 6°N–20°N); the Maritime Continent (MTC, 95°E–150°E, 10°S–6°N); the Philippines (PHP, 119°E–130°E, 5°N–20°N); southern China (SCH, 105°E–125°E, 20°N–33°N); northern China (NCH, 105°E–125°E, 33°N–45°N), Japan and Korea (JPK, 125°E–145°E, 30°N–40°N and 138°E–149°E, 40°N–46°N); northwestern China (NWC; 80°E–105°E, 35°N–45°N).

Figure 6.

Regional averages for precipitation indices: (a) Pav_JJA, (b) Pav_DJF, (c) Pav, (d) SDII, (e) R5d, and (f) R95T. The indices are calculated using the observational data: TRMM 3B42 (light gray), GPCP 1DD (dark gray), and APHRO (black), and using the simulated data: SP_YS (open blue bars), HP_YS (blue), HP_AS (red), and HP_KF (green).

[27] For the extreme precipitation indices (e.g., SDII, R5d, and R95T), there are relatively large differences among not only the simulations but among observations. TRMM 3B42 tends to show the highest value for the extreme indices among the three observations. APHRO tends to show the lowest value for SDII and R5d of the observations, but it does not show such a bias for R95T. For the model simulations, the SP_YS, HP_YS, and HP_KF models successfully simulate the extreme indices in terms of magnitude; consequently, their values are generally within those observed. For example, these models reproduce a large R5d value exceeding 200 mm in BGD and PHP. In contrast, the HP_AS model clearly underestimates R5d and R95T in almost all regions.

[28] Comparison between the SP_YS and the HP_YS in terms of the three precipitation indices shows that the SP_YS simulates larger values than the HP_YS for SDII in ICH, PHP, SCH, NCH and JPK, for R5d in MTC, PHP, SCH, NCH and JPK, and for R95T in MTC and PHP, with statistical significance at the 5% level. Here the SP_YS and the HP_YS models are identical except for horizontal resolution. Thus, the model with higher horizontal resolution (i.e., SP_YS) tends to simulate larger heavy precipitation over the regions surrounding the tropical western Pacific. These increases in heavy precipitation for SP_YS would come from more realistic representation of tropical cyclones [Murakami and Sugi, 2010] and a more active rainband of East Asian summer monsoon [Kusunoki et al., 2006, 2011] in the higher resolution model, although some parts of the differences in the extremes would simply come from the data resolution, i.e., spatial averaging of data causes smaller values in extreme precipitation [Kitoh and Kusunoki, 2008].

6. Future Changes

[29] In this section, the projections by the 20-km model and the 60-km models for the end of the 21st century (2075–2099) are compared to those for the present-day (1979–2003). The 60-km model members consist of 12 ensemble projections combining 3 different cumulus convection schemes and 4 different prescribed future SSTs.

6.1. Mean Precipitation

[30] Figure 7shows future changes in mean precipitation for the 20-km model and the 60-km model projections in JJA and DJF, and the annual mean. For the 20-km model, areas with statistically significant change (at the 10% level) are shaded in color. For the 60-km models, the 12 ensemble projections are averaged, and areas with statistically significant change (at the 5% level) are shaded in color.

Figure 7.

Future changes in precipitation indices between the present-day (1979–2003) and the future (2075–2099) for the 20-km model (SP_YS): (a) Pav_JJA, (b) Pav_DJF, and (c) Pav, and for the 60-km model ensemble simulations: (d) Pav_JJA, (e) Pav_DJF, and (f) Pav. The definition of these precipitation indices is shown inTable 2. In Figures 7a–7c, statistically significant grid points (at the 10% level) are shaded in color. In Figures 7d–7e, statistically significant grid points (at the 5% level) are shaded in color, and grid points where all 12 (>=10) experiments show the same sign of changes are closely (widely) hatched based on a common 1.25° latitude/longitude grid.

[31] To test the statistical significance of future change, we employed Student's t-test for analyses of mean values from paired samples [e.g.,Wilks, 2011]. In our AGCM experiments, the prescribed boundary SST for the future simulations (2075–2099, 25 years) was prepared by superposing the assumed increments on the observed monthly SST in the present-day (1979–2003, 25 years), as described insection 3.2. Accordingly, precipitation data obtained from experiments in these two periods are not independent of each other especially in the tropics, when considering their strong relation to SST distribution. Thus, they should be handled as paired samples in applying the t-test to assess differences in mean values. The differences between the corresponding years are examined in an one-samplet-test about the null hypothesis that the mean difference is zero [e.g.,Wilks, 2011]. Thus, we have 300 samples (= 25 years × 12 ensemble projections) for the 60-km models and 25 samples for the 20-km model, testing the null hypothesis that the mean difference is zero.

[32] Figure 7shows an overall agreement between the 20-km model results and the 60-km models' results. For the JJA mean (Figures 7a and 7d), the models generally project an increase in precipitation over the Asian summer monsoon region, and a decrease in precipitation over the Maritime Continent and inland Asia. For the DJF mean (Figures 7b and 7e), they generally project increased precipitation over the northern region including East Asia, Bangladesh, the ocean east of the Philippines, and the southern part of the Maritime Continent. In contrast, they project decreased precipitation over the area between the equator and 20°N in Asia, which coincides with the dry season and is unaffected by midlatitude synoptic disturbances. For the annual mean (Figures 7c and 7f), an increase in precipitation is generally projected over land in Asia. Thus, precipitation tends to be higher in the wet season but lower in the dry-season over tropical Asia, resulting in a greater seasonal contrast than the present-day.

[33] Comparing our results with the projection by the CMIP3 multimodel ensemble in Asia [e.g., IPCC, 2007a, 2007b], our results are generally similar to their ensemble mean changes, but some differences are found on a regional scale. For example, JJA precipitation over the central-to-southern part of the Maritime Continent is projected to decrease significantly in our experiments, but not so changed in the CMIP3 models. For DJF precipitation, our models (particularly the SP_YS) project an increase of precipitation significantly over the region from southern China to southern Japan as well as over the Plain of Hindustan including Bangladesh, while the CMIP3 models do not project an increase over these regions.

[34] Over the west of Western Ghats mountains, our projection indicates a significant reduction in precipitation. Rajendran and Kitoh [2008]reported the same trend over the region, in contrast to an overall increased trend in precipitation over most of the Indian peninsula, based on a projection using the earlier version of the 20-km AGCM [Mizuta et al., 2006]. Krishnan et al. [2012]suggested that a weakening of summer monsoon Hadley-type circulation over the Indian Ocean in response to global warming leads to a weakened large-scale monsoon flow, resulting in weaker vertical velocities and reduced orographic precipitation over the region. In contrast, the CMIP3 models show overall increase in precipitation over the Indian peninsula [e.g.,IPCC, 2007b; Turner and Annamalai, 2012]. The AGCMs employed in this study have much higher horizontal resolution than the CMIP3 AOGCMs, whose grid interval is of approximately 100 km to 400 km. This could allow the AGCMs to simulate the phenomena closely related to local topography and winds more realistically, and to provide better projection in this respect than the CMIP3 models.

6.2. Precipitation Extremes

[35] Figure 8shows future changes in the extreme precipitation indices for the 20-km model and the 60-km models projections. They are calculated for each year, and the statistical significances of their changes are assessed using annual values. For the future changes, an overall agreement is found between the 20-km model results and the 60-km models. Extreme precipitation is projected to generally increase in Asia. Compared with the case for mean precipitation (Pav), areas with significant positive changes are larger, particularly over land. Over most regions northward of approximately 20°N, including East Asia, Bangladesh, and northwestern China, the increases of the extremes are highly consistent among the 60-km ensemble projections. In contrast, over the tropics south of approximately 20°N, including South Asia and Southeast Asia, there is some uncertainty of the increase between the projections, although the uncertainty is relatively small over land in Southeast Asia.

Figure 8.

As for Figure 7, but for (a, d) SDII, (b, e) R5d, and (c, f) R95T. The definition of these indices is shown in Table 2.

[36] There are several studies of extreme precipitation changes based on the CMIP3 multimodel projections. Meehl et al. [2005] found a general increase in SDII in annual average over the tropics, northern Asia and the east coast of Asia. Similarly, Kusunoki and Arakawa [2012] indicated an increase in SDII in the wet season (June–July) over almost all of East Asia, and Li et al. [2011] showed an increase in extreme precipitation in July–August over land of China. On the other hand, Turner and Slingo [2009] shows that for extreme precipitation changes in Indian summer there is considerable diversity in the CMIP3 model projections, although the changes are generally positive. Thus, the general features found in our AGCM ensemble projections are consistent with those in the CMIP3 models.

[37] However, parts of our projections are different from those of the CMIP3 models. Looking at our projections carefully, there are decreased trends in the extremes over the area around the South China Sea, where substantial decreases in tropical cyclone existence are consistently projected by our 20-km and 60-km AGCM ensemble simulations [Murakami et al., 2012a, 2012b], as well as over the coastal region west of the West Garths. According to observational data analysis, Endo et al. [2009] reported that extreme precipitation decreased over stations in northern Vietnam from the 1950s to 2000s. Wu et al. [2005] showed a reduced frequency of tropical cyclones over the South China Sea in recent decades (June to October 1965–2003). It is interesting to note that these observed trends are qualitatively consistent with the trends projected by our models.

6.3. Regional Average Precipitation

[38] Figure 9 summarizes the future changes in the precipitation indices averaged over specific domains defined in Figure 5. The mean precipitation indices (Pav, Pav_JJA, and Pav_DJF) are generally projected to increase in NWC, NCH, JPK, SCH, and BGD, with agreement between the projections, particularly in the wet season. In contrast, the trends substantially differ among the projections in PHP, ICH, MTC, and especially IND, which are located in South Asia and Southeast Asia, although they show an increasing trend generally.

Figure 9.

Future changes (%) of precipitation indices for nine regional averages: (a) NCH, (b) JPK, (c) SCH, (d) PHP, (e) NWC, (f) BGD, (g) IND, (h) ICH, and (i) MTC. The future changes are shown as the percentage change ((F − P)/P) from the present-day (P: 1979–2003) to the future (F: 2075–2099). The 20-km model is indicated by a cross. The 60-km models are indicated by color line graphs and their averages are denoted by horizontal black lines. Models with different cumulus schemes are indicated by different colors: i.e., blue for YS, red for AS, and green for KF. CMIP3 MME, Cluster1, Cluster2, and Cluster3, which are the SST pattern prescribed in the future climate simulations, are denoted as MME, C1, C2, and C3, respectively. Note that the percentage change shows marked variation in the case of small denominators such as Pav_DJF, and that the vertical scale for Figure 9i (MTC) is different from that for the others.

Figure 9.

(continued)

[39] Extreme precipitation indices (i.e., SDII, R5d, and R95T) are projected to generally increase more than mean precipitation (Pav), although there is large diversity among the changes in the domains in South Asia and Southeast Asia. In particular, R95T shows a strongly positive trend on the ensemble mean, and shows a relatively larger positive trend in Cluster3 SST-forced experiments in the domains (i.e., ICH and PHP) near the western Pacific, probably due to the warmer SST over the western Pacific used in the experiments.

[40] Substantial increase rates are projected in BGD and SCH both in mean and extreme, with high agreement among the experiments. In BGD, for example, the ratios for Pav are 13.5–30.1% (average, 19.7%), and those for R5d are 11.4–25.6% (average, 18.6%). The large increase over these regions are previously pointed out by Kamiguchi et al. [2006], based on a global warming experiment using an earlier version of the 20-km AGCM. Our results based on ensemble simulations could give more confidence to their findings. In BGD, considering the projected large precipitation increase and the susceptibility to flooding over the region, the risk of floods will increase significantly.

[41] Dynamical down-scalings using RCMs have been performed to project mean and extreme precipitation changes.Kumar et al. [2011]reported that summer (June to September (JJAS)) precipitation is consistently increased in the all-India domain, but on a smaller regional scale some regions experience slightly decreased precipitation, based on three projections using a 50-km-mesh RCM nested by perturbed physics ensemble simulation of Hadley Centre Couple Model (HadCM3). On the other hand,Dobler and Ahrens [2011] projected a decrease in mean precipitation in Indian summer (JJAS) in contrast to an increase in precipitation intensity, based on RCM experiments for several emission scenarios. Similarly, Ashfaq et al. [2009] obtained a decreased trend of Indian JJAS precipitation based on ensemble projection (five member) employing a RCM. These large variations in the precipitation change in South Asian summer are also found among the CMIP3 AOGCM projections [Turner and Slingo, 2009] as well as our multiphysics and multiSST ensemble simulations.

7. Uncertainty From Different Future SST Patterns and Cumulus Schemes

[42] In South Asia and Southeast Asia, mean and extreme precipitation generally increase, but their changes show marked differences among the projections, suggesting some uncertainty in their changes over these regions. To clarify which factor (i.e., different SST changes or different cumulus parameterization schemes) contributes more to this uncertainty, analysis of variance (ANOVA) without replication [Storch and Zwiers, 1999] was applied to the 12 ensemble projections by the 60-km models. This statistical analysis decomposes the total variance of the 12 ensemble projections into three components: the variance due to the different SST changes, the variance due to the different cumulus convection schemes, and residual one, assuming that the effects from the two factors are mutually independent. TheF test [Storch and Zwiers, 1999] is used to assess the effect of each factor on variations among the ensemble projections statistically.

[43] Figure 10summarizes the result of the ANOVA based on the 60-km model ensemble projections, showing the proportion of the total variance. In this figure, filled bars indicate that the variances are affected significantly at the 5% level by the use of different SST changes (blue bars) or different cumulus schemes (red bars). For the mean precipitation indices (i.e., Pav, Pav_JJA, and Pav_DJF), variances arising from the use of different cumulus convection schemes (red bars) are generally larger than those arising from the use of different SST changes (blue bars) in most of the regions. Thus, the uncertainty in the projected changes is derived mainly from differences in the cumulus schemes employed in South Asia and Southeast Asia. However, for MTC the uncertainty arises mainly from the different SST changes, and for NWC and East Asia the uncertainty in Pav_DJF also arises mainly from the different SST changes. For the extreme precipitation indices (i.e., SDII, R5d, and R95T), their uncertainty is more strongly influenced by differences in the cumulus schemes than those for the mean precipitation indices.

Figure 10.

Proportion of total variance (%) that each factor accounts for, calculated by two-way analysis of variance (ANOVA) without replication, based on the percentage changes of each variable ((F − P)/P). Here, P (F) is an averaged value for 1979–2003 (2075–2099). Blue: prescribed SST changes; red: cumulus convection schemes; green: residual. (a) Pav_JJA, (b) Pav_DJF, (c) Pav, (d) SDII, (e) R5d, and (f) R95T. Filled bars indicate results that are significantly affected by the factor (SST changes or cumulus schemes) at the 5% level (F test).

[44] Figures 11a and 11bshows the future changes in Pav and SDII on the global scale, based on the 60-km model ensemble projections. Their future changes show marked differences among the projections in parts of the tropics and subtropics, with a slightly positive (negative) changes over the tropics (subtropics). Over land in South Asia and Southeast Asia, slightly positive changes are found.Figures 11c and 11d show the relative contribution ratio of variances among the ensemble projections. The ratio is defined as the variance resulting from the different cumulus schemes relative to the variance resulting from the different SSTs, based on the percentage changes. The grids shaded in blue (red) in the panels correspond to areas where the effect of the SST patterns (the cumulus schemes) is dominant. For Pav, differences in the SST changes make a larger contribution to the uncertainty in the projected changes over the tropical Pacific, including the Maritime Continent and parts of the tropical Atlantic. Over other tropical regions, including South Asia and northern Southeast Asia, differences in the cumulus schemes contribute more to the uncertainty. For SDII, differences in the cumulus schemes contribute more to the uncertainty over most of tropics except in the western and central tropical Pacific. These results suggest that the uncertainty in projected mean and extreme precipitation changes over South Asia and Southeast Asia is derived mainly from differences in the cumulus schemes, with an exception of the Maritime Continent where the uncertainty is derived mainly from differences in the SST change patterns.

Figure 11.

Future changes (F – P; mm/day) in (a) Pav and (b) SDII in the 60-km model ensemble simulations. Grid points where the future changes are statistically significant (at the 5% level) are shaded, and grid points where at least 10 experiments show the same sign of change are hatched. Relative contribution ratio (%) of variances among projections for (c) Pav and (d) SDII. The ratio is defined as the variance resulting from the different cumulus schemes relative to the variance resulting from the different SSTs, based on the percentage changes ((F – P)/P). The grid points shaded in blue (red) correspond to sites where the effect of the SST patterns (the cumulus schemes) is dominant. Note that unshaded grid points are not significantly affected by either factor (at the 1% level).

[45] Murakami et al. [2012a]revealed from the analysis of the same experiment data set that the differences in the SST spatial patterns cause more uncertainty in the future changes of tropical cyclone (TC) genesis frequency and TC numbers on ocean-basin scales compared with differences in the cumulus schemes. However, an exception is found over the northern Indian Ocean, where TC activity is affected mainly by the cumulus schemes [Murakami et al., 2012a, Figure 9], which is consistent with our results regarding precipitation changes.

8. Conclusion and Discussions

[46] This study focuses on projecting future changes in mean and extreme precipitation in Asia, and discusses their uncertainties. Time-slice experiments using a 20-km-mesh MRI-AGCM were performed both in the present-day (1979–2003) and the future (2075–2099). To assess the uncertainty of the projections, 12 ensemble projections (i.e., combinations of 3 different cumulus schemes and 4 different prescribed future SSTs) were conducted using 60-km-mesh MRI-AGCMs. For the present-day simulations, the models successfully reproduced the pattern and amount of mean and extreme precipitation, although the model with the Arakawa–Schubert (AS) cumulus scheme underestimated the amount of extreme precipitation.

[47] For the future climate simulations, an overall agreement was found between the 20-km and 60-km model projections. In South Asia and Southeast Asia, mean and extreme precipitation generally increase, but their changes show marked differences among the projections, suggesting these trends include some uncertainty over these regions. In East Asia, northwestern China, and Bangladesh, in contrast, mean and extreme precipitation show consistent increases among the projections, suggesting their increases are reliable for this model framework over these regions.

[48] To clarify which factor of the use of different SST pattern changes or different cumulus parameterization schemes contributes more to this projected uncertainty, analysis of variance (ANOVA) without replication was applied to the 12 ensemble projections by the 60-km models. The results reveal that the uncertainty in the precipitation changes in South Asia and Southeast Asia are derived mainly from differences in the cumulus schemes, with an exception in the Maritime Continent where the uncertainty originates mainly from differences in the SST change patterns. In particular, a large portion of the spreads of the extreme precipitation changes in Asia is originated more from the cumulus scheme differences.

[49] Our results suggest that mean and extreme precipitation changes in the Asian summer monsoon region are more sensitive to the choice of cumulus convection scheme than to the prescribed SST change pattern, despite the fact that there is large uncertainty in the SST change pattern over the tropical Pacific and this would bring large differences in tropical precipitation change [Yamaguchi and Noda, 2006; Xie et al., 2010; Ose and Arakawa, 2011]. A possible reason for this is that the prescribed four SST patterns are not so different over the Indian Ocean (Figure 1). As a matter of fact, inter-model spread of the SST change pattern over the Indian Ocean, based on 18 CMIP3 models, is much smaller than that over the Pacific Ocean and the Atlantic (not shown). Our findings are basically consistent withTurner and Slingo [2009], who suggested that differences in the cumulus parameterization employed may explain large diversity in extreme precipitation change in the Indian monsoon region, based on an examination of the CMIP3 model projections.

[50] Figure 12shows JJA precipitation changes in all twelve projections with the 60-km models. It is clearly found that the model with the KF scheme projects consistently suppressed precipitation in South Asia in contrast to enhanced precipitation over the equatorial Indian Ocean. Such systematic changes of the KF model are responsible for the large diversity in the precipitation changes over South Asia in our experiment framework. Can we find any systematic biases of the KF model in the present-day simulations? Looking at the present-day simulations carefully, only the KF model produces locally excessive precipitation over the equatorial central Indian Ocean (Figure 2). This bias may give an indication of the anomalous response of the KF model in a warmer climate. Nevertheless, the KF model shows quite high skill comparable to the YS model for reproducing precipitation climatology in the present-day over the South Asian domain (60°–100°E, 0°–30°N; not shown) as well as the Asian monsoon domain (60°–160°E, 10°S–50°N;Figures 2, 3, and 4) in terms of RMSD and the Taylor diagram. Therefore, it should be emphasized that a model incorporating different types of cumulus scheme could provide quite different projections on a regional scale, even though the model's present-day simulation results seem similar to one another.

Figure 12.

Future changes in JJA precipitation (mm/day) by the 60-km model ensemble projections.

[51] The latest model projections still have considerable variations for future changes in South Asian summer precipitation both for AOGCMs [Turner and Annamalai, 2012] and RCMs [e.g., Ashfaq et al., 2009; Dobler and Ahrens, 2011; Kumar et al., 2011]. This large uncertainty in the future projection may be largely responsible for differences in the cumulus parameterization scheme incorporated into a climate model. To reduce this uncertainty, it would be necessary to examine cumulus schemes carefully in the present-day climatology and their response to global warming.

[52] Last, we should refer to the AGCM experiment weaknesses. Our study is founded on AGCM time-slice ensemble experiments with multiphysics and multiSST. This experiment framework has advantages for being able to separate the projected uncertainty into two origins: model difference and prescribed SST difference, and to obtain realistic precipitation climatology and variability owing to constraint of the prescribed realistic SSTs. On the other hand, there are some criticisms for AGCM time-slice experiments applied for a study in Asian monsoon region due to their lack of ocean-atmosphere coupling [Kitoh and Arakawa, 1999; Douville, 2005]. To overcome these weak points, we plan to develop a high-resolution AGCM coupled with an ocean model which exchanges heat, momentum, and freshwater with the atmosphere, but tends to keep a specified SST by flux adjustment or relaxation methods. This new type of CGCM would provide more reliable information for Asian precipitation change in a warmer climate.

Acknowledgments

[53] This work was conducted under the framework of the “Projection of the Change in Future Weather Extremes using Super-High-Resolution Atmospheric Models” supported by the KAKUSHIN Program and the “Program for Risk Information on Climate Change” of the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan. The calculations were performed on the Earth Simulator. We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP's Working Group on Coupled Modeling (WGCM) for their roles in making available the WCRP CMIP3 multimodel data set. Support of this data set is provided by the Office of Science, U.S. Department of Energy. We are grateful to the two anonymous reviewers and the editor, S. C. Pryor, for their comments, which greatly improved the manuscript.

Ancillary