Journal of Geophysical Research: Atmospheres

Future projections of heat waves around Japan simulated by CMIP3 and high-resolution Meteorological Research Institute atmospheric climate models

Authors

  • Masuo Nakano,

    Corresponding author
    1. Japan Agency for Marine-Earth Science and Technology, Yokohama, Kanagawa, Japan
    • Corresponding author: M. Nakano, Japan Agency for Marine-Earth Science and Technology, 3173–25 Showa-machi, Kanazawa-ku, Yokohama, Kanagawa, 236–0001, Japan. (masuo@jamstec.go.jp)

    Search for more papers by this author
  • Mio Matsueda,

    1. Clarendon Laboratory, University of Oxford, Oxford, UK
    2. Meteorological Research Institute, Tsukuba, Ibaraki, Japan
    Search for more papers by this author
  • Masato Sugi

    1. Japan Agency for Marine-Earth Science and Technology, Yokohama, Kanagawa, Japan
    2. Meteorological Research Institute, Tsukuba, Ibaraki, Japan
    Search for more papers by this author

Abstract

[1] Future changes in heat wave characteristics around Japan are investigated using Coupled Model Intercomparison Project Phase 3 (CMIP3) model outputs, and high-resolution present-day (1979–2003) and future (2075–2099) climate simulations under the Intergovernmental Panel on Climate Change (IPCC) Special Report on Emissions Scenarios (SRES) A1B emission scenario. The high-resolution simulations are conducted using 20, 60, and 180 km mesh atmospheric general circulation models (AGCMs), as well as a 5 km mesh regional climate model (RCM) nested within the 20 km mesh AGCM. The CMIP3 models project that the frequency of heat wave days (HWF) will increase 22.8, 22.3, and 26.5 d yr–1 in northern, eastern, and western Japan, respectively. The multimodel ensemble spread of future changes in HWF averaged around Japan is large (4–58 d yr–1). The spread is affected by the climate sensitivity of the models and the simulated magnitude of the interannual variation of the daily maximum temperature in present-day climate. In the atmospheric model simulations, the 5 and 20 km mesh models can qualitatively simulate observed HWF distributions, which are affected by steep backbone mountain ranges in Japan. The 5 and 20 km mesh models project large (>30 d yr–1) increases in HWF in the coastal areas of Japan. The duration of heat wave days is projected to increase in areas with increasing HWF.

1 Introduction

[2] Heat waves, defined as prolonged periods of abnormally hot weather, have substantial negative impacts on human health, sometimes resulting in outbreaks of disease or death [Basu and Samet, 2002]. The extraordinary 2010 summer heat wave in Japan was the worst on record in this respect, resulting in more than 1700 deaths. Nakai et al. [1999] reported that the annual number of heat-related deaths in Japan depends exponentially on the annual number of days with maximum temperature (Tmax) exceeding 32°C. In addition to these negative impacts on human health, heat waves have substantial socioeconomic consequences due to increased crop prices (e.g., rice and lettuce).

[3] Observations show that global average surface temperature is rising, and climate models project a rise in mean summer temperature of approximately 3 K over eastern Asia by the end of the 21st century [Intergovernmental Panel on Climate Change, 2007]. The number of extremely hot days (Tmax > 35°C) and the number of tropical nights have both increased significantly in Japan over the past 80 years [Japan Meteorological Agency, 2010]. Future climate projections suggest that heat waves in Europe and North America will become more intense, more frequent, and more persistent in the 21st century [Meehl and Tebaldi, 2004; Fischer and Schär, 2010]. It is therefore important to accurately project future changes in the characteristics of heat waves in and around Japan.

[4] Mizuta et al. [2005] examined projected changes in the Heat Wave Duration Index (HWDI) for the whole of Japan using a 20 km mesh atmospheric general circulation model developed at the Meteorological Research Institute of Japan (MRI-AGCM3.0; [Mizuta et al., 2005, 2006]). HWDI was defined by Frich et al. [2002] as the maximum number of consecutive days with Tmax exceeding the long-term climatology of Tmax by 5 K or more. HWDI is widely used to study heat waves; however, it is also important to account for local interannual Tmax variability, because the implications of Tmax values of 5 K above normal vary with location. Koo et al. [2009] investigated future changes in the total number of heat wave days over the Korean Peninsula using a 27 km mesh regional climate model (RCM), where heat wave days are defined as those for which Tmax exceeds the present-day 95th percentile of local Tmax. Persistence of hot days should be considered alongside total number of days, however, since longer-lasting heat waves have greater impacts on human health [Trigo et al., 2009]. Nighttime temperature also has great impact on human health [Karl and Knight, 1997]. Quantifying uncertainty in future projections is also important for policy making on climate change [Reilly et al., 2001]. Neither of these studies addressed uncertainty in future projections because they did not perform ensemble experiments.

[5] In this study, we investigate future changes in heat wave characteristics using a heat wave definition proposed by Fischer and Schär [2010], which considers both the magnitude of interannual Tmax variability and the persistence of heat waves. First, future changes in heat wave characteristics around Japan are analyzed using the World Climate Research Programme (WCRP) Coupled Model Intercomparison Project Phase 3 (CMIP3) dataset [Meehl et al., 2007]. Second, to obtain high-spatial-resolution information on future changes in heat wave characteristics, present-day and future climate simulations are conducted using 20, 60, and 180 km mesh AGCMs and a 5 km mesh cloud-system-resolving RCM nested within the 20 km mesh AGCM. To evaluate the uncertainties in our future projections of heat wave characteristics that arise from internal variability in the model climate, we conduct multimember initial-value ensemble simulations using the 60 and 180 km mesh AGCMs. The remainder of this manuscript is organized as follows. Section 2 provides a definition of heat wave and describes the CMIP3 model outputs analyzed here and the model experiments performed in this study. Section 3 considers the model performance in reproducing heat wave characteristics in the present-day climate and outlines the future changes projected by CMIP3 models and by high-resolution model simulations. Finally, a summary and discussion are provided in section 4.

2 Model and Method

2.1 Model Experiments and Observation Data

[6] The 18 CMIP3 models analyzed in this study are listed in Table 1. The present-day and future climates are based on daily Tmax data of the 20th Century Climate in Coupled Models (20C3M) experiments (1980–1999) and of the IPCC Special Report on Emission Scenarios (SRES) A1B experiments (2081–2100), respectively. The initial-value ensemble experiments are conducted using most of the CMIP3 models. The results of all available initial-value ensemble experiments are used for analysis.

Table 1. Details of the CMIP3 Models Used for Analysis
LabelModel NameInstitute (Nation)Resolution of Atmospheric ModelNumber of Ensemble Members
20C3MSRES-A1B
  1. “TLx” and “Tx” refer to a horizontal truncation at global wavenumber x, using a triangular spectral truncation based on a linear Gaussian and Gaussian grid, respectively. “i × j” refers to the horizontal resolution in the meridional (i) and zonal (j) directions of grid models (units: degree). Values in parentheses indicate the horizontal resolution in the zonal direction at the equator. “Ly” denotes the use of y vertical levels.

aBCCR-BCM 2.0Bjerknes Centre for Climate Research (Norway)TL63 (270 km), L3111
bCGCM3.1(T47)Canadian Centre for Climate Modelling & Analysis (Canada)TL47 (360 km), L3153
cCGCM3.1 (T63)TL63 (270 km), L3111
dCNRM-CM3Météo-France/Centre National de Recherches Météorologiques (France)TL63 (270 km), L4511
eCSIRO-Mk3.0CSIRO Atmospheric Research (Australia)T63 (180 km), L1811
fCSIRO-Mk3.5T63 (180 km), L1831
gGFDL-CM2.0US Dept. of Commerce/NOAA/Geophysical Fluid Dynamics Laboratory (USA)2.5 × 2.0 (280 km), L2411
hGFDL-CM2.12.5 × 2.0 (280 km), L2411
iGISS-AOMNASA/Goddard Institute for Space Studies (USA)4 × 3 (440 km), L1211
jGISS-MODEL-ER5 × 4 (550 km), L2011
kINGV-ECHAM4Instituto Nazionale di Geofisica e Vulcanologia (Italy)T106 (120 km), L1911
lIPSL-CM4Institut Pierre Simon Laplace (France)3.75 × 2.5 (420 km), L1921
mMIROC3.2 (hires)Center for Climate System Research (The University of Tokyo), National Institute for Environmental Studies, and Frontier Research Center for Global Change (JAMSTEC) (Japan)T106 (120 km), L5611
nMIROC3.2 (medres)T42 (270 km), L2033
oMIUB-ECHO-GMeteorological Institute University of Bonn (Germany)T30 (380 km), L1933
pMPI-ECHAM5Max-Planck-Institut für Meteorologie (Germany)T63 (180 km), L3222
qMRI-CGCM2.3.2aMeteorological Research Institute (Japan)T42 (270 km), L3055
rNCAR-PCM1National Center for Atmospheric Research (USA)T42 (270 km), L1821

[7] The AGCM used for the climate simulations in this study, MRI-AGCM3.2 [Mizuta et al., 2012], is the atmospheric component of an Earth system model, MRI-ESM1 [Yukimoto et al., 2011]. This model has been developed at MRI in Japan. MRI-AGCM3.2 is based on spectral medium-range numerical weather prediction (NWP) models developed for operational use at the Japan Meteorological Agency (JMA). AGCM simulations were performed at three different resolutions: TL959L64, TL319L64, and TL95L64. Here “TLx” refers to a horizontal truncation at global wavenumber x (using a triangular spectral truncation based on a linear Gaussian grid) and “Ly” denotes the use of y vertical levels. The spectral truncations for the three model resolutions listed above correspond to horizontal spatial truncation scales of 20 km (TL959, hereafter AGCM-20km), 60 km (TL319, hereafter AGCM-60km), and 180 km (TL95, hereafter AGCM-180km). The 20 km truncation scale is typical of state-of-the-art global NWP models, whereas the 180 km truncation scale is typical of state-of-the-art global climate models. The model top is located at 0.1 hPa.

[8] Model simulations of the present-day (1979–2003) climate were conducted using observed interannually varying sea surface temperatures (SST) and sea ice concentrations (SIC) from the HadISST1 dataset [Rayner et al., 2003] as the lower boundary conditions. For the AGCM simulations of future (2075–2099) climate, SST and SIC are estimated by the multimodel ensemble mean projected by the CMIP3 models, to which the detrended interannual variations observed in HadISST1 have been added [Mizuta et al., 2008]. Future concentrations of greenhouse gases follow the IPCC SRES A1B emissions scenario.

[9] This study also uses a nonhydrostatic cloud-system-resolving RCM (NHM-5km) developed at MRI [Nakano et al., 2012] to obtain high-spatial-resolution information on future changes in heat wave characteristics. The model is based on the JMA's operational mesoscale model for short-term NWP [Saito et al., 2007]. Simulations were performed at a horizontal resolution of 5 km, with 50 vertical levels and the model top located at 22 km altitude. The model domain was centered at 33.0°N, 132.2°E, with a horizontal domain size of 669 × 594 grid points. This domain covers eastern Asia and includes the Japanese Islands and the Korean Peninsula. NHM-5km was driven by 6-hourly output from the AGCM-20km simulations and was initiated at 0000 UTC 17 May and integrated continuously until 24 UTC 31 October (a period of 168 days) for each year during the simulated time periods (1979–2003 and 2075–2099). It is well known that a phase gap occurs in synoptic-scale atmospheric states between the inner and outer models when the integration period is longer than several days, and in the case that the horizontal extent of the inner model domain is larger than several thousand kilometers [Kida et al., 1991; Yasunaga et al., 2005]. To avoid the development of a phase gap, the spectral nudging method is applied above a height of 7 km for large-scale wave components (wavelength >1000 km) of horizontal momentums and potential temperature.

[10] Four-member initial-value ensemble simulations were conducted for AGCM-180km and AGCM-60km to estimate the degree of uncertainty in future projections of heat waves. Ensemble simulations were not conducted for NHM-5km and AGCM-20km because of insufficient computing resources.

[11] Observations of Tmax from the Automated Meteorological Data Acquisition System (AMeDAS) administered by the JMA are used for comparison with the present-day simulations. The observational data represent a selection of 469 AMeDAS stations for which Tmax is available throughout the period 1979–2003.

2.2 Heat Wave Definition

[12] Our analysis of multiday heat waves follows that conducted by Fischer and Schär [2010]. First, the 90th percentile of Tmax (Tmax90) in the late 20th century is calculated for each calendar day from June through August at every grid point for each model resolution using a centered 15-day-long time window. For example, Tmax from 25 May to 8 June in 1980–1999 (CMIP3 models) and 1979–2003 (MRI models) is used to calculate the local Tmax90 for 1 June for the CMIP3 and MRI models, respectively. A heat wave event is defined as a spell of at least six consecutive days with Tmax exceeding the local Tmax90 according to the present-day climate simulations. Our analysis considers both the number of days meeting this heat wave criterion (frequency of heat wave days, HWF) and the mean duration of heat wave events (heat wave duration, HWD). Quantitative analyses of heat wave characteristics are performed in three subregions of Japan; northern Japan (NJ), eastern Japan (EJ), and western Japan (WJ) (see Figure 1a). The location of the districts and cities that appear in the manuscript is shown in Figure 1b.

Figure 1.

(a) Three subregions of Japan for which quantitative analysis was carried out. (b) The location of districts and megacities.

3 Results

3.1 Present-Day Climate Simulations by CMIP3 Models

[13] Figures 2a–2r show the model topography for the area surrounding Japan used in each CMIP3 model. Most of the CMIP3 models have horizontal resolutions coarser than 200 km. The highest-resolution models, INGV-ECHAM4 and MIROC3.2 (hires), have a horizontal resolution of 120 km (Table 1). The actual topography of the Japanese Islands (Figure 2w) shows backbone mountain ranges that divide the islands into the Japan Sea side and the Pacific Ocean side. Mountain ranges in central Japan are higher than 3000 m above sea level. It is clear that the CMIP3 models are unable to represent such steep mountain ranges.

Figure 2.

Model topography for (a–r) the CMIP3 models, (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, (v) NHM-5km, and (w) observed topography derived from GTOPO30 data compiled by the US Geological Survey (USGS). The labels a–r correspond with those in Table 1.

[14] Figure 3 shows the seasonal mean (June–August) percentiles of Tmax averaged over three subregions of Japan. Observed seasonal mean Tmax90 values (w) are 300.6, 303.7, and 304.9 K in NJ, EJ, and WJ, respectively. In spite of low spatial resolution, some CMIP3 models well simulate (simulation error is less than 2 K) seasonal mean percentiles of Tmax. CGCM3.1 (T47 and T63) (b and c) and CSIRO-Mk3.5 (f) well simulate observed seasonal mean Tmax90 in all the subregions. In addition, INGV-ECHAM4 (k), MIROC3.2 (hires and medres) (m and n), and MPI-ECHAM5 (p) well simulate observed seasonal mean Tmax90 in NJ. MIROC3.2 (hires) (m) and MRI-CGCM2.3.2a (q) well simulate seasonal mean Tmax90 in EJ. MIROC3.2 (medres) (n) and MPI-ECHAM5 (p) well simulate observed seasonal mean Tmax90 also in WJ. Also relatively small errors are found in MPI-ECHAM5 (p) in EJ and MIROC3.2 (hires) (m) and MRI-CGCM2.3.2a (q) in WJ. These errors are, however, a bit larger than 2 K. The other models underestimate seasonal mean Tmax90. The difference between seasonal mean Tmax90 and 10th percentile of Tmax (Tmax90-10) indicates the magnitude of interannual variation. Observed Tmax90-10 values (w) are 10.3, 8.7, and 7.1 K in NJ, EJ, and WJ respectively. CGCM3.1(T47) (b), CGCM3.1(T63) (c), and MPI-ECHAM5 (p) well simulate (error is less than 20%) Tmax90-10 in all the subregions. BCCR-BCM2.0 (a) well simulates Tmax90-10 in NJ and WJ. CSIRO-Mk3.5 (f), GFDL-CM2.0 (g), IPSL-CM4 (l), MRI-CGCM2.3.2a (q), and NCAR-PCM1 (r) well simulate Tmax90-10 in NJ. MIROC (medres) (n) well simulates Tmax90-10 in EJ and WJ. GISS-MODEL-ER (j) and MIUB-ECHO-G (o) well simulate Tmax90-10 in WJ. The other models underestimate Tmax90-10.

Figure 3.

Box-and-whisker plots of the seasonal mean (June–August) percentiles of daily maximum temperature averaged over northern Japan (NJ; top), eastern Japan (EJ; middle), and western Japan (WJ; bottom) for (a–r) the 20C3M (1980–1999; blue color) and SRES A1B experiments (2081–2100; red color) with the CMIP3 models and the present-day (1979–2003; blue color) and future (2075–2099; red color) climate simulations using (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, and (v) NHM-5km. (w) The percentiles derived from AMeDAS observations (1979–2003) are also shown. The horizontal line indicates the median. The box covers 10th–90th percentiles. The edges of the whiskers indicate the 1st percentile and 99th percentile values. The light blue hatch shows the observed range of the 10th–90th percentiles.

[15] The observed HWF (Figure 4w) from 1979 to 2003 is generally high in western Japan (>1.44 d yr–1) and low in northern and eastern Japan (<0.96 d yr–1). The observed HWF is high (>1.44 d yr–1) in northern Hokkaido, in Tohoku on the Japan Sea side, in Kanto, and in western Japan on the Pacific Ocean side. The observed HWF is low (<0.48 d yr–1) in southern Hokkaido, in Tohoku on the Pacific Ocean side, and in western Japan on the Japan Sea side.

Figure 4.

Frequency of heat wave days around Japan (a–r) simulated in 20C3M experiments (1980–1999) by the CMIP3 models, and those according to present-day (1979–2003) climate simulations using (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, and (v) NHM-5km. (w) Frequency of heat wave days derived from AMeDAS observations (1979–2003). Results are shown for the ensemble mean of the CMIP3 models and the 180 and 60 km mesh AGCM ensembles. The labels a–r correspond with those in Table 1.

[16] The CMIP3 models are, however, unable to simulate local features in the observed distribution of HWF (e.g., higher frequencies in Tohoku on the Japan Sea side than in Tohoku on the Pacific Ocean side). Some CMIP3 models are able to simulate general features of the observed spatial distribution (Figures 4a–4r). For example, although GFDL-CM2.0 (Figure 4g) and GFDL-CM2.1 (Figure 4h) overestimate HWF in eastern and western Japan, the models simulate the general features of the observed spatial distribution (i.e., a higher HWF in western Japan and a lower HWF in Hokkaido). CSIRO-mk3.0 (Figure 4e) performs well in simulating a higher HWF in western Japan but overestimates HWF in Hokkaido. These results indicate that the CMIP3 models have the potential to simulate the areal average of HWF.

[17] Figure 5 shows HWF and HWD averaged over the three subregions of Japan (see Figure 1a). The observed values of HWF are 0.7 (NJ), 0.9 (EJ), and 1.4 (WJ). In NJ, BCCR-BCM2.0 (a), CGCM3.1(T47) (b), CNRM-CM3 (d), and MRI-CGCM2.3.2a (q) quantitatively well simulate the observed HWF (error is less than 20%). In EJ, CSIRO-Mk3.0 (e), CSIRO-Mk3.5 (f), and MIROC3.2 (hires) (m) quantitatively well simulate the observed HWF. In WJ, CSIRO-Mk3.0 (e), MIROC3.2 (medres) (n), and MPI-ECHAM5 (p) quantitatively well simulate the observed HWF. The observed values of HWD are 0.6 (NJ), 0.8 (EJ), and 1.0 (WJ) d event–1. In NJ, BCCR-BCM2.0 (a), CNRN-CM3 (d), CSIRO-Mk3.0 (e), and MIROC3.2 (medres) (n) quantitatively well simulate the observed HWD. In EJ, GISS-AOM (i) quantitatively well simulates the observed HWD. In WJ, NCAR-PCM1 (r) quantitatively well simulates the observed HWD.

Figure 5.

Frequency of heat wave days (HWF; red) and heat wave duration (HWD; blue) averaged over the three subregions of Japan for (a–r) the 20C3M experiments (1980–1999) with the CMIP3 models and the present-day (1979–2003) climate simulations using (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, and (v) NHM-5km. (w) HWF and HWF derived from the AMeDAS observations (1979–2003). The top, middle, and bottom panels are for northern (NJ), eastern (EJ), and western (WJ) Japan, respectively.

[18] Table 2 shows the relationship between model resolution and the percentage of the number of models with good simulation for seasonal mean Tmax90, Tmax90-10, HWF, and HWD, respectively. The percentage is defined as a ratio of the number of models that well simulate each index for each subregion to three (number of subregions) times the total number of models for each resolution. Although the percentages of the number of models with good simulation for the models finer than 200 km are the highest for seasonal mean Tmax90 and HWF, the percentage is still small (<50%).

Table 2. Total Number of the CMIP3 Models for All of the Subregions in Which Bias Is Small (<2 K for Seasonal Mean Tmax90 and <20% for Other Indices) for Each Resolution
ResolutionTmax90Tmax90-10HWFHWD
  1. Bold underlined numbers indicate that the percentage of the number of models with good simulation is the largest among resolutions.

<200 km7/15 (47%)4/15 (27%)5/15 (33%)1/15 (7%)
200–300 km6/24 (25%)8/24 (33%)4/24(17%)4/24 (17%)
>300 km3/15 (20%)6/15 (40%)1/15(7%)1/15 (7%)

3.2 Future Climate Simulations by CMIP3 Models

[19] Figure 6 shows future changes in summertime (June–August) SST around Japan, as derived from the SST difference between SRES-A1B experiments (2081–2100) and 20C3M experiments (1980–1999). The minimum SST increase of about 1.5 K is projected by NCAR-PCM1 (Figure 6r), and the maximum SST increase of about 4 K is projected by MIROC3.2 (hires and medres) (Figures 6m and 6n). The multimodel ensemble mean SST increase is about 3 K (not shown). Most CMIP3 models project a larger SST increase around northern Japan than around southern Japan in the future climate.

Figure 6.

Future changes in summertime (June–August) SST around Japan simulated by the CMIP3 models. The changes are derived from the SST difference between the SRES-A1B simulation (2081–2100) and 20C3M simulations (1980–1999). The labels a–r correspond with those in Table 1. The rectangles in Figure 6m outline the areas used for calculating the areal average of future changes in the frequency of heat wave days (dashed rectangle) and of future changes in SST (solid rectangle), as shown in Figure 9.

[20] Figure 7 shows the future changes in the seasonal mean Tmax90 and Tmax90-10. NCAR-PCM1 (r) projected the largest decrease in Tmax90-10 (−2.8, –1.4, and −0.9 K in NJ, EJ, and WJ, respectively) and the largest increase in seasonal mean Tmax90 (4.6, 4.7, and 5.3 K in NJ, EJ, and WJ, respectively) in a warmer climate. Averaged future changes in seasonal mean Tmax90 projected by the CMIP3 models, excluding NCAR-PCM1, are 3.1, 2.8, and 2.7 K in NJ, EJ, and WJ, respectively. This result is consistent with projected patterns of future changes in SST. Averaged future changes in Tmax90-10 projected by the CMIP3 models, excluding NCAR-PCM1, are negligible (−0.04 K in all subregions).

Figure 7.

Future changes in the seasonal mean (June–August) 90th percentile of Tmax (Tmax90 (K); red) and difference between seasonal mean (June–August) 90th and 10th percentiles of Tmax (Tmax90-10 (K); blue) averaged over the three subregions of Japan for (a–r) the 20C3M experiments (1980–1999) with the CMIP3 models and the present-day (1979–2003) climate simulations using (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, and (v) NHM-5km. (Ave) CMIP3 ensemble mean of future changes in seasonal mean Tmax90 and Tmax90-10, except for NCAR-PCM1, is also shown.

[21] Figure 8 shows projected future changes in HWF and HWD. A model with a large increase in HWF tends to show a large increase in HWD (e.g., CNRM-CM3 (d), GFDL-CM2.1 (h), IPSL-CM4 (l), MIROC3.2 (hires) (m), NCAR-PCM1 (r)). Projected future changes in HWF and HWD can be explained by the projected future changes in seasonal mean percentile values of Tmax (Figure 3). CNRM-CM3 (d), CSIRO-Mk3.5 (f), GFDL-CM2.0 (g) and 2.1 (h), IPSL-CM4 (l), MIROC3.2 (hires) (m), MIUB-ECHO-G (o), and NCAR-PCM1 (r) project that the seasonal mean 50 percentile value of Tmax in a warmer climate will exceed the seasonal mean Tmax90 in present-day climate (Figure 3), leading to a relatively larger increase in HWF (Figure 8). The highest increase in seasonal mean Tmax90 with the largest decrease in Tmax90-10 in NCAR-PCM1 results in the largest future changes in HWF and HWD. On the other hand, CGCM3.1 (T47) (b), CSIRO-Mk3.0 (e), and GISS-MODEL-ER (j) project that future changes in the seasonal mean of percentile values of Tmax are small (Figures 3 and 7). In these models, future changes in HWF and HWD are also small (Figure 8). Averaged future changes in HWF (HWD) over the CMIP3 models, excluding the NCAR-PCM1, are 22.8, 22.3, and 26.5 d yr–1 (8.5, 8.0, and 9.1 d event–1) in NJ, EJ, and WJ, respectively. The standard deviation of projected future changes in HWF and HWD among the CMIP3 models suggests that the uncertainty in projecting future heat wave characteristics is not small.

Figure 8.

Future changes in the frequency of heat wave days (left ordinate in days per year; red) and heat wave duration (right ordinate in days per event; blue) averaged over the three subregions of Japan for (a–r) the 20C3M experiments (1980–1999) with the CMIP3 models and the present-day (1979–2003) climate simulations using (s) AGCM-180km, (t) AGCM-60km, (u) AGCM-20km, and (v) NHM-5km. (Ave) The CMIP3 multimodel ensemble mean of future changes in HWF and HWD, except for NCAR-PCM1; the error bar indicates one standard deviation of HWF and HWD simulated by the CMIP3 models, except for NCAR-PCM1. MRI model (s–v) ensemble average and standard deviation are also shown.

[22] Figure 9a shows the relationship between future changes in seasonal mean Tmax90 averaged over land areas around Japan (30°–48°N, 128°–146°E; dashed rectangle in Figure 6m) and future changes in SST averaged throughout the Japan region (20°–50°N, 120°–150°E; solid rectangle in Figure 6m) projected by the CMIP3 models. There is an obvious positive correlation between future changes in SST and future changes in seasonal mean Tmax90 around Japan (correlation coefficient r = 0.77). Projected future changes in SST are widely spread (1.5–4.2 K). Looking at around 3 K of future changes in SST (nearly CMIP3 mean), future changes in seasonal mean Tmax90 are also widely spread (1.9–3.5 d yr–1). From the definition of HWF, it is expected that large (small) future changes in seasonal mean Tmax90 relative to Tmax90-10 in present-day climate result in large (small) future changes in HWF. Figure 9b shows the relationship between future changes in HWF averaged over land areas around Japan versus future changes in seasonal mean Tmax90 relative to Tmax90-10 in present climate. There is also an obvious positive correlation between future changes in HWF and future changes in seasonal mean Tmax90 relative to Tmax90-10 (r = 0.94). Projected future changes in HWF are widely spread (4–58 d yr–1). Relatively small future changes in HWF are projected by CGCM3.1 (T47 and T63) (b and c) and MPI-ECHAM5 (p) in which Tmax90-10 is well simulated over all subregions of Japan. Although CGCM3.1 (T47) (b) and CNRM-CM3 (d) projected nearly same future changes in seasonal mean Tmax90 (2.0 and 1.9 K, respectively), underestimation of Tmax90-10 by CNRM-CM3 (d) resulted in larger future change in HWF. Therefore, not only climate sensitivity of the models but also simulated magnitude of Tmax90-10 in present climate affects future projection of HWF.

Figure 9.

Scatter diagram of (a) areal averages in future changes in seasonal mean Tmax90 (ordinate) (the analysis area is shown by the dashed rectangle in Figure 6m) versus areal averages in future changes in SST around Japan (the analysis area is shown by the solid rectangle in Figure 6m) and (b) areal averages in future changes in the frequency of heat wave days (ordinate) versus ratio of areal averages in future changes in seasonal mean Tmax90 to Tmax90-10 of present climate (the analysis area is shown by the dashed rectangle in Figure 6m). The labels a–r correspond with those in Table 1. Also shown are the correlation coefficient (R), linear regression curve, and root mean square error (RMSE) calculated using all data except the outlier NCAR-PCM1 (r).

3.3 Present-Day Climate Simulations by MRI-AGCM and NHM-5km

[23] Figures 2s–2v show the model topography for the area surrounding Japan at each model resolution. AGCM-180km is clearly unable to represent the complex topography (Figure 2s). The representation of topography is improved in AGCM-60km relative to AGCM-180km; however, the backbone mountain ranges that divide the Pacific Ocean side of the Japanese Islands from the Japan Sea side are poorly resolved (Figure 2t). AGCM-20km (Figure 2u) and NHM-5km (Figure 2v) are both able to realistically represent the backbone mountain ranges of Japan.

[24] AGCMs underestimate seasonal mean Tmax90 averaged over the three subregions of Japan in present-day climate (s–u in Figure 3). NHM-5km significantly improves the underestimation especially in NJ. Although AGCMs well simulated (error is less than 20%) Tmax90-10, it is a bit underestimated by NHM-5km (v in Figure 3).

[25] Whereas AGCM-180km well simulates observed HWF and HWD in NJ and EJ, AGCM-180km underestimates the HWF and HWD in WJ. AGCM-20km and NHM-5km underestimate both HWF and HWD in all of the subregions (Figure 5). As noted in section 3.1, the CMIP3 models are unable to reproduce local features in the observed distribution of HWF, which is affected by the complex topography of Japan. AGCM-180km simulates a high HWF in western Japan (Figure 4s), and AGCM-60km simulates a high HWF in western Japan, Kanto, and central Hokkaido (Figure 4t). The observed distribution of HWF (Figure 4w) is higher in Tohoku on the Japan Sea side than in Tohoku on the Pacific Ocean side. Moreover, HWF is higher in northern Hokkaido than in southern Hokkaido. However, neither AGCM-180km nor AGCM-60km successfully reproduces the observed spatial distributions in Hokkaido and Tohoku. All ensemble simulations indicate the same characteristics in the spatial pattern of HWF. AGCM-20km and NHM-5km are able to qualitatively reproduce the observed distribution of HWF throughout Japan. In particular, the simulated HWF in Tohoku and Hokkaido is significantly improved in the higher-resolution models than in the AGCM-60km, AGCM-180km, and CMIP3 models. The higher-resolution models correctly simulate a higher HWF in Tohoku on the Japan Sea side than on the Pacific Ocean side, and a higher HWF in northern Hokkaido than in southern Hokkaido (Figures 4u and 4v). These results suggest that a detailed representation of the backbone mountain ranges of Japan is critical for accurate simulations of the spatial distribution of HWF in Japan, particularly in Hokkaido and Tohoku. These results suggest that the models with the horizontal grid size of less than 20 km have the potential to simulate the local characteristics of heat wave although the models did not improve their performance to quantitatively simulate HWF and HWD averaged over the subregions (Figure 5).

3.4 Future Climate Simulations by MRI-AGCM and NHM-5km

[26] In future climate simulations with MRI-AGCM, future changes in SST and SIC estimated from the multimodel ensemble mean by the CMIP3 models are prescribed (see section 2.1). This experimental design is equivalent to an SST increase of about 3 K around Japan. Future changes in seasonal mean percentiles of Tmax averaged over the subregions simulated by MRI-AGCM and NHM-5km are almost the same (s–v in Figures 3 and 7). AGCM-180km projects smaller future changes in the HWF and HWD in EJ and WJ than the other horizontal resolution models. Areal averaged future changes in HWF (HWD) projected by MRI-AGCM and NHM-5km (ensemble mean of s–v in Figure 8) are 12.2, 14.5, and 19.3 d yr–1 (6.2, 7.1, and 8.3 d event–1) in NJ, EJ, and WJ, respectively. These projected values are comparable to those projected by the CMIP3 model (Figure 8).

[27] Since high-resolution models have the potential to simulate local heat wave characteristics in present climate as noted in section 3.3, possible local future changes in heat wave characteristics projected by the high-resolution models are worth investigating. Figure 10 shows simulated future changes in HWF. The models simulate increases of more than 5 d yr–1 across most of the analysis region. The spatial pattern of future changes shows good agreement not only among simulations using different horizontal resolutions but also among the ensemble members of AGCM-180km and AGCM-60km. HWD is projected to increase substantially, particularly over areas with large simulated future change in the frequency of heat wave days (Figure 11). The underestimation of HWF by the models in present-day climate simulations is of little consequence to these projections of future changes, as the magnitude of the present-day underestimation is much smaller than the simulated future changes in HWF.

Figure 10.

Projected future (2075–2099) changes in the frequency of heat wave days relative to present-day (1979–2003) climatologies according to (a–d) AGCM-180km, (e–h) AGCM-60km, (i) AGCM-20km, and (j) NHM-5km. Results are shown for each individual member of the AGCM-180km and AGCM-60km ensembles.

Figure 11.

Same as Figure 10 but for the duration of heat wave days (days per event).

[28] All of the models project a large increase in HWF in western Japan (including Nagoya and Osaka). AGCM-180km projects increases of more than 10 d yr–1 in Kyushu and Chugoku (Figures 10a–10d), and AGCM-60km projects increases of more than 10 d yr–1 in Kyushu and more than 20 d yr–1 in Chugoku (Figures 10e–10h). This large increase in HWF in western Japan is consistently projected by all ensemble members (Figures 10a–10h). AGCM-20km and NHM-5km project increases of more than 30 d yr–1 along the coastline of western Japan (Figures 10i and 10j), although this feature does not appear in the AGCM-60km and AGCM-180km results.

[29] All of the models except for AGCM-20km show larger increases in HWF (by more than 10 d yr–1) in western Hokkaido (including Sapporo) than in eastern Hokkaido (Figures 10a–10h and 10j). This finding is consistently projected by all the ensemble members (Figures 10a–10h). NHM-5km projects increases of more than 20 d yr–1 along the coastline of western Hokkaido (Figure 10j).

[30] The projected future increase in HWF is larger in the western and southern coastal areas of the Korean Peninsula than in other areas of the peninsula. AGCM-180km simulates an increase of more than 15 d yr–1, and AGCM-60km and the higher-resolution models simulate increases of more than 20 d yr–1. This relatively large increase in HWF in the western and southern coastal areas of the Korean Peninsula is consistently projected by all the ensemble members (Figures 10a–10h). AGCM-20km and NHM-5km also project increases of more than 30 d yr–1 on Jeju Island, for which AGCM-60km and AGCM-180km are unable to represent the topography (Figures 10i and 10j, respectively).

[31] AGCM-60km and higher-resolution models project large increases in HWF in Kanto (including Tokyo) and in Tohoku on the Pacific Ocean side (including Sendai). These large increases are consistently projected by all the members of the AGCM-60km ensemble (Figures 10e–10h). In Kanto, AGCM-60km projects an increase of more than 15 d yr–1 (Figures 10e–10h). AGCM-20km and NHM-5km project increases of more than 30 d yr–1, especially in coastal areas of Kanto (Figures 10i and 10j). AGCM-60km projects an increase of more than 15 d yr–1 near Sendai (Figures 10e–10h). AGCM-20km projects increases of more than 25 d yr–1 near Sendai. AGCM-20km and NHM-5km project larger future increases (by more than 10 d yr–1) in Tohoku on the Pacific Ocean side than on the Japan Sea side (Figures 10i and 10j).

[32] These simulated spatial patterns of future changes in HWF are explained by future changes in seasonal mean percentiles of Tmax. Figure 12 shows areas where the seasonal mean 50th percentile of Tmax in future climate exceeds the seasonal mean Tmax90 in present-day climate simulated by NHM-5km. The areas show good agreements with areas in which the projected future changes in HWF and HWD exceed 20 d yr–1 and 3 d event–1, respectively (Figures 10j and 11j). NHM-5km projects almost the same future changes in the seasonal mean percentiles of Tmax averaged over the three subregions of Japan (v in Figure 3), in terms of 90th percentile, 2.5 (NJ), and 2.6 K (EJ and WJ) (v in Figure 7). On the other hand, the simulated seasonal mean magnitude of present-day interannual variation of Tmax (Tmax90-10) in WJ is smaller than the other areas (v in Figure 3)—8.2, 6.5, and 5.3K in NJ, EJ, and WJ, respectively—and very small future changes are projected in all of the subregions (v in Figure 7). This small magnitude of interannual variation of Tmax results in large future changes in HWF and HWD in WJ.

Figure 12.

Areas where seasonal mean 50th percentile of Tmax in future climate exceeds seasonal mean 90th percentile of Tmax in present-day climate simulated by NHM-5km.

4 Summary and Discussion

[33] We investigated future changes in HWF and HWD projected by the CMIP3 models using the results of the 20C3M (1980–1999) and SRES-A1B (2081–2100) simulations. Some 20C3M simulations are able to roughly simulate the observed distribution of HWF and HWD, showing a difficulty in accurately simulating local details of the distributions. Moreover, the models finer than 200 km tend to well simulate seasonal mean Tmax90 and HWF compared to coarser-resolution models. However, the percentage of the number of models with good simulation is still small. The CMIP3 models project that HWF (HWD) will increase 22.8, 22.3, and 26.5 d yr–1 (8.5, 8.0, 9.1 d event–1) in northern, eastern, and western Japan, respectively. Although the multimodel ensemble spread of future changes in HWF averaged around Japan is large (4–58 d yr–1), the models which well simulate Tmax90-10 in present-day climate project relatively small future changes in HWF. Significant positive correlations between future changes in SST and future changes in seasonal mean Tmax90 around Japan and between future changes in HWF and future changes in seasonal mean Tmax90 relative to Tmax90-10 in present-day climate are found. These findings indicate that not only climate sensitivity of the models but also simulated magnitude of Tmax90-10 in present-day climate affects projection of heat wave characteristics.

[34] We also investigated future changes in heat wave characteristics in the Japan region using high-resolution present-day (1979–2003) and future (2075–2099) climate simulations. The simulations were conducted using 20, 60, and 180 km mesh AGCMs and a 5 km mesh RCM under the IPCC A1B emission scenario. The present-day climate simulations show that the 5 and 20 km mesh models perform well, in a qualitative sense, in simulating the spatial pattern of HWF, although the models tend to quantitatively underestimate HWF itself. In particular, the simulations indicate that the higher-resolution models that are capable of representing detailed mountain topography (such as the 20 and 5 km mesh models) are needed to accurately reproduce the spatial pattern of the frequency of heat wave days in Tohoku and Hokkaido. The projected patterns of future changes in HWF are generally consistent among ensemble members and among models with different spatial resolutions. The models project large increases in HWF in western Hokkaido, western Japan, Kanto, Tohoku on the Pacific Ocean side, and the western and southern coastal areas of the Korean Peninsula. The 20 and 5 km mesh models project increases of more than 30 d yr–1 along the coastlines in these regions. The models also project increases in HWD.

[35] Here, we discuss conceivable causes of model biases in future projections of heat wave characteristics since reductions of model biases and improvements of the models are important for more accurate future projections of heat wave characteristics. We showed that high-resolution CMIP3 models (finer than 200 km) tend to well simulate seasonal mean Tmax90. However, more than half of the high-resolution CMIP3 models failed to accurately simulate seasonal mean Tmax90. Although MRI-AGCMs have finer horizontal mesh than do CMIP3 models, the AGCMs have significant cold biases and could not show the improved performance. Since the radiation budget near the surface affects the simulation of surface temperature, it is expected that radiation processes, land surface processes, or moist processes (convective parameterization) cause the biases. Although Tmax90-10 is well simulated by MRI-AGCMs in which an observed SST with interannual variation is prescribed as the lower boundary condition, more than 60% of the CMIP3 models underestimate Tmax90-10. Therefore, it is worth investigating interannual variation of SST and assessing its biases. In addition, the model performance in simulating long-lasting large-scale weather phenomena that can cause heat waves (e.g., atmospheric blocking) and the intraseasonal variability and future changes in their characteristics need to be investigated.

[36] Given that the world's coastal areas are heavily populated, accurate projections of future changes in the characteristics of heat waves in these areas can enable us to accurately assess the impacts of climate change on human health, agriculture, and socioeconomics. In this study, we investigated possible future changes in HWF and HWD using high-resolution atmospheric models. Although our simulations did not employ an urban canopy model that represents the effect of urbanization, the high-resolution models projected large increases in HWF and HWD in megacities such as Tokyo, Osaka, and Nagoya. The use of urban canopy models may produce even larger increases in the frequency and duration of heat wave days in the future climate. Quantifying uncertainty in future projections is important for policy making on climate change [Reilly et al., 2001]. To estimate the degree of uncertainty of future changes in heat wave characteristics in a coastal area, it is required to conduct a large number of ensemble simulations using 20 km mesh or finer-resolution models. However, we were not able to conduct ensemble simulations using AGCM-20km and NHM-5km due to the limitation of computational resources. As previous studies have emphasized the urgent need for access to adequate high-performance computing facilities that are currently unavailable to the climate modeling community [Palmer, 2005; Nature, 2008; Shukla et al., 2009; Matsueda and Palmer, 2011], we also request multi-petaFLOPS computers for high-resolution climate modeling, which enables to estimate the uncertainty of future climate projections in coastal areas.

Acknowledgments

[37] We thank Mr. Osamu Arakawa at MRI for arranging the CMIP3 dataset. 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 of the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan. The numerical simulations were performed using the Earth Simulator. We also acknowledge the valuable comments from three anonymous reviewers.