Validation, parameterization dependence, and future projection of daily precipitation simulated with a high-resolution atmospheric GCM



[1] A relatively high-resolution (T106) atmospheric general circulation model (AGCM) was used to simulate the present-day climate with two different assumptions in cumulus parameterization. While the two runs show comparable performance for the annual mean precipitation, one shows much better agreement with satellite-based analysis data for the extreme daily precipitation than the other. Accumulation of convectively available potential energy is shown to be important for more realistic intensities of extreme precipitation. This demonstrates that the performance of an AGCM for extreme precipitation is strongly dependent on cumulus parameterization even when the resolution of the model is as high as 1°, but can be reasonably good with an adequate choice of cumulus parameterization. A time-slice doubled CO2 experiment was also conducted with the better version of the model. Though the global mean percentage change is larger for extreme precipitation than for the annual mean, this relation is found to vary regionally.

1. Introduction

[2] Extremely heavy precipitation, which may cause floods and consequent serious damages to natural and socio-economic systems, is considered to be highly important in the context of climate change assessment. Previous studies investigating extreme daily precipitation in various climate change simulations have consistently indicated that intensity or frequency of extreme precipitation increases in a future warmer climate over most part of the globe [e.g., Zwiers and Kharin, 1998]. However, there are two major limitations in these studies. First, the validation of climate models for the simulation of daily precipitation in present-day climate has been limited. Secondly, the investigation of the mechanisms for enhancement of extreme daily precipitation has been limited.

[3] The limited validation of models is partly because of the lack of observational data, since station measurement of daily precipitation is available only over limited land regions of the globe with dense population. Recently, analyzed data of daily precipitation with 1° resolution based on satellite observations has become available from the Global Precipitation Climatology Project (GPCP 1DD) [Huffman et al., 2001]. It provides spatially and temporally quasi-uniform quality of data over the whole globe including the ocean, which is advantageous for comparison with climate model results as shown by May [2004]. Note that the data over lower latitudes (40°S–40°N), where it is based on infrared data from geostationary satellites and microwave (SSM/I) data, is expected to be more reliable than the higher latitudes, where the TIROS Operational Vertical Sounder (TOVS) data is used.

[4] Another cause of the limited validation has been low spatial resolution of climate models, that is, a large gap between the scale of heavy precipitation occurrence and the grid scale of models. Recently, relatively higher resolution climate models (∼1°) are becoming available and they are expected to improve simulation of extreme daily precipitation compared to lower resolution models [e.g., Voss et al., 2002]. However, it should be noted that climate models have inherent uncertainty especially in various parameterizations and, because of this uncertainty, higher resolution models do not necessarily give better simulation of climate than lower resolution models.

[5] As for the mechanism of the enhancement of extreme precipitation, Trenberth [1999] conceptually argued that its principal cause should be the enhancement of atmospheric moisture content, which feeds enhanced moisture to all weather systems. Allen and Ingram [2002] showed a globally integrated view that the extreme precipitation is constrained by moisture availability and would increase faster than the mean precipitation, which is constrained by energy balance. However, these considerations have scarcely been examined in climate change simulations, especially on a basis of geographical distribution, which is more relevant to the impacts of extreme precipitation than the globally integrated statistics.

[6] In this study, relatively high resolution (T106) atmospheric general circulation model runs with two different assumptions in cumulus parameterization are validated against the GPCP analysis data to demonstrate that the simulation of extreme daily precipitation can be highly parameterization dependent. The model with the better performance is used to examine the change in extreme precipitation intensity in the doubled CO2 climate and to investigate the mechanism for enhanced extremes.

2. Model and Experiments

[7] The model used in this study is the CCSR/NIES/FRCGC AGCM, an atmospheric general circulation model collaboratively developed at the Center for Climate System Research of the University of Tokyo (CCSR), the National Institute for Environmental Studies (NIES), and the Frontier Research Center for Global Change of the Japan Agency for Marine-Earth Science and Technology (FRCGC). It is the atmospheric part of the CCSR/NIES/FRCGC coupled climate model called MIROC (K-1 Model Developers [2004] provide its description), except that the distribution of aerosols is prescribed in the AGCM used here instead of calculated interactively as in the coupled model, and the second indirect effect of aerosols is switched off. The cumulus parameterization used is based on a prognostic closure of the Arakawa-Schubert scheme [Pan and Randall, 1998]. In conventional closure assumptions based on buoyancy, including this scheme, cumulus convection is considered to occur whenever atmospheric stratification is unstable with respect to moist convection. Emori et al. [2001] introduced an empirical cumulus suppression treatment in which cumulus convection is suppressed when the cloud-mean ambient relative humidity is smaller than a certain critical value, RHc. We have applied this treatment to the model used in this study.

[8] The resolution of the model used in this study is T106 spectral truncation in horizontal (approximately 1.1° in equivalent grid) and 56 layers in vertical. Relatively high vertical resolution is applied in the boundary layer and around the tropopause. The model top is approximately 40 km.

[9] Three runs were conducted of 20 model years for each. The first two runs were driven by observed SST and sea-ice distributions for 1979–1998 (HadISST) [Rayner et al., 2003] with a constant CO2 concentration of 345 ppmv (an approximate present-day value). The first is with cumulus suppression treatment (RHc = 0.8; hereafter denoted as CS) and the second is without it (RHc = 0; denoted as NOCS). The third run is a time-slice doubled CO2 run (CO2 concentration is 690 ppmv), whose driving SST is the sum of the observed values and a seasonally varying ‘warming pattern’. The warming pattern was derived from a transient climate change experiment with the CCSR/NIES1 coupled ocean-atmosphere climate model [Emori et al., 1999] according to Intergovernmental Panel on Climate Change (IPCC) IS92a scenario without aerosol forcing. The sea-ice data is adjusted to the SST by setting sea ice concentration to zero where SST is above the freezing temperature of the sea. This run was conducted only with the cumulus suppression.

3. Results

3.1. Validation and Parameterization Dependence

[10] For the geographical distribution of climatological annual mean precipitation, the performance of CS and NOCS is comparable. Table 1 summarizes the root mean square error (RMSE), the square of correlation coefficient (R2), and the ratio of global mean values (Ratio), of models against analysis data (GPCP 1DD). The periods for the climatological average is 1979–1998 for the two models and 1997–2003 for GPCP 1DD. The model CS is a little worse in RMSE and slightly overestimating the global mean, but a little better in R2, than NOCS. The other climatological variables such as temperature and moisture are also not notably different between the two models (figures not shown).

Table 1. Comparison of Annual Mean Precipitation Fields Between Models and Observation
SubjectReferenceRMSE [mm · day−1]R2 [−]Ratio [−]

[11] Figure 1 shows the distribution of the annual 4th largest daily precipitation (approximately 99th percentile, hereafter denoted simply as 99th percentile) as an index of extreme daily precipitation. For this figure, all data are interpolated into T106 grid before the analysis. In the analysis data (Figure 1c), heavy extreme precipitation over 40 mm/day is mainly found over ITCZ, SPCZ, Asian monsoon region and tropical Indian Ocean. This is well reproduced in the CS result (Figure 1a). On the other hand, in the result of NOCS (Figure 1b), extreme precipitation intensity is generally much weaker than the analysis data over lower latitudes (40°S–40°N) and, especially, it fails to represent the peak over the subtropical western Pacific. These are also indicated in the statistics summarized in Table 2. Considering the expected better reliability of daily GPCP data over the lower latitudes (40°S–40°N), the results for the lower latitudes as well as the global domain are shown in Table 2. Overall superiority of CS over NOCS is clear for the extreme daily precipitation, regardless of this selection of the domain. Moreover, the frequency of daily precipitation greater than 1 mm/day over the tropics is unrealistically high in NOCS (figure not shown). The frequency averaged over 10°S–10°N is 41% for GPCP, 51% for CS and 70% for NOCS. Thus, tropical daily precipitation in NOCS is characterized by too high frequency and too weak intensity.

Figure 1.

The annual 4th largest daily precipitation event (approximately 99th percentile): (a) model CS, (b) model NOCS and (c) analysis data.

Table 2. Comparison of the Annual 4th Largest Daily Precipitation (Approximately 99th Percentile) Fields Between Models and Observationa
SubjectReferenceRMSE [mm · day−1]R2 [−]Ratio [−]
  • a

    The values outside and inside of the parentheses are for the global domain and for the lower latitudes (40°S–40°N), respectively.

CSGPCP8.2 (9.7)0.77 (0.60)1.01 (1.01)
NOCSGPCP12.2 (14.8)0.57 (0.49)0.71 (0.62)

[12] However, CS has too intense extreme precipitation over tropical land areas. A further analysis revealed that this is a result of ‘grid-point storms’, in which instability is released by resolved motion and condensation because the suppression of parameterized cumulus convection is unrealistically strong over those areas. This is an evident problem of the empirical treatment of cumulus suppression and should be resolved by further improvement of the parameterization.

3.2. Mechanism of Parameterization Dependence

[13] Figure 2 shows scatter plots of daily convectively available potential energy (CAPE) vs. daily precipitation over the tropics (10°S–10°N) in arbitrarily chosen single month (January 1997). The CAPE for analysis data is calculated from the ECMWF 40-year re-analysis [Simmons and Gibson, 2000]. For this plot, all data are interpolated into a 2.5° grid before the analysis. The CS result (Figure 2a) is similar to the analysis data (Figure 2c) in that many points with large CAPE and small precipitation exist. That is, in the real tropical atmosphere, as well as in the model CS, precipitation does not always occur whenever there is large CAPE. Modeled CAPE reaches larger values than the analysis, implying that the modeled cumulus suppression is too strong, although quantitative comparison is difficult since there is ambiguity in the calculation of modeled and analyzed CAPE. In the NOCS result (Figure 2b), on the other hand, whenever CAPE is large, precipitation is also large, which proved to be unrealistic. In this case, since there is no chance of accumulation of large CAPE, maximum precipitation intensity is not as large as that of CS nor that of the analysis data. The correlation coefficient between daily CAPE and daily precipitation is 0.26 for CS, 0.75 for NOCS and 0.24 for the analysis data.

Figure 2.

Scatter diagram of daily CAPE (abscissa) and daily precipitation (ordinate) over the tropics (10°S–10°N) in January 1997: (a) model CS, (b) model NOCS and (c) analysis data.

[14] It was also observed that stronger precipitation in CS leads to more organized convective activity with stronger coupling with motion than in NOCS. For example, the number of occurrence of tropical cyclones (defined similarly to Sugi et al. [2002]) over the northwestern Pacific is 21.1 in CS and 1.2 in NOCS compared to 27.4 observed per year. Moreover, equatorial waves coupled to the tropical convection are more realistic in CS than in NOCS, which will be described in our separated paper.

3.3. Future Projection

[15] Since the model CS has superior validation against the analysis data for extreme daily precipitation, we used this model for a time-slice doubled CO2 climate change simulation. The percentage change in global mean 99th percentile precipitation is 11.7% and is significantly larger than that in global annual mean precipitation of 4.3% [cf. Allen and Ingram, 2002]. Figure 3a is the percentage change of annual mean precipitation. This shows the increase in precipitation over the tropics and mid- to high-latitudes and the decrease over the subtropics, which is a common structure obtained from various climate models [IPCC, 2001]. The percentage change of the 99th percentile daily precipitation (Figure 3b) has a overall pattern that is quite similar to the mean precipitation change (Figure 3a). Moreover, they have the same magnitude in many parts of the globe including most of the extra-tropical Northern Hemisphere land areas. There are some areas where the mean is decreased but the extreme is increased, for example the subtropical oceans especially in the Southern Hemisphere, the northern North Atlantic, and parts of Amazon and south Africa (18% area of the globe). Additionally, over some areas, the relative increase in the extreme is larger than that in the mean, including the eastern and northern Europe and parts of the tropics. It is these areas that contribute to the larger enhancement of the extreme than that of the mean on the global mean basis. Over the Arctic Ocean and some part of tropical ocean (e.g., the central Pacific), the relative change in the extreme is significantly smaller than that in the mean, where increase in relatively weak daily precipitation contributes to the increase in the mean.

Figure 3.

(a) Percentage change in annual mean precipitation from present-day to 2 × CO2; (b) As for (a) except for the change in the annual 4th largest daily precipitation event (approximately 99th percentile). White areas represent negative values.

4. Concluding Discussion

[16] In this study, we have demonstrated that the performance of climate models to reproduce daily precipitation intensity is strongly dependent on cumulus parameterization. Even if mean precipitation fields of two models are very similar, they can give totally different characteristics in daily precipitation intensity. We have also shown that, with appropriate cumulus parameterization, a model of approximately 1° horizontal resolution can successfully reproduce the intensity of daily precipitation over most part of the globe as estimated from satellite data, whose resolution is also 1°. This is not trivial as the observed 1° resolution data is actually a result of aggregation of any smaller scale phenomena in reality, which are either parameterized or ignored in a model. One of the key factors in this seems to be a realistic accumulation of CAPE and the consequent organization of precipitation systems. It should be noted that the extreme daily precipitation treated here is limited to relatively ‘common’ extremes that occur a few times a year and not multi year return values, because of the relatively short length of the validation data (7 years). It should also be noted that the cumulus suppression treatment used here is empirical and should be replaced by a more process-based scheme in the future work.

[17] The time-slice doubled CO2 future projection was performed with the model with the better performance (i.e., with cumulus suppression). Though the percentage change is larger for extreme precipitation than for annual mean on a global mean basis, consistent with the theory of Allen and Ingram [2002], this relation is different for different areas when the geographical pattern is examined. This is a natural result considering that the energy balance to constrain mean precipitation does not close for the atmospheric column over a limited area and that moisture availability to constrain extreme precipitation may depend not only on temperature. The investigation of mechanisms for this geographical distribution of the relationship between the extreme change and the mean change of precipitation is remained for future work.


[18] The authors thank the K-1 project members and Simon Brown for support and discussion and two anonymous reviewers for valuable comments. This work was partially supported by the Research Revolution 2002 (RR2002) of the Ministry of Education, Sports, Culture, Science and Technology of Japan and by the Global Environment Research Fund (GERF) of the Ministry of the Environment of Japan. The model calculation was made on the Earth Simulator. The GFD-DENNOU Library was used for the drawings.