GEOPHYSICAL RESEARCH LETTERS, VOL. 30, NO. 20, 2078, doi:10.1029/2003gl018297, 2003 Convective scheme and resolution impacts on seasonal precipitation forecasts D. W. Shin, T. E. LaRow, and S. Cocke Center for Ocean-Atmospheric Prediction Studies, Florida State University, Tallahassee, Florida, USA Received 31 July 2003; revised 5 September 2003; accepted 16 September 2003; published 31 October 2003. [1] An investigation is carried out to assess the role of different convective schemes and resolutions in seasonal quantitative forecasts of precipitation. The model performances are evaluated by changing convective schemes and the resolution impacts are examined by increasing the model horizontal resolution from T63 to T170 and finally T255. The predicted rainfall amounts are compared against the TRMM satellite estimate. Every forecast captures prominent rainfall features reasonably well. However, there are pros and cons in each of the forecasts. Predicted seasonal rainfall patterns and intensities from changing convective schemes exhibit larger variability and higher impact on predictability than those from increasing resolution. The impact of higher resolution with any currently implemented cumulus parameterization turns out to be smaller in seasonal precipitation forecasts than in short- to medium-range forecasts. INDEX TERMS: 3354 Meteorology and Atmospheric Dynamics: Precipitation (1854); 3337 Meteorology and Atmospheric Dynamics: Numerical modeling and data assimilation; 3374 Meteorology and Atmospheric Dynamics: Tropical meteorology; KEYWORDS: seasonal precipitation, cumulus parameterization, horizontal resolution. Citation: Shin, D. W., T. E. LaRow, and S. Cocke, Convective scheme and resolution impacts on seasonal precipitation forecasts, Geophys. Res. Lett., 30(20), 2078, doi:10.1029/2003gl018297, 2003. 1. Introduction [2] There have been substantial arguments regarding which physical parameterization and horizontal resolution are most suitable for weather and climate forecasts. In this effort, previous studies have only covered the impact of different cumulus parameterizations [e.g., Das et al., 2002] or that of higher resolution [e.g., Jha et al., 2000] especially on precipitation forecasts, claiming, such as, that one convective scheme is better suited than another, or that a higher resolution forecast resolves organized rainfall patterns better. However, there are almost no studies that deal with both impacts together in detail. Climate modelers have to find whether their models have sufficient spatial resolution to resolve the physical processes affecting climate. This fact motivates the authors in this study to explore their roles simultaneously for seasonal precipitation forecasts. Seasonal precipitation forecasting is a very difficult problem due to a number of inherent modeling complexities. This short contribution is intended to describe a comparison of seasonal precipitation forecasts with different convective Copyright 2003 by the American Geophysical Union. 0094-8276/03/2003GL018297 CLM 8-1 schemes at different horizontal resolutions. Does changing the convective schemes and horizontal resolutions have a strong impact on seasonal rainfall patterns and intensities? The authors also wish to determine whether there is a useful skill in the seasonal rainfall over the global tropics in a recently upgraded version of the Florida State University Global Spectral Model (FSUGSM). It is believed that this study gives a preliminary outlook for the roles of convective schemes and high resolution in seasonal precipitation forecasts. It must be stated that the performance of convective schemes is model-dependent and this study should not be seen as an overall comparison of different schemes. The important point is the range of response. 2. Model and Experiments [3] The FSUGSM [Cocke and LaRow, 2000] has recently been equipped with five different state-of-the-art cumulus parameterization schemes. They are (1) NCEP/SAS (National Center for Environmental Prediction/Simplified Arakawa-Schubert [Pan and Wu, 1994]); (2) NCAR/ZM (National Center for Atmospheric Research [Zhang and McFarlane, 1995]); (3) NRL/RAS (Naval Research Laboratory/Relaxed Arakawa-Schubert [Rosmond, 1992]); (4) MIT (Massachusetts Institute of Technology [Emanuel and Zivkovic-Rothman, 1999]); and (5) GSFC/RAS (Goddard Space Flight Center/Relaxed Arakawa-Schubert [Moorthi and Suarez, 1992]). [4] Seasonal precipitation forecasts are first made by changing the above five convective schemes in the FSUGSM at a resolution of T63 (1.86 ) with 17 vertical levels. The higher resolution experiments are next carried out by increasing the model horizontal resolution from T63 to T126 (0.94 ), T170 (0.70 ), and finally T255 (0.47 ) with each convective scheme. A total of 20 (5 cumulus schemes by 4 different resolutions) three-month integrations are performed from the same initial condition for February 1, 2002, with sea surface temperature prescribed according to the observed monthly mean climatology. The predicted rainfall amounts are verified against the satellite rainfall retrievals from the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI) 2A12 rainfall algorithm. 3. Results [5] The seasonal mean TRMM rainfall estimate is shown in Figure 1a. It is an average value from February to April (FMA), 2002. The global area average (30 Sto30 N) value is indicated in the upper right corner (3.11 mm d 1 ). The corresponding seasonal CMAP (Climate Prediction Center Merged Analysis of Precipitation) 21-year climatology is shown in panel (b) to address a common precipitation
CLM 8-2 SHIN ET AL.: CONVECTIVE SCHEME AND RESOLUTION IMPACTS Figure 1. An observed seasonal mean (February to April, 2002) rainfall (mm d 1) from the TRMM is compared with the CMAP climatology. The area average (30 S to 30 N) value is indicated in the upper right corner. climatology. While the pattern correlation coefficient between them is 0.76, the equitable threat score (ETS) for a higher threshold value (>10 mm d 1) is only 0.10. This implies that the climatology produces a good rainfall pattern but not a good intensity estimate. Most of the intense rainfall events cannot be well resolved in the climatology. [6] Twenty panels in Figure 2 illustrate seasonal predicted rainfall maps valid for FMA, 2002, from all experi- Figure 2. Seasonal precipitation forecasts from different resolutions and convective schemes. The verification map is shown in Figure 1a.
SHIN ET AL.: CONVECTIVE SCHEME AND RESOLUTION IMPACTS CLM 8-3 Table 1. Correlation Coefficients of Seasonal Precipitation Forecasts in Terms of Different Resolutions and Cumulus Parameterizations (30 S to 30 N, February to April, 2002) T63 0.61 0.58 0.60 0.62 0.50 T126 0.62 0.59 0.62 0.69 0.52 T170 0.56 0.57 0.61 0.71 0.52 T255 0.54 0.58 0.63 0.70 0.53 ments. This figure concisely shows the impact of different convective schemes and resolutions. The corresponding satellite-retrieved rain was shown in Figure 1a. The T63 experiments with different cumulus schemes are shown in panels (a) to (e), the T126 in panels (f ) to ( j), the T170 in panels (k) to (o), and the T255 in panels ( p) to (t). The areaaveraged values from the FSUGSM are between 1.0 to 2.0 mm d 1 higher than the observed value. Much of this rain occurs over the ocean. This might be related to the vertical moisture diffusion treatment in the FSUGSM. However, every forecast captures the prominent rainfall features such as the ITCZ (Inter-Tropical Convergence Zone) and SPCZ (Southern Pacific Convergence Zone) reasonably well. [7] Different convective schemes predict distinguishable seasonal precipitation patterns and intensities. Each scheme exhibits its own strength over different locations. For example, the MIT scheme has a better forecast skill over the Indian Ocean compared to the others. The NRL, MIT, and GSFC schemes produce reasonably good forecasts over the tropical land areas. Skill scores in term of correlation coefficients are shown in Table 1. After interpolating higher resolution forecasts onto the T63 resolution, the skill scores are computed in order to ensure consistency in comparison. The best score within each resolution is written with bold numerals. The highest correlation coefficient for each convective scheme is highlighted with a gray box. The best score is consistently obtained by the MIT scheme for all resolutions in our experiments. As can be seen in Figure 2, too much rainfall occurs over the Indian Ocean in all convective schemes except the MIT scheme. This false alarm rain reduces the skill scores for those schemes. The correlation coefficient for the MIT T170 forecast (0.71) is lower than that of the climatology (0.76) shown in Figure 1b. This indicates that the FSUGSM has some difficulty in predicting the proper global rainfall pattern. This problem can be partially reduced by the bias correction approach [Shin and Krishnamurti, 2003]. However, if a categorical skill measure is employed, a different conclusion can be reached. Table 2 shows the ETS for seasonal precipitation forecast using threat greater than 10 mm d 1 threshold. All scores are higher than the climatological score (0.10). This implies that intense rainfall events are better predicted by the numerical models. Even with this Table 3. Same as Table 1 But for Day One Precipitation Forecast T63 0.30 0.27 0.32 0.41 0.32 T126 0.45 0.43 0.48 0.53 0.41 T170 0.46 0.46 0.51 0.54 0.41 T255 0.49 0.50 0.53 0.55 0.42 skill measure, the MIT scheme also proved to be the most skillful for seasonal precipitation forecasts at least within the FSUGSM. It must be noted here that the excessive rainfall amount is not cured by changing cumulus schemes. This problem might be, we believe, mainly related to the treatment of vertical diffusion and land surface parameterization in the FSU model. [8] It is commonly regarded that if a model resolution is increased, the short- to medium-range forecast gives a better skill and resolves the location and timing of mesoscale precipitation systems well. A higher horizontal resolution model usually appears to preserve the organization of convection. The organization of convection furthermore appears to have a large implication for overall circulation. This feature has been shown in our experiments as well. Table 3 shows correlation coefficients for the day one forecast. The skill scores increase as the horizontal resolution increases. The impact of higher resolution reduces as forecast lead time increases. There is not much impact with higher resolutions on the ten-day-average forecast (Table 4). The skills of some cumulus schemes appear to be not resolution dependent. Overall, we can conclude that while the impact of higher resolution is noticeable in the short- to medium-range forecast, the impact is much smaller in the seasonal forecast. The horizontal grid distance quantitatively influences the rainfall amount, although the large-scale structure remains unchanged. The rainfall amount increases as the grid interval is reduced. It is not easy to find a consistent higher resolution benefit from the skill scores in Tables 1 and 2. But if we examine the forecast maps (Figure 2) carefully, several higher resolution advantages can be drawn. The higher resolution forecasts produce more organized precipitation patterns and stronger intensities. Especially, widespread ITCZ patterns in the T63 become narrowed as the resolution increases. Double ITCZs over the eastern Pacific shown in the TRMM estimate are better captured in the higher resolution in most of the convective scheme simulations. The model with a finer resolution is able to simulate mesoscale organization of rainfall over the land masses. [9] In order to characterize the sensitivity of the model to change of resolution or cumulus parameterization, we show the variance of ensembles using same convective scheme but differing resolutions (though interpolated to T63 resolution) versus ensembles of same resolution but Table 2. Same as Table 1 But for ETS (>10 mm/d) T63 0.13 0.15 0.16 0.27 0.13 T126 0.15 0.12 0.14 0.27 0.12 T170 0.12 0.11 0.17 0.32 0.12 T255 0.10 0.12 0.18 0.28 0.11 Table 4. Same as Table 1 But for 10-Day Accumulated Precipitation Forecast T63 0.52 0.52 0.52 0.55 0.50 T126 0.57 0.55 0.57 0.63 0.53 T170 0.55 0.56 0.58 0.63 0.52 T255 0.54 0.54 0.59 0.64 0.54
CLM 8-4 SHIN ET AL.: CONVECTIVE SCHEME AND RESOLUTION IMPACTS Figure 3. Variance maps in terms of different cumulus schemes ((a) (d)) and different horizontal resolutions ((e) (i)). Unit is of mm 2 d 2. using different convective schemes. Figures 3a to 3d show spatial plots of the variance for using the same resolutions but different convective schemes for each resolution ensemble, and Figures 3e to 3i are for using the same convective scheme ensembles but different resolutions. The global area-averaged variance is given in the upper right corner of each plot. It is interesting to note that for increasing resolution, the overall variance for the convective scheme ensembles tends to decrease somewhat, indicating that the simulations may be in more agreement at the higher resolution. In general, it appears that the model is more sensitive to change in convective scheme than change in resolution. Thus, forecast uncertainty may be more related to the cumulus parameterization than the model resolution for seasonal predictions. One striking result is that the MIT scheme appears to be considerably less sensitive to model resolution than any of the other schemes. Both of the RAS schemes (NRL and GSFC) exhibited the largest sensitivity. Further study is needed to understand more clearly why this is happening. 4. Discussion and Summary [10] The FSUGSM was used to examine the impact of different convective schemes and resolutions on seasonal precipitation forecasts. A total of twenty three-month precipitation forecasts were made by a combination of five cumulus schemes and four different horizontal resolutions. TRMM was the verification data. Intense rain events were well resolved not in the precipitation climatology but in the numerical seasonal forecasts. Although every convective scheme has its own strength over different locations, the MIT scheme was able to provide relatively better results compared to other schemes at least within the FSUGSM. Higher resolution benefits were shown to be smaller in seasonal precipitation forecasts than in short- to mediumrange forecasts. The model was more sensitive to changing convective schemes than changing resolutions in predicted seasonal rainfall patterns and intensities. [11] The number of experiments done in this study is still too few to finally conclude our findings. Experiments with different seasons and years are required to have statistically significant results. However, this study provides a fundamental insight into the roles of convective schemes and resolutions in seasonal precipitation forecasts. In order to achieve an improved seasonal forecast of precipitation, even though there are plenty of uncertainties, the impact of other model physical components (such as vertical diffusion and land surface parameterization) and higher vertical resolution must be carefully tested as well. [12] Acknowledgments. Computations were performed on the IBM SP4 at the FSU. COAPS receives its base support from the Applied Research Center, funded by NOAA Office of Global Programs awarded to Dr. James J. O Brien. References Cocke, S., and T. E. LaRow, Seasonal predictions using a regional spectral model embedded within a coupled ocean-atmosphere model, Mon. Weather Rev., 128, 689 708, 2000. Das, S., A. K. Mitra, G. R. Iyengar, and J. Singh, Skill of mediumrange forecasts over the Indian monsoon region using different parameterizations of deep convection, Weather Forecast., 17, 1194 1210, 2002. Emanuel, K. A., and M. Zivkovic-Rothman, Development and evaluation of a convective scheme for use in climate models, J. Atmos. Sci., 56, 1766 1782, 1999. Jha, B., T. N. Krishnamurti, and Z. Christidis, A note on horizontal resolution dependence for monsoon rainfall simulations, Meteorol. Atmos. Phys., 74, 11 17, 2000. Moorthi, S., and M. J. Suarez, Relaxed Arakawa-Schubert: A parameterization of moist convection for general circulation models, Mon. Weather Rev., 120, 978 1002, 1992.
SHIN ET AL.: CONVECTIVE SCHEME AND RESOLUTION IMPACTS CLM 8-5 Pan, H.-L., and W.-S. Wu, Implementing a mass flux convection parameterization scheme for the NMC Medium-Range Forecast model, paper presented at Tenth Conference on Numerical Weather Prediction, Am. Meteorol. Soc., Portland, Oreg., 1994. Rosmond, T. E., The design and testing of the Navy operational global atmospheric system, Weather Forecast., 7, 262 272, 1992. Shin, D. W., and T. N. Krishnamurti, Short- to medium-range superensemble precipitation forecasts using satellite products: 1. Deterministic forecasting, J. Geophys. Res., 108(D8), 8383, doi:10.1029/2001jd001510, 2003. Zhang, G. J., and N. A. McFarlane, Sensitivity of climate simulations to the parameterization of cumulus convection in the Canadian Climate Centre general circulation model, Atmos. Ocean, 33, 407 446, 1995. D. W. Shin, T. E. LaRow, and S. Cocke, Center for Ocean-Atmospheric Prediction Studies, Florida State University, Tallahassee, FL, USA, 32306-2840. (shin@coaps.fsu.edu)