Volume 34, Issue 1
Atmospheric Science
Free Access

Differences in rain rate intensities between TRMM observations and community atmosphere model simulations

Yi Deng

Yi Deng

Institute for Geophysics, University of Texas at Austin, Austin, Texas, USA

Search for more papers by this author
Kenneth P. Bowman

Kenneth P. Bowman

Department of Atmospheric Sciences, Texas A&M University, College Station, Texas, USA

Search for more papers by this author
Charles Jackson

Charles Jackson

Institute for Geophysics, University of Texas at Austin, Austin, Texas, USA

Search for more papers by this author
First published: 10 January 2007
Citations: 9


[1] Precipitation related latent heating is important in driving the atmospheric general circulation and in generating intraseasonal to decadal atmospheric variability. Our ability to project future climate change, especially trends in costly precipitation extremes, hinges upon whether coupled GCMs capture processes that affect precipitation characteristics. Our study compares the tropical-subtropical precipitation characteristics of simulations by the NCAR CAM3.1 atmospheric GCM and observations derived from the NASA Tropical Rainfall Measuring Mission (TRMM) satellite. Despite a fairly good simulation of the annual mean rain rate, CAM rains about 10–50% more often than the real world and fails to capture heavy rainfall associated with deep convective systems over subtropical South America and U.S. Southern Plains. When it rains, there is a likelihood of 0.96–1.0 that it rains lightly in the model, compared to values of 0.84–1.0 in TRMM data. On the other hand, the likelihood of the occurrence of moderate to heavy rainfall is an order of magnitude higher in observations (0.12–0.2) than that in the model (<0.02). Comparison of regionally aggregated PDFs of the rain rate shows that CAM underestimates the probability of NOT raining, overestimates the probability of light rain and almost completely misses the tails of the PDFs. The model compensates for the lack of heavy precipitation through raining more frequently within the light rain category, which leads to an annual rainfall amount close to what is observed. CAM captures the qualitative change of rain rate PDF from a “dry” oceanic to a “wet” oceanic region, but it fails to simulate the change of precipitation characteristics from an oceanic region to a land region where thunderstorm rainfall dominates.

1. Introduction

[2] On climatic timescales, latent heating associated with precipitation is a major component of the global atmospheric energy budget and a key source of atmospheric variability [Peixoto and Oort, 1992]. On shorter timescales, heavy precipitation can trigger flooding, which on average cost the United States over $4 billion annually from 1990 to 1997. If the planet warms, the hydrologic cycle is expected to intensify, which will increase the probability of intense rainfall events [Intergovernmental Panel on Climate Change, 1996, p. 7]. Our ability to predict changes in the probability of extreme precipitation events depends on whether coupled atmosphere-ocean global climate models (GCMs) can simulate realistic precipitation characteristics. Because climate model development has traditionally focused on correctly simulating precipitation means, there is considerable uncertainty that models can also capture the probability distribution of observed rain rate. The ability of mesoscale models to simulate extreme precipitation events is improving [Mass et al., 2002; Bryan et al., 2003; Roebber et al., 2004; Lynn et al., 2005], but increased skill becomes available only when models begin to resolve convective systems explicitly. Until these resolutions become possible for long-term climate simulations, global climate models must rely on parameterizations of convective precipitation.

[3] Studies using coupled GCMs and regional climate models with parameterized precipitation agree that there will likely be a general global increase in precipitation intensity with enhanced greenhouse gas forcing [Semenov and Bengtsson, 2002; Watterson and Dix, 2003; Wehner, 2004; Meehl et al., 2005; Chen et al., 2005]. But the rain rates considered in the above studies are averaged over a period of one day, one pentad and sometimes even one month. Because the heaviest rainfall may last only minutes to hours, greater attention needs to be paid to the ability of models to simulate realistic rain rates on short timescales (≤1 hour). Here we compare the probability density functions of observed and simulated hourly rain rates at the spatial resolution of a typical GCM (a few hundred kilometers). Beyond the first moments (means), we also compare the percentage time rain occurs, the 90th percentile of rain rates (as a measure of extreme events), and the fraction of the rain falling in broad categories (light and moderate to heavy).

[4] Through systematic comparisons of multiple rain rate parameters, our study aims to identify key differences between the precipitation characteristics of satellite observations and a typical GCM simulation.

2. Data and Model

[5] Observations of precipitation are the 12/1997–12/2005 combined instantaneous rain rate retrieval from the Tropical Rainfall Measuring Mission (TRMM) 3G68 product. The data are nearly-instantaneous rain rates averaged onto a 0.5° × 0.5° longitude-latitude grid covering the global tropics from 40°S to 40°N. The microwave remote-sensing technique used by TRMM measures the volume emission of radiation from the atmospheric column, which implies an averaging time of 10 to 15 minutes. The orbit of the TRMM satellite allows sampling roughly once per day in the tropics. The data have been averaged to the coarser T42 model grid for model-observation comparison.

[6] NCAR CAM3.1 is used to make an 11-year control simulation of the present-day-climate. The model is forced with observed monthly sea surface temperature (SST) and sea ice extent from 03/1990 to 02/2001. The model uses standard T42 horizontal resolution (roughly 2.8° × 2.8°) and 26 vertical levels. The Zhang and McFarlane scheme [Zhang and McFarlane, 1995], the default for CAM3.1, is used to parameterize cumulus convection. Hourly-averaged rain rates are used in the following analysis.

[7] Another issue relevant to our results is the definition of “zero” rain rate. For TRMM data, we deem that there is no rain falling when the retrieved rain rate is less than 10−4 mm/hr. In the model output, a significant amount of “drizzle” events with rain rate < 0.05 mm/hr exists. Because there is a gap between these near-zero events and the start of a more continuous spectrum of light rain, we attribute these events to numerical noise and define “zero” rain rate in the model to be rain rate < 0.05 mm/hr.

[8] To check the sensitivity of the results to the length of the time averaging window, we also compare TRMM observations to instantaneous CAM rain rates (that is, at the model time resolution of ∼17 minutes). The results are quite similar to those based on the model's hourly-averaged rain rates. In addition, an ensemble of ten realizations of the model was generated by perturbing the initial conditions. The spread among these ten simulations is extremely small for all of the statistics presented below.

3. Results

[9] The size of the satellite data set limits the stability of the estimated rain rate PDF for a single grid point, and it is difficult to visualize the two-dimensional spatial distribution of PDFs. Therefore, four parameters of the rain rate PDFs are used to quantify differences between the TRMM data and the CAM3.1 simulation. They are: (1) annual-mean rain rate, (2) percentage time rain occurs, (3) 90th percentile of the rain rate, (4) likelihood of observing light or moderate to heavy rainfall. In addition, aggregated PDFs of the rain rates are shown for selected regions.

3.1. Annual Mean Rain Rate

[10] Figures 1a and 1b show maps of the annual mean rain rate in TRMM and CAM, respectively. The model captures major rainy zones, such as the ITCZ, SPCZ, and the Amazon Basin. However, as indicated by the difference between CAM and TRMM (Figure 1c), annual-mean rain rates are generally larger in the model by an average of 0.06–0.3 mm/hr. Regions with negative model biases include the southern U.S., subtropical South America, and the Atlantic ITCZ. When stratifying the data by season (not shown), both positive (especially over the Indian subcontinent) and negative biases appear to be largest during the boreal summer (June-July-August), which implies that the largest discrepancies arise in regions where monsoonal circulations or thunderstorm rainfall is important. This points to potential deficiencies in the model's parameterization of deep convective precipitation. Another notable model bias is the excessive rainfall over the Arabian Peninsula. Zhang and Mu [2005] argue that this is a result of a positive feedback between summer time convection and lower tropospheric moisture through surface evaporation.

Details are in the caption following the image
Distribution of annual mean rain rate in (a) TRMM, (b) CAM and (c) CAM-TRMM; units are in mm/hr. Distribution of annual percentage time rain occurs in (d) TRMM, (e) CAM and (f) CAM-TRMM. Distribution of 90th percentile of rain rate in (g) TRMM, (h) CAM and (i) CAM-TRMM; units are in mm/hr.

3.2. Annual Percentage Time Rain Occurs

[11] The total number of rainy days per month/year is widely used by hydrologists to describe the wetness of an area. Because TRMM overpasses typically occur only about once per day, this statistic per se is not available from the observations, but it is possible to estimate the percentage time rain occurs from the TRMM samples. The distributions of these percentages are given in Figures 1d and 1e. Figure 1f is the difference (CAM-TRMM). The geographic patterns are similar to the observed and simulated annual mean rain rate, but there are substantial quantitative differences between the observations and the model. For the maritime continent region, for example, the observed percentage ranges between 60% and 70%, whereas in the model it rains 70–90% of the time. The bias of this quantity (Figure 1f) is mostly positive and can be as high as 50% around places like the Indian subcontinent and Pacific Ocean west of Central America. From Figures 1a–1f we can see that it rains too frequently in the model. It is worth pointing out that the percentage time rain occurs does not contain information about the temporal evolution of precipitation at a particular location, as would be apparent in a time series of rain rate.

3.3. 90th Percentile of the Rain Rate

[12] We use the value of 90th percentile of the rain rate PDFs to characterize heavy precipitation in both the observations and the model simulation. Maps of these values are shown in Figures 1g and 1h. Since the percentage time rain occurs is much higher in the model than in the real world, these percentiles are derived excluding data with zero rain rate. That is, these are the 90th percentiles of the conditional rain rate. The global maximum of the 90th percentile (∼4 mm/hr) appears over subtropical South America with a second maximum (∼2.75 mm/hr) over the U.S. Southern Plains (Figure 1g). Both regions are also characterized by the highest lightning flash rate, as discussed by Cecil et al. [2005]. Therefore, those heavy rainfalls come mostly from intense convective thunderstorms. CAM 3.1 does a rather poor job of capturing those extremes (Figure 1h) and the largest negative biases (Figure 1i) are found over these two regions. This is in contrast to the moderate positive biases over Amazon and the Indian subcontinent where monsoonal rainfall is important. Such contrasts further point to likely model deficiencies in the parameterization of convective precipitation. Generally speaking, in most regions of the tropics and subtropics the model is raining too gently.

3.4. Likelihood of Observing Light or Moderate to Heavy Rainfall

[13] The likelihood of the occurrence of light rain (0.1–1 mm/hr) or moderate to heavy rain (2–5 mm/hr) is estimated from the ratio of the number of rainfall events with the light or moderate to heavy rain rate to the total number of rainfall events. Figures 2a and 2b show the annual fraction of rain that falls in the light category (0.1–1 mm/hr) in TRMM and CAM, respectively. In rainy regions over the tropical oceans, the observed likelihood of light rain is ∼0.85, while the model simulation shows distinctly higher values of 0.96–1.0. The positive bias in CAM is typically ∼0.1 over large areas of the tropical oceans and subtropics (Figure 2c). Several disparate regions, notably the stratus areas in the eastern ocean basins, the deserts, the Amazon and the Indian subcontinent have negative biases, with the model producing too little light rain. Such features are consistent with the differences between thunderstorm and monsoonal rainfall characteristics found in the 90th percentile of the rain rate. The fraction of rain that falls in the heavy category (2–5 mm/hr) shows the exact opposite (Figures 2d–2f). In the TRMM data the likelihood of observing moderate to heavy rainfall can be as high as 0.12–0.2 in places like subtropical South America and the U.S. Southern Plains. The model gives values less than 0.02 over almost the entire tropical-subtropical area. Together, these results reveal that in the model there is generally too little heavy rainfall and too much light rainfall.

Details are in the caption following the image
Ratio of the number of rainfall events with a rain rate in the range 0.1–1 mm/hr to the total number of rainfall events in (a) TRMM, (b) CAM and (c) CAM-TRMM. Ratio of the number of rainfall events with a rain rate in the range 2–5 mm/hr to the total number of rainfall events in (d) TRMM, (e) CAM) and (f) CAM-TRMM.

3.5. Regionally-Aggregated PDF of the Rain Rate

[14] Regionally-aggregated PDFs of the rain rate have been examined over eight climatologically consistent regions: the Pacific ITCZ, SPCZ, tropical Indian-Pacific Ocean, Amazon, tropical Africa, Indian monsoon region, Arabian Peninsula and southwest U.S. The choice of these regions is based on the observed annual rainfall amount, i.e. the annual mean rain rate is similar within the range of each individual region. Since the model biases that exist across these regions are strikingly similar, we use the Pacific ITCZ (130°E–90°W, 0°–12°N) as an example to convey the major points.

[15] Figure 3a shows the observed and simulated rain rate PDF over the Pacific ITCZ. TRMM is shown in blue; CAM in red. Three key differences can be identified. First, the probability of not raining is significantly smaller in the model than in the real world. This is consistent with the results presented in 4. Second, precipitation is more likely to fall within the light rain category (∼0.1–1 mm/hr) in CAM3.1 than in TRMM. Note the scale of the ordinate is logarithmic, so the differences are quite large. Third, the model fails to produce the exponential tail of high rain rates beyond about 2 mm/hr. This problem is most severe in the Pacific ITCZ, SPCZ, tropical Indian-Pacific Ocean and Indian monsoon regions (not shown). The model compensates for the lack of heavy precipitation through raining more frequently within the light rain category. The net result is an annual mean rain rate close to what is observed, but serious underestimation of the frequency of heavy rain rates.

Details are in the caption following the image
(a) Regionally aggregated PDF of rain rate for ITCZ in TRMM (blue line) and CAM (red line). Change of rain rate PDF from “ocean” (solid line) to “land” (dashed line) in (b) TRMM and (c) CAM. Change of rain rate PDF from an “ocean (wet)” (solid line) to an “ocean (dry)” (dashed line) region in (d) TRMM and (e) CAM.

[16] We also consider qualitative changes in rain rate between different climate zones to determine whether the model-observation differences are consistent. Any inconsistencies would undermine the applicability of a model to infer qualitative effects of climate change on regional rainfall characteristics. Figures 3b and 3c compare the observed and simulated rain rate PDFs for two regions: the subtropical eastern Pacific (165°W–155°W, 25°N–35°N, solid line, “ocean”) and the U.S. Southern Plains (94°W–84°W, 25°N–35°N, dashed line, “land”). In the TRMM data, the probability of high rain rates is much larger over the U.S. Southern Plains, where thunderstorm rainfall dominates. In CAM, by contrast, the probability of heavy rainfall is lower over that region. These results are in contrast to the changes of rain rate PDF from part of the Pacific ITCZ (150°E–180°E, 0°–10°N, solid line, “ocean (wet)”) to an adjacent subtropical area (150°E–180°E, 15°N–25°N, dashed line, “ocean (dry)”) which are compared in Figures 3d and 3e. In this case, both areas are oceanic and the difference represents a change from a “dry” to a “wet” zone with similar environments. Though the tails of the PDFs are missing, CAM does simulate the overall pattern well. The obvious difficulty CAM has in terms of simulating the change from the subtropical eastern Pacific to the U.S. Southern Plains confirms the model's weakness in reproducing the characteristics of thunderstorm rainfall. A counterpart analysis on regions in the Southern Hemisphere revealed qualitatively similar results (not shown).

4. Summary and Discussion

[17] Comparison of TRMM satellite observations of rain rate PDFs with simulations by CAM reveal important errors in the model simulations. The model substantially underestimates the occurrence of large rain rates nearly everywhere. The lack of heavy rainfall is compensated by too-frequent occurrence of light rain. These errors could be attributed to the inability of global models to resolve convection. However, Emori et al. [2005] demonstrated that an AGCM's simulation of extreme precipitation is strongly dependent on the cumulus parameterization even when the resolution is relatively high at T106. So the model errors discussed here at least partly arise from inadequacies in the model's convective parameterization. In the future, advances in computing technology will likely make possible global-scale weather/climate models that resolve convection; but until that happens, models will depend on convective parameterizations. Therefore, it is important to continue to develop, test, and improve convective parameterization schemes.

[18] Dai and Trenberth [2004] reported that the premature initiation of convection in NCAR Community Climate System Model (CCSM2) prevents the accumulation of convective available potential energy (CAPE) and the occurrence of intense convection. The extended duration of daytime convection causes the model to rain too frequently at reduced intensity. Zhang and Mu [2005] showed that this problem can be partly solved by relating convection to tropospheric large-scale forcing instead of CAPE and by including a relative humidity threshold to suppress convection when the boundary layer air is too dry. Similar ideas are adopted by Emori et al. [2001, 2005], where they achieved a better simulation of extreme precipitation after adding a treatment of cumulus suppression based on relative humidity.


[19] This material is based upon work supported by the National Science Foundation under grant OCE-0415738 and the John A. and Katherine G. Jackson School of Geosciences, University of Texas at Austin. K. Bowman was supported by NASA Global Precipitation Mission Program grant NAG5-4753 to Texas A&M University. The authors thank the TRMM Science Data and Information System (TSDIS) for providing access to TRMM data and thank NCAR for providing CAM.