Analysis of precipitation climatology simulated by high resolution coupled global models over the South America

The results of coupled high resolution global models (CGCMs) over South America are discussed. HiGEM1.2 and HadGEM1.2 simulations, with horizontal resolution of ~90 and 135 km, respectively, are compared. Precipitation estimations from CMAP (Climate Prediction Center – Merged Analysis of Precipitation), CPC (Climate Prediction Center) and GPCP (Global Precipitation Climatology Project) are used for validation. HiGEM1.2 and HadGEM1.2 simulated seasonal mean precipitation spatial patterns similar to the CMAP. The positioning and migration of the Intertropical Convergence Zone and of the Pacific and Atlantic subtropical highs are correctly simulated by the models. In HiGEM1.2 and HadGEM1.2, the intensity and locations of the South Atlantic Convergence Zone are in agreement with the observed dataset. The simulated annual cycles are in phase with estimations of rainfall for most of the six regions considered. An important result is that HiGEM1.2 and HadGEM1.2 eliminate a common problem of coarse resolution CGCMs, which is the simulation of a semiannual cycle of precipitation due to the semiannual solar forcing. Comparatively, the use of high resolution in HiGEM1.2 reduces the dry biases in the central part of Brazil during austral winter and spring and in most part of the year over an oceanic box in eastern Uruguay.


INTRODUCTION
The climatic system has its fluctuations determined mainly by the complex fluxes from the ocean and atmosphere that occur in a wide range of spatial and temporal scales.According to Shaffrey et al. (2009) the development of high resolution global models is indispensable to simulate the energy transfer to smaller scales and to capture the nonlinear interactions between wide range of scales, and between the different components of climatic system.Therefore, there are strong reasons to increase the resolution of the atmospheric and oceanic components of the coupled climatic models.For the atmosphere, high resolution simulations have already shown significant improvement in representing the storm trajectories (storm tracks) and precipitation distribution over Europe, where the orographic effects are important (Pope and Stratton, 2002;Junge et al., 2006).The resolution increase in the oceanic models has the objective of improving the representation of important currents, like Gulf and North Atlantic currents, and in the atmosphere it can assist in more realistic cyclones and storm tracks representation.
As reported by Vera et al. (2006), currently the CGCMs have many problems to simulate the precipitation over South America (SA).The CGCMs from IPCC-AR4 showed large discrepancies in the representation of both intensity and location of rainfall in the SACZ (South Atlantic Convergence Zone) during summer, as well as in its seasonal evolution.Moreover, most of the models analyzed by Vera et al. (2006) do not reproduce the observed precipitation maximum over Southeastern of South America (SESA) during the cold season.According to Cavalcanti et al. (2002) the CPTEC-COLA global model underestimated (overestimated) the precipitation in the tropical (subtropical) sectors of the convergence zones.Besides, over the Andes Mountains and Northeastern Brazil this model overestimated precipitation, while large deficit of rainfall was obtained in the interior of the SA continent, including the Amazon basin.Analyzing high resolution CGCMs, Shaffrey et al. (2009) showed improvements in the representation of orographic forcing and a marked reduction in the errors of precipitation along the tropical Pacific Intertropical Convergence Zone (ITCZ) and over the tropical Indian Ocean comparing with previous coarser resolution models.The distribution of precipitation still presents large-scale errors which unchanged with the increasing of resolution, although the magnitudes of the errors are slightly reduced.
In this context, it is important to know whether higher resolution CGCMs can improve rainfall climatology in many parts of world.Therefore, models validation is of great importance, serving as a deciding factor about the cost and benefit of using high resolution CGCMs.The present work evaluates the performance of two versions of high resolution coupled global models, HiGEM1.2(90 km) and HadGEM1.21(135 km), in simulating the climatology of precipitation over the SA.The CGCM simulations were compared with different data sets of precipitation.

Model
The atmosphere-ocean coupled global environment model HiGEM1.2 is the first high resolution model developed in the United Kingdom from a partnership of Natural Environment Research Council (NERC) and Met Office Hadley Centre (Shaffrey et al., 2009).Two simulations with different horizontal resolution were evaluated: HadGEM1.2 with 1.25° × 1.875° in latitude and longitude for the atmosphere and 1° × 1° (increasing 1/3°m eridionally near the equator) for the ocean; and HiGEM1.2 with 0.83° × 1.25° in latitude and longitude for the atmosphere and 1/3° × 1/3° for the ocean.
The models have three components: atmospheric, oceanic and sea-ice.The main features are resumed as follow and Shaffrey et al. (2009) provide further details.In the atmospheric component the models have a nonhydrostatic dynamical core with semi-Lagrangean transport; they are grid point models formulated in Arakawa C grid, and included an iterative scheme for aerosols.The models use the second version of the U.K. Met Office Surface Exchange Scheme (MOSES-II; Cox et al., 1999;Essery et al., 2001;Martin et al., 2006) to represent the land surface processes; it allows tiling of land surface heterogeneity using nine different surface types.According to Martin et al. (2006) a separate surface energy balance is calculated for each tile and area-weighted grid box mean fluxes are computed, which results in more realistic fluxes than when a single surface type is considered.In addition, vegetation leaf area varies seasonally, providing a more realistic representation of seasonal changes in surface fluxes.The models have 38 vertical levels up to 39 km, means that the stratosphere is not well resolved.The oceanic component has spherical latitude-longitude grid, with 40 unevenly spaced vertical levels that have higher resolution near the surface to better represent the mixing layer and the ocean-atmosphere interaction processes.For the sea-ice component, part of it is treated inside the oceanic model, and a small part is solved by the atmospheric model.The time step of coupling between the three model components is one day.

Data
The simulated seasonal spatial patterns are compared with standard coarse resolution (2.5° by 2.5°) precipitation data derived from Climate Prediction Center -Merged Analysis of Precipitation (CMAP; Xie and Arkin, 1997).It is interesting to analyze the model results in the context of precipitation estimation variability.For this, the coarse resolution (2.5° by 2.5°) Global Precipitation Climatology Project (GPCP; Quartly et al., 2007) and fine resolution (0.5° by 0.5°) Climate Prediction Center (CPC; Chen et al., 2008) are included to analyze the annual cycle.Both CMAP and GPCP data include numerous precipitation estimations from satellite data over the land and oceans, and from rain gauge data over the land.The CPC is a gauge-only precipitation estimation available only over the continent.These datasets are available from 1979-September/2009 for CMAP, from 1979-April/2009 for GPCP and from 1979-2005 for CPC.The 850 hPa horizontal wind fields from NCEP (National Center for Environmental Prediction) reanalysis 1 (Kalnay et al., 1996), with 2.5° of horizontal resolution, were also used to validate the low level circulations.
The seasonal climatology of CGCMs were evaluated over the SA, while the annual cycles considered six regions: Amazon -AMZ, Northeast -NDE, Center-West -RCO, Southeast -SDE, Southeastern South America -SESA and eastern Uruguay -Argentina -LUR.This oceanic region is evaluated because it is a core of cyclogeneses that controls a different annual cycle of precipitation (Reboita et al., 2010a).All these subdomains are depicted in Figure 1.

RESULTS AND DISCUSSIONS
The seasonal precipitation and 850 hPa wind field climatology  are shown in Figures 2 and 3.In these Figures the main features of the seasonal cycle of precipitation in CMAP are well represented by the CGCMs as the maximum southward displacement of ITCZ during the austral summer (DJF)-autumn (MAM) and its northward movement during the austral winter (JJA)-spring (SON), although with larger than observed amplitude.The correct CGCMs positioning of ITCZ in SON (Figure 3) and DJF (Figure 2) periods is an improvement compared with the results of Seth et al. (2010), which examined nine models from the Coupled Model Intercomparison Project (CMIP) version 3 and obtained the ITCZ shifted southward of the equator and with weaker amplitude than observed.In austral spring (Figure 3) and summer (Figure 2), the simulations present a fracture in the ITCZ over the tropical Pacific, a common characteristic of many global models that is attributed to an adjustment in the mass flux, i.e., in response to the intense rainfall in a region the model compensates it generating subsidence that dries the air in a near region (Gandu and Silva Dias, 1998;Cavalcanti et al., 2002).
In Figures 2 and 3, the presence of SACZ in the CMAP and CGCMs is indicated by a region of large precipitation rate from Southern Amazon to the subtropical South Atlantic Ocean, with its oceanic branch well evident over the coast of Southeastern SA (Kodama, 1992;Carvalho et al., 2004).The acting of SACZ is usually confined to the austral spring and summer seasons, when there are larger solar heating and greater availability of moisture.The 850 hPa wind fields show that low level northwesterly winds transport moisture from the Amazon to the subtropical SA (Figures 2a-2c and 3d-3f), which together with the western periphery of the South Atlantic subtropical high provide the necessary conditions to develop intense convection at these times of the year.In austral summer (Figure 2a-2c) and spring (Figure 3d-3f) both CMAP and CGCMs simulations present intense precipitation associated with the convergence zones (SACZ and ITCZ) over the continental land northward of 30°S, while the dryer areas situate southward of 35°S.In austral autumn (Figure 2d-2f) and winter (Figure 3a-3c), due to the frontal systems activity the wet area lies over the southsoutheastern Brazil, and rainfall is also intense in Northern SA associated with the southward displacement of the ITCZ.During austral winter, both CGCMs (Figure 3b, 3c) simulate the large dry area covering from Northeastern to Southeastern Brazil (north of 20°S); the extension of this area simulated by HiGEM1.2 is similar to CMAP.In austral autumn and winter (Figures 2 and 3), there is intense precipitation in the eastern coast of Northeastern Brazil in CMAP and CGCMs simulations.This pattern, according to Kousky (1980), is mainly attributed to the convergence between trade winds and nocturnal land breeze.The greatest precipitation rate in the cold season covers mainly SESA in the simulations, which is in agreement with CMAP.Most of this rain is due to the passage/development of extratropical cyclones and its associated frontal systems, that in this area are more frequent during austral winter (Gan and Rao, 1991;Reboita et al., 2010b).The seasonal positioning and displacement of the SACZ and subtropical highs in the Pacific and Atlantic Oceans are correctly simulated by HiGEM1.2 and HadGEM1.2.However, in all seasons the simulations tend to place the subtropical highs northward of that of NCEP reanalysis.
The CGCMs represent coherently the rainfall spatial patterns observed in CMAP (Figures 2 and 3).However, they overestimate the precipitation rate, mainly in the western boundary of higher topography regions, as in Southern Chile.This is a common characteristic of many global models that block the western flow and force upward motion with consequent increase of rainfall.The overestimation of precipitation over the Andes Mountains by CGCMs was also obtained by Cavalcanti et al. (2002) and Stern and Miyakoda (1995).Such an overestimation was attributed to the Gibbs error associated with deficiency of the spectral representation over the region.However, this error also occurs in the grid-point CGCMs.Comparatively, Figures 2 and 3 show that the use of high resolution in HiGEM1.2 is contributing to reduce the excessive rainfall simulated by HadGEM1.2 over the Andes Mountains (from 25° to 10°S mainly during summer and spring).Besides, over most of the Amazon both simulations have small biases (≤ ±1 mm day −1 , Figure 4).This differs from other global models, such as CPTEC-COLA (Cavalcanti et al., 2002), ECMWF (Brankovic and Molteni, 1997) and NCAR-CCM3 (Hurrel et al., 1998) that simulate an excessive deficit of rainfall over the Amazon region during the summer.
The seasonal maps (Figures 2-4) show that HiGEM1.2 has many similar spatial patterns and errors of HadGEM1.2.However, in some regions the use of high resolution in HiGEM1.2contributes to reduce the biases, as for example over center-western Brazil during winter, northern Amazon and eastern of southern Brazil and Uruguay during spring  The mean annual cycles for six regions are presented in Figure 5.In this Figure, the observed annual cycle is the mean value from three data sets (CMAP, GPCP and CPC) and its maximum and minimum value are also presented.The observed precipitation shows a well defined annual cycle of precipitation over AMZ, with maximum of 10 mm day −1 in the wet season and minimum of 1 mm day −1 in the dry season, i.e., with annual amplitude of ~9 mm day −1 .The CGCMs simulations represent correctly both the phase and intensity of the annual cycle, being slightly moister (~1 mm day −1 ) during the wet season.Therefore, both HadGEM1.2 and HiGEM1.2 did not present the dry biases which is appointed as a common error of climate models over Amazon (Cavalcanti et al., 2002;Marengo et al., 2003;Seth and Rojas, 2003;Seth et al., 2007), although a small underestimation (~1-2 mm day −1 ) occurs from January-May in the northern part of this basin (see Figure 4).The periods of June-August and December-March are, respectively, the dry and rainy seasons, and HiGEM1.2 and HadGEM1.2 eliminate a common problem in this region that is the simulation of a semiannual precipitation cycle induced by the semiannual cycle of solar forcing (Seth et al., 2007;Bonam et al., 2002).
The simulated and observed annual cycles of precipitation in RCO (Figure 5b) and SDE (Figure 5e) regions are similar, with the rainy (dry) presenting a maximum (minimum) in the period of November-March (May-September).HiGEM1.2 and HadGEM1.2 are in phase with the observed precipitation, but they are slightly wetter during most of the year.In RCO (SDE) HiGEM1.2reduces the HadGEM1.2overestimation of rainfall from May-November (July-November).
In NDE, rainfall presents a different cycle during the year, with the wet season peak occurring in January-March due to the southward displacement of the ITCZ, and with a later decrease to ~1.0 mm day −1 in August (Figure 5d).The main features of this annual cycle are simulated by HiGEM1.2 and HadGEM1.2, but both are wetter (dryer) than the estimated precipitation during part of rainy (dry) season.In this region, HiGEM1.2anticipates the rainy season peak to February (1 month before observed data), while HadGEM1.2 shows a flat peak with two months (February-March) and a smaller wet bias than HiGEM1.2during January.
The systematic errors larger than +2 mm day −1 occur from January-March in the NDE (Figure 5d) and SDE (Figure 5e) regions.In NDE, this is mainly due to the influence of the ITCZ, while in SDE it is attributed to more intense simulated SACZ (see Figure 4a-4e).The larger biases in NDE and SDE rainy seasons are not related with  In the SESA (Figure 5c) region, the observed mean annual cycle of rainfall presents small amplitude (~3 mm day −1 ), with dry period from June-August and wet one from October-April.Both HadGEM1.2 and HiGEM1.2simulated this pattern, but they overestimated rainfall intensity mainly from May-December.The HadGEM1.2 precipitation is closer to the observed precipitation than HiGEM1.2.In this area, the higher resolution of HiGEM1.2 is contributing to increase its wet biases from September-December.However, the time evolution of precipitation annual cycle simulated by HiGEM1.2 and HadGEM1.2 is similar to the observed, representing an improvement compared with nine CMIP models (Seth et al., 2010).In this area, CMIP3 models underestimated winter precipitation (which is attributed to the frontal passage) and they do not represent the phase of the annual cycle.
The areas used to define regions RCO, SDE and SESA (Figure 1) are similar to that used by Seth et al. (2010).Comparatively, in these three regions both HiGEM1.2 and HadGEM1.2 simulate the rainfall annual cycle (both phase and amplitude) in larger agreement with estimated precipitation, which does not occur with CMIP3 models (Seth et al., 2010).For example, in a region similar to SDE, CMIP3 models present large underestimation (2-3 mm day −1 ) of rainfall from September-November.Over the SESA region, CMIP3 models are very dry (−3 mm day −1 ) during austral winter and do not simulate the phase of the rainfall annual cycle (Seth et al., 2010).In the AMZ and RCO regions, Seth et al. (2010) showed that all nine CMIP3 models have a delay in both onset and demise of the rainy season by approximately one month, a problem that does not occur in Figure 5a, 5b.
Over the ocean, in LUR (Figure 5f) the observations (CMAP and GPCP only) show a different rainfall annual cycle that peaks in April (austral autumn) and has a minimum in December (austral summer).In this region, except by the peak in April, both simulations present time evolution of monthly rainfall similar to the observation.However, HiGEM1.2 rainfall intensity is closer to the observation than HadGEM1.2 during a larger part of year.In both SESA and LUR regions there is a different behavior of the models (Figure 5e, 5f) that would be attributed to the high resolution in the HiGEM1.2once this also occurs in the atmosphericonly models (Figures not shown).Figures 4 and 5e, 5f indicate that in these regions of transient systems development the HiGEM1.2high resolution acts to increase (decrease) the rainfall over continent (ocean), but the physical mechanism associated with this feature is not clear at moment.
The small (~ 0.5 mm day −1 ) systematic errors over SESA, mainly of HiGEM1.2,indicate substantial improvement compared with regional model RegCM3 (Reboita et al., 2010a) where rainfall was underestimated up to 50%.In LUR, the higher resolution is contributing to increase rainfall in larger part of year (May-November) making HiGEM1.2closer to the estimated precipitation.This region presents the larger differences among observed precipitation, with amplitude up to 1.5 mm day −1 from June-December.This is related to uncertainties in satellite estimation of rainfall over the ocean.In others regions, Figure 5 shows that the differences between the observed precipitations are in general smaller than 1 mm day −1 .

CONCLUSIONS
In general, the seasonal characteristics of the precipitation and low level circulation spatial patterns over the SA are correctly simulated by the two CGCMs, although both HiGEM1.2 and HadGEM1.2 overestimated rainfall intensity, mainly over the convergence zones (ITCZ and SACZ).Moreover, HadGEM1.2 and HiGEM1.2reproduce the intensity and location of SACZ and its seasonal evolution, indicating an important improvement compared with previous coarse resolution CGCMs (Vera et al., 2006;Seth et al., 2010).
Considering different subdomains (AMZ, NDE, RCO and SDE), the simulated precipitation annual cycles are in phase with the observed ones, although there are some errors in intensity.In other regions (SESA and LUR), even without a defined annual cycle, the simulated precipitation is in phase with the observations.The AMZ is highlighted as the region with smallest biases among all regions and together with SESA and LUR can be considered as a fairly important improvement to justify higher resolution CGCMs.Previous coarse resolution CGCMs (Vera et al., 2006;Seth et al., 2010;Cavalcanti et al., 2002) were unable to simulate some features of the precipitation annual cycle in these regions.Comparatively, the use of high resolution in HiGEM1.2reduces the dry biases in the central part of Brazil (RCO and SDE) during austral winter and spring and over LUR in most part of the year.Besides annual cycle, another important aspect under investigation is the impact of high resolution in the precipitation variability and extreme events.These results will be present elsewhere.

Figure 1 .
Figure 1.South America topography (shaded, in m) and location of six sub domains for annual cycle analysis.

Figure 5 .
Figure 5. Mean annual cycle of precipitation rate (mm day −1 ) in 6 subdomains simulated by HadGEM (blue line) and HiGEM (red line), and for analyses (black line).The vertical line indicates maximum and minimum values from analyses.