The study of seasonal climate, more specifically precipitation, with regional climate models (RCM) has become increasingly popular over the past years (Fennessy and Shukla, 2000; De Sales and Xue 2006; Xue et al., 2007; Rockel et al., 2008). The use of general circulation model (GCM) results as lateral boundary conditions (LBCs) for RCM simulations is called dynamic downscaling (Castro et al., 2005; De Sales and Xue, 2006). Pioneer studies in this area include Dickinson et al. (1989) and Giorgi and Mearns (1991). The goal of dynamic downscaling is to obtain a finer and more realistic representation of regional-scale features associated with topography and land surface features. GCMs, in most cases, do not have sufficient horizontal and vertical resolution to resolve local climate features induced by smaller-scale variations in topography and land cover, which in turn affect precipitation.
This is especially important over South America (SA) where the continent–ocean interaction circulations (e.g. SA monsoon system) and topography-induced circulations (e.g. SA low-level jet) have important regional-scale features and play a crucial role in the formation and distribution of tropical and subtropical precipitation. The use of RCM to investigate SA's seasonal precipitation has shown encouraging results (Chou et al., 2000; Misra et al., 2003; De Sales and Xue, 2006; Seth et al., 2007). The work by De Sales and Xue (2006) indicated that, by dynamically downscaling GCM's outputs through one-way nested RCM simulations, one can reduce the root mean squared error (RMSE) in seasonal mean predictions of precipitation and surface air temperature when compared to GCM alone. The study showed that the only region that did not experience improvement from the downscaling was the eastern Amazon basin.
A less-studied topic associated with dynamicdownscaling techniques is the quantitative evaluation of the downscaling impact at different spatial scales. In the works of Castro et al. (2005) and Rockel et al. (2008), the power spectrum decomposition of column-average moisture flux convergence, kinetic energy, and precipitation were used to assess the added value by dynamic downscaling for a series of simulations over North America. They concluded that the downscaling did not retain the value of the large-scale features over and above that which were already present in the GCM results. In another North American downscaling study, however, Xue et al. (2007) found that the RCM has proper downscaling ability only under certain conditions, such as proper domain boundary location among others. These studies suggested that the purpose of the downscaling was not to add skill to the large scale, but to add value to the smaller-scale features, which have a greater dependence on the land surface characteristics.
For the present study, we use the 80-km resolution NCEP ETA RCM nested in the T62 NCEP GCM for a series of 3-month simulations of the austral summers and winters over South America. We focus the analysis on the precipitation downscaling ability at different scales as simulated by the GCM and RCM. The impact of the Andes Mountain Range on the RCM results is also investigated through a sensitivity test.
The La Plata basin is the focus region in this study. It is analogous to the Amazon basin in terms of its biological diversity, but far exceeds the latter in its economic importance to south-eastern and central South America in terms of hydroelectricity and food production (Vera et al., 2006). The basin comprises almost all the southern part of Brazil, as well as parts of Uruguay, Paraguay, and an extensive part of northern Argentina. The total human population depending on the basin is estimated to be approximately 67 million (Organization of American States). Therefore, the ability to forecast the inter-annual variability of seasonal precipitation is critical for this region.
We apply the intensity-scale verification technique (ISVT) of precipitation (Casati et al., 2004; Casati, 2010) to assess the role of dynamic downscaling on the simulation quantitatively. This verification method evaluates the forecast skill score (SS) and energy of precipitation as a function of different spatial scales and precipitation intensities. The scale components are obtained by a 2-D Haar discrete wavelet transform, whereas different precipitation intensities are selected by thresholding. The ISVT's SS is a generalisation of the Heidke SS based on the mean square error (Wilks, 2006). The precipitation energy is assessed through the Frequency Bias Index (Jolliffe and Stephenson, 2003), and it is used in this study to evaluate the inter-annual difference in the models' results.
Unlike the method utilised by Castro et al. (2005) and Xue et al. (2007), the intensity-scale approach can simultaneously provide information about the downscaling performance of both the size and intensity of precipitation features. Furthermore, results obtained from the wavelet-based approach are bound to be more robust. In fact, because of their locality, Haar's wavelet transforms are less susceptible to Gibb's phenomena often triggered by the Andes steep topography. In addition, wavelets are also more suitable than Fourier transforms to deal with spatially discontinuous fields, such as precipitation, especially in limited-area, non-global domains.
The ISVT has been successfully used in nowcasting and probability forecasting studies (Mittermaier, 2006; Casati and Wilson, 2007; Csima and Ghelli, 2008). This method has not been used to evaluate dynamic-downscaling performance at seasonal time scales. As mentioned above, the method allows the model skills to be diagnosed as a function of the spatial scale of the forecast error and intensity of the rainfall events. Therefore, it is an efficient tool to evaluate large-scale and mesoscale features, as well as intense and weak events in downscaling studies.
Brief descriptions of the models, detailed experimental design, as well as a concise explanation of the ISVT are presented in the next section. The impact of dynamic downscaling on the seasonal average precipitation is examined in Section 3.1, while in Section 3.2 we assess its role in the precipitation inter-annual variability. The effect of the Andes Mountain Range topography on the RCM simulation is described in Section 4. The final discussions and conclusions are presented in Section 5.
2. Models and experimental design
2.1. Atmospheric models
The version of the NCEP GCM utilised in this study was similar to the one used by De Sales and Xue (2006). The model was set up on a triangular 62-truncation horizontal resolution and 28 vertical levels. The model includes Chou (1992) and Chou and Suarez (1994) radiation scheme, Hong and Pan (1996) non-local planetary boundary scheme, and the Moorthi and Suarez (1992) convection scheme.
As for the higher-resolution RCM, we utilised the NCEP limited-area ETA model. The ETA is a state-of-the-art atmospheric model used for research and operational purposes. This model evolved from the earlier Hydrometeorological Institute and Belgrade University model with step-like mountain vertical coordinates (Mesinger et al., 1988; Janjic, 1990). The model's code has since been upgraded to include more advanced schemes such as the Arakawa-style horizontal advection scheme (Janjic, 1984), a radiation scheme based on Lacis and Hansen (1974) and Fels and Schwartzkopf (1975); a Kolmogorov–Heisenberg-type closure scheme to represent turbulence in the planetary boundary layer and in the free atmosphere. In terms of precipitation, the model utilises the Betts–Miller–Janjic scheme for deep and shallow moist convection (Betts, 1986; Janjic, 1994), and a grid-scale precipitation scheme based on Zhao and Carr (1997).
For this study, the ETA model was set up on an 80-km horizontal resolution and 38 vertical levels grid covering most of the South American continent and the adjoining Atlantic and Pacific Oceans. Further detailed information on the GCM and ETA model can be found in the works by Kanamitsu et al. (2002a) and Black (1994).
Both models were modified to include a more sophisticated land surface processes parameterisation scheme (Xue et al., 2001, 2004). The Simplified Simple Biosphere model version 1 (SSiB-1, Xue et al., 1991) was used as the land surface processes model in all simulations. In addition to simulating processes, such as runoff, direct vegetation and bare soil evaporation, and photosynthesis-controlled canopy transpiration, this model also ensures energy, water, and momentum conservation at the atmosphere–land surface interface.
The SSiB-coupled versions of GCM and ETA models have been extensively tested (Xue et al., 2004, 2006, 2007). Information regarding the ETA/SSiB-1 coupling, vegetation classification, and vegetation parameters for the SSiB-1 can be found in Xue et al. (2001). Hereafter, for simplicity, the coupled versions of the NCEP GCM/SSiB-1 and ETA/SSiB-1 are referred to as GCM and RCM respectively.
2.2. Experimental design
The austral summer (December through February) and winter (June through August) of 1988 and 1997 were selected for this study because they exhibited very large inter-annual differences in precipitation over tropical and subtropical South America. The winter of 1997 and the summer of 1997–1998 (referred to as JJA97 and DJF97–98, respectively, hereafter) witnessed the strongest El Niño of the twentieth century (Oceanic Niño Index (ONI) equal to 1.7 and 2.4, respectively). Similarly, the winter of 1988 and summer of 1988–1989 (JJA88 and DJF88–89, hereafter) were characterised by a strong La Niña event (ONI equal to − 1.2 and − 1.7, respectively). The investigation of such intense El Niño-Southern Oscillation events would be very instructive in demonstrating the GCM and RCM abilities to simulate precipitation inter-annual differences.
Five-member ensemble integrations were performed with the GCM and RCM for each season studied, starting from slightly different initial conditions. Table I lists the initial and final dates of each ensemble member. The results presented herein are actually based on the five-member ensemble mean of each experiment. The GCM integrations were carried out first and results were saved every 6 h. The GCM integrations were initialised from NCEP-DOE AMIP-II global reanalysis (Kanamitsu et al., 2002b) at 00 UTC of each starting date as shown in Table I.
Table I. Initial and final dates of ensemble members used in the GCM and RCM simulations
29 Nov 1988 to 01 Mar 1989
29 May 1988 to 01 Sep 1988
30 Nov 1988 to 01 Mar 1989
30 May 1988 to 01 Sep 1988
01 Dec 1988 to 01 Mar 1989
01 Jun 1988 to 01 Sep 1988
02 Dec 1988 to 01 Mar 1989
02 Jun 1988 to 01 Sep 1988
03 Dec 1988 to 01 Mar 1989
03 Jun 1988 to 01 Sep 1988
29 Nov 1997 to 01 Mar 1998
29 May 1997 to 01 Sep 1997
30 Nov 1997 to 01 Mar 1998
30 May 1997 to 01 Sep 1997
01 Dec 1997 to 01 Mar 1998
01 Jun 1997 to 01 Sep 1997
02 Dec 1997 to 01 Mar 1998
02 Jun 1997 to 01 Sep 1997
03 Dec 1997 to 01 Mar 1998
03 Jun 1997 to 01 Sep 1997
For each of the GCM integrations, a downscaling integration was performed with the ETA model, starting from the same initial date and initial conditions as the GCM. The RCM's LBCs were updated every 6 h of simulation from the GCM simulations output. The initial conditions for soil temperature and wetness, initial snow cover, as well as daily sea surface temperature (SST) and sea ice concentrations for all experiments were also taken from the NCEP-DOE AMIP-II global reanalysis.
To validate the simulations, the model results were compared against the National Oceanic and Atmospheric Administration (NOAA) NCEP Climate Prediction Center (CPC) observed daily precipitation analysis (Shi et al., 2001; Silva et al., 2007). This dataset contains daily precipitation over land for the entire South American continent. Figure 1 shows the distribution of rain gauge for the time period covered in this study. The minimum number of stations for the daily precipitation analysis is 250. If the number of stations is less than the minimum, then the analysis is not performed for that day. The CPC daily precipitation analysis is available on a 1° by 1° horizontal grid. All model fields were interpolated to the CPC analysis grid for comparison purposes. As CPC precipitation data only covers land points, our analysis is limited to precipitation over the continent.
2.3. The intensity-scale verification technique
The ISVT (Casati et al., 2004; Casati, 2010) was used to assess the models' performances in terms of seasonal and inter-annual precipitation simulations. This technique assesses the forecast skill and bias of precipitation features as a function of the spatial scale and intensity. Unlike decomposition methods that use Fourier transforms, the ISVT is based on a categorical distribution approach, which is a more suitable and robust method to deal with spatially discontinuous and sparse fields, such as precipitation, which are also characterised by highly skewed intensity distributions. Furthermore, discrete wavelet decompositions are orthogonal, which allows for the decomposed verification statistics to be additive (Casati, 2010).
Thresholding is initially used to convert the observed and modelled precipitation fields into binary fields, Io and Im respectively. These fields are transformed into binary images on a precipitation/no-precipitation basis for a given precipitation rate threshold (τ). The difference between binary modelled and observed precipitation is defined as the binary error field (Z = Io − Im). The 2-D Haar-wavelet transform is then used to decompose the binary error field into the sum of components at different spatial scales (l). For a given threshold, the mean squared error (MSE) of the binary field is defined by the average of all the squared differences over all the grid-points
where is the MSE of the lth spatial scale component of the binary error field for threshold τ the overbar indicates averaging of the squared values over all the grid-points. For each precipitation threshold and spatial scale, the SS is calculated as
where MSErandom, τ represents the MSE of a random forecast calculated on the basis of the bias and base rate for each threshold and equipartitioned across all scales (Casati, 2010); and MSEbest is the MSE associated with a perfect simulation (MSEbest = 0). The double overbars represent temporal aggregation over the 3-month length of simulation as described in Casati (2010).
Generally, intensity-scale decompositions of SS tend to show low skill for very intense rainfall rates and very small spatial scales, which represent small-scale convective storms and are more difficult to simulate. In contrast, frontal and non-convective large precipitation systems are more often properly simulated, thus yielding higher scores. Negative scores are associated with the model's performances that are not better than a random prediction.
The observed and modelled precipitation energies are evaluated in a similar fashion as MSEτ, l from the thresholded precipitation binary fields decomposed into different spatial scale components by the same wavelet transform. For instance, PEτ, l = I2 is the precipitation energy associated with the lth spatial scale component of the observed binary field obtained for threshold τ. The same notation is used for the model energy. Precipitation energy is used in this study to evaluate the precipitation inter-annual differences between the summer and winter seasons of 1997 and 1988 through the precipitation energy relative difference (ERD) as defined below.
where the double overbars represent temporal aggregation over the length of simulation. The precipitation ERD decomposition is proportional to the number of precipitation events exceeding the threshold τ and is associated with features at scales l. The ERD thus provides insightful information on the intensity-scale nature of the mechanisms associated with the inter-annual variability for the summer and winter precipitation. By comparing ERD from observation and simulation, the downscaling ability in inter-annual variation at different spatial scales and precipitation intensity can be validated.
We applied the ISVT over two areas of interest, i.e. southern and northern South America, area 1 and area 2 respectively, as depicted in Figure 1. Each area is 32° latitude by 32° longitude wide, which yielded a total of 6 Haar-wavelet components, l = 1, 2, …6, corresponding to 1°, 2°, 4°, 8°, 16°, and 32° resolutions respectively. The 1° component is beyond the GCM's resolving capability, and, therefore, it is not included in the analysis. It is important to note that the 32° component corresponds to the average over the entire area of interest. The decision to separate the results into two areas was based on the fact that precipitation in southern (subtropical) and northern (tropical) South America are associated with different physical processes of different spatial scales. While in area 2, convective storms are the prevalent source of precipitation, a comparable mix of frontal and convective precipitation occurred in area 1.
3. Analyses of results
In this section, we discuss the validation results from seasonal mean and inter-annual variability perspectives. Summer (winter) seasonal mean precipitation in this paper is referred to the average between DJF88–89 and DJF97–98 (JJA88 and JJA97) daily model forecasts, while summer (winter) seasonal difference indicates DJF97–98 minus DJF88–89 (JJA97 minus JJA88) for models and observation.
3.1. Quantitative assessment of dynamic downscaling for seasonal precipitation
Observed and modelled summer seasonal average precipitations are shown in Figure 2(a)–(c). CPC data show a typical wet-season rainfall pattern over South America with precipitation totals above 6 mm day−1 over the central portion of the continent, extending northward towards the Amazon River mouth, and south-eastward towards south-east Brazil. Areas with average precipitation above 4 mm day−1 extended southwards into the La Plata basin. Dry areas include north-east Brazil, extreme northern South America, and the southern Andes Mountain Range.
The GCM shifted the intense precipitation area north-eastward, overestimating the rainfall over eastern Brazil. On the other hand, the global model underestimated the rainfall over the La Plata basin and western Amazon basin. Figure 2(b) also shows an area of intense rainfall (>10 mm day−1) along the eastern slopes of the northern Andes for the GCM, which only marginally appeared in the CPC product.
The RCM downscaling improved the results over the subtropical section of the continent, especially along the eastern slopes of the Andes and southern Brazil, where the regional model increased the total rainfall for the summer season (Figure 2(c)). However, the regional model was not able to simulate the rainfall over northern South America very well, which was better represented by the GCM. Both models captured the dry areas over the south-west well. Chou et al. (2005) reached similar results based on 4-month long simulations (November to February) with a 40-km resolution version of the ETA model and a different GCM for the 2002–2003 wet season.
The lack of rainfall in the RCM over northern South America is also noticeable in a number of studies (Berbery and Collini, 2000; Rojas and Seth, 2003; Chou et al., 2005; De Sales and Xue, 2006; Seth et al., 2007). It has been suggested to be caused by mass imbalance led by the one-way nesting configuration (De Sales and Xue, 2006), as well as deficient parameterisation of convective processes and biases in the initial soil moisture distribution over the tropical areas (Berbery and Collini, 2000). Seth et al. (2007), in a set of multi-annual downscaling simulations, also found a similar dry bias over the Amazon region. They concluded that the bias was associated with weak moisture transport from the ocean due to improper representation by the regional model of the sea-level pressure over the equatorial Atlantic, which in turn reduced the strength of the trade winds in the region. In addition, Seth et al. (2007) also showed that the Amazon dry bias is very sensitive to the RCM's convective scheme.
A detailed investigation of the mechanisms behind this issue is beyond the scope of this article and is not discussed further in this paper. Nevertheless, the intensity-scale verification analysis of the RCM SSs, discussed next, illustrates the nature of the precipitation associated with the model performance over northern South America.
Winter seasonal average precipitations are shown in Figure 2(d)–(f). The observation indicates that most of the precipitation was concentrated over the northern sections of the continent between 10°S and 10°N. Significant precipitation was also measured over the La Plata basin region, southern Andes, and the north-eastern littoral of Brazil. Most of the continent interior centre received only less than 1 mm day−1 of precipitation (Figure 2(d)).
The global and regional models were able to capture the main features of the wintertime precipitation. The GCM overestimated the rainfall over in the north-eastern South America, while generally underestimated it in the La Plata basin (Figure 2(e)). The downscaling results were more comparable to the observation over south-eastern and north-eastern South America. The regional model also produced some precipitation along the eastern slopes of the central Andes, which is in the CPC product but is absent in the GCM results (Figure 2(f)). The wintertime differences between GCM and RCM, especially the rainfall increase over the La Plata basin, are similar to the results in Chou et al. (2005), which describe 40-km ETA model simulations for May to August of 2002.
On the basis of Figure 2, the average summer precipitation MSEs over area 1 for the GCM and RCM were respectively 3.33 and 1.56 mm2 day−2. As for area 2, the average MSEs were 10.34 and 16.89 mm2 day−2, respectively. These numbers represent a decrease in MSE of approximately 53% over southern South America and an increase of about 63% in MSE over northern South America due to downscaling. Winter precipitation MSE results for the GCM and RCM were respectively 1.24 and 0.85 mm2 day−2 over area 1 (31% decrease) and 4.14 and 6.53 mm2 day−2 over area 2 (58% increase).
To provide a quantitatively and detailed measure of the dynamic-downscaling performance, we utilise the ISVT. The intensity-scale decompositions of summer precipitation SS for the GCM and RCM over southern South America (area 1) are shown in Figure 3(a) and (b). Lighter grey areas indicate higher scores; negative scores are blackened and indicate the portion of the spectrum for which the model's performance is not better than random prediction. Scores equal to or higher than 0.8 appear white.
In general, for both models, the scores decrease as precipitation thresholds increase and spatial scales decrease. Large thresholds at small scales are associated with intense localised convective storm, which are more difficult to simulate at the model resolutions utilised in this study. As the spatial scale increases, the scores for a given rainfall threshold also increase. The higher scores are found for very large and weak rainfall thresholds, which include frontal and non-convective synoptic rainfall features. At 32° spatial scale, both models show reasonably skilful results (above 0.8) for all rainfall intensities. As previously mentioned, this scale represents the SSs for the average precipitation over the entire area 1.
The black shaded areas on the SS spectra indicate the spatial scales and thresholds for which the models show no skill in terms of precipitation simulation. These include thresholds higher than 10 mm day−1 at scales smaller than 8° for the GCM and 4° for the RCM. Therefore, the no-skill spectrum for the RCM is smaller than that for the GCM, indicating an improvement in simulating intense summer rainfall events over area 1 with the downscaling.
Figure 3(c) and (d) exhibits the summer precipitation's intensity-scale decompositions of SS over northern South America (area 2). Similar to area 1, higher scores are associated with weak rainfall thresholds and large rainfall events. However, RCM's scores at area 2 exhibit a significant decrement at weaker rainfall thresholds when compared to that of GCM. The intensity-scale SS spectrum shows that the RCM had some difficulties in simulating the average rainfall over area 2, especially at higher-intensity thresholds. There was still some reduction in the no-skill (negative scores) region of the spectrum by the downscaling.
To clearly delineate the comparison between models' performances, Figure 4(a) and (b) display the difference between RCM and GCM scores for the summer simulations over the two areas of interest. To highlight the major differences between models, only differences higher than 0.1 and smaller than − 0.1 are shown. For area 1, the downscaling improvement was concentrated for spatial scales between 4° and 16° at different thresholds. Such scales correspond to nearly 2 to 8.5 times the GCM's horizontal resolution, which is approximately 1.875°. The RCM's largest improvement was for rainfall events at 8° spatial scale. Meanwhile, improvement was also observed in high-precipitation events around 4° and in middle-to-high precipitation events around 16°.
On the other hand, SS differences over northern South America indicate that the downscaling improved the simulation results for only a small fraction of the rainfall spectrum. The improvement associated with downscaling was mostly restricted to intense rainfall events (above 4 mm day−1) of spatial scale of 4°, which correspond to about two times the GCM's resolution. For most of the spectrum, however, the GCM exhibits higher forecasting skills, especially for weak rainfall thresholds, below 4 mm day−1. The GCM's scores over area 2 were significantly higher for nearly all rainfall rates. This result agrees with Figure 2(c), which shows a large precipitation bias over the Amazon basin in the RCM results. The SS decomposition clearly indicates that the biases were due to the RCM's poor simulation of mid-intensity to weak precipitation thresholds events (between 0.1 and 4.0 mm day−1), which comprise a large part of the rainfall over the region.
Winter precipitation SS decompositions were also calculated for the same two areas (Figure 5). Observation of Figure 5(a) and (b) shows a shrinking of the no-skill portion of the spectrum after downscaling over area 1. This is more evident for rainfall events at scales between 4° and 8° at almost every precipitation intensity event. The same occurred for precipitation rates at and above 12 mm day−1, which the GCM could not produce at all over this area in the 1988 simulations. Figure 5(c) shows that, over area 2, such an improvement was not as evident by the RCM. Overall, GCM exhibits higher score in this area during winter (Figure 5(c) and (d)).
The difference between winter scores of RCMs and GCMs over area 1 (Figure 6(a)) shows that downscaling improvement was more widespread spectrum wise, than that during summer. Most of the positive differences were not concentrated at the centre of the spectrum as for summer, but reached nearly all rainfall rates and scales. The largest differences occurred for scales between 4° and 8°. Again, the SS decomposition corroborates with the winter average maps, which indicated better results by the RCM over area 1. Unlike seasonal average maps, the SS decomposition explicitly indicates which rainfall events actually benefited from downscaling; in this case, mostly events of sizes between about 4° and 8°.
Figure 6(b) shows the difference between RCM and GCM winter precipitation scores over area 2. The most distinct differences are located at very large and very small rainfall spatial scales, for which the GCM produced better simulations. Downscaling improvements are restricted to rainfall features at 4° of size for thresholds between 6 and 10 mm day−1. Figure 2(f) indicates a negative bias in winter rainfall simulated by the regional model over area 2. While both observation and GCM exhibit large areas of seasonal rainfall over 6 mm day−1, the RCM barely shows signs of rainfall over 4 mm day−1. Winter SS decomposition suggests that the dry bias resulted from RCM's deficiency in simulating large- and small-scale rainfall features.
3.2. Assessing RCM's downscaling of inter-annual precipitation differences
The simulated and observed average precipitation difference over land between DJF97–98 and DJF88–89 is shown in Figure 7. The CPC product exhibits a typical El Niño/La Niña summer/winter precipitation difference pattern in South America (Zhou and Lau, 2001). Negative differences are found over northern South America, including in the Amazon basin. Some locations exhibited over − 4 mm day−1 in precipitation difference. In contrast, positive differences were found over north-western South America, between 10°S and the equator, the La Plata basin, as well as over central-eastern South America between 20°S and 10°S. The La Plata basin has the largest positive differences.
The GCM and RCM precipitation inter-annual difference is shown in Figure 7(b) and (c) respectively. The shaded areas in the figure indicate that the differences were consistent over the 95% confidence levels, based on a 2-tailed Student's t-test. The GCM captured the negative difference north of the equator and the positive difference along the coastal Ecuador. The model also gives some indication of negative difference in the Amazon basin. However, it does not reproduce the differences over the La Plata basin. Moreover, the GCM results exhibit positive differences over the north-eastern tip of South America, which is inconsistent with observations.
The downscaling significantly improved the spatial pattern and intensity of the precipitation inter-annual difference in the La Plata basin. The average precipitation differences (DJF97–98 minus DJF88–89) over area 1 (Figure 1), which includes the La Plata basin, are 0.08 and 0.55 mm day−1 for the GCM and RCM respectively. The same average equals to 1.65 mm day−1 for the CPC product. In general, the RCM produced better results than the GCM over eastern and subtropical South America. In contrast, the RCM did not capture the large area of negative differences over the Amazon basin and northern South America. The average precipitation differences over area 2 (Figure 1), which consists of the Amazon basin and northern South America, for the CPC, GCM, and the RCM are − 1.20, − 1.11, and − 0.06 mm day−1 respectively.
Figure 7(d) shows the observed precipitation difference between JJA97 and JJA88. While both models were able to simulate the negative difference in rainfall in northern South America, only the RCM was able to simulate the precipitation difference in the south-east. The winter average precipitation differences over area 2 from the CPC observation, GCM, and RCM were − 1.33, − 1.71, and − 1.24 mm day−1, respectively. On the other hand, over area 1 the average precipitation differences were 1.19, 0.46, and 1.08 mm day−1, respectively.
As discussed in Section 2.3, the energy of precipitation is directly related to the amount of rainfall events during a given period. Therefore, the intensity-scale decomposition of precipitation ERD can be used to assess quantitatively the intensities and spatial scales of precipitation features that contributed to the total difference between two given time periods. As the spatial scale is intimately related to physical processes behind a precipitation event, the comparison between observation and model precipitation ERD can provide an assessment on the model's ability to simulate such processes. Only differences larger than ± 0.1 are shown to highlight the major intensity-scale differences.
Figure 8(a) shows the observed ERD of precipitation, as defined in Equation 3, between DJF97–98 and DJF88–89 over area 1. Positive (negative) values indicate a larger number of rainfall events in DJF97–98 (DJF88–89). The observation data exhibits the largest difference between the two summers at larger rainfall thresholds. This indicates that intense rainfall events contribute more to the overall inter-annual differences.
The GCM shows very weak ERD (Figure 8(b)), indicating that the model could not simulate the difference in intense rainfall between the two summers. On the other hand, the RCM improved the results by correctly simulating the partition of precipitation ERD towards higher rainfall thresholds. The 32° spectral component shows that the RCM simulated the average rainfall difference in area 1 better by increasing the number of intense rainfall events, which corroborates with the average difference shown in Figure 7(c).
For area 2, the overall negative values indicate a larger number of intense precipitation events in the summer of 1988 over that region (Figure 8(d)). The largest observed differences occurred at very intense rainfall thresholds throughout the spatial scale spectrum. Even at very small scales, the differences were significant. The global model simulated the increase in intense rainfall events during DJF88–89 over area 2 generally well (Figure 8(e)). The downscaling results also show differences for intense precipitation events. However, the RCM's differences are rather weak (Figure 8(f)). In terms of the average over the entire area, the downscaling underestimated the mean precipitation difference, which agrees with the results shown in Figure 7(c).
Winter precipitation ERD over area 1 can be found in Figure 9. Observation shows an increase in JJA97 rainfall for most of the decomposition spectrum. The largest differences occurred at high rainfall thresholds. Both models exhibit an overall positive difference between JJA97 and JJA88 in area 1 (Figure 9(b) and (c)). The GCM was not able to simulate the rainfall events with intensities above 8 mm day−1 in 1988 over this area; therefore, it was impossible to compare the precipitation ERD for this threshold. The RCM, on the other hand, captured the precipitation ERD at such high rainfall rates well.
The winter precipitation ERDs, in area 2, were similar to the summer results. In general, there were a larger number of intense rainfall events in 1988 when compared to 1997. Weaker rainfall events also played a role on the overall JJA88 higher rainfall totals. The GCM captured the increase in intense convective events well but tended to overestimate the difference at very high thresholds, which led to the larger-than-observed seasonal rainfall difference (Figure 7(e)). The RCM exhibits positive differences for thresholds larger than 12 mm day−1, compared to observation and GCM, indicating downscaling deficiency in simulating the inter-annual difference of very strong convective events over this area (Figure 9(f)).
The above-mentioned inter-annual difference analysis indicates that, in general, the downscaling approach improved the results over southern South America, whereas the GCM provided better results over northern South America. While the poor results over the northern latitudes by the RCM can probably be explained by the model's deficiency in reproducing the inter-tropical convergence zone (ITCZ) rainfall and its inter-annual variability as indicated by the seasonal average precipitation results, and corroborated by the intensity-scale analyses of precipitation SS; the RCM improvement over subtropical South America, both in terms of higher scores and better simulation of precipitation inter-annual variability, as shown by the ERD, is thought to be directly linked with the RCM's more realistic representation of the topography and land surface characteristics (De Sales and Xue, 2006).
4. RCM topography and precipitation simulation
The precipitation SS and ERD decompositions proved that the dynamic downscaling improved the precipitation simulation over the La Plata basin. It also showed that the same was not true for the Amazon basin. The works of Nogués-Paegle and Berbery (2000) and Marengo et al. (2004) showed that the South American low-level jet (SALLJ) is an important source of moisture to the La Plata basin region. The SALLJ can be characterised as a narrow stream that channels the moisture flux in the lower troposphere between the Amazon and the La Plata basins, resulting from the deflection of easterlies winds by the Andes Mountain Range (Marengo et al., 2004).
De Sales and Xue (2006), using a similar downscaling configuration, showed that the RCM produced better simulations of strength and position of the SALLJ seasonal average when compared to the GCM's results alone, which led to improved simulation of seasonal average rainfall at the SALLJ's exit region. In this study, we further explore the possible cause of this improvement by exploring the impact of the Andes topography on the RCM's precipitation inter-annual differences over the La Plata basin.
Figure 10 shows the cross sections of meridional moisture flux inter-annual differences for both summer and winter seasons. The moisture flux differences by The European Centre for Medium-range Weather Forecasts (ECMWF) reanalysis (Uppala et al., 2005), hereafter ERA-40, agree with the CPC rainfall observation and show an increase in the northerly flow associated with the SALLJ for the summer and winter of 1997 (Figure 10(a) and (d)), which contributed to more precipitation over the La Plata basin for DJF97–98 and JJA97. A second core in moisture flux difference appears over the Atlantic in both seasons. The differences are larger over the continent (>4 × 10−2 kg m s−1 kg−1). While the GCM could not simulate the moisture flux differences, the RCM simulations captured the position of the two difference maxima well, despite producing weaker inter-annual differences.
The GCM and RCM used in this study have different moist physical processes, radiation, and planetary boundary layer (PBL) schemes, which control the formation of precipitation, and thus the complete causes for the RCM's improvements cannot be explained solely by the topography difference. However, we can speculate that the Andes topography played an important role in RCM's precipitation results over the La Plata basin. To verify the impact of Andes Mountain Range topography on the RCM's precipitation simulation in the La Plata basin, a sensitivity test was performed in which the topography height (Zs) west of 60°W was lowered by a factor (f) according to following rule:
Topography heights lower than 3500 m were not changed. This rule was chosen so that the resulting topography resembled that from the GCM model. Figure 11(b) shows the difference in topography height between lowered topography experiments (RCM-tp) and the original RCM setup. Next, we describe the results of these experiments on summer and winter seasonal rainfall over the continent.
Figure 11 shows the RCM-tp's summer and winter seasonal precipitation. Overall, the RCM-tp produced less precipitation over western Amazon River basin, southern and central South America than original topography for both seasons. The dry bias over most of north-west is still present in the RCM-tp summer results (compare Figures 2(c) and 11(a)), which suggests that this bias is probably not associated with the model's topography. The summer and winter rainfall amounts along the eastern slopes of the central Andes are significantly reduced in the RCM-tp results. Such a difference is more remarkable for winter between 5°S and 20°S, where precipitation is absent in the low-topography simulations (Figure 11(b)). The lower Andes topography also reduced the rainfall over south-eastern South America and over western Amazon basin during winter. During summer, the RCM-tp produced significantly weaker moisture flux over land and over South America's Atlantic coast than the original topography results (Figure 12(a)). RCM-tp northerly jets were not only weaker but also shallower than those with original topography. Similar weaker moisture flux results were also obtained for the RCM-tp winter simulations (Figure 12(b)).
It should be highlighted, however, that the precipitation reduction over the Plata basin in the RCM-tp experiments cannot be solely explained by the moisture flux reduction associated with the lower Andean topography. Other mechanisms such as baroclinic instability and topography-induced cyclogenesis, which are intimately related to the topography height, could also play an important role on the precipitation reduction at the lee side of the Andes. The investigation of these mechanisms is beyond the scope of this study.
To examine the impact of the Andes topography on the RCM's precipitation downscaling performance over the area 1 in more detail, we compare the precipitation SS decompositions between RCM-tp and RCM. Figure 13 shows the precipitation SS difference between RCM-tp and RCM for summer and winter over the La Plata basin. Summer differences indicate worsening of the scores for rainfall events larger than 4°, especially at mid-intensity thresholds. Very small scale features do not seem to be strongly affected by the Andes topography height, except for high-intensity thresholds for which the RCM-tp produced higher scores.
For the winter season, RCM-tp worsened the precipitation scores for small-scale events at high thresholds, while RCM-tp large-scale rainfall events exhibited higher scores than those with original topography (Figure 13(b)). These results indicate that the proper representation of the Andes Mountain Range is a crucial contributor to this RCM's improvement in the La Plata basin. Higher Andes topography produced stronger moisture-laden SALLJ from tropical latitudes towards the La Plata basin, thus promoting rainfall events there. Precipitation SS decompositions showed that a stronger SALLJ improved the simulation of mid- to large-scale rainfall events over the basin during summer, and improved the simulation of small-scale precipitation events during winter. This result suggests that the SALLJ's role as a source of moisture for the La Plata basin's rainfall differs significantly between seasons.
The number of rainfall events, as indicated by the precipitation energy over the La Plata basin, was also impacted by the lower Andes topography in RCM-tp. Precipitation energy spectra indicated that nearly all rainfall spatial scales exhibited a reduction in the number of events in RCM-tp runs (not shown). The reduction in precipitation events was not, however, homogeneous throughout the rainfall intensity rates. Figure 14 shows the impact of the Andean topography height on the precipitation energy per intensity rate. Summer results exhibited a reduction in the number of rainfall events for all thresholds. The largest reduction occurred for rainfall thresholds between 2 and 8 mm day−1. Overall, RCM results were more comparable to observation than RCM-tp for all rainfall thresholds.
As for the winter results, RCM-tp tended to reduce the number of weak rainfall events more than the strong ones. The precipitation energy distribution indicates that original topography RCM runs overestimated the number of weak rainfall thresholds over the La Plata basin. RCM-tp results were more comparable to observation for those thresholds (Figure 14(b)). However, for stronger rainfall (over 3 mm day−1), the original topography RCM generated better rainfall counts, leading to better overall average precipitation simulation. These results suggest that the SALLJ has a different role on the formation of strong and weak rainfall system over the La Plata basin during summer and winter respectively.
5. Discussion and conclusions
The dynamic-downscaling method was used to investigate the simulation of the seasonal precipitation and inter-annual precipitation difference over South America. The NCEP regional circulation model (RCM) ETA, at an 80-km horizontal resolution, was nested in the T-62 NCEP GCM on a one-way nesting setup for a series of 3-month-long simulations. The austral summer (December through February) of 1988–1989 and 1997–1998 and winter (June through August) of 1988 and 1997 were selected for this study because these two years exhibited very large inter-annual differences in precipitation over tropical and subtropical South America. NCEP-DOE AMIP-II global reanalysis was used as an initial condition for all model runs.
The ISVT was used to quantitatively assess the seasonal precipitation and precipitation inter-annual difference between the two summers and winters at different spatial scales and precipitation intensities over two areas of interest. One area was located over southern South America, while the second was centred over north-western South America. Area 1 encloses the La Plata basin, which is of great economic importance to south-eastern and central South America in terms of hydroelectricity and food production; whereas area 2 is dominated by the Amazon basin, which is characterised by its abundant biodiversity and crucial role on South America's climate.
The summer precipitation intensity-scale analysis for the La Plata basin indicated that, in general, the RCM produced higher scores for some rainfall features with scales mainly around 8°. The RCM's largest improvements were for rainfall thresholds higher than 4 mm day−1. Also, the RCM reduced the no-skill portion of rainfall spectrum, indicating improvement in the simulation of intense localised events. For winter, the RCM improvement was more widespread spectrum wise. The largest improvements were associated with rainfall thresholds at almost every category at 4° to 8° spatial scales. On the other hand, over the Amazon basin, the GCM generated better results for all precipitation rates during both summer and winter experiments, except for intense rainfall events at 4°. According to the SS intensity-scale decomposition, the prevalent dry bias over northern South America in the RCM runs resulted from the model's deficiency to simulate large spatial scale weak rainfall events in that area.
On the basis of multi-decade downscaling simulations over South America, Seth et al. (2007) concluded that, where large-scale SST-forced variability was strong and GCM performed well, RCM had difficulties in improving the large-scale precipitation. On the other hand, in regions where local physical processes were important and GCM performed less well, downscaling showed potential to improve the GCM results. Although our study covers only four seasons, the results presented herein generally agree with those in Seth et al. (2007). ISVT showed that the RCM's largest improvements occur for spatial scales associated with local physical processes (400 to 800 km). On the other hand, for larger spatial scales, the RCM overall results did not show consistent improvement when compared with the GCM runs, expect over the La Plata basin during winter, where RCM's improvement also included some large spatial scales.
In terms of inter-annual differences between two summers (DJF97–98 and DJF88–89) and two winters (JJA97 and JJA88), the RCM significantly improved the precipitation difference predictions in some parts of South America. Overall, downscaling produced better results than the GCM over eastern, north-western, and subtropical South America for both summer and winter. In contrast, the RCM underestimated the negative differences for northern South America in both seasons.
The intensity-scale analysis of precipitation ERD made evident which rainfall rates and spatial scales contributed more to the overall rainfall inter-annual differences. For both areas and for either season, the precipitation differences between the wet and dry years were mainly associated with differences in the number of intense precipitation events. This was especially evident for the summer season. Weaker rainfall events contributed more to winter rainfall differences over area 1. Analogous to the SS decomposition analysis, in general, the RCM produced better results over southern South America by simulating the increase in intense precipitation events in the wet year correctly, when compared with the dry one for both seasons.
The comparison between meridional moisture flux cross sections across central South America indicated that the enhanced precipitation in JJA97 and DJF97–98 in the RCM simulations over the La Plata basin was caused by augmented moisture advection associated with the SALLJ during those months compared to JJA88 and DJF88–89. The moisture transport analyses corroborate the inter-annual precipitation difference analysis, and raised a question about the role of the Andean mountains topography on the RCM simulations.
A sensitivity test was carried out to examine the impact of the Andes Mountain Range on the downscaling results. Topography elevations along the western coast of South America were lowered to similar heights to that generated by the GCM. In general, the lower topography experiment (RCM-tp) produced less precipitation over western Amazon basin, southern, and central Brazil than the original topography simulations for both summer and winter seasons. The dry bias over northern South America was still present in the summer results with low topography, which suggests that this bias is probably not associated with the model's topography. For both seasons, the lower topography experiments produced weaker and shallower SALLJ, which contributed to the rainfall reduction in the La Plata basin.
The intensity-scale distribution of precipitation SS and precipitation energy over the La Plata basin also exhibited strong sensitivity to the Andes topography. In the summer, RCM-tp deteriorated the scores at mid- to large-scale rainfall features over the area (Figure 13(a)), while in winter it lowered the SSs for small-scale and intense rainfall events. Furthermore, the sensitivity test reduced the total number of rainfall events throughout the spectrum of scales; however, such reduction was not uniform with respect to rainfall rates. The RCM-tp test reduced the number of mid-to-high intensity rainfall features more during summer, but especially tended to reduce the number of weak rainfall events during winter.
The precipitation SS and ERD decomposition analyses presented in this paper have indicated that the ISVT can be a valuable tool in the analysis of dynamic-downscaling simulations, by allowing the models' performances to be diagnosed as a function of the spatial scale and intensity of the rainfall events. In general, the results showed that the downscaling can be useful in improving the simulation of South America's seasonal rainfall for certain scales and intensities.
The authors would like to thank Dr Barbara Casati for the invaluable discussions and insightful advices. This research was sponsored by the NOAA's grants NA05OAR4310010, NA07OAR4310226, and NA08 OAR4310591. The model simulations were carried out at National Center for Atmospheric Research (NCAR's) supercomputers.