A common problem with global climate models is the fact that their grids do not always resolve important topographic features which determine the spatial variability of rainfall at regional scales. Here we present and compare simulations of rainfall for the relatively mountainous New Guinea region from six relatively coarse resolution climate models and the corresponding results using a higher resolution model (the Conformal Cubic Atmospheric Model—CCAM). While the large-scale climatological mean rainfall from both the coarse models and CCAM tend to be similar, unsurprisingly, the CCAM results better reflect some of the important topographic effects. However, the results for projected changes (under the A2 emissions scenario) to rainfall for later this century reveal some important differences. The coarse-scale results indicate relatively smooth patterns of projected change consistent with the representations of the underlying topography, but over New Guinea, there is little agreement on the sign of the change. The CCAM projections show greater spatial detail and better agreement among the six members. These indicate that West Papua and the relatively wet northern and southern mountain slopes may get wetter during December to February—the peak of the Austral monsoon season, and the highland regions may actually become drier during June to August—the dry season. These results are consistent with the theoretical concept that warmer temperatures may lead to increases over already wet regions and decreases over the relatively drier regions—the so-called “rich-get-richer” mechanism. They also highlight the fact that the climate of mountainous regions can be relatively complex and indicate potential difficulties that can arise when attempting to synthesize regional-scale projections from coarse-scale models.
 In an analysis of results from phase 3 of the Coupled Model Intercomparison Project (CMIP3) [Meehl et al., 2007], Smith et al.  noted that model projections for the western Pacific monsoon region suggested a warmer, wetter, somewhat less windy climate as a result of enhanced greenhouse gas concentrations. However, due to the relatively coarse horizontal resolution of the models, it was noted that the climate change signal could quite possibly manifest itself differently at smaller scales—particularly for relatively small island countries which are crudely (if at all) represented on model grids with horizontal resolutions of the order of several hundred kilometers. The existence of (unresolved by models) large mountain peaks and valleys can introduce local circulations such as mountain and valley breezes, katabatic winds, and rain shadow effects, while detailed coastal boundaries and contrasting land surface types can also be important in shaping diurnal cycles of rainfall [Holland and Keenan, 1980; Liberti et al., 2001; Wu et al., 2003; Mori et al., 2004; Byon and Lim, 2005; Sakurai et al., 2005; Biasutti et al., 2012]. This is particularly true in the case of New Guinea where rainfall varies considerably over relatively small scales [Zhou and Wang, 2006].
 One method for better capturing the topographic effects is to simply use higher resolution models, but this is not always practical given the resources needed to perform an appropriate suite of experiments—particularly when they involve coupling with a dynamic oceanic model. A partial solution is to use the results from relatively coarse resolution global coupled model simulations (GCMs) to either host (or “drive”) a higher resolution atmospheric model. The Commonwealth Scientific and Industrial Research Organisation (CSIRO) stretched-grid Conformal Cubic Atmospheric Model (CCAM) [McGregor and Dix, 2008] provides this facility and, given prescribed sea surface temperatures (SSTs), can provide global climate simulations at relatively high horizontal resolutions. A preliminary analysis of the results at a 60 km horizontal resolution over the south Pacific region is described in Nguyen et al. , who found that the CCAM results showed a general improvement over those of the host models for the wider southeast Pacific region. In particular, it was noted that both mean rainfall and the representation of features such as the South Pacific Convergence Zone (SPCZ) were improved. Much of this can be attributed to the fact that the SSTs used in the CCAM simulations are bias corrected. It was also noted that, on the large scale, the CCAM projected changes in rainfall were generally similar to those of the host GCMs. Other studies have also demonstrated the improvement in the representation of climatological features for a mountainous region such as Tasmania when CCAM has been run at both 60 and 10 km resolutions [ACE CRC, 2010a; Grose et al., 2012].
 Here we analyze the results from six host GCMs from the CMIP3 suite and the corresponding CCAM. Unlike Nguyen et al. , we focus on New Guinea and utilize high-resolution satellite-based estimates of rainfall that reflect the relatively complex topography. The analysis assesses the representation of present-day rainfall patterns and the pattern of projected changes for later this century.
 The CCAM results analyzed here have been generated on a global, relatively fine scale (60 km horizontal resolution, 18 levels in the vertical) grid. CCAM is an atmosphere-only model and, in this case, relies only on the host GCM SSTs. However, the host SSTs (and sea-ice values) are bias corrected on a monthly basis to ensure that, for simulations of present-day climate, the SSTs match present-day observations. For these experiments, the bias adjustments were calculated by comparing the host SSTs over the period 1971–2000 with observed values [Reynolds, 1988]. The same adjustments were then applied to the host projected SSTs which were then used to generate the CCAM climate projections. As a consequence, CCAM simulations can be thought of as relatively simple SST-only simulations with the only forcing being the host-simulated change in the SSTs. Because the SSTs are prescribed, any temporal variability (both long term and at the interannual scale) is determined by the host GCM [Nguyen et al., 2012]. This means, for example, that any El Nino Southern Oscillation (ENSO)-like behavior will only be present if it exists in the host GCM.
 Despite the relative simplicity of this approach, computing resources limit the number of GCMs, emission scenarios, and time periods that can be sampled. For this study, the SST results from six CMIP3 GCMs (CSIRO-Mk3.5, ECHAM5/MPI-OM, GFDL-CM2.0, GFDL-CM2.1, MIROC3.2 (medres) and UKMO-HadCM3) were used to drive CCAM (see Table 1). Previous assessments suggest that these models perform relatively well, particularly with regard to SST variability and rainfall over the Australian region [Suppiah et al., 2007; Smith and Chandler, 2010; Colman et al., 2011].
Table 1. Details of Models Whose Results Have Been Analyzed in This Study
 In assessing the performance of the GCMs at simulating present-day climate, the so-called “climate of the 20th century” CMIP3 results were sampled. For the climate change projections, only the results from the A2 emissions scenario were used. Finally, we have only analyzed time slices of 20 or 30 years duration at the end of the 20th and 21st centuries, as opposed to an entire climate run of 100 years or more.
 Figure 1 shows a map of the island of New Guinea (which comprises the Indonesian West Papua region plus Papua New Guinea (PNG)) and illustrates both the complex topography according to the ETOPO2v2 global gridded 2 min database (National Geophysical Data Center, National Oceanic and Atmospheric Administration, U.S. Dept. of Commerce, http://www.ngdc.noaa.gov/mgg/global/etopo2.html) and the location of stations where long-term reliable observations have been kept. The observations can be accessed from the Bureau of Meteorology Pacific Climate Change Portal (http://www.bom.gov.au/climate/pccsp/). The land mass is effectively divided into two by a ESE-WNW trending mountain cordillera that reaches over 4800 m above sea level (asl) at several peaks. It has long been noted that the mountain peaks can experience lower rainfall totals compared to their flanks with the consequence that waterlogging hinders agriculture on the outer flanks of the mountains, while the drier, sunnier intermontane valleys support relatively dense populations [Prentice and Hope, 2007].
 The relative sparseness of the stations in relation to the varied topography means that the observations are of limited use in comparisons involving grid-scale averages over tens or hundreds of kilometers. Instead, we make use of the satellite-based dataset Tropical Rainfall Measuring Mission (TRMM) (product 3B43), Huffman et al. . The data, which are on a global grid with a horizontal resolution of 25 km, have been assessed over several regions including Africa [Adeywa and Nakamura, 2003; Nicholson et al., 2003] and Australia [Ebert et al., 2007; Fleming et al., 2011]. We focus on monthly average data over the 13 year period 1998 to 2010.
 Figure 2 compares the representation of topographic features along a south-to-north transect and demonstrates the very coarse nature of CMIP3 model representations for New Guinea. The actual maximum elevation lies close to 4500 m asl, but the maximum CMIP3 value is only about 1200 m asl, and is much broader. Furthermore, the GCMs do not resolve the existence of multiple peaks which define the highland valleys. These are important since the regional rainfall patterns appear strongly influenced by these features. Figure 3 shows the associated mean (1998 to 2010) rainfall values according to the TRMM data along the same transect for the months of January and July. According to these data, there is a distinct maximum over the southern inland slopes, followed by a minimum over the highland valley, followed by another maximum over the northern inland slopes. These features are apparent in both January and July and reflect the fact that the highland valleys experience a rain shadow effect associated with the surrounding peaks.
 It is therefore not surprising that the GCMs have difficulty representing the spatial details of the rainfall distribution. Figure 4 compares mean rainfall for the December to February (DJF) and June to August (JJA) seasons according to TRMM data with the ensemble mean (for the period 1971 to 2000) results from the six GCMs and the six associated CCAM simulations. Even though the means are based on data from 13 years (1998 to 2010), the TRMM data indicates considerable spatial variability over the land. While much of West Papua appears to remain relatively wet during both seasons, there is a definite seasonal cycle over much of PNG, involving relatively wet conditions during the monsoon season (November to April) followed by relatively dry conditions for the remainder of the year. The other major feature of interest is the shift of the rainfall maximum from north of New Britain (see Figure 1) during DJF to the south during JJA.
 According to the GCM results, the seasonal cycle over most of the region is one of a wet DJF followed by a relatively dry JJA. Only in the easternmost regions is this pattern reversed. The ensemble mean indicates almost zero influence of the topography since the maxima are located in the westernmost and easternmost regions of the entire grid. This lack of topographical influence is consistent with the low envelope of topography shown in Figure 2. While the maxima over the mountains are not well captured, the CCAM patterns are generally much more realistic since they do reflect topographic effects. In both DJF and JJA, there is evidence for local minima along the highest points and local maxima over the surrounding slopes. The shift in the rainfall maximum over New Britain is reasonably captured as is seasonal cycle of a relatively dry New Guinea in JJA compared to DJF. The easternmost reversal pattern is also reasonably simulated. However, the CCAM values appear to be underestimates over central and west New Guinea.
 How different are the CCAM results to the GCM results and the TRMM data? Here we present a preliminary assessment based on the results for DJF and JJA.
 The data from each of the GCM and CCAM simulations were interpolated to the same (25 km) grid as the TRMM data. In the case of the CCAM data, the results from each of the six simulations are very similar so only the results from one simulation are shown. This is expected since the CCAM results are based on SSTs which have been bias corrected so that they are effectively identical for each simulation. The mean (1971 to 2000) seasonal rainfall values were calculated, and the median value, the spread of 50% near-median values, and minimum and maximum values across the grid were calculated and are displayed as box-whisker plots in Figures 5a and 5b. For example, Figure 5a shows that the TRMM data indicates a median grid value of about 8 mm per day for DJF with half the values falling between about 7 and 9 mm per day. It also indicates that the TRMM data includes a maximum grid value of about 17 mm per day and a minimum grid value of about 4 mm per day and that the data are highly skewed. In comparison, we see that the CCAM results are similar to TRMM with perhaps a slight tendency to underestimation. The GCM results are all different but tend to overestimate the median, underestimate the maximum, but overestimate the minimum values. In only one case (CSIROMk3.5) is the distribution skewed similarly to the TRMM data.
 Figure 5b shows the same results for the JJA season. In this case, the CCAM median is slightly less than the TRMM value, but the maximum and minimum values are comparable. Again, both the TRMM and CCAM data are similarly highly skewed. The GCM median values are also comparable to both TRMM and CCAM values, but the maxima are again underestimates, while the minima are similar. All bar two simulations (GFDL-CM2.1 and MIROC-M) exhibit a similar skewness to the TRMM data. Overall, it appears that on the large scale, the simulated median (and mean) values are all reasonably simulated by the models, but the largest differences are seen in the extreme values, as would be expected as a result of different horizontal resolutions.
 Figures 6 and 7 compare the projected (A2 scenario) percentage changes in DJF and JJA mean rainfall (2071 to 2100 compared with 1971 to 2000) from the GCMs, while Figures 8 and 9 show the corresponding results from the CCAM simulations. Each pattern of change is different from GCM to GCM, and each CCAM result is different from that of the corresponding GCM.
 Figure 10 compares the projected percentage changes for both seasons from both the GCM results and the CCAM results. These are calculated by forming the ensemble means for present and future conditions, prior to calculating the differences (as opposed to calculating the percentage differences from each ensemble member, and then averaging). There is little in common when comparing the GCM results and the CCAM results for either DJF or JJA. The GCM results tend to indicate wetter conditions to the north and east and drier conditions to the south and west. The CCAM results are more complex, suggesting wetter conditions over much of New Guinea during DJF but drier conditions during JJA.
 The differences are made somewhat clearer in Figure 11, which shows where the ensemble members tend to agree on the sign of the projected changes. The GCM results imply a consistent tendency towards wetter conditions to the north and east during DJF, but the changes for JJA are far less consistent with almost no agreement over most of the land. On the other hand, while the CCAM projections also tend to favor increases over most of the land during DJF, they also reflect the presence of highland regions where the results consistently indicate decreases. This feature is much more evident for JJA, where the CCAM results indicate that consistent drier conditions are more extensive than consistent wetter conditions over the land. A comparison with Figures 1 and 3 indicates that, during DJF, the relatively wet regions which occur over the mountain slopes get wetter, but during JJA, the highest regions (which are sometimes relatively dry) get drier. A key feature of the results is that the CCAM projected changes imply more consistent signals compared to the GCMs.
 Figure 12 summarizes the projections from each of the GCM simulations and associated CCAM simulations. In each case, the grid statistics (minimum, mean, and maximum) of the 30 year means were calculated, and the percentage changes plotted as a function of the present-day values. In the case of DJF, the CCAM present-day values for minima cluster around 2.5 mm per day, while the various GCM values are spread around 5.0 mm per day. However, the projections from both sets are consistent since only one (GCM) result out of 12 projections indicates an increase (of only a few percent). In the case of the grid means, both the CCAM and GCM present-day values are similar (about 8 mm per day), but there is no evidence for a consistent signal either way. Finally, in the case of the maxima, the CCAM present-day values tend to cluster around 17 mm per day compared to about 13 mm per day for the GCMs, and there is unanimous agreement for an increase. A similar pattern is seen in the results for JJA, where there is unanimous agreement for a decrease in the minima, little evidence for a change in the mean, but unanimous agreement for an increase in the maxima.
Discussion and Conclusions
 The results suggest that, compared to the relatively coarse-scale GCM simulations, the CCAM simulations provide a more realistic representation of rainfall over the relatively mountainous New Guinea region. This occurs because (a) the CCAM simulations are forced with bias-corrected SSTs, and (b) they take place with a more realistic representation of the topography. CCAM may also be advantaged by the fact that it uses a higher resolution grid which may affect its representation of sub-grid-scale mechanisms relevant to the simulation of convection. However, there is little evidence of this over regions where the topography is not a factor. It is worth noting that there is little difference between the area averaged CCAM rainfall values and the multi-GCM area average values. This may indicate that the major differences arise solely from a better representation of the topography.
 The projected changes (2071 to 2100 versus 1971 to 2000, A2 emissions scenario) for rainfall indicate that all models tend to indicate a decrease in the grid minima, little change in the grid medians, but increases in the grid maxima. However, there are significant differences when comparing the spatial patterns of change. The ensemble mean patterns for percentage changes to rainfall (Figure 10) and the patterns of consensus among the model simulations (Figure 11) suggest the following synthesis of the CCAM results:
West Papua and the relatively wet northern and southern mountain slopes are likely to get wetter during DJF.
During JJA, the highland regions appear likely to become drier.
Elsewhere, the GCMs or the CCAM results are both uncertain, or else both sets of results differ. In these cases, the projections must remain uncertain.
 There are some similarities to the results from a similar study into rainfall projections over Tasmania, situated in the midlatitudes, where it was found that, while there was no significant change to overall annual rainfall, the west coast showed an increase in winter rainfall contrasting with a decrease in summer rainfall [ACE CRC, 2010b]. The high central plateau showed a decrease across all seasons, while the lower coastal regions tended to show increases, further highlighting the importance of grid resolution in regions of high topography.
 Previous analyses of the full suite of CMIP3 GCM results [e.g., Brown et al., 2012] have pointed towards strong increases in rainfall coinciding with tropical convergence zones lying over northern Indonesia and Indochina during Northern Hemisphere summer (June–August or JJA) and over southern Indonesia and Papua New Guinea during Southern Hemisphere summer (December to February or DJF), [Meehl et al., 2007]. An analysis of the GCM results over the Western Pacific monsoon region by Smith et al.  suggested that these were reasonably consistent with a mechanism which causes relatively dry areas to become drier and relatively wet areas to become wetter, the so-called “rich-get-richer” mechanism [Chou et al., 2009]. Observations also indicate that this mechanism can also explain recent trends in large-scale rainfall patterns [Zhou et al., 2011; Wang et al., 2012]. The CCAM projections are consistent with this mechanism since, spatially, they indicate increases over the relatively wet mountain slopes during DJF (Figures 8 and 9) and decreases over the relatively dry mountains during JJA. This is also reflected in the fact that the grid minima decrease in both seasons, whereas the grid maxima increase. However, overall, the large-scale average rainfall appears to be relatively unchanged—particularly over the land masses.
 In summary, while rainfall projections are still subject to uncertainty, the CCAM results suggest that a plausible scenario for New Guinea is one where DJF becomes wetter; JJA becomes drier, with the possibility that the highland regions become drier overall. The other implication of the CCAM results relates to the expectations concerning changes in the intensity of rainfall events [e.g., Tebaldi et al., 2006]—namely, that intense events may become more intense, but dry spells may become more severe. While these have yet to be analyzed, the results imply an increased potential for very intense rainfall (possibly flooding) during DJF and possibly more prolonged dry spells in the highlands. These need to be confirmed by further analyses of rainfall projections at the daily time scale.
 Finally, while GCM projections may be useful at the large scale, the results suggest that these should not be construed as applying at the smaller scales, especially where topography is significant but largely unresolved. The CCAM results indicate that, while slightly wetter conditions may be expected at the large scale, it is quite possible that some regions will experience drier conditions. They also indicate that further analysis may be required to better estimate the likely magnitude and impacts of these changes not only over New Guinea but also other mountainous regions.
 This research was supported by the Pacific Climate Change Science and Adaptation Program (PCCSAP), a program supported by AusAID, in collaboration with the Department of Climate Change and Energy Efficiency, and delivered by the Bureau of Meteorology and the Commonwealth Scientific and Industrial Research Organisation (CSIRO). We acknowledge the modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI), and the World Climate Research Program's Working Group on Coupled Modelling for their roles in making available the CMIP3 multimodel dataset. Support of this dataset is provided by the Office of Science, U.S. Department of Energy. More details on model documentation are available at the PCMDI website (www-pcmdi.llnl.gov). The authors would like to thank Jo Brown and Ramasamy Suppiah for valuable comments on an early version of this paper and colleagues within the PCCSAP program for their comments and suggestions during the course of this project.