the Iberian Peninsula revisited. Part II: Max and Min Temperature

. In the recent past, the increase of computation resources led to the development of regional climate models with increasing domains and resolutions, spanning larger temporal periods. A good example is the World Climate Research Program Coordinated Regional Climate Downscaling Experiment for the European domain (EURO-CORDEX). This set of regional models 20 encompass the entire European continent, for a 130-year common period until the end of the 21 st century, while having a 12 km horizontal resolution. Such simulations are computationally demanding, while at the same time, not always showing added value. This study considers a recently proposed metric in order to assess the added value of the EURO-CORDEX Hindcast (1989-2008) and Historical (1971-2005) simulations, for the maximum and minimum temperature over the Iberian Peninsula. This approach allows an evaluation of the higher against the driving lower resolutions rela tive to the performance of the whole or partial 25 probability density functions, by having an observational regular gridded dataset as reference. Overall, the gains for maximum temperature are more relevant in comparison to minimum temperature, partially owed to known problems derived from the snow-albedo-atmosphere feedback. For more local scales, areas near the coast reveal higher added value in comparison with the interior, which displays limited gains and sometimes notable detrimental effects with values around -30%. At the same time, the added value for temperature extremes reveals a similar range, although with larger gains in coastal regions and in locations from the 30 interior for maximum temperature, contrasting with the losses for locations in the interior of the domain for the minimum temperature.

Abstract. In the recent past, the increase of computation resources led to the development of regional climate models with increasing domains and resolutions, spa nning larger temporal periods. A good example is the World Climate Research Program -Coordinated Regional Climate Downscaling Experiment for the European domain (EURO-CORDEX). This set of regional models RCM downscaling experiment, where models are forced by the IPCC Coupled Model Intercomparison Project -Phase 5 (CMIP5) GCMs. Most studies focus their assessment for the precipitatio n variable. However, in the recent past, these simulations have been extensively evaluated, and important gains were found for variables such as temperature (Vautard et al., 2013(Vautard et al., , 2020Kotlarski et al., 2014;Soares and Cardoso 2018;Ciarlo et al 2020;Herrera et al., 2020;Soares, 2021, Careto et al 2021) . Vautard et al (2013) evaluated temperature for the Hindcast simulations finding an overestimation of extreme temperatures and heat waves for summertime over the Mediterranean, and an underestimation over Scandinavia. These differences were attributed to the convection and microphysics schemes, which affect the flux partitioning and consequently the temperature. Moreover, despite the improvements, namely over coastline areas, the use of higher resolution did not show clear improvements in the representation of heatwaves. More recently, Vautard et al (2020) assessed the Historical EURO-CORDEX models reporting colder, humid, and more windy biases relative to observations, with joint liability across GCMs and RCMs. The authors also report ed that no model 70 stands out for all metrics considered. Also, Kotlarski et al (2014) assessed the summertime and wintertime temperature, precipitation, and mean sea level pressure from the Hindcast simulations for the 12 km and 50 km resolution, reporting a cold and humid bias over most of Europe and seasons, together with a summer dry and warm bias in southern Europe. Yet the models were still able to capture the space-time variability of the European climate. Herrera et al (2020), assessed the Hindcast models for the Iberian Peninsula against three observational datasets. While the models can reproduce the spatial pattern and variability, t here is 75 a dependence on the observational dataset considered. Nevertheless, the results reveal a higher agreement for temperature than for precipitation, decreasing for the extremes. The historical period simulations were also assessed for specific regions (Smiatek et al., 2016;Lhotka, 2018;Cardoso et al., 2019). Smiatek et al (2016) assessed precipitation and temperature for an alpine region, reporting a cold bias together with a more humid summer and winter. The authors describe that the models reveal no significan t gains in comparison with previous experiments. On the other hand, Lhotka (2018) investigated the capability of EURO-CORDEX 80 models in simulating heatwaves over central Europe, reporting an overall difficulty by models in reproducing these extremes. For Portugal, the temperature was evaluated in Cardoso et al (2019), where models can correctly describe the main orographic and coastal-related gradients, however, a cold bias is present for most RCMs. Nevertheless, the authors built a multi-model ensemble, which was able to outperform the individual models.
A common issue across all RCMs is related to the snow-albedo-temperature feedback, through a misrepresentation of the surface 85 energy balance (García -Díez et al., 2015;Minder et al., 2016;Terzago et al., 2017). Uncertainties in the snow depth, melt and cover could have a potential impact on surface air temperatures around 0 ºC. Biases in the albedo representation leads to positive feedback, thus enhancing too cold temperatures during winter.
To assess the added value of the EURO-CORDEX high-resolution models against their forcing counterparts, Soares and Cardoso (2018) were the first to propose a new metric in order to gauge the quality of a simulation relative to their Probability Density 90 Functions (PDFs). They used a Distribution Added Value (DAV) to characterise the gains or losses of higher resolution simulations for precipitation for several Hindcast simulations, by comparing the high 12 km and the low 50 km resolutions, with the stationbased dataset ECAD (Klein Tank et al 2002, Klok & Klein Tank 2009) as reference. The added value was shown, particularly for extreme precipitation, although not always the highest resolution displayed the most significant gains. Ciarlo et al (2020) used a similar proba bility density function based metric to assess the added value of all available EURO-CORDEX and CORDEX-CORE 95 (Gutowski et al., 2016) simulations for precipitation. Although added value of the high -resolution simulations is found, particularly for the tail end of the distributions, the authors also reported a significant uncertainty derived from the observational datasets. More recently, Careto et al (2021) revisited the added value from the EURO-CORDEX by implementing a similar methodology to this work over the Iberian Peninsula, considering the highest resolution observational dataset to date. In Di Luca et al (2013), sites located near the coast showed relevant gains. Cardoso and Soares (2021) assessed the temperature with the DAVs metric, for the 100 lower (50 km) and higher resolution (12 km) Hindcast simulations, with the E-OBs regular dataset as a reference. The authors reported a difficulty in obtaining added value, partly owed to the assimilation of observations into the ERA-Interim. Still, gains are reported for the Mediterranean and British Isles, while losses are found for regions of complex terrain. Nevertheless, th e authors found added value, particularly for maximum temperature extremes.
The DAV metric relies on comparisons from high/low-resolution models, with observations, regarding their PDFs, thus evaluating the gains or losses of RCMs against their driving ERA-Interim or GCM counterparts. DAVs is a versatile metric being able to be applied to either the entire PDF or to PDF sections, such as those related to extremes. In this study, DAVs will be used to a ssess the performance of maximum and minimum temperature for th e Iberia Peninsula for all available Hindcast (1989Hindcast ( -2008 and Historical  EURO-CODEX simulations. This assessment uses the Iberian Gridded Dataset (IGD, Herrera et al., 2019) as references. The IGD is a high-resolution dataset, built from an extensive station network, covering the Iberian Peninsula for the 110 1971-2015 period. Here, a new and unprecedented way of assessing the EURO-CORDEX simulations against observations at a similar scale is considered. The next section introduces the data and a de scription of the methods considered. The results and discussion are presented in the following section. Finally, the main conclusions are drawn in the last section.  (see table S2 for each GCM resolution) or ERA-Interim grid (0.75 o ). For temperature, a constant lapse rate of -6.5 o C/km was considered for the interpolation. First, for each dataset, the effect of each orography was removed by adiabatically adjusting the temperatures to mean sea level. Afterward the horizontal interpolations, the temperature is adiabatically adjusted on ce again, to the orography from the target resolution, considering the same lapse rate. Soares and Cardoso (2018) recently proposed a Distribution Added Value (DAV) metric which allows one to assess the performance of downscaling low-resolution GCMs or reanalysis in terms of Probability Density Functions (PDFs). The DAV metric is based on a PDF skill score proposed by Perkins et al (2007) which measures the similarity between two PDFs. The first step is to build a PDF for each dataset. For maximum and minimum temperature, bins with a width of 1 o C were chosen and the number of events is added for each bin. Next, each bin is divided by the sum of all events, thus resulting in an empirical re lative frequency distribution. With this approach, one can more accurately compare the results across the different da tasets as changes 150 are more directly identified (Gutowski et al., 2007). Next, the Perkins score is given by the sum of the minimum bin value between the model and the observational PDF:

Distribution Added Value (DAV)
Where n is the PDF's number of bins, m denotes the high or the low-resolution model and obs is the observational PDF. For both temperatures, the limits are bounded between -50 o C and 60 o C, encompassing all of the observational temperature range. With the 155 scores for all simulations, the DAVs are given by the normalised difference between the high -resolution model Perkins skill score and the lower resolution GCM or reanalysis Perkins skill score: with the subscript ℎ depicting the high resolution (RCM) and the low resolution (GCM or reanalysis). Thus, DAV represents the percentage of gains or losses with respect to PDFs between the high and low resolutions models. If for a particular bin there is 160 no data from either model or observations, then by definition the contribution of that bin would be 0. Moreover, an advantage of this metric is in its ability to be computed not only for the entire PDF but also for PDF sections, thus enabling an assessment for extremes. In this work, the extremes are also evaluated, where for maximum temperature, values above the observational 90 th percentile are considered, while for minimum temperature, the values below the observational 10 th percentile are taken into account.
Therefore, a new relative frequency PDF is built only from the extremes.

165
For the assessment of the DAVs of temperatures, two approaches are considered. A first region al approach for the entire Iberian Peninsula, where the PDFs are built by pooling together all data, thus returning a representative result for the whole domain . The second approach is composed by a spatialisation of the DAVs, where all data from within each individual low-resolution grid cell is pooled together, thus computing individual percentiles and PDF sets.  where GERICS, IPSL, and MPI reveal percentages above 7.8 %. In fact, in Cardoso and Soares (2021) the IPSL model also has higher gains for the Iberian Peninsula. The same models which display smaller DAVs, below 3% at the annual scale, also reveal lower percentages for winter, with 4 RCMs having a small detrimental effect. These negative values for winter are enough to condition the yearly DAVs, since in spring and summer there is a higher added value for 6 and 9 models, respectively.

180
From these, GERICS, IPSL and MPI standout with gains above 19 % for both seasons, reflecting in more noticeable DAVs at the annual scale. As for autumn, almost all models reveal some added value, although not as large as for spring or summer. For ICTP, this RCM also displays a similar performance to the driving simulation. In fact, this model also revealed neutral DAVS for spring, although gains around 10 % for summer, and small losses for winter. Another model with limited gains, but throughout the year , is the DHMZ, which despite the positive seasonal DAVs, at the annual scale a small detrimental effect is shown. The PDFs in Fig. 185 S1 from the high-resolution models have a close representation relative the IGD. Particularly for the winter season, the already tightly packed PDFs, reveal a more difficulty for the higher resolution in obtaining added value in comparison to the other s easons.
These results are in line with Herrera et al (2020), which evaluated the annual and seasonal means and variability for a se t of Hindcast simulations. For temperature, the authors obtained small spatial biases and small standard deviation ratios between the models and the IGD. . These biases, through the snow-albedo-feedback causes an underestimation (overestimation) of temperature at higher (lower) altitudes, not only for winter, but also for springtime due to an extended (reduced) snowmelt period (Minder et al., 2016).
In the Iberian Peninsula there is not many locations with a significant amount of snow. In fact, the PDFs of the RCMs for TASMIN in Figure S2 reveal a slight shift to the right for all seasons, indicating an overestimation of minimum temperature. Therefore, the issues related to snow and the smaller differences between all PDFs at the low-resolution anticipates an overall difficulty in 210 obtaining added value for the minimum temperature. Accordingly, 8 out of 13 RCMs in Fig. 2b for the annual scale reveal some losses, specially CNRM53 with -6.9 %. These detrimental effects come from the losses of the individual seasons. While for TASMAX, summer had some of the highest scores in Fig. 2a, only 3 models show added value in Fig. 2b. The other 10 RCMs reveal detrimental effects, where from these, 5 reveal losses up to -11 %. For the other seasons, the values are not so different, with the majority of RCMs displaying losses up to -11.7 % for CNRM53 in winter. In fact, this RCM together with ETHZ, GERICS and MPI reveal the worst performance throughout the year in comparison to the driving simula tion. Despite the overall lower values for TASMIN, 5 models are still able to display at least two seasons with added value. Only DHMZ reveals a different result by obtaining positive percentages at the annual and seasonal scale. Figure 2c shows the TASMAX extremes, where only data above the observational 90 th percentile set for the maximum temperatures are considered. The overall picture is different from Fig. 2a. At the annual scale 5 RCMs display DAVs ranging from -3.3 % to 220 1.1 %, while the other 8 models reveal added value, particularly DMI, KNMI, and SMHI. For winter, 9 models show a very noticeable added value, with percentages above 17.8 %. As before, the ICTP RCM stands out due to the negative DAVs of -6.6 %. Similar to winter, the summer also shows higher gains for 7 RCMs, ranging from 14.  With this approach the high-resolution orography is incorporated into the ERA-Interim. This fact results in substantially improving the low-resolution PDF, hence the lower values. However, since the ERA-Interim is downscaled, not only spurious values are introduced, but also enhanced due to the orography change, increasing the uncertainty associated with this second approach. DAVs for TASMAX considering the whole PDF are shown in Fig. 3a, where the gains are mostly focused on coastal sites, particularly over the Mediterranean. Previous works such as Di Luca et al (2013), Vautard et al (2013) and Cardoso and Soares (2021) also show added value in coastal sites, namely for the Mediterranean coast, mostly owed to a better representation of the land-sea boundary and associated differential warming, and due to improvements in the representation of breezes. The overall results correlate well with those shown in Fig. 2a i.e., models with most points with added value will inevitably reveal gains at the

275
While the maximum temperature revealed some widespread gains, the picture for TASMIN in Fig. 3b is quite different, still correlating well with the results for the regional overview (Fig. 2b). The DAVs for minimum temperature reveal losses in the interior for all time scales, while a positive added value occurs in coastal sites. At the seasonal scale, RCMs such as CCLM, CNRM63 DMI for summer or CNRM53 for winter, which reveals detrimental effects in locations from the interior also show noteworthy losses in Fig. 2b. An exception is the MOHC RCM for summer, which despite the high losses for the southern region 280 of the Peninsula, the larger gains in points over the north and Balearic Islands are able to even out the results obtained at a regional scale.
The TASMAX extremes (Fig. 3c) reveal more variability in comparison to Fig. 3a. Still, a good agreement to the Iberian Peninsula DAVs (Fig. 2c) is found for almost all seasons. At the annual scale, the lower values for the 5 RCMs identified in Fig. 2c are due to losses located over Portugal and northern Spain, while the other models reveal larger gains in the central regions of Spain. For 285 winter, Fig. 2c revealed most models with maximum DAVs, except in 4 RCMs. This high added value for this season comes from the widespread positive percentages over the domain and some coastal points (Fig. 3c). The models with lower values in the regional overview, reveal a detrimental effect for most of the peninsula, masking the added value for points located near the coast.
For spring, although there is a widespread positive added value for the entire do main, the results show the opposite in Fig. 2c.
These differences highlight the fact that the results in Fig. 2 are not a spatial mean from the DAVs in Fig. 3 and care must be taken 290 in the comparison. For summer, the spatial variability for all RCMs is higher in comparison to winter and spring. For this case, larger added values are found over a north-south stripe located at the centre of the peninsula, surrounded by a mix of negative and positive DAVs. The same pattern can be seen for autumn fo r 9 RCMs, although limited to the southern half of the domain. Similar to spring, autumn also showed lower DAVs in Fig. 2c. These lower values in the regional DAVs are derived from the losses over corroborating the results in Fig. 2c.
The next panel shows the results for the TASMIN extremes (Fig. 3d). In Fig. 2d, the annual, spring, summer, and autumn revealed losses. However, in Fig. 3d, the models do not reveal such detrimental effects, with neutral DAVs in the interior and larger gains for locations near the coast. The exceptions are GERICS, MPI , and SMHI with noteworthy losses at the annual scale, winter, and autumn in agreement with Fig. 2d. As for winter, 8 RCMs reveal noticeable added value not only for coastal sites but also for points scattered in the interior. In fact, these same 8 RCMs also showed larger gains for the Iberian Peninsula DAVs in Fig. 2d.

Historical (1971-2005)
This  The overall results for RCMs driven by the same GCM are similar for most cases, while no connection is found between models 325 forced by different GCMs. This lack of connection was also reported in Careto et al. (2021) for the EURO-CORDEX precipitation.
Still, the lower DAVs for some GCMs groups might be relate to the quality of the driving simulation itself, namely across the lateral boundary forcing zone. Following Brands et al. (2013)  For temperature, the effect of the orographic correction with a constant lapse rate in the interpolation may also be a relevant factor

355
Contrary to the results shown in the Hindcast, the TASMIN for the EURO-CORDEX Historical simulations mostly reveal positive values (Fig. 4b). For this case, both the low and high-resolution models are affected by the same problems related to snow (García -Díez et al., 2015;Terzago et al., 2017), namely at the annual scale and the colder seasons. The inter-model variability for TASMIN is higher in comparison to TASMAX in Fig. 4a, with values ranging from -17.3 % to 66 %. At the annual scale, most models reveal added value, except for CNRM driven RCMs, and for some models forced by either ICHEC2, IPSL_LR, and MOHC.
Seasonally, the pairs with notable gains for the annual scale, also shows higher added values. From these, highlight to NCC driven models, with larger gains for spring and summer, although winter and autumn also reveals added value around 20%. Moreover, 4 GCM groups displays added value for at least one season, with percentages above 30 %. On the other hand, three RCMs forced by CNRM and other three by ICHEC2 have noticeable losses, namely CNRM-CNRM53 for winter and spring, or ICHEC2-MOHC for summer.

365
From the PDFs in Fig S4, one can infer the large gains found for TASMIN and for TASMAX in Fig. 4. In Fig. S4, the GCMs tend to underestimate maximum temperature, while minimum temperature is overestimated. Moreover, the variability from the GCMs is lower in comparison to the observations. For the high-resolution models, these issues are partly corrected, hence the added value ( Fig. S3). However, for maximum temperature, the RCMs still tend to underestimate TASMAX. As for TASMIN, the issues related to snow-albedo-feedback are more evident for the high-resolution models (Minder et al., 2016) enhanced by the higher definition 370 of topographic features. On the contrary, the RCMs reveal a closer mean and standard deviation relative to the IGD. Nevertheless, these issues might cause an overall underestimation of the mean for TASMIN. Cardoso et al (2019) for Portugal and Vautard et al (2020) for the Iberian Peninsula also found a cold bias for mean tempera ture from the Historical RCMs. These biases were also present in the forcing models, matching the results found here.
The next panel shows the results for the TASMAX extremes (Fig. 4c). The added value is limited in comparison to Fig. 4a, with more models revealing detrimental effects at the annual scale and for at least 3 seasons. In fact, 12 GCM -RCM pairs reveal this pattern, sometimes with losses surpassing -10 %, capped at -24.7 % for IPSL_MR-IPSL at the summer season. Yet, 26 pairs reveal positive DAVs for at least 3 seasons, with a focus on either winter or summer, whose values range approximately from 5 % to 30 %. For the Hindcast simulations (Fig. 2c), only the winter and summer seasons have added value, whereas the values for spring and summer were limited and sometimes negative. MOHC driven models also reveal large gains for summer, well above 20 % in 380 most cases. Moreover, in some cases, the results for the TASMAX extremes differ, where noticeable added value occurs for seasons whose models had a worse performance in Fig. 4a. This behaviour is particularly evident for models driven by ICHEC or NCC GCMs. Also, pairs that have higher gains in Fig. 4a, reveal either an absence or a loss of value, occurring namely for MPI1 and MPI3 driven models for the winter and for the summer season for RCMs forced by MPI2.
in the Hindcast simulations (Fig. 2d). These losses are associated with the known problems derived from the snow-albedoatmosphere feedback, however, in this case, the low-resolution GCMs are a lso somewhat affected. Still, 15 out of 53 GCM-RCM pairs reveal losses at the annual scale higher than -10 %, whose effects are also extended for the individual seasons. Only a few RCMs still display added values, on the order of 10 %. From these, highlight to models forced by CNRM for the summer season, and also for models driven by NCC for winter.
As with the Hindcast simulations and in Careto et al (2021) for the EURO-CORDEX precipitation variable, a second method was implemented (Fig. S5), where all high and low-resolution models are interpolated into the IGD resolution. The interpolation of the IGD reduces the variability, thus the probability of extremes is lower. By interpolating the low-resolution GCMs with an orography correction, the results are substantially improved, thus increasing the score of the GCMs and reducing the DAVs.

400
The next set of figures displays the DAVs results from Fig. 4 but extended for a spatial overview (Figs. 5 to 8). Overall, as with the Hindcast (Fig. 3), the added value is more focused over the coastal regions, primarily owed to the better representation of landocean boundaries (Di Luca et al., 2013;Vautard et al., 2013;Cardoso and Soares, 2021 ). Still, all figures reveal a good agreement with the DAVs shown in Fig. 4 by considering the entire Iberian Peninsula. Moreover, as with Hindcast, the results from Fig. 4 are not a mean obtained from the spatial DAVs in Figs

415
The spatialisation of TASMIN is shown in Fig. 6, revealing a similar pattern as in Fig. 5, with gains in the coast and detrimental effects in the interior, although displaying a stronger contrast. Depending on the strength of these detrimental effects, the results for Fig. 4b might reveal neutral or even losses. For instance, the large losses found for CNRM-CNRM53 for all time scales or IPSL-MR-IPSL for spring effectively impacted the regional overview. Iberia. This is particularly relevant for the TASMIN extremes. Fig. 4d display s most pairs with a detrimental effect for the entire domain. However, higher gains are still found for the coastal regions in almost all cases. The importance of the spatial overview is highlighted here. While a particular model reveals losses for the whole domain, this fact might not be true at a more local scale.

4 Summary and Conclusions
In this work a Distribution Added Value metric proposed by Soares and Cardoso (2018) is used to gauge the added value of higher resolution simulations confronting each low-resolution driving GCM or Era -Interim, with the IGD observations over the Iberian Min temperature Frequency 0-100 simulations for the variables maximum and minimum temperatures. Overall, the high -resolution models are able to provide some added value, particularly for TASMAX. On the other end, minimum temperature reveals some difficulties in obtaining added value for the Hindcast simulations, mainly when the whole PDF is considered and also for the extremes in the Historical simulations, partly owed to problems associated with the snow-albedo-atmosphere feedback, derived from uncertainties related to snow cover, depth, and melt (García -Díez et al., 2015;Minder et al., 2016;Terzago et al., 2017). These uncertainties substantially affect the PDFs from the RCMs around 0 o C, by overestimating the frequency of events in comparison with the observations. This poor representation has a more significant impact on the Hindcast simulations, where the ERA-Interim reanalysis does not reveal 445 these kinds of issues owed to the incorporation of temperature observations (Prömmel et al. 2010). Nevertheless, the finer details from the downscaling allow a more spatial variability of temperature, which in the end could result in to added value. This fact is Max temperature Frequency 90-100 Still for the TASMAX extremes winter and summer reveals an evident added value, contrasting with the neutral or negative DAVs for spring and autumn. As for the TASMIN extremes, depending on the season, the DAVs depend more on the season. For instance, the 10 th percentile for winter is too low, not incorporating the problems around 0 o C, thus revealing some added value. The opposite occurs for spring and autumn which revealed more neutral values. As for summer the temperatures are too high for snow in most part of the territory, yet the models still revealed losses.
For the Historical simulations, no connection is found between each GCM downscaling group, while the results for multiple RCMs forced by the same GCM reveal a similar range. Each driving simulation has its own resolution and performance, which could Min temperature Frequency 0-10

465
orographic correction considering a constant lapse rate might in the end have a more significant impact in the performance of the individual GCMs and RCMs. Moreover, other factors such as the parametrizations of smaller scale processes and the representation of feedback systems can have a major impact in how a variable represent reality. Another factor that may play a major role is related to how well do GCMs represent storm tracks. If a GCM is not able to properly represent storm tracks, then the downsca ling 470 RCMs will inherent these issues.
Similar to the Hindcast simulations, the gains of the Historical sp atial DAVs are also more focused in coastal regions. Models which reveal more substantial gains and neutral or slightly positive DAVs in the interior, tend to have higher values at the Iberian Peninsula scale. Although, care must be taken when comparing th e spatial and regional DAVs as both follow slightly different approaches. Still both methods correlate well, primarily for the whole PDF case. As for the extremes, the different threshold s for each individual point makes it difficult for a direct compariso n.
A second methodology was also implemented, following the secondary results in Careto et al (2021). In this case, all data is interpolated with an orographic correction to the resolution from the observations. The interpolation from the high-resolution observations to each of the low-resolution model grids degrades the observational PDF. However, by downscaling the driving lowresolution with an orographic correction considering a constant lapse rate, unrealistic values can be generated, not only due to the 480 interpolation, but also derived from the higher resolution orography. This method substantially improves the representation of temperature at the cost of not considering land-atmosphere feedbacks, thus resulting in a larger uncertainty. In fact, for this case, the low-resolution scores are improved, resulting in an overall lower DAV.