Advantages of using a fast urban canopy model as compared to a full mesoscale model to simulate the urban heat island of Barcelona

As most of the population lives in urban environments, the simulation of the urban climate has become a key problem in the framework of the climate change impact assessment. However, the high computational power required by these simulations is a severe limitation. Here we present a study on the performance of a Urban Climate Model (UrbClim), designed to be several orders of magnitude faster than a full-fledge mesoscale model. The simulations are validated with station data 5 and with land surface temperature observations retrieved by satellites. To explore the advantages of using a simple model like UrbClim, the results are compared with a simulation carried out with a state-of-the-art mesoscale model, the Weather Research and Forecasting model, using an Urban Canopy model. The effect of using different driving data is explored too, by using both relatively low resolution reanalysis data (70 km) and a higher resolution forecast model (15 km). The results show 10 that, generally, the performance of the simple model is comparable to or better than the mesoscale model. The exception are the winds and the day-to-day correlation in the reanalysis driven run, but these problems disappear when taking the boundary conditions from the higher resolution forecast model.


Introduction 15
According to the United Nations, more than 50% of the world population lives in cities, and this percentage is expected to increase in the coming decades.The urban environment is known to modify the local climate in several different ways.The main one is the so-called Urban Heat Island (UHI) effect, that consists on temperatures being several degrees higher over the urban area with respect to its rural surroundings.Due to anthropogenic climate change, the frequency of heat waves is expected 20 to undergo a widespread increase (Meehl and Tebaldi, 2004) in the following decades.This raises concerns about the vulnerability of people living in urban areas.
were particularly damaging in the Euro-Mediterranean arch (Robine et al., 2008).The seasonal mortality excesses were indeed similar in Spain (13,7%), France (11,8%) and Italy (11,6%), although temperature anomalies were at least twice as large in France than in the southern countries (Ballester et al., 2011).This larger sensitivity to environmental conditions is exacerbated by urban pollution in old people living in cities with pre-existing or chronic cardiovascular and respiratory diseases 65 (McMichael et al., 2006).
Taking into account all these considerations, the city of Barcelona emerges as a particularly vulnerable area within the continent.Barcelona is located in northeastern Spain, surrounded by the Mediterranean Sea in the south and east, a small 500m mountain range in the northwest, and two rivers in the southwest and northeast (figure 1).Its Mediterranean climate (Csa in the Köppen 70 classification) is shaped in summer by the local wind breeze regime, whose diurnal evolution exhibits a clockwise rotation from southerlies in the morning to winds blowing roughly parallel to the southwest-northeast shoreline in the late afternoon (Redaño et al., 1991).
The main goals of the paper are: -Evaluation of a simulation of the urban climate of the city of Barcelona with the UrbClim 75 model, by comparing it with station and satellite data.
-Analysis of the sensitivity of the simulation to the boundary conditions, comparing two simulations nested in global datasets with different horizontal resolution (70 km and 15 km).
-Comparison of the UrbClim simulations with a benchmark simulation carried out with a stateof-the art mesoscale model, focusing both on model skill and computational resources demanded.

Surface stations
As a first approach in the evaluation of the model performance, we have used data from a set of meteorological stations, 4 of them belonging to the Spanish Meteorological Agency (AEMET) and 85 7 to the Catalan Meteorological Service (SMC).All are well maintained automatic stations, that deliver meteorological data with 10 or 20 min frequency.In the present work, only hourly data were used.The locations of these stations, as well as their names, are displayed in figure 1a, together with the topography.
Station number 5 (El Prat de Llobregat) is chosen to be representative of a rural location near the 90 city.This station is located in the middle of cereal fields, 300 m from the Llobregat river and 650 m from the closest urban area.Station number 6 (el Raval) is instead chosen as the reference urban station.This station is located on the roof of a building, at the city centre of the city, 8.5 Km away temperature inversions during night time hours.Thus, the differences between these two stations are considered to be representative of the UHI effect in the city of Barcelona.

Satellite data 100
The spatial pattern of the simulations is evaluated through data from the MODerate Resolution Imaging Spectoradiometer (MODIS) of the National Aeronautics and Space Administration (NASA) of the United States.Following previous works (Schwarz et al., 2011;Zhou et al., 2015), MODIS datasets MOD11A2 and MYD11A2 (version 5) were downloaded and processed.These correspond to the Terra and Aqua satellites respectively, and are 8-day aggregations of the daily MOD11A1 and 105 MYD11A1 datasets, using only the clear-sky days.The variable considered is Land Surface Temperature (LST), which is derived from the infrared radiance and emissivity estimated from land cover types.A more detailed description of the algorithms is available in Wan (2008).
The LST data were processed considering only the data flagged as "good quality, not necessary to examine more detailed QA" in the Quality Flag provided with the data, and with no cloudy days 110 during the 8 day period.MODIS and UrbClim LST data were interpolated to a 0.01 deg.regular grid for direct comparison.Finally, only the images with less than 14% missing values were used (this does not include the data over the sea which are always missing).This process left a total of 15 values for most gridpoints (supplementary figure 1) over the whole period.

The UrbClim model 115
The UrbClim model is designed to reproduce the main features of the urban climate requiring the minimum amount of computational power, so that it is possible to perform long runs at a resolution of hundreds of meters.A detailed description of the model is available in De Ridder et al. (2015).
UrbClim models the lower 3 km of the atmosphere, and consists of a 3-D boundary layer model and land-surface scheme with urban physics.The boundary data needs to be read from a lower 120 resolution model.Given that UrbClim generates very small internal variability, the stability of the simulation is not compromised by the difference in the resolutions of the UrbClim and driving models, as it normally occurs with conventional mesoscale models.Nonetheless, this resolution jump can sometimes affect the quality of the simulation if the driving model does not accurately reproduce the local climate.

125
The land use data, that are needed to represent the surface properties, are taken from the CORINE dataset (http://www.eea.europa.eu/publications/COR0-landcover).This dataset is publicly available online, and was produced by the European Environmental Agency at a resolution of 100 m.
The land-surface scheme is a standard soil-vegetation-atmosphere model based on De Ridder et al.
(1997) with some extensions.In the original scheme, the urban canopy was represented as a simple 130 impermeable slab.In the updated version, described in detail in (De Ridder et al., 2015), this is extended in many ways.However, it can still be considered a simple urban canopy model compared

UrbClim
The UrbClim simulations cover the five warmest months of year 2011, i.e. from May to September.
The domain is represented by a horizontal grid with 121x121 points at a resolution of 250 m, with 19 vertical levels up to 3000 m (figure 1a).The driving model data is updated every 3 hours.Two 140 simulations have been studied, labeled as UC-ERA and UC-FC.The former is driven by the ERA-Interim reanalyses (Dee et al., 2011), while the latter is driven by the Integrated Forecast System (IFS) version 37r2 global forecast model of the European Centre for Medium-Range Weather Forecasts (ECMWF).In 2011, this model ran with a spectral resolution of T1279 ( 15 km), in contrast with the T255( 70 km) of ERA-Interim.Thus, it is able to provide more local details, which can be 145 important give the aforementioned mesoscale-driven weather of Barcelona.

WRF
The Weather Research and Forecast model is an open-source, non-hydrostatic limited area model (Skamarock et al., 2008).Thanks to its availability, it has a large community of users.These contribute to the development of WRF, which is leaded by the National Center for Atmospheric Research 150 (NCAR).One particularity of this model is that a large amount of parameterization schemes, dynamical options, and sub-modules, available to the user to choose among them.These options are set up in a namelist file that must be edited for each simulation.The version of WRF used is the 3.6.1.
In the present work, WRF was configured to run in three nested domains 1b, with horizontal resolutions of 9, 3 and 1 km and 40 vertical levels.The 250m of UrbClim were not reached because the 155 computational cost was not affordable.However, the simulations were carefully configured to make them comparable with UrbClim: They were nested in the same dataset (ERA-Interim) and used the same land use (CORINE).WRF land use is by default taken from the United States Geological Survey (USGS) dataset.Thus, the CORINE land classes were mapped to the USGS 33 classes following table 7.1 of Chrysoulakis et al. (2014).Despite being nested in a reanalysis, the regional models tend to generate their own internal variability.There are two approaches to solve this: using nudging, or restarting the model frequently.
In this case, based on the experience of previous works (Menendez et al., 2014;García-Díez et al., 2015), daily 36 hours simulations have been carried out and concatenated leaving 12 hours as spinup.These simulations cover the same time span as UrbClim, May to September 2011.Thus, 153

Time series
Table 1 shows the standard scores of daily mean 2 meter temperature for the UC-ERA, UC-FC and WRF simulations and the 11 stations.The largest errors are found in UC-ERA, which generally overestimates daily temperatures by up to +2 • C in some stations.This overestimation is associated with the misrepresentation of the sea breeze, which has larger effect on maximum temperatures (see below).
UC-ERA also overestimates the day-to-day variability, having higher Root Mean Square Error (RMSE) than the other runs.Instead, UC-FC and WRF show similar, smaller scores, which indicate the good performance of these simulations.
As UrbClim is hardly able to generate internal variability, these results can be interpreted as a comparison between Urban Canopy+PBL models driven by ERA-Interim (70 km), ECMWF forecast ( 16km) and WRF (1 km).Thus, differences in the results show the added value of the higher resolution in the ECMWF forecast model and WRF.However, note that the extra resolution of WRF (about 15 times higher than ECMWF) is not clearly improving the results.This is consistent with previous studies suggesting diminishing returns for added value in this resolution ranges (García-Díez et al., 2015).The average magnitude of the UHI during the night is found to be 2.5 • , which is large enough to have direct impacts on human health during heat wave episodes (Ye et al., 2012).During daytime hours, the UHI is found to decrease down to -0.5 • .Note that this is in very close agreement with the values derived from observational data in (Moreno-garcia, 1994), despite it used two different reference points.The measurement of the UHI with only two points has some limitations, as it may be sensitive to very local features such as the land use in the vicinity of the stations.However, the representativeness of these points has been carefully checked with high resolution satellite images.In addition, the agreement with previous studies increases our confidence in the results here presented.
UC-ERA tends to overestimate temperatures in both stations after 10 UTC and particularly during daytime hours, but errors in both stations cancel each other, and therefore the UHI magnitude is generally well represented with biases smaller than 0.5 • .The UHI average daily cycle is similar in UC-FC and UC-ERA, but UC-FC does not show any warm bias, and accurately reproduces the observed temperatures of the individual stations.
In the case of WRF, we initially considered the nearest gridpoint to the rural and urban stations, and biases in the three panels were found to be clearly larger than those in UrbClim (not shown).This problem was found to be related to the land use of the gridpoints, which were not representative of the land use of the stations.Indeed, the gridpoint representing the rural station was found to be classified as urban in the land cover map used by WRF.In order to address this problem, we considered a more representative, adjacent gridpoint to represent the rural station, which is used throughout the paper.
Results show that biases in WRF for the individual stations are large and negative during the morning hours, before 15 UTC, but comparable in magnitude to those in UC-ERA.In addition, although the UHI at noon is correctly reproduced by WRF, it exhibits large bias maxima of -1.5 • and -1 • at 7 and 17 UTC, respectively.Regarding the wind speed (figure 3), the intensity of the sea breeze is clearly underestimated in the run driven by the ERA-Interim reanalysis, with a bias of up to -2.5 m s −1 at noon in the rural station.This problem is likely to be related with the coarse resolution of the ERA-Interim driving run, which is not able to resolve the sharp daytime, thermally-driven pressure gradient between the continent and the sea.The lack of sea breeze in turn explains the nearly constant daily cycle of the rural minus urban difference in wind speed in UC-ERA.The wind regime is clearly better reproduced in the other simulations.UC-FC accurately reproduces the daily wind cycle in both the urban and rural stations, while WRF overestimates the wind speed by up to 1 m s −1 during daytime hours.Regarding the urban minus rural difference, WRF is the model that better captures the hourly evolution of the wind speed.UC-FC does catch well the overall magnitude of the difference, but without reproducing the secondary minima and maxima of 6 and 8 UTC.
It is interesting to highlight the day-to-day variability of the observed and simulated times series, which are here depicted in figure 4 for the month of May.The whole period is not shown for clarity, but the same conclusions are applicable for the other months.The daily evolution in the UHI is well represented in UC-FC and WRF, while biases of the order of up to 4 • C at noon are found in UC-ERA for some specific days.However, the largest Mean Absolute Error (MAE) is found for WRF (1.11 • C), due to the systematic underestimation of the UHI during daytime hours (figure 2).The best MAE score is found in UC-FC (0.80 • C), which shows regular skill with almost no large errors in specific days.

Spatial pattern
The evaluation of the spatial variability simulated by the urban climate model is a challenging issue due to the lack of reliable, high-resolution observations.Figure5 shows the average daily minimum temperatures for the UC-FC and WRF, for the 5 months considered.Although both models are able to resolve the main features of the UHI of Barcelona, the surrounding cities and the airport,  Unfortunately, the scarcity of surface observations did not allow us to evaluate the spatial patterns at the screen level, and therefore we evaluated the spatial variability of the model by analysing the MODIS satellite LST, as described in section 2.2.
During the night, both UrbClim and WRF have been found to overestimate LST over urban areas and, thus, the LST UHI (figure 6).This is surprising, given the small error found in the validation   of the screen level UHI.Other studies (Zhou et al., 2015) also found small errors when comparing MODIS and UrbClim LST in and around the city of London.advectionAsmentioned in the introduction, measuring LST over urbanised areas is challenging, due to the uncertainties associated with the measurement of both the radiation and the emissivity.The bias outside the urban areas is found to be small (figure 6), and the spatial patterns are reasonably similar.Determination of emissivity 245 over urban areas is notoriously difficult and subject to a large uncertainty, which could explain at least part of model deviation for LST.The spatial Pearson correlations between the observed and simulated fields are 0.74±0.06for UC-FC and 0.69±0.07for WRF, where the confidence bounds were computed with bootstrapping (1000 samples).Thus, UrbClim correlation is higher, but the difference is not statistically significant, as the confidence bounds overlap.It is worth mentioning that

Computational resources
In this section, the computational resources consumed by UrbClim and WRF are compared.The comparison is not fully trivial because UrbClim does not currently support running in parallel.This 255 can be seen as an important drawback.However, UrbClim does not require a long spin-up, and therefore the simulations can be parallelised just by splitting the time period in subperiods and run the corresponding simulations simultaneously in different machines or nodes.
For a direct comparison, both models were run in the local cluster of the Institut Català de Ciències del Clima (IC3), while the main UrbClim runs used in the paper were carried out in the VITO cluster.

260
The IC3 cluster is made of 48 homogeneous server blades, having each of them two "quad core" processors, 48GB of memory, 146GB of disk space and fast network interconnect (Infiniband).The blade model is Sun Blade X6270 (see http://www.oracle.com/us/products/servers-storage/servers/blades/sun-blade-x6270-m2-ds-080923.pdf for a full description) equipped with Xeon (Nehalem) X5570 processors.With this settings, a WRF simulation of 36 hours took 2.5 hours to finish (using 265 an average of 10 simulations), including the preprocess carried out with the WRF preprocessor (WPS).This preprocess was run in serial, in 1 core, while WRF was run in 16 cores, this is, two blades.Thus, the total serial equivalent wall-time was 40 hours, assuming perfect scaling (the real value will be somewhat below).WRF was compiled using the Intel fortran compiler version 14.0.1 with the Intel MPI Library for Linux OS, Version 4.1 Update 3.

270
Regarding UrbClim, for this test, it has been compiled with the same compiler and run in the same cluster.A 36-hours simulation with UrbClim took 0.3 hours to finish (average of 10 simulations) running in one core.Thus, UrbClim running at 250 m resolution is found to be 133 times faster than WRF at 1 km resolution.This enables downscaling large climate change ensembles for a big collection of cities.

4 Conclusions
In the present work, we have evaluated the performance of a boundary-layer urban climate model (UrbClim) for the warm season in the city of Barcelona.We were particularly interested in the study of the urban heat island (UHI) effect, given that it represents a major source of health problems in summer for vulnerable people living in urban environments (e.g.heat stress, temperature-related 280 mortality, pollution, vector-borne diseases).We have analysed the effect of the model resolution in the driving simulation (UC-ERA and UC-FC), and compared these runs with the output of a regional climate model (WRF).All these simulation have been evaluated against observations from meteorological stations and satellite data (MODIS), in order to analyse the temporal and spatial variability of the UHI effect, respectively.

285
The main conclusions of our work can be summarised as follows: -The UHI in the city of Barcelona reaches 2.5 • C at night.This is relevant for the study of climate impacts, given that it increases the stress to the vulnerable population and for the health care systems under extreme conditions.
-UrbClim correctly reproduces the UHI of Barcelona when it is nested to the coarse dataset 290 of ERA-Interim, with some systematic biases.When it is nested to a higher resolution model (ECMWF IFS), UrbClim additionally reproduces well the temperature evolution of the individual rural and urban stations used for the calculation of the UHI.
-WRF reproduces the UHI intensity nearly as well as UrbClim, but it provides less detailed spatial information consuming much larger computational resources.

295
-The realism of the spatial pattern of LST is similar in UrbClim and WRF, when it is validated against MODIS data, even though significant biases are found in both models.
In conclusion, the choice between UrbClim and WRF for the simulation of the urban environment largely depends on the type of variable and process that is analysed.WRF has the advantage of providing a more detailed and complete description of atmospheric winds and rainfall, which may 300 be required in some applications (e.g.pollutant dispersion).Apart from this, UrbClim has been found to be an optimal tool for the numerical description of the UHI of Barcelona, providing an accurate description of the temperature field that is generally better than that in WRF.However, it must be taken into account that, if nesting UrbClim in a low resolution model, there will be inaccuracies caused by the misrepresentation of the wind, specifically the sea breeze daily cycle.The sea breeze 305 is important for reproducing the climate of Barcelona in summer, where the influence of mesoscale processes is strong.Note that this problem is a particularity of Barcelona, as it has not been found in other European cities where UrbClim nested in ERA-Interim has been successfully tested (De Ridder et al., 2015;Lauwaet et al., 2016;Zhou et al., 2015).
From these results, it is reasonable to infer that the skill of UrbClim, and probably of other similar 310 urban climate models, is constrained by the performance of the driving model, and particularly for variables that are important for the UHI, this is, wind speed and cloudiness.

Figure 1 .
Figure 1.a) Topography of the UrbClim domain and locations of the meteorological stations used.The two stations used as references to compute the Urban-Rural difference are highlighted with red stars.b) Three WRF domain edges (red squares) and UrbClim domain edges (black contour), together with the topography of the WRF domains with 10, 3.3 and 1.1 km resolution.

Figure 2 .Figure 2
Figure 2. Average daily temperature cycle in the urban (left) and rural (middle) stations.The difference urban -rural is shown in the panel on the right.

Figure 3 .
Figure 3. Same as figure 2, but for wind speed.

Figure 5 .
Figure 5. Daily minimum temperature averaged over the period May-September 2011 in UC-FC (left) and WRF (right).

Figure 6 .
Figure 6.Land surface temperature averaged during nighttime hours over the period May-September 2011 in MODIS (left), UC-FC (center) and WRF (right).
250 the MODIS LST appears to have an effective resolution coarser than 1 km, given that the spatial patterns are smooth and do not resolve many detailed features.11 Geosci.Model Dev.Discuss., doi:10.5194/gmd-2016-10,2016 Manuscript under review for journal Geosci.Model Dev.Published: 9 February 2016 c Author(s) 2016.CC-BY 3.0 License.

Table 1 .
Scores of daily mean 2 meter temperature for the UC-ERA, UC-FC and WRF simulations and the 11 stations.