Simulating the surface energy balance over two contrasting urban environments using the Community Land Model Urban



A single-layer urban canopy model Community Land Model Urban (CLMU) is evaluated over two contrasting urban environments of Toulouse (France) and Melbourne (Australia). For the latter, three measurement sites are available characterized by a varying amount of vegetation, which supports a detailed assessment of the representation of urban vegetation in CLMU. For Toulouse, observed roof, wall and road surface temperatures allow for a detailed evaluation of the anthropogenic heat parameterization. Overall, CLMU performs well in simulating the canyon and urban surface temperatures, anthropogenic heat flux and urban energy balance, with an overall better performance for the dense old city centre of Toulouse in comparison to the more vegetated sites in Melbourne. Results for the latter sites reveal that the pervious road fraction provides a reasonable approximation of vegetation in the urban canyon while the tile approach often results in an underestimation of latent heat fluxes. A detailed analysis of the radiative, turbulent and anthropogenic heat fluxes as well as surface temperatures for Toulouse point to a complex interaction between urban surfaces and canyon properties. Decoupling the roof from the urban canyon to the atmosphere aloft is shown to be important. Our findings suggest that more evaluation is necessary for contrasting urban geometries in order to obtain a better understanding of the interaction between the roof surface on the one hand and canyon air and air aloft on the other hand. The results simultaneously reveal a trade-off in errors between surface temperatures, radiative and turbulent fluxes and anthropogenic heat which again stresses the importance of the intended model application. Also, our results suggest that model complexity should, perhaps, relate to the site complexity. These results provide a robust basis for the construction of additional sensitivity experiments, tailored towards the intended application for urban climate mitigation studies.

1. Introduction

Continuous efforts are addressing the role of land cover changes on global and regional climate through biogeophysical feedbacks (i.e. indirect changes) as well as the transformation of natural land to surfaces that serve human needs (e.g. the conversion of tropical forest to agriculture) (Pielke et al., 2002; Kalnay and Cai, 2003; Pitman et al., 2009; Lawrence and Chase, 2010). Yet to date, the effect of changes in natural land use to built-up areas (urbanization) is only rarely addressed in climate science (Best, 2006; Christensen et al., 2007; Oleson et al., 2008a), despite the fact that at present approximately 50% of the global population resides within urban areas, a number that is expected to increase to 69% by 2050 (UN, 2012). The only way to address the future impacts of urban areas – where most people's activities take place – and climate change is to include these artificial surfaces in global and regional climate models (Best, 2006).

Representing urban surfaces in global or regional climate models (or higher resolution mesoscale and numerical weather prediction models) first of all requires the spatial characteristics and thermal, radiative and morphological properties of these areas. The latter vary widely because of different cultural factors and resources (Shepherd and Menglin, 2004; Jackson et al., 2010). Also, the complexity and particularity of urban processes may explain why until recently most modern land surface models have not formally included urban parameterizations dedicated to the representation of the urban surface energy balance (Shepherd and Menglin, 2004; Best, 2006; Bonan, 2008). Finer grid resolutions and an increase in computer resources have increased interest in modelling cities and have resulted in the development of various urban parameterizations (see references in Grimmond et al., 2010). According to Masson (2006), these urban parameterizations can be classified in three categories: (1) empirical models, (2) vegetation models with and without drag terms adapted to include the urban canopy and (3) single- and multilayer urban canopy models. Here, we focus on the Community Land Model Urban (CLMU), a single-layer urban canopy model that was recently implemented in the Community Land Model 4.0 (CLM4). CLM4 is the land surface model of the fully coupled Community Earth System Model (CESM), a global climate model that provides state-of-the-art computer simulations of the Earth's past, present and future climate states. The rationale behind CLMU is that the model should be simple enough to be compatible with structural, computational and data constraints of a land surface model coupled to a global or regional climate model, yet complex enough to represent the physically based processes known to be important in the urban climate (Oleson et al., 2008a).

Before one can use this fully coupled approach in which CLM4 provides inputs of energy and moisture to the lowest atmospheric level, the land surface model (and urban parameterization more specifically) needs to be evaluated separately. Otherwise, it is impossible to disentangle the imperfections of both the atmospheric (or other) model components and the land surface scheme (Masson, 2006). As the urban parameterization has only recently been implemented in CLMU, the model is presently only evaluated for Vancouver (Canada) and Mexico City (Mexico) over a short-term period (1 and 2 weeks, respectively) (Oleson et al., 2008a) and for Preston (Melbourne, Australia) as one of the 33 urban land surface models in the International Urban Energy Balance Models Comparison Project (PILPS-urban) (Grimmond et al., 2010, 2011). The latter project presents the first systematic intercomparison of urban land surface models with an overall aim to understand the complexity required to model energy and water exchanges in urban areas. Although this exercise is extremely useful in its kind, the intercomparison is done in a generic and anonymous way and does not provide a more detailed analysis of the underlying reason(s) for the deficiencies of a specific urban land surface model.

The research presented here is a first step towards a high-resolution (1–10 km) regional climate initiative focusing on Australia in which urban mitigation strategies will be tested with CLMU. For example, Coutts et al. (2010) suggested a number of strategies that could be taken into account in a strategic long-term plan to manage an estimated population growth of 1 million by 2030 (Coutts et al., 2007b): a wide-scale ‘cool roof’ program, which encourages highly reflective and high thermal emittance surfaces, increase in pervious surfaces, energy efficient buildings and revegetation of urban areas in combination with e.g. water sensitive urban design strategies (Coutts et al., 2013). Before the effect of such mitigation strategies (under possible climate change scenarios) can be addressed, the behaviour of CLMU in simulating the present-day observed urban climate should be evaluated. Thus, the objective of this study is to evaluate CLMU across two distinctly different urban landscapes, Toulouse (France) and Armadale, Preston and Surrey Hills, all part of the larger Melbourne metropolitan area (Australia). First of all, two types of simulations are performed for both locations: URB and URB_VEG. In the former, the vegetative fraction is treated as a pervious fraction within the urban canopy while in URB_VEG, the vegetation is explicitly treated by CLM4 without a direct interaction with the urban canyon. In addition, a number of sensitivity experiments are conducted to quantify the role of the anthropogenic heat flux and the characteristics of vegetation in an urban environment. The availability of long-term observations (1 year or more) for the Toulouse and Preston sites permits the consideration of seasonal variability which was not possible with the Mexico City and Vancouver evaluations (Oleson et al., 2008a). In addition, observations campaigns performed at these locations provide the capacity for more in-depth assessment of specific features of the model. For example, the availability of observed canyon temperatures and roof, wall and road surface temperatures in Toulouse allow for a detailed evaluation of the anthropogenic heat parameterization in CLMU. Also, the availability of three sites in the city of Melbourne, all characterized by a different vegetation fraction, supports a proper assessment of the representation of urban vegetation in CLMU.

This article is organized as follows. In Section 'Site description and methodology', an overview is given of the data and model used in this study, including a description of the characteristics of the Toulouse and Melbourne sites (Section 'Site specifics'), general features of CLMU (Section 'Model description'), a more detailed description of the hydrological properties of the model (Section 'Hydrological properties of the urban canyon') and the configuration of the baseline simulation (Section 'Configuration of baseline simulations') as well as the evaluation strategy (Section 'Evaluation strategy'). Section 'Sensitivity studies' describes in more detail the sensitivity studies that address the role of the anthropogenic heat flux and the effect of biophysical properties on the latent heat flux (Sections 'The parameterization of the anthropogenic heat flux' and 'Biogeophysical properties affecting the latent heat flux', respectively). Section 'Results' discusses the results for the sites in Melbourne (Section 'Sites in Melbourne') and Toulouse (Section 'Toulouse') and provides an in-depth analysis of the sensitivity studies (Sections 'Effect of anthropogenic heat' and 'Effects on evapotranspiration'). Finally, a discussion and conclusions are presented in Section 'Discussion and conclusions'. More details on the flux error estimation used in the evaluation strategy are provided in the Appendix A1.

2. Site description and methodology

2.1. Site specifics

In this study, two urban settings with distinctly different characteristics in terms of urban geometry, thermal characteristics and amount of vegetation are used. The observation site of Toulouse (France) is located in the old city centre where measurements were conducted during the Canopy and Aerosol Particle Interactions in TOulouse Urban Layer (CAPITOUL) experiment (Masson et al., 2008). In this neighbourhood, vegetation is very scarce (8%) and buildings are typically 4 to 5 stories (Pigeon et al., 2008). The area is classified as an old core thermal climate zone (TCZ) according to Stewart and Oke (2009). For the Melbourne region, three measurement sites are available, each characterizing a slightly different urban density and vegetation fraction (Coutts et al., 2007a,2007b). The first urban site is a highly developed medium-density residential site with housing and flats, located to the southeast of Melbourne in Armadale (Compact housing TCZ). The second urban site is located north of the Melbourne Central Business District (CBD) in Preston and consists of moderately developed low-density housing (Regular housing TCZ). This site has been used in phase 2 of the urban model intercomparison project (Grimmond et al., 2011). The third urban site is also a moderately developed suburban residential housing site, but of lower density, located east of Melbourne in the well-vegetated suburb of Surrey Hills (regular housing TCZ). A quantification of the urban morphological characteristics for the three Melbourne sites and Toulouse is presented in Table 1.

Table 1. Morphological characteristics of the Toulouse and Melbourne sites. The values between brackets denote the CLMU surface dataset configuration for the mixed urban/vegetated simulations. In this case, the ‘other’ fraction is regarded as pervious canyon floor. Subscript i refers to the layer number with i = 1: surface. All values are compiled from Coutts et al. (2007a), Pigeon et al. (2008) and Grimmond et al. (2011)
 SymbolArmadalePrestonSurrey HillsToulouse
Latitude (° N) −37°51′−37°43′−37°49′43°36′
Longitude (° E) 145°1′145°0′145°5′1°26′
UCZ 3552
Canyon height-to-width ratioH/W0.560.420.411.4
Roof fractionWroof0.46 (0.69)0.445 (0.71)0.39 (0.68)0.54 (0.59)
Pervious road fractionfprvrd0.63 (0.01)0.68 (0.005)0.75 (0.03)0.17 (0)
Vegetation fraction (no grass) 0.20.2250.280.06
Other 0.010.0050.030
Roof thickness (m)Δzroof0.11410.11410.11410.09
Wall thickness (m)Δzwalls0.14890.14890.14890.3
Impervious road thickness (m)Δzimprvrdi,1-4: 0.95i,1-4: 0.95i,1-4: 0.95i,1-4: 0.25
Building height (m)H8.
Max. height roughness elements (m) 16121620

For Armadale and Surrey Hills, no detailed inventory of the thermal properties is available. But since all Melbourne sites are primarily residential with similar characteristics as Preston, the same thermal properties are used (Table 2). The albedo values for Armadale and Surrey Hills are retrieved from Coutts et al. (2007a) while the emissivity values are kept constant for all three sites (0.973). For Toulouse, the thermal and radiative characteristics of the walls, roads and roofs are adopted from Pigeon et al. (2008) (see Table 2). Bare soil emissivity is hard-coded in the model and is set to 0.97 (Oleson et al., 2010c).

Table 2. Thermal input data for Preston and Toulouse required for the urban model. Thermal parameters of the soil for the pervious road are determined from soil texture. Subscript i refers to the layer number with i = 1: surface. All values are compiled from Coutts et al. (2007a), Pigeon et al. (2008) and Grimmond et al. (2011)
Roof thermal conductivity (W m−1 k−1)λroof,ii,1: 6.53i,1-10: 1.15
  i,2-8: 0.025i,11-15: 0.2
  i,9-14: 0.23 
  i,15: 0.16 
Wall thermal conductivity (W m−1 k−1)λwall,ii,1-4: 0.61i,1-15: 1.15
  i,5-10: 0.43 
  i,11-14: 0.024 
  i,15: 0.16 
Impervious road thermal conductivity (W m−1 k−1)λimprvrd,ii,1: 1.17i,1-2: 0.82
  i,2-3: 0.3i,3: 2.1
  i,4: 0.42i,4: 0.4
Pervious road thermal conductivity (W m−1 k−1)λprvrd,iSoil textureSoil texture
Roof volumetric heat capacity (MJ m−3 K−1)croof,ii,1: 2.07i,1-10: 1.58
  i,2-8: 0.0071i,11-15: 2.2
  i,9-14: 1.5 
  i,15: 0.67 
Wall volumetric heat capacity (MJ m−3 K−1)cwall,ii,1-4: 1.25i,1-15: 1.58
  i,5-10: 1.4 
  i,11-14: 0.0013 
  i,15: 0.67 
Impervious road volumetric heat capacity (MJ m−3 K−1)cimprvrd,ii,1: 1.14i,1-2: 1.74
  i,2-3: 1.05i,3: 2.0
  i,4: 1.29i,4: 1.4
Pervious road volumetric heat capacity (MJ m−3 K−1)cprvrd,iSoil textureSoil texture
Maximum internal building temperature (K)Tib,maxEvolves freelyEvolves freely
Minimum internal building temperature (K)Tib,minEvolves freely292.15
Soil texture (%) 65% sand, 10% clay33% sand, 33% clay

2.2. Model description

The CLMU is the urban parameterization in the latest version of the CLM4 (Oleson et al., 2010c; Lawrence et al., 2011), the land surface model of the CESM. Overall, CLM4 represents the land surface heterogeneity as a nested subgrid hierarchy in which grid cells are composed of multiple land units at the first subgrid level (glacier, lakes, vegetation, wetland and urban), and columns at the second subgrid level (e.g. for urban: roof, sunlit and shaded walls, pervious and impervious canyon floor (Figure 1 in Oleson et al., 2010a). A third subgrid level [plant functional type (PFT) level] includes the representation of up to 22 possible PFTs. All biogeophysical processes are simulated for each subgrid land unit, column, and PFT independently, while the same atmospheric forcing is used to force all subgrid units within one grid cell. The surface variables and fluxes are afterwards obtained by averaging the subgrid quantities weighted by their fractional areas (Oleson et al., 2010c). The city representation in CLMU is based upon the urban canyon configuration (Oke, 1987) in which the complexity of urban surfaces is reduced to a single urban canyon with an infinite length, integrated over all possible directions with equal probability and consisting of a canyon floor of width W surrounded by two facing buildings of height H (Figure 2 in Oleson et al., 2008a). The canyon system consists of roofs, sunlit and shaded walls, and pervious and impervious canyon floor. Each of these urban columns is divided into 15 layers for temperature and hydrology calculations and an additional five layers for snow where appropriate.

According to Grimmond et al. (2010, 2011), the model can be categorized as a model of ‘medium complexity’, having one or two simple characteristics. In CLMU, the storage heat flux ΔQs is calculated as the residual from the net shortwave and longwave radiation and latent and sensible heat fluxes (see Equation 49 in Oleson et al. 2008a), which is typified as a ‘simple’ characteristic in Grimmond et al. (2011). For further details on the formulations and parameterizations of the biogeophysical processes in CLMU, we refer to Oleson et al. (2010b).

2.3. Hydrological properties of the urban canyon

The roof and the impervious road columns of CLMU are hydrologically inactive except for their capacity to intercept, store and evaporate a limited amount of liquid precipitation and snow. The sunlit and shaded wall columns are hydrologically inactive. The impervious road column is intended to represent surfaces that do not allow water infiltration (e.g. roads, parking lots, sidewalks) while the pervious road column is intended to represent surfaces that may have active biophysical hydrology (e.g. residential lawns and parks). In this respect, the PILPS-urban project (Grimmond et al., 2010, 2011) suggests a threefold classification to denote whether vegetation is modelled in an urban surface or not: (Vn) not considered, (Vs) modelling vegetation following a tile approach in which the vegetation fraction does not interact with the urban canopy and (Vi) fully integrated into the urban surface. From their inventory it is clear that most models implement the Vn or Vs approaches. At present, CLMU implements the Vi approach, however, processes such as interception, evapotranspiration and shadowing are not yet explicitly integrated in CLMU because of their complexity and increase in computing resources. Hence, CLMU allows the urban vegetation to be defined as a separate vegetation tile (Vs-type), thereby building on all capacities of CLM4 (e.g. representation of different vegetation types). On the other hand, CLMU also has the ability to allow for evaporation in the urban canopy via a pervious canyon floor fraction that interacts directly with the urban canopy layer properties. As this approach does not account for the vegetation effects such as shadowing and absorption of radiation and momentum drag by an explicit representation of trees (see e.g. the vegetated urban canopy model of Lee and Park, 2008) this approach can be classified as a semi-integrated vegetation approach (Visemi).

2.4. Configuration of baseline simulations

As several of the parameterizations in CLMU are based on the town energy balance (TEB) model (Masson, 2000; Masson et al., 2002), which has been tested for Toulouse by Pigeon et al. (2008), we closely follow their simulation strategy allowing for a direct comparison of the results. Hence, CLMU is run in an offline mode for Toulouse and the three sites in Melbourne, meaning that the model is not coupled to an atmospheric model (no land–atmosphere feedback processes) but is forced with atmospheric data observed above the canopy layer. This removes a potential source of error produced by the atmospheric model and makes it possible to run longer simulations. Nevertheless, it is important to recognize that observed forcing data are not without error, typically related to the instruments and their calibration, post-processing of data, prevailing meteorological conditions etc (Grimmond et al., 2011). The atmospheric forcing parameters to run the model are surface pressure (Pa), incident short- and longwave radiation (W m−2), atmospheric temperature (K), wind speed (m s−1), precipitation (mm s−1) and relative or specific humidity (% or kg kg−1). A more detailed description of the instrumentation used to obtain the forcing data can be found in Section 3a of Coutts et al. (2007b) and Sections 2.3a and b in Pigeon et al. (2008) for Melbourne and Toulouse, respectively. For Melbourne, the precipitation forcing is the same for all the sites as precipitation rates were not measured directly at the sites but were taken from the nearby observation station at Monash University.

In Pigeon et al. (2008), TEB was used jointly with ISBA (Interaction Soil–Biosphere–Atmosphere model Noilhan and Planton, 1989) using the Vs approach, while in Grimmond et al. (2011) CLMU was used without explicitly modelling the vegetation (Visemi approach). Therefore, both approaches are used here: (1) 100% urban (hereafter referred to as URB), in which the vegetative fraction is assigned as the pervious fraction in the canyon and (2) mixed urban/vegetation (hereafter referred to as URB_VEG), in which the vegetation fraction is explicitly treated by CLM4 without a direct interaction with the urban canyon (Table 1). For the URB_VEG simulation, additional input is required to describe the vegetation characteristics. In Toulouse, the site is characterized by broadleaf deciduous and evergreen temperate (BDT) trees and grass, while the Melbourne sites generally have broadleaf evergreen temperate (BET) and deciduous temperate trees and grass (Coutts et al., 2007b). As in a next step of this research, this model will be applied for high-resolution regional climate simulations for the larger Melbourne area, generic gridded information for the PFTs is necessary. Hence, the default URB_VEG simulations use the default CLM4 values for monthly leaf area index and stem area index (LAI and SAI) and top and bottom monthly vegetation height for BET-temperate, BDT-temperate and C3 grass. This land surface data is derived from Moderate Resolution Imaging Spectroradiometer (MODIS) images from 2001 to 2003 and is available on a resolution of 0.05° (Lawrence and Chase, 2007). Irrigation was limited for the Melbourne sites during the observation period due to restrictions of domestic water use (Coutts et al., 2007b).

In order to be consistent with the approach followed by Pigeon et al. (2008) for Toulouse, the anthropogenic heat released by traffic (QFtraffic) is prescribed using the hourly QFtraffic values described in Pigeon et al. (2007). The additional heat flux from heating buildings is parameterized directly as a function of the temperature difference between the roof and wall surface temperatures and the interior building temperature. As in Pigeon et al. (2008), no cooling is applied so that only a minimum building temperature TiB,min of 292.15 K is used. Thus, this approach does not take into account human metabolism, a reasonable assumption as the latter is generally <1% of the total anthropogenic heating for a city (Sailor, 2011). For the Australian sites, the total anthropogenic heat flux (QF) is low throughout the year, with a maximum hourly value of 16 W m−2 in winter, based on an inventory approach by Coutts et al. (2007b) that was similar to the approach suggested by Sailor and Lu (2004). For these sites, we opt to prescribe the total anthropogenic heat flux in order to minimize possible parameterization errors with respect to QF and to allow for a more in-depth analysis of the latent heat flux and its possible sources.

Spin-up of the model is achieved following the criteria mentioned in Chen et al. (1997); Stockli et al. (2008). CLMU is run repeatedly over a full year of data until the mean yearly latent and sensible heat fluxes were within 0.1 W m−2 of those from the previous spin-up cycle. This corresponds to a change in temperature and volumetric soil moisture of the upper surface layer of <0.01 K and 0.001 mm3 mm−3, respectively. For the impervious fractions such as roofs, walls and impervious roads, spin-up was achieved after one cycle. For the pervious fractions, more cycles are needed until the surface temperature and soil moisture are stable.

2.5. Evaluation strategy

For all four sites, the evaluation of the model results was done using observed outgoing short- and longwave radiation, net radiation and sensible and latent heat fluxes for four 30-d period, each representing a season (Table 3). Owing to the limited amount of vegetation, the evaluation for Toulouse focuses on outgoing short-and longwave radiation and the sensible heat flux. As the storage heat flux is calculated as the residual term, both in the observations and in the model and hence accumulates all the measurement errors and missing terms (e.g. horizontal advection), this flux is not considered in this study. However, it is possible to evaluate the surface temperatures for canyon walls, roads and roofs that are state variables of the model and are directly measured. The evaluation of these variables directly reflects the ability of the model to reproduce the energy balance of each surface. The measured temperatures for roads are corrected for the occurrence of traffic during the observation period (Pigeon et al., 2008) and will be compared to the surface temperatures of the impervious fraction of the road in CLMU, as the measurements were done in an urban canyon without vegetation/pervious fraction. As CLMU (or TEB) only produces a single wall and road temperature representative for all canyon directions that are equally probable, a weighted averaging method was developed by (Moscicki, 2007) that appropriately combines the observations from multiple wall and road surfaces while accounting for variations in both their canyon geometry and their orientation. As such, the wall temperatures are compared with the arithmetically averaged sunlit and shaded wall temperatures from CLMU. Furthermore, urban canyon temperature observations are also available for Toulouse, measured in the street canyon along the building where the flux tower was set, mounted on a boom 2 m below the top of that building. In addition, the uncertainty estimates developed in the Appendix A1 will be used as a framework of model performance against observations and to decide whether differences between the URB and URB_VEG and the default and sensitivity experiments are significantly different (e.g. when |URB − URB_VEG| > δ).

Table 3. Offline simulation periods (denoted by the full black circles) for the sites in Melbourne and Toulouse
 15 January to 15 February 200415 April to 15 May 200415 July to 15 August 200415 October to 15 November 2004
  1. a

    Due to data limitations, the period 15 June to 15 July 2004 is used here.

  2. b

    These periods coincide with the TEB simulation periods in Pigeon et al. (2008).

Surrey Hills a 
 15 July to 15 August 200415 October to 15 November 200415 January to 15 February 200515 April to 15 May 2004

Comparison statistics used in this study follow the ones used in Pigeon et al. (2008), Oleson et al. (2008a) and Grimmond et al. (2010). They include observed and modelled standard deviation, root mean square error (RMSE), the mean bias error (MBE), the coefficient of determination (R2) and both systematic (RMSES) and unsystematic (RMSEU) components. Ideally, the unsystematic error should be larger than the systematic error as the former error originates more from the model's inability to cope with the variability in the observations compared to having problems with model physics or parameter values (systematic error). For more details on the exact equations used to calculate the statistics refer to Grimmond et al. (2010).

3. Sensitivity studies

Some sensitivity simulations are conducted to help understand the results of the model with respect to a range of boundary conditions: the anthropogenic heat (see Section 'The parameterization of the anthropogenic heat flux') and the characteristics of vegetation in an urban environment (see Section 'Biogeophysical properties affecting the latent heat flux').

3.1. The parameterization of the anthropogenic heat flux

In Oleson et al. (2008b) CLMU is used to investigate the relationship between heat island intensity and city size based on a single North American site. In that study, the effect of the total anthropogenic heat is tested in terms of its effect on the urban heat island intensity (simulation with and without QF) while the validity of the magnitude of the anthropogenic heat flux itself is not evaluated based on estimates of QF. Similarly, Oleson et al. (2010a) presents a slightly revised scheme in CLMU to calculate the anthropogenic heat in order to address urban heat island characteristics in a global climate model. Here, the total global anthropogenic heat from the urban model for the present-day climate simulation is compared with the global anthropogenic heat estimates released from non-renewable resources derived by Flanner (2009). Thus, the results from the default URB simulation provides the opportunity to evaluate the CLMU urban heating parameterization based on estimates of QF made for Toulouse by Pigeon et al. (2008). Moreover, two additional runs were performed to test the sensitivity of the model with respect to QF: (1) a simulation without QF (hereafter referred to as URB-noQF) and (2) a simulation in which the total anthropogenic heat flux estimated by Pigeon et al. (2007) is prescribed (hereafter referred to as URB-QFpre). The effects of these different treatments of anthropogenic heat on outgoing short- and longwave radiation, sensible heat and surface temperatures are discussed in Section 'Effect of anthropogenic heat'

3.2. Biogeophysical properties affecting the latent heat flux

As the observation sites in Melbourne are characterized by a large fraction of vegetation compared to the old city core of Toulouse, some additional analysis were performed to test the impact of vegetation properties and other hydrological properties of the urban canyon on the latent heat flux. Peel et al. (2005) suggest that the broad definition of vegetation-based biophysical parameters that are often used in regional climate simulations may not represent the specific vegetation of the Australian sites adequately. Although they show that their vegetation model does not substantially improve, they do find a significant impact on the latent heat fluxes. In the default URB_VEG simulation, the trees in the Melbourne sites are described as the BET trees PFT, using the default BET-temperate values of CLM4. Nevertheless, the Preston site is largely dominated by eucalypt, the main tree group on the continent. Hence, a sensitivity analysis is done using more eucalypt-specific values for the broadleaf evergreen trees PFT class, as shown in Table 4. In addition, Coutts et al. (2007b) states that for Armadale and Surrey Hills more exotic (deciduous) trees are present. For these sites, the vegetation fraction of the sensitivity simulation is divided into 30% BET-temperate (eucalypt) and 70% BDT-temperate compared to 100% BET-temperate trees with their default characteristics in URB_VEG.

Table 4. The default and eucalypt-specific biophysical values for vegetation roughness length for momentum (same as for heat) (Z0m,v), minimum stomatal resistance rs,min, maximum vegetation height for BET and BDT temperates trees (ztop,BET and ztop,BDT, respectively), maximum leaf area index (LAImax,BET), leaf visible and near-IR reflectances and transmittances math formula, math formula, math formula, math formula) and leaf inclination factor χL used for the broadleaf evergreen temperate (BET) tree plant functional type class. Other parameters such as stem visible and near-IR reflectances and transmittances are kept constant as no Eucalypt-specific values were available. The default values are from the PFT tables in Oleson et al. (2010c) while the eucalypt-specific values are compiled from Sinclair and Thomas (1970), James and Bell (2000) and Peel et al. (2005)
Z0m,v1.22.41CLM4 calculates the Z0m,v indirectly via the height of the canopy and a constant Rz0m [Tables 2.2 and 5.1 in Oleson et al. (2010c)]. These values are changed to 16 m (Coutts et al., 2007b) and 0.15 m, respectively in order to obtain a value as mentioned in the table
rs,min250The inverse of the maximum stomatal conductance in Peel et al. (2005) is used to define a minimum stomatal resistance
ztop, BET3516 (12)The maximum height of the BET temperate trees is defined as the maximum height of the roughness elements at Armadale and Surrey Hills (Preston) (see Table 1)
ztop, BDT2016 (12)The maximum height of the BDT temperate trees is defined as the maximum height of the roughness elements at Armadale and Surrey Hills (Preston) (see Table 1)
LAImax, BET2.63.2The default monthly LAI values are taken from Lawrence and Chase (2007) while the seasonal cycle of the eucalypt-specific LAI follows the same seasonal cycle but with a maximum value as defined in Peel et al. (2005)
math formula0.10.44See Sinclair and Thomas (1970)
math formula0.450.6See Sinclair and Thomas (1970)
math formula0.050.03See Sinclair and Thomas (1970)
math formula0.250.36See Sinclair and Thomas (1970)
χL0.1−0.38The mean leaf inclination (math formula) is a function of χL as shown in Equation 3.14 of Oleson et al. (2010c). The value here is obtained using math formula = 72° (James and Bell, 2000)

The changes in biophysical values are listed in Table 4. In addition, the two root distribution parameters ra and rb that are used to determine the root fraction for each soil layer (see Equation 8.21 in Oleson et al., 2010c) are set to 11 and 3, respectively, to get a minimum root fraction at the maximum root depth as described in Peel et al. (2005). The parameters described in Table 4 also correspond to the parameters that have the strongest effect on latent heat fluxes as suggested in Loridan et al. (2010). Nonetheless, in their approach they optimize the parameters with respect to the performance of some targeted fluxes (in this case Qe) instead of using the best available and realistic information on a specific vegetation type. The effect of the specific vegetation properties on latent heat fluxes for the Melbourne sites is presented in Section 'Effects on evapotranspiration'

In contrast to the URB_VEG simulation, the latent heat flux from the URB simulation is on the one hand governed by the extent and properties of the pervious (bare soil) fraction in the urban canyon (fprvrd) that is hydrologically fully active and on the other hand the amount of liquid water or snow that is available from the impervious surface, viz. the amount that can be intercepted, stored and evaporated from the roof and impervious road. With respect to the former, the amount of evaporation that is released by fprvrd is a function of the wetness of the soil column which allows all the soil moisture to be potentially available for evaporation. The wetness of the soil depends on its hydraulic properties which in turn depend on the soil texture through the saturated hydraulic conductivity ksat (for more information we refer to Oleson et al., 2008a). The soil texture for the Australian sites is defined as sandy loam (Grimmond et al., 2011). In the default URB simulation, this texture is defined as 10% clay and 65% sand which is near the midpoint of the sandy loam soil texture class composition. As a sandy loam soil covers a larger spectrum of sand/clay ratios, the sensitivity of the latent heat from the pervious fraction of the soil for the URB simulation is shown by using the four outer combinations of the sandy loam texture (hereafter referred to as SOILTEX1: 50% sand – 0% clay, SOILTEX2: 70% sand – 0% clay, SOILTEX3: 80% sand – 15% clay and SOILTEX4: 50% sand – 20% clay). Furthermore, the amount of latent heat from the impervious surfaces (roof and impervious road) depends on the fraction of the surface that is wet, with a maximum amount of water an impervious surface can hold set by default to wpond,max = 1 kg m−2 (Oleson et al., 2008a). The liquid water in excess of this storage capacity is routed to surface runoff. In order to test the sensitivity of the model to the maximum amount of water a surface can hold the default value of wpond,max is perturbed by −50 and +50%. The effect of the sensitivity experiments related to soil texture and water ponding on the latent heat flux per surface fraction is discussed in Section 'Effects on evapotranspiration'

4. Results

The following section evaluates the modelled outgoing short- and longwave (and net) radiation and sensible and latent heat fluxes with the same observed parameters for the measurement sites in Melbourne and Toulouse. All statistical scores and diurnal cycles have been computed for both the URB and the URB_VEG simulation, for all observation sites and available seasons and only take into account the time steps when valid observations were available (denoted by Nobs in Tables 5). For Toulouse, mean diurnal cycles and relevant statistics are presented for observed and modelled canyon temperatures and surface temperatures of roofs, walls and roads as well.

Table 5. Summary statistics of simulated and observed outgoing shortwave (K↑) and longwave (L↑) radiation and sensible (Qh) and latent heat flux (Qe) for Preston. Here, Nobs is the number of valid observations used for each seasons, σobs and σmod the standard deviations of observations and model, respectively (W m−2), MBE is the mean bias error (model – obs) (W m−2), RMSE the root-mean-square error (W m−2) with its systematic (RMSES) and unsystematic (RMSEU) component and R2 is the coefficient of determination. Bold indicators denote the best performance between URB and URB_VEG, while italic RMSES show systematic errors smaller than unsystematic errors
  1. The asterisks (*) refer to the fluxes for which |MBEURB − MBEURB_VEG| are larger than δ (see Appendix A1).

Winter (Nobs = 595)        
Spring (Nobs = 620)
Summer (Nobs = 689)
Fall (Nobs = 789)

4.1. Sites in Melbourne

First, we apply the daily-differencing uncertainty estimation described in Appendix A1 to decide whether the modelled Qh and Qe from the URB and URB_VEG simulations are significantly different for the sites in Melbourne. As the magnitude of the uncertainty δ scales linearly with the net radiation, the highest uncertainty is obtained for summer with a temporal mean δ for Preston of 10.1 and 3.4 W m−2 for Qh and Qe, respectively. Values for the other seasons are lower over all sites with mean uncertainties for Qh (Qe) typically around 0.7–1.8 (0.2–0.6) W m−2, 1.2–2.2 (0.4–0.7) W m−2 and 6.2 (2.1) W m−2 for winter, fall and spring respectively. For Qh and Qe at most sites and seasons, except for Qh in spring and fall in Preston and summer in Armadale, the difference |URB − URB_VEG| is larger than δ so that URB and URB_VEG are considered significantly different. Hence, the main results for both the URB and URB_VEG configuration are discussed below.

For Preston, URB outperforms URB_VEG for shortwave outgoing radiation (K↑) over all seasons (Figure 1) with a small MBE between −0.6 and 1.0 W m−2 and a RMSE between 1.7 and 5.5 W m−2 (Table 5). Table 5 also shows that RMSES is smaller than RMSEU for K↑ in URB over all seasons. In general, the results for the outgoing longwave radiation (L↑) closely follow the conclusions made by Grimmond et al. (2011) (their Section 3.1.2): the bias for L↑ is larger than for K↑, RMSES is larger than RMSEU over all seasons and the model generally performs better during the night than during the day for both URB and URB_VEG (Figure 1), with a maximum MBE/RMSE in summer of 20.9/35.6 and 12.6/30.1 W m−2, respectively (Table 5). The use of the URB_VEG configuration seems to release less L↑ over the 24-h period, improving the MBE but deteriorating the RMSE.

Figure 1.

Observed and modelled monthly average diurnal cycle of outgoing shortwave (a, e, i and m) and longwave (b, f, j and n) radiation and sensible heat (c, g, k and o) and latent heat flux (d, h, l and p) for Preston in spring (left column), summer (second column), fall (third column) and winter (right column). Open circles/error bars represent the observations, grey/red the URB simulation and blue the URB_VEG simulation. Hourly variability over the considered time is represented by error bars (observations), grey-shaded area and dashed-dotted lines (URB and URB_VEG, respectively).

Overall, the good representation of K↑ and daytime overestimation of L↑ results in an underestimation of net radiation (Q*) over all seasons, with a slightly better performance for URB than for URB_VEG in winter and spring (not shown). Sensible heat (Qh) is overestimated in all seasons but winter (Figure 1), and is better represented by URB (compared to URB_VEG) in winter and spring. Only for fall the statistics reveal a significant improvement of the diurnal variability of Qh for URB_VEG (Table 5). The differences in latent heat flux (Qe) between URB and URB_VEG are significant for all seasons, as the difference in MBE between the two model configurations is higher than the uncertainty estimate (Table 5 and Figure 1). Apart from winter and spring, the bias in Qe is negative in both the URB/URB_VEG simulations, but smaller in fall compared to summer. For both seasons (and especially summer), there is a steep drop in latent heat flux in URB_VEG after 9 am. Overall, the unsystematic error is higher than the systematic error for both turbulent fluxes over all simulations.

Similar results are obtained for Armadale and Surrey Hills (Figures 2 and 3). Again, the difference |URB − URB_VEG| is larger than δ which results in a significant difference in performance for Qh and Qe between URB and URB_VEG. For both sites, K↑ is overall underestimated while L↑ is generally overestimated. For the latter case, night-time L↑ is underestimated while daytime L↑ is overestimated, resulting in a small overall MBE of −2.6 and −4.9 W m−2 for URB and URB_VEG respectively. The modelled sensible heat flux in Armadale is generally as in Preston while better results are obtained for Surrey Hills, with a significant improvement for URB_VEG compared to URB (Figure 3). The better representation of Qh originates from a better partitioning of the turbulent fluxes in sensible and latent heat for Surrey Hills, especially in winter. This could be due to the fact that this site is characterized by the highest vegetation fraction which, in case of URB_VEG, seems to be better represented by the explicit representation of vegetation in CLM. For Armadale, the summer Qe of the URB_VEG simulation is again characterized by a steep drop after 9 am, while the URB simulation shows some improvement (Figure 2(d)).

Figure 2.

Observed and modelled monthly average diurnal cycle of outgoing shortwave (a, e, i and m) and longwave (b, f, j and n) radiation and sensible heat (c, g, k and o) and latent heat flux (d, h, l and p) for Armadale in summer (left column) and fall (right column). Open circles/error bars represent the observations, grey/red the URB simulation and blue the URB_VEG simulation. Hourly variability over the considered time is represented by error bars (observations), grey-shaded area and dashed-dotted lines (URB and URB_VEG, respectively).

Figure 3.

Observed and modelled monthly average diurnal cycle of outgoing shortwave (a, e, i and m) and longwave (b, f, j and n) radiation and sensible heat (c, g, k and o) and latent heat flux (d, h, l and p) for Surrey Hills in fall (left column) and winter (right column). Open circles/error bars represent the observations, grey/red the URB simulation and blue the URB_VEG simulation. Hourly variability over the considered time is represented by error bars (observations), grey-shaded area and dashed-dotted lines (URB and URB_VEG, respectively).

4.2. Toulouse

For Toulouse, differences in MBE for Qh between the URB and URB_VEG simulation are smaller than the uncertainty derived in Appendix A1, probably due to the low amount of vegetation at this site. Hence, only the results for the URB simulation are discussed here. For all seasons, the net radiation is very well modelled in terms of the mean daily cycle and variability (Figure 4). This results in a low mean bias and RMSE over all seasons, similar to the results obtained by Pigeon et al. (2008) (Table 6). This good representation of Q* relates to a trade-off between the errors in outgoing short- and longwave radiation. Overall, K↑ is underestimated with the largest bias in spring and summer, while L↑ is overestimated with a maximum MBE of 10.6 W m−2 in summer. This results in a small underestimation of Q* for all seasons except spring, in which the underestimation in K↑ is slightly larger than the overestimation of L↑. Sensible heat fluxes are generally overestimated with a MBE between 10.7 and 25.4 W m−2 in fall and spring, respectively (Table 6). For Q* and Qh (K↑ and L↑) in most seasons, RMSES is smaller (larger) than RMSEU suggesting some room for improvement with respect to the model physics and parameter values. Finally, the magnitude of QF is relatively well reproduced for the transition seasons spring and fall while it is underestimated (overestimated) in summer (winter). In addition, all seasons except summer show a fairly poor temporal correlation with the observed QF values (Figure 5 and Table 6). This is further discussed in Section 'Effect of anthropogenic heat'

Figure 4.

Observed and modelled monthly average diurnal cycle of outgoing shortwave (a, e, i and m) and longwave (b, f, j and n) radiation, net radiation (c, g, k and o) and sensible heat flux (d, h, l and p) for Toulouse in spring (left column), summer (second column), fall (third column) and winter (right column). Open circles/error bars represent the observations, grey/red represents the URB simulation and the light blue and dark blue represent the QFpre and noQF model simulations, respectively. Hourly variability over the considered time is represented by error bars (observations) and grey-shaded area (model).

Table 6. Similar as for Table 5 but for Toulouse and net radiation (Q*), sensible heat (Qh) and total anthropogenic heat (QF) and the model simulation. The values in brackets in fall and winter for MBE and RMSE denote the values obtained by Pigeon et al. (2008) with TEB
Winter (Nobs = 301)     
MBE2.66.5−2.6 (5)18.1 (16)20.8 (18)
RMSE9.27.45.8 (7)32.1 (52)29.3 (28)
Spring (Nobs = 295)
Summer (Nobs = 686)
Fall (Nobs = 459)
MBE−3.95.2−1.2 (4)10.7 (−11)4.8 (2)
RMSE6.36.96.0 (8)31.4 (44)23.8 (23)
Figure 5.

Observed and modelled monthly average diurnal cycle of the total anthropogenic heat flux in spring (a), summer (b), fall (c) and winter (d). Hourly variability over the considered time is represented by error bars (observations) and grey-shaded area (model).

As the longwave radiation is primarily a function of surface temperature, the model performance with respect to the surface temperatures for all impervious surfaces is shown in Figure 6 and Table 7. The surface temperatures for the URB simulation in winter and fall for walls and roads (Twall and Troad, respectively) are well represented with a MBE/RMSE between 1.7–1.8/1.8 K and −0.5 to 0.2/1.2–1.5 K, respectively. Thus, CLMU performs slightly better (worse) for Troad (Twall) compared to the values obtained with TEB (Pigeon et al., 2008). The urban surfaces in summer are generally too warm which could explain the largest overestimation of L↑ for that season (Figure 4). For example, the bias in Troad and Twall are higher than the values obtained by Moscicki (2007) with a difference in MBE/RMSE of 0.56/0.52 and 0.96/0.28 W m−2, respectively (Table 7).

Figure 6.

Observed and modelled (URB, URB-Qfpre and noQF) monthly average diurnal cycle of the urban canyon temperature (a, e and i) and surface temperatures for roofs (b, f and j), walls (c, g and k) and roads (d, h and l) for summer (left column), fall (middle column) and winter (right column) in Toulouse. In addition, the observed air temperature measured aloft the urban canyon in shown in orange in (a), (e) and (i). Hourly variability over the considered time is represented by error bars (observations) and grey-shaded area (model). For clarity, the hourly variability is not shown for URB-QFpre, URB-noQF and the observed air temperature.

Table 7. Comparison between modelled (URB, URB-QFpre and URB-noQF) and observed surface temperatures for roofs, walls and roads and urban canyon temperatures for Toulouse. The results for URB_VEG are not shown as they are the same as for the URB simulation (road temperatures refer to the impervious fraction of the road which is the same in both simulations [see Table 1)]. The values in brackets in fall and winter (summer) for MBE and RMSE denote the values obtained by Pigeon et al. (2008), Moscicki (2007) and Bueno et al. (2011). Values in bold refer to the simulation with the best performance
  Troof (K)Twall (K)Troad (K)Tcanyon (K)
Winter (Nobs = 1408)
URBMBE3.3 (1.2)1.8 (0.8)0.5 (−2.2)0.2
 RMSE3.9 (2.2)1.8 (0.9)1.2 (2.4)0.5
Summer (Nobs = 1143)
URBMBE5.9 (3.81)1.5 (0.54)1.8 (1.24)0.9 (0.21)
 RMSE6.6 (4.28)1.7 (1.42)2.8 (2.28)1.2 (0.96)
Fall (Nobs = 1387)
URBMBE2.5 (1.0)1.7 (0.0)0.2 (−1.4)0.2
 RMSE3.0 (2.1)1.8 (0.5)1.5 (1.6)0.4

Roof surface temperatures are overestimated in all seasons, with a MBE between 2.5 and 5.9 K for fall and summer, respectively. These values are larger than the results presented in Pigeon et al. (2008) and Moscicki (2007) with a difference up to 2.1 K in winter and summer (Table 7). The daily cycle and variability of the urban canyon temperatures (Tcanyon) are overall well simulated with the smallest (largest) bias in fall (summer) (Table 7 and Figure 6). Here, the MBE/RMSE strongly depends on the time of day, as e.g. at night during summer MBE/RMSE is only 0.3/0.4 K compared to 0.9/1.2 K during the day (Figure 6).

The larger bias for the roof surface temperatures in CLMU and the larger overestimation of Tcanyon compared to TEB (Bueno et al., 2011) could point to the different approach used by CLMU and TEB with respect to the roofs: in CLMU, above-roof air is mixed with the canyon air while in TEB, the roof surface interacts directly with the atmosphere above (Masson, 2000; Oleson et al., 2010c). We hypothesize that in CLMU, cooling of the roof fraction is generally inhibited by the higher Tcanyon (compared to air temperature above the urban canopy layer) which in turn suppresses a cooling of the other urban canyon surfaces via the iterative calculation of the sensible heat flux. In order to address this hypothesis, an additional simulation is performed in which the roof is decoupled from the urban canyon (hereafter referred to as URBroof) as sketched in Figure 7. In this approach, new aerodynamic resistances to sensible heat and water vapour (rah,roof and raw,roof, respectively) are defined between the roof surface and the atmosphere, while the roof aerodynamic resistance for momentum is set to 0.15 m s−2, the default used in TEB (Masson, 2000). Consequently, the canyon potential temperature θs and specific humidity qs are governed by the wall and road surfaces only. As this approach is more similar to the original TEB scheme, the evaluation of surface temperatures and turbulent fluxes is again compared with the results obtained by Moscicki (2007) and Pigeon et al. (2008).

Figure 7.

Sketch of modified sensible and latent heat fluxes for the urban canopy. T, H and E represent, respectively, the surface temperature, sensible and latent heat fluxes from all the urban surfaces (roof, sun and shaded wall and (im)pervious road). There are no latent heat fluxes from the walls as they are hydrologically inactive. The original CLMU urban canopy scheme and a more detailed explanation of the abbreviations can be found in Oleson et al. (2010c).

Figure 8 depicts the monthly average diurnal cycle of temperature differences between URBroof-URB for the urban canyon and surfaces. For all seasons, the URBroof simulation decreases Troof with almost 1 K at night and up to 5.5 K during daytime in summer. This results in a MBE of 2.4, 3.6 and 1.6 K for winter, summer and fall, respectively, with a good model performance for daytime roof surface temperatures and an overestimation at night. These values correspond more closely to the values obtained by Pigeon et al. (2008) and Moscicki (2007) as shown in Table 7. Owing to its connection with the atmosphere aloft, Troof is cooler during daytime, although this does not alter the daytime canyon temperature. At night, the cooler roof surface air does no longer interact with the canyon properties but this has little to no effect on Tcanyon, Twall and Troad (Figure 8). The changes in surface temperatures are also reflected in urban surface energy balance changes. On average, the lower surface temperatures in URBroof decrease L↑ which in turn results in more net radiation. This consequently leads to higher sensible heat fluxes, especially during daytime in spring and summer when the temperature gradient between the roof surface and the atmospheric air aloft is larger.

Figure 8.

Monthly average diurnal cycle of temperature differences between URBroof and URB for the urban canyon and the roof, wall and road surface in summer (upper panel), fall (middle panel) and winter (lower panel).

4.3. Sensitivity analysis

4.3.1. Effect of anthropogenic heat

In order to address the importance of the anthropogenic heat flux and assess the way in which CLMU deals with this additional source of energy, two additional simulations are performed as described in Section 'The parameterization of the anthropogenic heat flux' Figure 5 shows that the URB simulation well reproduces the magnitude of the anthropogenic heat flux in spring and fall, although the diurnal cycle is too dynamic as it has a strong response on the diurnal cycle of the outdoor air temperature. Especially in the hours before sunrise this results in an overestimation of QF. On the other hand, the diurnal cycle of anthropogenic heat in winter is overall overestimated (except for the afternoon hours) while summer QF is underestimated throughout the day because of the fact that air temperatures do not fall below the TiB,min threshold. Here, the magnitude of the anthropogenic heat flux is solely caused by the prescribed traffic heat flux based on the inventory from Pigeon et al. (2007).

A first sensitivity simulation fully prescribes the total observed anthropogenic heat flux as derived in Pigeon et al. (2007) (URB-QFpre). In Oleson et al. (2008a), the waste heat generated as a byproduct of heating/cooling buildings was modelled as a sensible heat flux into the urban canopy layer. However, it was found that if this flux is large enough, the numerical solution may become unstable because the urban canopy air itself has no heat capacity and the heat capacities of the roofs and walls are relatively small. In order to circumvent this, the waste heat is presently added to the net heat flux of the canyon floor, viz. the pervious and impervious road. Thus, the prescribed waste heat is treated similar as the waste heat parameterization of CLMU, viz. added as an additional heat flux to the road fraction of the urban surface. This increases the road surface temperatures, leading to an increase in MBE up to 4.9 K in winter (Table 7). On the other hand, it decreases the bias in Troof/Twall to 0.9/−0.4 and 1.3/0.7 K in winter and fall, respectively, and slightly increases MBE in summer. The general net effect of changes in surface temperatures for the different surface fractions is a decrease (increase) of L↑ that is highest in winter (summer) (Table 8). As the shortwave outgoing radiation does not change, this results in an overall deterioration of the modelled net radiation in all seasons except summer. For the latter, higher road surface temperatures (Figure 6 and Table 7) with similar wall and roof surfaces temperatures increase L↑ and hence decrease Q*. The effect on the sensible heat flux generally follows the bias in modelled QF, with a better representation of Qh in winter (MBE = −2.8 W m−2) and fall (MBE = 0.5 W m−2) (Table 8). For summer, the higher QF values, slightly compensated by a lower net radiation, result in a larger overestimation of Qh (Table 8).

Table 8. Similar as for Table 6 but for the URB-noQF and URB-QFpre simulations. Statistics for QF are excluded as QF is not taken into account (URB-noQF) or is similar to the observed QF as plotted in Figure 5 (URB-QFpre)
Winter (Nobs = 301)
Spring (Nobs = 295)
Summer (Nobs = 686)
Fall (Nobs = 459)

As a second sensitivity test, a simulation is performed without additional anthropogenic heat (URB-noQF). In summer, the results described in Section 'Toulouse' revealed a low anthropogenic heat flux only originating from the prescribed QFtraffic as the urban canopy temperatures did not go below the TiB,min threshold of 292.15 K. Hence, excluding this additional source of energy mainly affects Troad with a decrease in MBE of almost 1 K. Nevertheless, a small improvement can be seen for the other surface temperatures via their interaction with a slightly better represented Tcanyon. In addition, although the waste heat parameterization for the URB simulation performed best for QF in fall (Table 6), better results are obtained for Twall in fall with the URB-noQF simulation. Simultaneously, the overestimation of L↑ in summer is reduced resulting in a better representation of Q* and Qh (Table 8 and Figure 4(f)–(h)). For the other seasons the URB-noQF simulation overall results in lower surface temperatures throughout the urban canyon with some of the lowest biases over all simulations depending on surface type and season (Table 8 and Figure 6). As fall and winter are characterized by a larger amount of anthropogenic heat, the removal of this additional heat source underestimates Tcanyon (especially at night) while Troof (Troad) has a smaller (larger) bias, more similar to the results obtained by Pigeon et al. (2008). Wall temperatures are very well represented for fall while underestimated in winter (Table 8). These changes in surface temperatures reduce the MBE of L↑ to −6.8 and −3.0 W m−2 which in turn result in an increase in MBE for Q* up to 12.9 and 9.8 W m−2 for winter and fall, respectively. This increase in available net radiation slightly compensates the missing anthropogenic heat but ultimately reveals a strong underestimation of Qh, with a MBE of −37.2 and −56.7 for fall and winter, respectively (Table 8 and Figure 4(l) and (p)). For spring the URB-noQF actually provides better results for Qh with a MBE/RMSE of −0.7/34 W m−2 (Figure 4(d)).

4.3.2. Effects on evapotranspiration

The modelled latent heat flux Qe from the URB and URB_VEG simulations for the sites in Melbourne revealed a varying model performance depending on the season considered (see Section 'Sites in Melbourne'). In order to have a better understanding of the contribution of each individual surface to the total latent heat flux, Figure 9 shows the various latent heat fluxes weighted per surface fraction for Armadale in summer, Surrey Hills in fall and Preston in spring. Each of these seasons are characterized by the largest difference in Qe between the URB and the URB_VEG simulation (Table 5 and Figures 2 and 3).

Figure 9.

Observed and modelled monthly average diurnal cycle of latent heat fluxes per surface fraction (W m−2) for the default URB (left) and URB_VEG (right) simulations for Armadale in summer (upper panels), Surrey Hills in fall (middle panels) and Preston in spring (lower panels). The dashed line represents the sum of the latent heat flux for all pervious fractions (pervious road, BET tree and grass).

For the URB simulation, the pervious fraction contributes the most to the total latent heat flux as every soil layer of this fraction contributes to the evaporative flux which allows more water to be available during dry conditions (as was the case for this study period). The Qe originating from the roof fraction is characterized by a rapid decrease of evaporative flux during the course of the day as the amount of water that can be stored is limited according to wpond,max. As the ensemble average diurnal cycle of precipitation reveals highest precipitation in the morning (not shown), this amount of water evaporates rapidly in the beginning of the day. This results in a sharp decrease of Qe from these surfaces after 9 am. But due to the relative small contribution of this surface type this has only limited influence on the total latent heat flux.

The URB_VEG simulation produces less latent heat compared to the URB simulation for all sites. For Armadale, characterized by 33% vegetation, the BET fraction contributes the most to the total latent heat flux, followed by evaporation from the roof. As the latter falls back to 0 W m−2 around midday, this causes the strong decrease in total Qe (Figure 9). In contrast, βt, a soil moisture function limiting transpiration (Oleson et al., 2010c) is approximately 1 throughout the whole period (not shown) which indicates that the soil column is wet and hence allows for transpiration. For Surrey Hills, characterized by a larger fraction of vegetation, the URB_VEG simulations presents a better partitioning of the turbulent energy between sensible and latent heat. For example, for fall, the sum of the latent heat fluxes from the pervious road, BET trees and grass in URB_VEG is significantly smaller than the latent heat from the pervious road in URB and better represents the observed Qe during daytime (Figure 9). On the other hand, the small pervious road fraction in the URB_VEG simulation and the absence of transpiration at night reduces the total latent heat flux for this simulation at night. This is better represented in the URB simulation, although the daytime latent heat flux is too high due to a strong evaporation from the pervious road (80% of the total Qe) and the roof (15% of the total Qe) as depicted in Figure 9. Results for Preston are similar to the results for Armadale. The amount of latent heat originating from the pervious road fraction in URB is larger than the sum of the latent heat fluxes from the pervious road, BET trees and grass in URB_VEG. Hence, the drop in latent heat flux from the roof surface fraction after 9 am is more pronounced in the total latent heat flux (Figure 9) from the URB_VEG simulation.

The results using the eucalypt-specific biophysical parameters as described in Table 4 are summarized in Table 9. The higher albedo for the eucalypt leaves (Table 4) increases outgoing shortwave radiation for the Preston site, consequently increasing the MBE up to 10.0 W m−2 in summer (Table 9). Combined with a negligible effect on L↑ this in turn decreases net radiation (not shown) and thus reduces the amount of available energy. As differences in latent heat between the default URB_VEG and the eucalypt URB_VEG are negligible (for all seasons lower than the uncertainty derived in Section 'Sites in Melbourne') this results in a significant better representation of Qh in spring and summer while MBE values worsen for winter and fall (Table 9). For the other sites, the effect of eucalypt-specific values is less pronounced due to the smaller fraction of eucalypt in the tree-covered fraction. For Surrey Hills, most variables profit from a more detailed representation of the vegetation characteristics while this is less the case for Armadale. A number of counterbalancing processes can explain why the effect in latent heat flux for the eucalypt-specific simulations is rather small. The eucalypt-specific minimum stomatal resistance is larger than the default value which limits the trees to transpire when less water is available. Together with a reduced root fraction, this leads to a decreasing evaporative flux. On the other hand, the LAI is increased which enhances the total transpiration from vegetation. Hence, the net effect is almost nil. For Armadale and Surrey Hills, the small increase of latent heat flux is especially due to the fraction of BDT-temperate trees that are characterized by a higher transpiration rate (Table 9).

Table 9. Sensitivity on K↑, L↑, Qh and Qe for the three sites in Melbourne simulated by CLMU using the eucalypt-specific URB_VEG and the default URB_VEG simulation. ΔMBE and ΔRMSE are defined as |URB_VEGeucalypt|−|URB_VEG|. Decreases in ΔRMSE and ΔMBE (improvement in model performance for URB_VEGeucalypt) are denoted in bold
 PrestonSurrey HillsArmadale
  1. The asterisk (*) refers to the flux differences that are larger than δ (see Appendix A1).


As the URB simulation performed well in Armadale in summer, this specific site and season is selected to test the sensitivity of the modelled latent heat flux to different soil textures and the maximum amount of water that a surface can hold (URB simulation only). Although almost half of the urban canyon surface in Armadale is defined as pervious and is thus influenced by the soil texture, latent heat fluxes are relatively insensitive to changes in % of sand and clay (not shown). Increasing the clay fraction while maintaining the sand fraction creates a larger soil matric tension (SOILTEX1 vs SOILTEX4), thus reducing the amount of water available for evaporation. On the other hand, increasing the percentage of sand with a similar fraction of clay (SOILTEX1 vs SOILTEX2) will increase the soil hydraulic conductivity and has a small but positive effect on the latent heat flux. Increasing both clay and sand fractions (SOILTEX3) counterbalance each other by increasing the soil matric potential and the hydraulic conductivity with a small decrease in latent heat flux. A lower sand and clay fraction (SOILTEX4) compared to the default soil texture values of URB result in smaller hydraulic conductivity values throughout the soil column (not shown) and higher matrix potentials which result in the largest decrease of latent heat fluxes in comparison to the other experiments. The impact of a change in the default wpond,max value on Qe is limited for all surface fractions. For the impervious surfaces roof and impervious road, an increase (decrease) in maximum ponding depth results in an increase (decrease) of water that is available for evaporation and hence an increase (decrease) of latent heat. As all layers of the pervious fraction contribute to the evaporative flux and the soil is never saturated so that no run-off occurs (not shown), there is no effect on the latent heat flux. All changes in Qe due to soil texture and the maximum amount of water that a surface can hold are considered insignificant as they are smaller than the uncertainty δ derived for Qe in Armadale (3.4 W m−2, see Section 'Sites in Melbourne').

5. Discussion and conclusions

In this study, an evaluation of the urban parameterization scheme in CLM4 is performed against long-term measurements from three sites in Melbourne and data from the CAPITOUL experiment in Toulouse. For all sites, the model has been run offline mode and is forced by meteorological parameters observed above the urban canopy layer. Overall the model performs well in simulating the surface energy balance for all sites, with similar results for CLMU compared to the results with TEB for Toulouse as elaborated in Pigeon et al. (2008). Differences between the URB and URB_VEG simulations for the sites in Melbourne are significant while this is not the case for Toulouse. In general, CLMU performs better for the dense old city centre of Toulouse in comparison to the more vegetated sites in Melbourne. Nevertheless, for the latter, differences can be seen between seasons, sites and fluxes, with an overall performance that sits in the upper range of the results obtained for the model cohort of the PILPS-urban project (Grimmond et al., 2010, 2011). Overall CLMU well represents outgoing shortwave radiation for the medium-density suburban sites of Melbourne and overestimates outgoing longwave radiation resulting in an underestimation of net all-wave radiation. For Toulouse, an underestimation of K↑ is compensated by an overestimation in L↑ resulting in a good representation of Q*. This supports the finding of Grimmond et al. (2011) that an urban canopy model can perform well for the wrong physical reasons. Thus, it is important to consider the results in the framework of the intended model application (see e.g. Baklanov et al., 2009): a user who wants to apply CLMU to study e.g. mitigation and adaptation strategies should thus not only address the initial flux of interest but also the implication of these strategies to the other fluxes as well. Another difference between the two contrasting urban environments is the behaviour of CLMU with respect to the sensible heat flux. For the sites in Melbourne, Qh is generally overestimated while this is not the case for the default URB simulation in Toulouse. In addition, even though the three sites in Melbourne correspond to slightly different urban densities with a relatively high amount of vegetation, the results do not show a systematic improvement or deterioration using explicitly modelled vegetation (URB_VEG) or a semi-integrated vegetation based on a pervious road fraction in the urban canyon (URB). For example, turbulent fluxes in Preston are generally better represented by the URB_VEG simulation, while for Armadale, better results are obtained for the URB simulation. In this respect, no conclusive statement can be made with respect to the benefits/disadvantages of both schemes.

Results from the default URB simulation and sensitivity simulations with respect to QF for Toulouse depict the complex urban canyon interactions between incoming radiation, urban surface and canyon properties, turbulent heat fluxes and the anthropogenic heat flux. In this framework, the properties of the bulk urban air mass through which all urban surfaces (including the roofs) interact are crucial. First of all, the similarity between Tcanyon for both the URB and URB-QFpre simulations suggests that the heat conduction from the interior of the building towards the outside roof and wall surfaces is potentially too high even though the magnitude of the modelled anthropogenic heat flux compares well with the observed QF. For example, the roof and wall surface temperatures are better simulated with URB-QFpre while the road surface temperatures are largely overestimated due to the QF parameterization in CLMU. Hence, this could suggest that the heat conductivity for walls and roofs is too high due to the lack of e.g. insulation properties for these surfaces. This was also suggested by Bueno et al. (2011) who showed that TEB and EnergyPlus consistently over-predict exterior wall and roof temperatures in winter when no insulation is applied. In summer, differences between Troof and Twall between URB and URB-QFpre are negligible even though an additional 20 W m−2 is added to the urban canyon floor for the latter simulation. This increases Troad by 2 K, which in turn increases the overestimation of L↑ and hence decreases the available net radiation. As Tcanyon does not change, this results in an increase of sensible heat flux. Only when the anthropogenic heat is totally removed from the urban canyon in summer (URB-noQF) the sensible heat flux decreases via a better representation of the road surface temperatures.

At present, the interaction of the roof with the urban canyon or with the atmosphere aloft is still ambiguous. Daytime roof surface temperatures are generally higher than canyon surface temperatures, while nocturnal roof surface temperatures are lower, with a similar behaviour as observed recently from thermal imagery for the Sperrstrasse in Basel (Salmond et al., 2012). Our study reveals that the URBroof simulation results in a better representation of daytime Troof while the nocturnal positive bias remains. Thus, our hypothesis that a cooling of the roof surface is generally inhibited by a higher Tcanyon which in turn suppresses a cooling of the other urban surfaces is only partially valid. A coupling of the roof to the atmosphere above does enhance a cooling during the day, without a significant change in canyon properties. On the other hand, night-time roof surface temperatures do not change much while Tcanyon slightly increases leading to slightly higher road and wall temperatures. Moreover, our URBroof simulation for Toulouse has shown improvement with respect to roof surface temperatures but a deterioration with respect to net radiation and sensible heat. In this respect, a similar URBroof simulation was performed for Preston resulting in insignificant changes of sensible heat fluxes (not shown). This suggest that the lower urban roof heights in Preston allows for a better mixing of urban canyon properties and the air aloft, hence resulting in a negligible effect of URBroof on the turbulent heat fluxes. Previous studies performed by Oleson et al. (2008a) and Masson et al. (2002) allow for a similar comparison. Both studies evaluated the modelled roof surface temperatures from CLMU and TEB, respectively, for Vancouver (Voogt and Grimmond, 2000; Grimmond and Oke, 2002) and Mexico City (Oke et al., 1999). For Mexico City (characterized by H/W of 1.18) CLMU performed better compared to TEB for both night- and daytime roof surface temperatures. For Vancouver, CLMU modelled daytime roof temperatures are within 1–5 K of observed roof temperatures and are warmer than TEB by about 6 K which better agrees with observations. These findings suggest that more evaluation is necessary for contrasting urban geometries in order to obtain a better understanding of the interaction between the roof surface on the one hand and air aloft and canyon air on the other hand. This is further supported by the recent work of Salmond et al. (2012) who state that roof properties are subject to a number of general and site-specific controls. As such, their results in comparing bottom–up versus inertial sublayer fluxes were inconclusive.

In addition, our results reveal a trade-off in performance between surface temperatures, radiative and turbulent fluxes and the anthropogenic heat. With respect to the latter, the CLMU parameterization of urban heating results in a too dynamic daily cycle of QF, while prescribing QF to the road fraction leads to an artificial increase of the road surface temperatures. As QF is also influenced by the conduction of heat through the urban surfaces and thus their thermal properties, a future study could investigate both the sensitivity of the model with respect to more detailed building information (e.g. including a simple air-cavity as insulation layer (Bueno et al., 2011) as well as an improved anthropogenic heat parameterization scheme and the interaction of both through the urban canyon properties. On the other hand, depending on the intended application of the model, one could wonder whether a time-consuming acquisition of urban material characteristics is feasible and relevant (Grimmond et al., 2011). The input data should, nevertheless, fairly reflect the characteristics of the city surface and the response of the model to specific combinations of perturbations should be known before it is used e.g. in more applied mitigation and adaptation studies.

On the basis of the results in this study and the findings of Grimmond et al. (2010) and Grimmond et al. (2011) the question arises whether single urban canopy models are appropriate for urban forms who do not really exhibit a true urban ‘canyon’. The results from this study suggest that model complexity should, perhaps, relate to the site complexity. For example, CLMU performs better for Q* and Qh in the old city core of Toulouse compared to the medium-density sites in Melbourne. In addition, sites with a suburban character and abundant vegetation could benefit from a more detailed vegetation scheme compared to a more complex urban canyon approach, a strategy that is currently adopted by the CESM community in global climate simulations. Whether or not vegetation is explicitly resolved (or represented as a pervious fraction) in the urban canyon or is treated as a separate tile could then again depend on the type of application and the scale of interest. But as stated by Grimmond et al. (2011), this could be further addressed by additional comparison studies as done in PILPS-Urban, but for a wider range of urban forms.


This work is funded by the Flemish regional government through a contract as a FWO (Fund for Scientific Research) post-doctoral position and a FWO mobility grant. The CLMU simulations are supported by the Australian National Computing Infrastructure (NCI) National Facility at the ANU. Furthermore, we would like to thank Erik Kluzek from NCAR for his technical support on CLMU. In addition, we acknowledge the support of Andrew Richardson in estimating the uncertainties of the eddy covariance flux measurements and the suggestions made by an anonymous reviewer.

Appendix: A1 Flux error estimation

Knowledge about the uncertainty of turbulent flux measurements is essential for a statistical evaluation of modelled versus observed fluxes. It is widely recognized that eddy covariance measurements are noisy and that this uncertainty is largely due to random measurement error (Hollinger and Richardson, 2005). Basically, one wants to know the actual flux F but what is really measured is x = F + δ + ϵ, where δ is the random measurement error whose characteristics are generally unknown and ϵ any systematic error (Richardson et al., 2006). The typical errors relate to instruments and their calibration, different meteorological conditions under which the measurements are taken and especially in urban areas, the representativeness of the turbulent and radiant footprint (Grimmond et al., 2011). To the best of our knowledge, there is no proper framework yet to quantify flux uncertainties from eddy covariance measurements at urban sites. Until now, these efforts are restricted to natural surfaces (e.g. grassland, forest) and are merely transferred to urban surfaces (Grimmond et al., 2011). Thus, in order to estimate the flux uncertainty related to the flux measurements for the urban sites of Toulouse and Melbourne, we follow a similar strategy as developed for the natural sites.

One way to derive random flux uncertainties is to use simultaneous flux measurements from two neighbouring towers, viz. multiple independent observations in one place (Finkelstein and Sims, 2001). Even if two flux towers would be available, the heterogeneity of an urban surface makes it hard to perform such simultaneous measurements over identical urban landscapes. Hence, we follow the alternative ‘daily-differencing’ approach developed by Hollinger and Richardson (2005) and Richardson et al. (2006) in which space is traded for time so that flux measurements on two successive days with similar environmental conditions are used as analogues for the two-tower paired observations. The following criteria were used to define similar environmental conditions: time lags can only be 1 d, observations are made at the same time of day and differences between air temperature and wind speed should remain below 3 K and 1 m s−1, respectively. These rather stringent thresholds reduce the amount of observations available for the analysis considerably and hence only the long-term time series of Preston and Toulouse are used here, with 4900 (20%) and 3000 (16%) valid paired measurements, respectively.

Assuming that the paired observations (x1, x2) are independent and identical distributed measurements of the same quantity F, with x1 = F + δ1 and x2 = F + δ2 where δi is a random variable with variance σ2 (δ). The random error in the measured values (x1, x2) can be quantified by determining σ(δ) = σ(x1 − x2)/2 (Richardson et al., 2006). The variation in standard error for both sensible (Qh) and latent (Qe) heat fluxes does not show a clear relationship with air temperature, wind speed and vapour pressure (not shown) but scales linearly with the net radiation Q*, similar as was found by Hollinger and Richardson (2005) and Richardson et al. (2006). For Qh and Preston (Toulouse), the uncertainty increases linearly with 0.059 ± 0.008 (0.046 ± 0.007) per W m−2 absolute value of Q* while for Qe in Preston, the slope of the increase is slightly smaller with 0.02 ± 0.002 per W m−2 absolute value of Q*. For both Qh and Qe, the slope coefficients are significantly different from zero (P < 0.001). These results show that although Preston and Toulouse are relatively different in terms of morphological and thermal characteristics, the linear relationship between δ and Q* for Qh are very similar. In addition, the linear increase for both Qh and Qe are slightly smaller than the values suggested by Hollinger and Richardson (2005) for forested and grassland sites. Results for Armadale and Surrey Hills show similar magnitudes for the random error uncertainties, although the analysis was found to be less robust due to relatively short observation period and hence the low number of observations that satisfy the above-mentioned environmental thresholds [1000 (10%) and 800 (11%) data points, respectively].