Evaluation of an urban canopy model in a tropical city: the role of tree evapotranspiration

A single layer urban canopy model (SLUCM) with enhanced hydrologic processes, is evaluated in a tropical city, Singapore. The evaluation was performed using an 11 month offline simulation with the coupled Noah land surface model/SLUCM over a compact low-rise residential area. Various hydrological processes are considered, including anthropogenic latent heat release, and evaporation from impervious urban facets. Results show that the prediction of energy fluxes, in particular latent heat flux, is improved when these processes were included. However, the simulated latent heat flux is still underestimated by ∼40%. Considering Singapore’s high green cover ratio, the tree evapotranspiration process is introduced into the model, which significantly improves the simulated latent heat flux. In particular, the systematic error of the model is greatly reduced, and becomes lower than the unsystematic error in some seasons. The effect of tree evapotranspiration on the urban surface energy balance is further demonstrated during an unusual dry spell. The present study demonstrates that even at sites with relatively low (11%) tree coverage, ignoring evapotranspiration from trees may cause serious underestimation of the latent heat flux and atmospheric humidity. The improved model is also transferable to other tropical or temperate regions to study the impact of tree evapotranspiration on urban climate.


Introduction
Due to the accelerated worldwide urbanization since the late 20th century, large expanses of natural surfaces have been converted to urban landscapes with implications for albedo (usually lower), surface roughness (higher), evaporation (less), transpiration (less), heat storage capacity (higher) and anthropogenic heat emissions (higher). These changes of underlying surface properties have led to more complex surface energy distribution in urban areas. In addition to impacting the local and regional climate (Grimm et al 2008), they may further bring adverse effects to human health (Patz et al 2005).
The urban canopy layer (UCL) is the layer between the surface and the mean height of buildings and trees. Its characteristics are dominated by the energy and mass exchange between individual surface facets and the canyon air (Oke 1976). With the recognition of the important roles played by urban areas in surface energy balance (SEB), various numerical models have been developed to investigate the energy exchange over cities, as well as the transfer of momentum and moisture between UCL and the overlying atmosphere. One type of such models is the physically based urban canopy models (UCMs), which explicitly consider the impacts of urban building morphology when calculating the SEB. The UCMs can be further categorized into two types (Fernando 2012): single-layer (Kusaka et al 2001) and multilayer (Martilli 2002) schemes. The single-layer urban canopy model (SLUCM) simply represents the urban geometry as two-dimensional street canyons with infinite length but considers the three-dimensional nature of urban morphologies, the shadowing from buildings, and reflection of radiation in the canopy layer. These models calculate the prognostic variables such as surface skin temperature and heat fluxes produced from urban facets (i.e. building roofs, walls and roads). Chen et al (2004) and Miao et al (2009) have coupled SLUCM into mesoscale atmospheric models (e.g. Weather Research and Forecasting, WRF model) to offer feasible lower boundary conditions to the overlying atmosphere. The coupled WRF/SLUCM system has been successfully applied in various cities around the world  for studying various microclimatic phenomena such as the urban heat island effect (e.g. Miao et al 2009, Li et al 2013, Li et al 2016. The SLUCM has achieved satisfactory performance in predicting net radiation and sensible heat flux but is still inadequate in predicting latent heat flux (e.g. Loridan et al 2010, Li et al 2013. Comparing 33 urban SEB models, Grimmond et al (2011) concluded that the underestimation of latent heat flux exists not only in SLUCM but also in most other urban land surface schemes. This mainly results from the uncertain mechanism of complex water vapor transport and the inadequate representation of urban hydrological processes, urban vegetation and soil properties. The underestimation of latent heat flux usually leads to overestimation of sensible heat flux and surface skin temperature, which ultimately jeopardizes the SLUCM's accuracy. In order to improve the simulation accuracy of latent heat flux, Loridan et al (2010) assessed the fitness of a coupled Noah land surface model/SLUCM parameterization scheme and suggested that a vegetation class like 'cropland/grassland mosaic' or 'grassland' with a low stomatal resistance could help improve the insufficient representation of urban evaporation in the SLUCM. Miao and Chen (2014) and Yang et al (2015) included several additional urban hydrological processes (i.e. the anthropogenic latent heat, urban irrigation, evaporation from paved surfaces, and the urban oasis effect) in the coupled Noah/SLUCM to enhance the urban evaporation components. Their evaluation of the model performance in Beijing, Phoenix, Vancouver and Montreal showed that the latent heat flux simulated by the new model improved substantially.
The latent heat flux is especially important since it impacts soil moisture dynamics, surface runoff, and the overlying atmosphere (Koster 2015). Accurate information about the latent heat flux and evapotranspiration is essential for estimating urban ecosystem water requirements and stress regimes, which is of utmost importance during a drought. However, urban latent heat flux and evapotranspiration measurement and estimates are sparse in both time and space (Gentine et al 2016). Accurate prediction of latent heat flux will therefore widely benefit the scientific community, urban designers and policy-makers (e.g. Mitchell et al 2008). It has been demonstrated that including vegetation effects in urban SEB models is crucial to obtain accurate prediction of the sensible and latent heat fluxes Grimmond 2015, 2016). In previous urban SEB models, the urban vegetation was usually represented by 'grassland' and the impact of urban trees was neglected in UCMs. Lee and Park (2008) included trees in their SLUCM but did not consider sub-surface moisture transport. Krayenhoff et al (2014Krayenhoff et al ( , 2015 incorporated trees into a multilayer UCM and added tree-building interaction, but they did not fully take the evaporative impact of trees into account. Ryu et al (2016) included the hydrothermal processes associated with trees in SLUCM and adequately incorporated the interaction of trees, ground and walls. Their simulation results in three urban/suburban sites in Basel, Switzerland showed improved prediction performance of latent heat flux. The abovementioned models with tree effects revealed the important role of urban trees in SEB modeling, especially in simulating latent heat flux.
In the light of previous studies and the need of accurate prediction of latent heat flux for urban climate modeling, we will conduct an evaluation study in a tropical city Singapore using the coupled Noah/ SLUCM model with physically-based urban hydrological processes and tree evapotranspiration. Some previous evaluation studies have been carried out in Singapore , but neither evaluated the role of tree evapotranspiration. The aims of our study are: (1) evaluate the performance of Noah/SLUCM in a tropical residential neighborhood; (2) improve the capability of this model with urban hydrological processes and tree evapotranspiration; and (3) assess the importance of urban trees in the urban SEB in Singapore.

Local climate and data collection
Singapore is a tropical city state located at the southern tip of the Malay Peninsula. It is a highly urbanized coastal city with a very high population density of 7821 persons per km 2 . As a 'City in the Garden' , about half of its land is covered by managed vegetation and young secondary forest (Yee et al 2011). Because of its geographical location near the equator, Singapore has a typical equatorial wet climate characterized by perennial high temperature, high humidity and abundant precipitation. Alternating northeast and southwest monsoon winds blowing throughout the island divides its annual climate into a northeast monsoon season (usually December to early March of the following year), a southwest monsoon season (usually June to September) and two inter-monsoon periods (Fong and Ng 2012).
A 23.7 m high eddy-covariance (EC) flux tower is located in Telok Kurau (1°18 0 51.46 00 N, 103°54 0 40.31 00 E, elevation: ∼10 m a.s.l), a low-density residential area in Southeast Singapore. The study area corresponds to 'compact low-rise' or Local Climate Zone 3 according to Stewart and Oke (2012). Given a mean building height of 9.86 m, the sensors are expected to be located with the inertial sublayer (above 2-5 times the height of the surface roughness according to Roth 2000) where fluxes are constant with height and representative of the underlying surface. Within a 1 km radius of the tower, 39% of the plan area is covered by mainly 2-3 story high residential buildings, 12% roads, 34% other impervious surfaces and 15% vegetation (including 11% tree crowns and 4% grassland) with little directional variation (figure S1 in Velasco et al 2013). Average flux footprints, which encompass the area on the ground which contributes to the measurement, extended to between 400 m (daytime) to 1000 m (nighttime) from the tower (

Coupled Noah/SLUCM model
The coupled Noah/SLUCM is utilized in this study. The heat fluxes associated with the different urban facets (building roofs and walls, roads and other impervious surfaces) are simulated by SLUCM, while those from the vegetation fraction are handled by the Noah LSM (Chen and Dudhia 2001). The urban SEB equations are (Oke 1988) where K ↑ and K ↓ are upwelling and downwelling shortwave radiation, L ↑ and L ↓ are upwelling and downwelling longwave radiation; Q Ã is net radiation, Q F is anthropogenic heat flux, Q H is sensible heat flux, Q E is latent heat flux and Q S is heat storage. More details of the coupled Noah/SLUCM model can be found in Chen et al (2011). This coupled Noah/SLUCM model was run offline driven by the 30 minute meteorological and radiation data collected from the EC flux tower, including wind speed, air temperature, relative humidity, ambient pressure, precipitation, downward shortwave and longwave radiation. Other input parameters of urban geometry, building materials and anthropogenic heat used in this study are given in table 1.

Enhanced hydrological processes
In order to mitigate the deficiency in modeling the latent heat flux in urban LSM models, Miao and Chen (2014) and Yang et al (2015) proposed to include some enhanced hydrological processes in the coupled Noah/SLUCM model. These processes are (1) anthropogenic heat flux release, (2) the urban oasis  (4) a modified evaporation scheme over urban impervious surfaces. Because the oasis effect is location specific and there is no routine urban irrigation in Singapore, these two processes are not considered in the present study. Instead, a sensitivity study of the oasis effect will be conducted in the supplementary material (section S3 available at stacks.iop.org/ERL/12/094008/mmedia). Considering the other hydrological processes, the latent heat flux can be calculated from two combined sources, urban and non-urban, according to Yang et al (2015), as where f urb is the urban (impervious) fraction, f veg is the vegetation fraction, Q Eurb is the latent heat flux from the urban sources calculated by the original SLUCM, Q Eveg is the latent heat flux produced by non-urban sources. C H is the exchange coefficient, E p is the potential evaporation rate, Q ALH is anthropogenic latent heat flux, which is calculated as (Yang et al 2015) where Δ is the slope of the saturation vapor pressure curve (kPa K À1 ), g is the psychrometric constant (66.1 Pa K À1 ), b is the moisture availability parameter (a value of 1.5 is used here following Yang et al 2015). Q ALHMAX , Q ASHMAX and Q AHMAX are daily maximum anthropogenic latent, sensible and total heat flux, respectively. f ALH is the diurnally-varying coefficient of anthropogenic sensible and latent heat.Q AHMAX and f ALH are both taken from Quah and Roth (2012) and Li et al (2013).
Many studies have demonstrated that urban impervious surfaces can partly store precipitation and supply evaporation over a given period of time (e.g. Kawai andKanda 2010, Ramamurthy andBou-Zeid 2014). The default evaporation scheme for impervious surfaces only considers the evaporation on the ground surface during precipitation, but excludes water availabilities from ponded surface water due to impeded drainage after precipitation. Based on the hypothesis that water availabilities in road, roof and wall exponentially decrease to 0 in 24 h since last precipitation, Miao and Chen (2014) proposed a new evaporation scheme over urban impervious surfaces in calculating Q Eurb (referred to as the new impervious evaporation scheme hereafter). It is noteworthy that this new impervious evaporation scheme is only applicable in the cases of consecutive precipitation and poor drainage, and therefore in this study we only apply it during the early NE monsoon season.

Tree evapotranspiration
To account for the tree evapotranspiration effect, a new term is added to equation (3): where f tree is the tree fraction and Q Etree is the latent heat flux produced by trees, which is calculated using the Penman-Monteith equation (Bosveld andBouten 2001, Ballinas andBarradas 2016) where λ v is latent heat of vaporization of water (J kg À1 ), ET is the mass water evapotranspiration  Environ. Res. Lett. 12 (2017) 094008 rate (kg s À1 m À2 ), r is the density of air (kg m À3 ), C P is the specific heat capacity of air at constant pressure (J kg À1 K À1 ), VPD is the vapor pressure deficit (kPa), r A is the bulk surface aerodynamic resistance for water vapor (s m À1 ), and r C is the canopy surface resistance (s m À1 ). The vapor pressure deficit is calculated by where e s and e a are saturation vapor pressure (kPa) and vapor pressure (kPa), respectively, and RH is relative humidity. The bulk surface aerodynamic resistance r A is calculated according to Allen et al (1998) as where k is the von Karman's constant (= 0.41), u is the wind speed at height z m (m s À1 ), z m is the height of wind measurement (m), z h is the height of humidity measurement (m), z d is the zero plane displacement height (m), z om is the aerodynamic roughness length for momentum transfer (m) and z oh is the aerodynamic roughness length for the transfer of heat and vapor (m). The zero-plane displacement height z d is estimated as 0.67 z r , where z r is the canopy height (m). In this study, z oh is calculated as 0.1 z tree , where z tree is the mean height of trees (m), while z om is calculated as 0.1 z oh . The canopy surface resistance r C is calculated following Chen and Dudhia (2001).

Numerical experiment design
The impact of different processes on the performance of the coupled Noah/SLUCM model was evaluated through three numerical experiments (table 2). Sim_1 used the original model without any additional effects. Sim_2 considered the enhanced hydrological processes including (1) anthropogenic latent heat release, and (2) the evaporation from impervious surfaces (Miao and Chen 2014). Sim_3 further included the tree evapotranspiration in addition to those enhanced hydrological processes considered in Sim_2.  the daytime refers to 0800-1800 LT, and nighttime refers to 2000-0600 LT every day.

Radiation and SEB validation
3.2.1. Upwelling radiation K ↑ and L ↑ Figure 1 compares the observed and simulated upwelling radiation for each experiment and season.
Since the additional processes included in Sim_2 and Sim_3 do not alter the albedo of the underlying surfaces, no variations of the simulated K ↑ are observed between the three experiments. In general, the simulated K ↑ all agree well with the corresponding observations. Only a slight overestimation in the early NE monsoon season is observed, possibly due to the water accumulation over paved surfaces resulted from consecutive rainfalls in this season. This surface wetness change leads to a lower albedo, while the model albedo keeps constant for all cases. The simulated L ↑ is generally overestimated during daytime (by about 10%) while slightly underestimated just before sunrise (figure 1). It is generally more complex in modeling L ↑ than K ↑ since L ↑ depends strongly on the surface temperature, which in turn is difficult to reproduce given the heterogeneity of materials and geometries within the urban fabric (Grimmond et al 2011). In Sim_3, the error of the simulated L ↑ during daytime decreases by about 20% compared with the other two experiments, suggesting a better performance after considering tree evapotranspiration effects, although the simulated L ↑ is still overestimated during daytime. This suggests that the tree evapotranspiration contributes more to the reduced surface temperature than the other hydrological processes considered.

Net radiation Q Ã
The net radiation Q Ã is calculated by equation (1), while the downwelling components are from the input forcing data, i.e. the measured data from the EC flux tower. Due to the aforementioned overestimation of L ↑ , Q Ã is always underestimated. The simulation performance of Q Ã is generally excellent in terms of the statistic metrics and the performance does not vary much across different seasons (figure 2). The index of agreement (IOA) remains in the interval of 0.97-0.99 during the entire day and during daytime, but decreases slightly during nighttime (table S1). Although the effect is only marginal, the introduction of the hydrological processes and tree evapotranspiration leads to a continuous reduction of root-meansquare error (RMSE), indicating their positive role in improving the simulated Q Ã . It is noteworthy that the reduction of the systematic RMSE (RMSEs, table S1) suggests an improvement in the model's prediction skills after considering the tree effect.
3.2.3. Latent heat flux Q E Q E is usually the least well-modelled energy flux component in existing urban SEB models (Grimmond et al 2011). The same situation can also be noticed in this study. The original model (Sim_1, see table S1) gives a RMSE of 52.65 W m À2 and an IOA as low as 0.52 during the entire period. Incorporating urban hydrological processes into the coupled Noah/SLUCM model has achieved satisfactory results with improving modeling accuracy of Q E in many cities such as Beijing, Phoenix and Montreal in previous studies Chen 2014, Yang et al 2015). In our case, after considering hydrological processes in our study, though an increased average value of 12.68 Wm À2 and a reduction of RMSE to 51.07 Wm À2 are achieved during the entire period (Sim_2, see table S1), there is still a notable discrepancy between the modeled and observed Q E . This situation may be explained by the different geography locations of the tropic city Singapore and mid-latitude cities. The mid-latitude cities are located in relatively dry areas with a high atmospheric demand and limited water available at the surface, whereas in Singapore the atmospheric demand is lower, while water in the urban fabric is more abundantly available.
In previous coupled Noah/SLUCM model simulations, the vegetation part of an urban site is generally represented by grassland without considering trees. In our case, the addition of tree evapotranspiration in Sim_3 sees a marked improvement in the accuracy of simulated Q E . During the entire period, the RMSE and mean biased error (MBE) drop to 38.71 W m À2 and À11.99 W m À2 , respectively, and the IOA increases to 0.75 (table S1). From Sim_1 to Sim_3, the RMSEs reduces continuously (table S1), manifesting the improved model skills by including the tree evapotranspiration.
The model performance in predicting Q E has a strong dependence on the seasonality. Sim_1 and Sim_2 give the poorest simulation results in the dry period, with the mean value of 2.54 W m À2 and 4.84 W m À2 , respectively, compared with the observed value of 34.39 W m À2 (table S1). Q E is usually highly impacted by precipitation. Though there was only 2.2 mm precipitation during the dry period, the mean value of observed Q E was still as high as 78% of that in other seasons (44.14 W m À2 in SW monsoon season, and 44.11 W m À2 in early NE monsoon season). There should be other processes contributing to Q E during the dry period. One such process is the tree evapotranspiration, which has been demonstrated in Sim_3 ( figure 3(d)). With added tree evapotranspiration, the RMSEs of Q E decreases from 46.94 W m À2 (Sim_1) to 20.55 W m À2 (Sim_3), indicating that the systematic error has been greatly reduced and most of the discrepancy between the simulated and observed Q E can be explained by random processes (e.g. due to uncertainties from input parameters). Another possible process is the urban irrigation. Although no regular urban irrigation is conducted in Singapore, in extremely dry period, it is reasonable to assume that some manual, sporadic irrigations may be carried out by local residents to mitigate the absence of water.  Environ. Res. Lett. 12 (2017) 094008 hydrological processes (Sim_2), the model performance improves accordingly. Nevertheless, Q H is still overestimated until the tree effect is considered in Sim_3. The model shows notable improvement in Sim_3 in the early NE monsoon season (figure 4(c)), with RMSE reducing to 31.46 W m À2 from 34.87 W m À2 and 31.59 W m À2 in Sim_1 and Sim_2, respectively (table S1). The RMSEs and RMSEu also decrease with added hydrological processes and tree evapotranspiration (table S1).
A consistent lag in the simulated Q H is found in all the seasons. Hysteresis starts from 0800 LT with peak value at 1300 to 1330 LT, slightly later than the observed peak at 1230 LT. This phenomenon is not typical as the long-term (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013) average Q H of Singapore peaks at 1300 LT, and 1400 LT for dry seasons . The diurnal cycle of predicted Q H corresponds more to the long-term statistics rather than the period under investigation here. Since the major goal of this study is to improve the prediction of Q E , no efforts were made to fine tune the parameters related to Q H . The observed lag in Q H needs further investigation, but it is beyond the scope of the current study.
3.2.5. Heat storage flux Q S As both the simulated and observed ΔQ S are calculated as the residual of the net radiation and the other energy components (see equation (2) 3.3. Further analysis of Q E 3.3.1. Taylor diagram Figure 6 shows normalized Taylor diagrams for Q Ã , L ↑ , Q H and Q E for all simulations, observations and seasons. Observations are indicated by the star symbol at (1.0, 1.0, 0.0), correlation coefficients are plotted on polar axes, normalized standard deviations on y-axes and normalized RMSE are given by the inner grey circles.
The performance of Q Ã shows no visible difference across all simulations, while the predicted L ↑ is sensitive to the seasonality, with the best performance during the dry period and worst during the early NE monsoon season. The correlation coefficients of Q H are in a narrow range from 0.9 to 0.95, with small variability in normalized standard deviation and normalized RMSE. A wide range of model performance is evident for Q E . Correlation coefficients for all simulations are less than 0.7 except for Sim_3 during the dry period (indicated by the green circle symbol).  The normalized standard deviations are scattered from 0.0 to 0.8 and RMSE values are around 0.8 except for Sim_3 during the dry period (indicated by the green circle symbol, which is below 0.7). Overall, taking the tree evapotranspiration into account greatly improves the model performance in predicting Q E . It can be seen from the diagram that the model with the tree evapotranspiration generally has the best performance during the dry period and early NE monsoon season with the new impervious evaporation scheme (indicated by the green and red circle symbols, respectively). During the dry period (green circle symbol), the simulated Q E shows the highest correlation and lowest RMSE, while during early NE monsoon season (red circle symbol) the standard deviation of the simulated Q E is the closest to the observational data.

Model performance during dry period
As revealed by the Taylor diagrams (figure 6), the Q E simulation during the dry period with added tree evapotranspiration shows best performance. To examine Q E simulation improvement during the dry period in detail, figure 7 compares the daily averaged Q E from both observations and simulations during the daytime (figure 7(a)) and nighttime ( figure  7(b)). An evident increase of all values is noted in Sim_3 during daytime with a magnitude 3 times larger than that in Sim_2, which suggests that urban tree evapotranspiration dominants the dry period's latent heat release. However, during nighttime, the model performance of Sim_3 is as bad as that of Sim_2, since no solar radiation is available for evapotranspiration at that time.

Conclusions and discussions
This study evaluated the performance of the coupled Noah/SLCUM model in a tropical suburban site during a 11 month period. Some urban hydrological processes (anthropogenic latent heat flux and evaporation from urban impervious surfaces) as well as the tree evapotranspiration were implemented in the model to improve the prediction skills of latent heat flux Q E . Comparisons with the field measurements showed that the added hydrological processes can help improve the accuracy in predicting Q E . However, there still remained much discrepancy between the simulated and observed Q E until the tree evapotranspiration was included in the model. The inclusion of the tree evapotranspiration had been shown to greatly reduce the MBE and RMSE of simulated Q E , particularly leading to much lower systematic RMSE, an indication of the improved prediction skills. During an unusual dry spell in early 2014 when there was little contribution from precipitation, the tree evapotranspiration was found to contribute most of Q E . The improved model demonstrated little effect during nighttime, when the solar radiation needed for transpiration is missing. The results demonstrated that tree evapotranspiration plays a very important role in latent heat flux in an urban environment like Singapore, and ignoring the tree effect will result in underestimation of latent heat flux.
Similar studies in Vancouver, Canada (Krayenhoff et al 2014, Krayenhoff et al 2015 and Basel, Switzerland (Ryu et al 2016) have demonstrated the role of trees in improving latent heat flux prediction. It is therefore expected that the proposed improved model from the present study can be transferred to other tropical and temperate urban environments to produce improved latent heat flux prediction. For neighborhoods without the necessary observational data as input to the model (which is very common), the online simulation with the well-calibrated coupled WRF/UCM can be performed instead. The improved model can be applied to better predict a variety of important real-world issues facing urban dwellers, such as urban heat island mitigation, outdoor thermal comfort, building energy usage (when coupled with a building energy model), or water usage management during drought conditions (Vahmani and Ban-Weiss 2016).
(CREATE) programme. The Center for Environmental Sensing and Modeling (CENSAM) is an interdisciplinary research group of the Singapore-MIT Alliance for Research and Technology (SMART). We thank Mr. YANG Jiachuan and Dr. WANG Zhihua at Arizona State University for providing the source code and their valuable suggestions.