Urban meteorological modeling using WRF: a sensitivity study

This study explores the sensitivity of high‐resolution mesoscale simulations of urban heat island (UHI) in the Chicago metropolitan area (CMA) and its environs to urban physical parameterizations, with emphasis on the role of lake breeze. A series of climate downscaling experiments were conducted using the urban‐Weather Research and Forecasting (uWRF) model at 1‐km horizontal resolution for a relatively warm period with a strong lake breeze. The study employed best available morphological data sets, selection of appropriate urban parameters, and estimates of anthropogenic heating sources for the CMA. Several urban parameterization schemes were then evaluated using these parameter values. The study also examined (1) the impacts of land data assimilation for initialization of the mesoscale model, (2) the role of urbanization on UHI and lake breeze, and (3) the effects of sub‐grid scale land‐cover variability on urban meteorological predictions. Comparisons of temperature and wind simulations with station observations and Moderate Resolution Imaging Spectroradiometer satellite data in the CMA showed that uWRF, with appropriate selection of urban parameter values, was able to reproduce the measured near‐surface temperature and wind speeds reasonably well. In particular, the model was able to capture the observed spatial variation of 2‐m near‐surface temperatures at night, when the UHI effect was pronounced. Results showed that inclusion of sub‐grid scale variability of land‐use and initializing models with more accurate land surface data can yield improved simulations of near‐surface temperatures and wind speeds, particularly in the context of simulating the extent and spatial heterogeneity of UHI effects.


Introduction
Urban heat island (UHI) effects, defined as elevated local temperatures in urban areas compared to rural areas, are common in developed areas, and are not observed solely in strongly urbanized environments. Suburban areas, for example, also exhibit UHI effects at smaller spatial scales (e.g. from localized groups of buildings, parking lots, etc.). The UHI modifies local meteorology, exacerbates local air pollution, reduces visibility, impacts agriculture, increases water usage, and exacerbates morbidity during urban heat waves (Hunt et al., 2012). Land features and land-use, including terrain (Sharma and Huang, 2012), streets and buildings, building materials, vegetation, and anthropogenic heat (AH) sources (among many other factors) exert strong controls on UHI effects. Some examples are the proximity to ocean, sea, or lake (Chen et al., 2011b;Salamanca et al., 2011;Sharma et al., 2016), type of vegetation and land-cover (Georgescu et al., 2011; et al., 2014, 2016), presence of streets, buildings, parking lots, residential versus commercial development, city morphology (e.g. horizontal vs vertical structure of buildings), surface albedo (Vahmani and Ban-Weiss, 2016), types of construction materials (Santamouris et al., 2011), and sources of AH [e.g. waste heat from air-conditioning (A/C) (Salamanca et al., 2014), power plants (Zevenhoven and Beyene, 2011), transportation and manufacturing (Sailor, 2011), etc.].
All of the above sources of UHI effects are commonly encountered in the Chicago metropolitan area (CMA), which covers the City of Chicago and its suburbs, making it an ideal case study for examining the performance of physically based simulation models of UHI effects. The CMA is located on a relatively flat terrain, with high-density residential development along the shore of Lake Michigan (Figures 1 and 2(a)). It is the third largest metropolitan area in the United States and the largest in the Great Lakes region (based on US Census Bureau). It covers 12 counties in Illinois, two in Wisconsin, and two in Indiana. The CMA is one of the most populous and heavily urbanized areas of the United States. According to US Census Bureau, 40.6% of the Illinois population lives in Cook County, Figure 1. The domain and geographical features of interest to this study with four two-way nested domains for the WRF model with grid size of 27, 9, 3, and 1 km as d01, d02, d03, and d04, respectively, overlaid on a terrain height map. The orientation indicated by north arrow shown here applies to all subsequent figures depicting aerial view.
which contains much of the CMA. The population of Cook County alone (∼5.2 million) exceeds the individual population of 29 US states. The total population of the CMA is about 8.6 million. The CMA is a major industrial centre in the United States. The surrounding land is fertile, and has been extensively drained and developed for agriculture. Abundance of flora and fauna provides ecosystem services to local and regional residents. With expansion of suburban communities in recent years, urban areas are growing parallel to the lakeshore and into previously suburban, agricultural, or undeveloped areas. According to recent estimates from Chicago Metropolitan Agency for Planning (CMAP: http://www.cmap. illinois.gov/data/land-use/inventory), over 44% of the land area in the CMA is classified as developed land-use; the rest is either agricultural land or vacant/wetlands/open space. Of the developed land, nearly two-thirds is residential. Land-use characteristics in the CMA are also rapidly changing in some areas. For example, early 245 km 2 of land identified as agricultural in 2001 were converted to another use by 2005. Increasing urban and suburban development in the CMA affects the intensity and spatial extent of UHI effects, and poses challenges for energy supply, transportation, ecological conservation, water resource management (water supply and stormwater), agriculture, and human quality of life (Susan, 2007). The rapid conversion of agricultural to urban/residential land calls attention to the need to regularly update land-cover and land-use data sets used in high-resolution mesoscale models. In many modeling studies, however, land-cover data sets have not been kept up to date, and do not accurately reflect the study period (Sertel et al., 2010;Törmä et al., 2015).
Like much of the Midwest region, the CMA experiences large daily and seasonal variations in air temperature. Summers are hot and humid, and winters are cold and windy. In recent decades, a statistically significant warming trend has been observed over the Midwest [US Global Change Research Program (USGCRP), 2013]. Associated with such warming are increased incidence of heat waves: which we will define as a period of time when maximum temperature is above 32.2 ∘ C (90 ∘ F). Heat waves can have important impacts on human health. For example, the severe heat wave in 1995 in the CMA was responsible for about 750 deaths (Karl and Knight, 1997;Chuang et al., 2013). Elevated night temperatures over multiple days, which are affected by UHI effects, are also implicated in human health impacts of heat waves (Semenza et al., 1996).
Also of meteorological significance is the lake breeze of nearby Lake Michigan, which helps regulate coastal temperatures . The effects of the lake breeze on the CMA has been investigated via measurements and numerical modeling in several previous studies (Keeler and Kristovich, 2012;Meir et al., 2013;Conry et al., 2014Conry et al., , 2015. During the afternoon, land areas in the CMA become hot relative to Lake Michigan, creating a high-pressure ridge over the lake and lower pressure due to UHI over the land, causing hot air over the CMA to rise. This vertical transport results in cooler near-surface air flowing from the lake as the lake breeze. Previous studies have shown that the lake breeze in Chicago can penetrate 15-30 km inland, moving cool humid air from Lake Michigan into the city, at times even penetrating to the outer suburbs of the CMA. A stronger lake breeze reduces the UHI, substantially suppressing heating over urban areas, with the most significant impact occurring near the lakeshore (Atkinson, 1989;Laird et al., 2001).
Depending on the specific characteristics of urban areas, with growing urbanization in a warming climate, the urban diurnal temperature range (DTR) (T max -T min ) is likely to decrease more than in rural areas, amplifying the UHI effect . In general, diurnal extreme temperatures (T max and T min ) are both rising, but the minimum temperature is rising faster, so the DTR is decreasing, and this decrease is larger in urban areas than in rural areas (Fernando et al., 2012). The amplification of UHI may even lead to local meteorological regime shifts with relatively warmer day-and night-time temperatures than in the past and with deeper but weaker nocturnal boundary layers (Emmanuel and Fernando, 2007;Fernando, 2008).
Notwithstanding the lack of quality observational and simulation data that have limited research progress to date, considerable advances have been made on understanding and modeling of UHI effects (Kalkstein and Davis, 1989;Karl et al., 1995;Kunkel et al., 1996;Kalkstein and Greene, 1997;Palecki et al., 2001;Brazel et al., 2007;Pullen et al., 2008;Grimmond and Athanassiadou, 2009;Basara et al., 2010;Grimmond, 2011;Chen et al., 2012;Best and Grimmond, 2014;Park et al., 2014;Georgescu, 2015). These studies have demonstrated that high-resolution mesoscale models need careful selection of urban parameter values to accurately simulate the effects in each region. Regions in the United States like Houston (Chen et al., 2011b), Baltimore-Washington metropolitan area (a) (b) Figure 2. The innermost domain d04 (a) with land-use classification using high-resolution NLCD data. Black dashed line is used for a vertical cross-section analysis for UHI and lake breeze ( Figure 9). It also shows different stations used for lake breeze analysis in Figure 10. Here, 'a' refers to the Lake Michigan station; 'b' the station D8777 close to the lakeshore; 'c' the Chicago O'Hare airport; and 'd' the Aurora station; and (b) meteorological stations (Table S1) used to evaluate statistics for different land-use classes. (Li et al., 2013a), Phoenix (Grossman-Clarke et al., 2010;Chow et al., 2012;Shaffer et al., 2015), and New York City (Gutiérrez et al., 2015) have already been well studied and urban parameter sets (Wang et al., 2011) have been modified to reflect the unique conditions in each urban centre.

Synopsis
High-resolution climate modeling using physically based simulation models has not been carried out for the CMA. As noted above, the CMA presents a near ideal case study for evaluating the performance of such models in reproducing UHI effects and the occurrence of lake breeze. This article addresses this opportunity by evaluating the sensitivity of a high-resolution climate model implemented over the CMA and adjoining rural areas shown in Figure 2(a). For our study, we analyse the sensitivity of simulated UHI effects to several important factors: (1) initialization of soil moisture and soil temperature, (2) the spatial resolution of the land-use representation, and (3) the choice of the different urban parameterizations. In addition, we analyzed the complex interaction between land-use, UHI effects, and lake breeze.

WRF implementation
Numerical simulations were performed using the Weather Research and Forecasting (WRF) model, version 3.6 (Skamarock et al., 2005), which is a non-hydrostatic, compressible model. WRF was nested across several domains, and Figure 1 shows details of terrain and the extent of each domain. The model grid was configured so that the outermost domain covers the Laurentian Great Lakes, and the innermost domain covers the CMA and adjoining rural and agricultural areas. The grid spacing (grid points) of these four two-way nested domains were 27 km (99 × 99), 9 km (155 × 166), 3 km (190 × 190), and 1 km (319 × 379), respectively. The model was implemented with 40 pressure-based terrain following vertical levels from the surface to 100 hPa, with the first 17 levels being in the lower 1.5 km and the first model level at 21 m from ground. Time-varying, large-scale, 3-h National Centers for Environmental Prediction (NCEP) North American Regional Reanalysis (NARR) simulations at 32-km resolution (http://rda.ucar.edu/datasets/ds608.0/) were applied as lateral boundary conditions to the outermost domain, with dynamical downscaling to interior domains. Owing to the proximity of the Lake Michigan and the influence of the Great Lakes in the outermost domain (Figure 1(a)), the lake and sea surface temperatures were updated at 3-h intervals using NCEP archives (ftp://polar.ncep.noaa.gov/pub/history/sst) to capture the possible lake effects in the WRF model. High-resolution static surface fields, such as terrain height and land-use categories, were also input to the corresponding nested domain boundary conditions. Hourly instantaneous outputs were used to analyse the UHI and lake breeze sensitivity studies. Henceforth, unless otherwise stated, our discussion refers to the innermost 1-km domain over the CMA. Figure 2(a) shows the land-use categories used in the innermost domain of the WRF model based on 30-m 2006 National Land Cover Data set (NLCD 2006) (Fry et al., 2011) with urbanization over much of the CMA, and adjoining regions covered primarily by agricultural land.
WRF has multiple parameterizations for its physics components: microphysics, convection, radiation, boundary layer, and surface. Previous studies identified the following combination of physical parameterizations over Chicago (Smith and Roebber, 2011;Sharma et al., 2014Sharma et al., , 2016Conry et al., 2015). We utilized a four-layer Noah land surface model (LSM: Chen and Dudhia, 2001). Sub-grid scale cumulus convective parameterization was turned on only for the two outermost domains (27 and 9 km), invoking the Kain-Fritsch scheme (Kain, 2004). We selected WRF single-moment three-class simple ice scheme (WSM3; Hong et al., 2004) for microphysics as it is a sufficiently accurate scheme to model precipitation during this study period. We selected the Dudhia scheme (Dudhia, 1989) for shortwave and the Rapid Radiative Transfer Model for longwave radiation parameterizations (Mlawer et al., 1997). The Monin-Obukhov similarity scheme was used for the surface-layer, and Mellor-Yamada-Janjic scheme (Janjic, 1994) for the planetary boundary layer.
The uWRF setup for studying sensitivities to urban canopy models (UCMs) included (1) a BULK parameterization scheme which represents zero-order effects of urban surfaces to reproduce the resultant urban effect associated with the land-use and vegetation characteristics over urban grid points by means of modified roughness length, surface albedo, heat capacity (1.0E6 J m −3 K −1 for roof and building walls and 1.4E6 J m −3 K −1 for ground), and thermal conductivity (0.75 J m −1 s −1 K −1 for roof and building walls and 0.4 J m −1 s −1 K −1 for ground) (Chen and Dudhia, 2001;Liu et al., 2006;Chen et al., 2011a); (2) a single layer urban canopy model (SLUCM) parameterization which uses a two-dimensional street canyon to explicitly parameterize canyon radiative transfer, turbulence, momentum, and heat fluxes. It accounts for different urban surfaces: roof, wall, and road. The SLUCM also has the capabilities of studying the impact of AH fluxes due to A/C, transportation, and human metabolism (Kusaka et al., 2001;Kusaka and Kimura, 2004;Lee et al., 2011); (3) building effect parameterization (BEP) which treats urban surfaces three-dimensionally, and considers heat in the buildings and the effects of vertical and horizontal surfaces on temperature, momentum, and turbulent energy (Martilli et al., 2002); and (4) coupled BEP with a multi-layer building energy model (BEP + BEM) which accounts for energy exchanges happening between the interior of building and the outer atmosphere (Salamanca and Martilli, 2010) for the CMA. The urban parameterization schemes are discussed in Salamanca et al. (2011) and the model setup and methods are the same as employed in Sharma et al. (2014Sharma et al. ( , 2016. Based on Akbari and Rose (2008), we modified urban parameters for the uWRF model to reflect current urban conditions in the CMA (Table 1). Estimates of other relevant parameters are obtained from CMAP (http://www.cmap.illinois.gov) or were recommended in discussions with the City of Chicago. We used the NLCD 2006 (Fry et al., 2011) to represent the modern-day land-use and land-cover (LULC) within the Noah LSM and National Urban Database and Access Portal Tool (NUDAPT; Ching et al., 2009) data set for the highly urbanized city of Chicago. These data were aggregated to the respective grid resolution of each domain. For SLUCM, we used AH profile averaged over CMA for summers based on empirical estimated values by Sailor and Lu (2004) (Figure S1). SLUCMAH requires that AH be specified as an input, while BEP and BEP + BEM compute AH using a parameterization scheme. The model explicitly accounts for sources of AH (sensible and latent). To account for AH, we treated A/C usage as a primary cause. Therefore, our model has A/C switched on 24 h a day with 75% efficiency. The target indoor temperature is 25 ∘ C with a comfort range of 0.5 ∘ C. The peak heat generated by other equipment in buildings is estimated as 28 W m −2 from 0800 to 1800 LST and 7 W m −2 during night (Note LST = UTC-5).

Sources of observational data
The uWRF simulations were evaluated using multiple Mesowest urban and rural stations located in the innermost domain that provided near-surface meteorological point data (http://mesowest.utah.edu). The location and land-use classification of stations are shown in Figure 2(b) and Table S1, and observed 2-m temperatures and 10-m wind speed and direction were used for comparison. Although our mesoscale simulations are high resolution (1-km innermost domain), they are still much coarser than the footprint of in situ point measurements, so observations should capture more spatial variability than the model. It is also important to note that point measurements are not fully representative of simulated values for a 1 × 1 km grid cell that contains them and classification of observational stations as 'urban' or 'rural', is at times poor (Stewart and Oke, 2012). This should be kept in mind when considering model evaluation results.
For spatial comparisons, we utilized Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua satellite land surface (skin) temperature observations to compare with the simulations (Leroyer et al., 2011;Zhang et al., 2011;Li et al., 2013b;Hu et al., 2014). The MODIS product used was MYD11A1 version 5, which provides daily level 3 global 1-km pixel grid based observations.

Choice of time-period for the study
Historically, over one-third of the summer and spring days show lake breeze over the CMA (Lyons, 1972;Laird et al., 2001). The period from 16 to 18 August 2013 was relatively hot with no precipitation over the CMA, but with strong lake breeze under quiescent synoptic conditions (Conry et al., 2015;Sharma et al., 2016). We used the criteria developed by Laird et al. (2001) to determine the lake-breeze period from observations with which we separated the lake breeze from the synoptic flow by requiring a change (a) wind conditions between morning and afternoon, and (b) a positive temperature gradient between the Lake Michigan and CMA. This case was used to examine the effects of land data assimilation, sub-grid scale land-cover variability, and their influence on the interaction of UHI effects with the lake breeze. Simulations were conducted from 0700 LST on 15 August 2013 to 1900 LST on 18 August 2013, with the first 12 h for spin up and 3 additional days for analysis. In this article, 'daytime' refers to average conditions in the late afternoon from 1400 to 1700 LST and 'nighttime' refers to average conditions in the very early morning before sunrise from 0200 to 0500 LST over the 3-day period.

High-resolution land data assimilation for UHI simulations
A state-of-the-art high-resolution land data assimilation system (HRLDAS v3.6; Chen et al., 2007) was used to initialize state variables of the Noah LSM for our high-resolution experiments (Table 2). HRLDAS is an uncoupled model that is run offline for each nested domain that integrates fine-scale static land characteristics, e.g.: land-use, terrain height, time-averaged vegetation fraction; observed and derived near-surface meteorological variables. To initialize the uWRF model for each nested domain, HRLDAS was spun up for the preceding year (1 August 2012 to 15 August 2013).

Sensitivity to sub-grid scale land-use variability
At 30-m resolution (the scale of high-resolution NLCD data), a 1-km grid cell contains multiple possibilities of different land-use classes, which presents an opportunity to include considerable sub-grid scale variability within a 1-km aggregate land-cover representation. To illustrate the effects of these more sophisticated schemes, we compared a number of possible land-use classes between MODIS_30 s resolution (∼900 m) and NLCD data for our innermost d04 domain. Using MODIS_30 s grid cells generally resulted in a single land-use classification for each 1-km grid cell. However, with NLCD data, we observed substantial sub-grid land-use variability in most 1-km cells (see Figure 6 and related discussion in Section 3.2). This variability was low in the core CMA (since most of the region was classified as urbanized), however, for the outer CMA, a typical 1-km grid cell contained seven to eight different land-use classes.

Interaction of lake breeze and urbanization
To study the interaction of UHI with the lake breeze, two sensitivity experiments for SLUCMAH were performed: an experiment with default NLCD land-use classification and another with NLCD urban land-use classification modified to croplands (agricultural). Land surface data assimilation (HRLDAS v3.6) was also incorporated in these model runs.
To measure the extent/penetration of lake breeze, we also compared simulated 10-m wind speed with observations by choosing four stations as shown in Figure 2 (2) the second station 'b' D8777 is close to the lakeshore (Figure 10

Sensitivity of different urban parameterizations
We hypothesize that the sensitivities related to the initialization of soil moisture and temperature (with and without HRLDAS) and the details in the land-use representation (mosaic vs dominant category) along with the choice of urban parameterization will help understand and reduce uncertainties in the numerical modeling of near-surface meteorology. Thus, this study performs the sensitivity on different urban parameterizations, viz., BULK, SLUCM with and without AH, BEP and BEP + BEM using 2-m temperature, and 10-m wind speed analysis for the above land-use modeling approaches ( Table 2). The analyses were divided into urban and rural categories. Urban stations were further classified into low-intensity, medium-intensity, and high-intensity land-use classes (refer to Tables S2-S5 for statistics on urban land-use classes). Since the UHI effect is most prominent at night, we aggregated the hourly data into 12-h periods for day (0600-1700 LST) and night (1800-0500 LST) for statistics presented in Tables 3 and 4.

Results
In this section, we discuss the results related to four sensitivity studies outlined in Section 2.1.
3.1. High-resolution land data assimilation for UHI simulations Figure 3 shows the uWRF model initialization with NARR data for soil moisture and soil temperature. Simulations without HRLDAS (referred to as 'noHRLDAS') did not resolve soil moisture from LULC variations, which was almost uniform with ∼0.1 m 3 m −3 . HRLDAS on the other hand captured the heterogeneity of soil moisture; soil moisture over urban areas was typically very low (0.05 m 3 m −3 ) compared to agricultural areas (0.25 m 3 m −3 ) (Figure 3(a) and (b)). Similarly, heterogeneity was simulated in soil temperature when using HRLDAS. The case of noHRLDAS did not show realistic soil temperature over urban areas, whereas HRLDAS showed physically realistic soil temperatures that were higher over urban areas and lower over agricultural areas (Figure 3(c) and (d)). Findings that are illustrated in Figure 3 are for the top land surface layer (0.1 m below surface layer), and similar patterns were also observed for the lower layers. Figure 4 presents the model comparison of 2-m air temperatures overlaid by 10-m wind vectors for different permutations of SLUCM urban parameterizations with and without AH and HRLDAS for nighttime and daytime. SLUCMAH with HRLDAS showed higher 2-m air temperatures during nighttime over the agricultural region and lower temperature over the CMA in comparison to the case with noHRLDAS (Figure 4(a)-(c)). High humidity in agricultural areas led to higher nighttime temperatures because water vapour reduces the loss of longwave radiation from the surface to atmosphere (Ruckstuhl et al., 2007). This was consistent with sensitivity experiments with initial soil temperatures and soil moisture (Figure 3). During the daytime, however, the overall temperatures decreased in comparison with the simulation with noHRLDAS (Figure 4(d)-(f)) due to reduction of artificial heating caused by higher soil moisture and lower soil temperature with HRLDAS. It was also noted that in the noHRLDAS simulations the lake breeze penetrated deep into the urban areas in comparison to HRLDAS simulations, because there was an overall warming over the non-urban domain due to less moisture and higher soil temperature (Figure 4(e)). The sensitivity of AH with HRLDAS showed that AH produced 3 ∘ C nighttime warming over the CMA (Figure 4(a), (g), (i)). During the daytime, the warming due to AH was not substantial, however, as the lake breeze penetrated over the CMA and reduced the impact of AH (Figure 4(d) and (h)). Figure  S2 compares CMA-averaged simulated surface fluxes using three different approaches of SLUCM: HRLDAS noAH, HRLDAS + AH, and noHRLDAS + AH. Figure 5 shows the sensitivity of BEP + BEM scheme to HRLDAS. Therein the 2-m temperatures over urban areas were increased during the nighttime and reduced during the daytime with HRLDAS.

Sensitivity to sub-grid scale land-use variability
This analysis led us to two important and complementary findings. First, at a resolution ∼1 km and below, the model did not have much sub-grid variability for 30 s resolution data and a dominant land-use approach could be used without any substantial loss of information from the primary data (MODIS). Second, when using NLCD data, there was a large impact of sub-grid scale variability at 1-km resolution (Figure 6), and modeling with the conventional dominant land-use classification approach would frequently lead to a loss of information from the primary land-cover data sets and also potentially spurious results from the uWRF simulations due to misclassification of the land-use in each cell. Thus, accounting influence of multiple land-use classes was appropriate within each grid cell for studying UHI effects study.
To test the sensitivity to land-use categories within a grid cell (Table 2), we compared the traditional approach of using a dominant land-use class for each grid cell (dominant approach) with a new approach using multiple classes to represent the sub-grid variability of the land-cover (mosaic approach) (Li et al., 2013b). Note Table 3. Statistical comparison of the simulated and observed rural 2-m temperatures ( ∘ C) and 10-m wind speeds (m s −1 ) for 16-18 August 2013. The criterion for HR calculation is 1 ∘ C for temperature and 1 m s −1 for wind speed.  that the mosaic approach was used only with SLUCM and the rest of the urban parameterizations employed the dominant approach. In this experiment, eight most prominent land-use classes were used within a grid cell to represent the impact of sub-grid land-use variability, by employing SLUCMAH with HRLDAS parameterization. For urban areas during daytime, the sensible heat flux was lower for the mosaic approach than for the dominant approach as mosaic accounted for non-urban land-use component within urban areas that released less sensible and more latent heat flux during daytime in the atmosphere ( Figure S3). Latent heat flux was generally lower for urban than rural ( Figure S3). In particular, latent heat flux increased with the mosaic approach in urban areas due to better representation of non-urban classes. For detailed discussion on surface energy fluxes using dominant and mosaic approaches, refer to Figure S3. Values in the plots are the water vapour mixing ratios (kg kg −1 ). During the day (1600 LST), the UHI and lake breeze developed close to the ground (Figure 9(a)). The lake breeze showed relatively high wind velocity over the lake, but the velocity was lower as it advected over the urban areas due to higher surface friction. During the night (0400 LST), the land surface eventually became cooler than the lake and a land breeze developed (Figure 9(c)). When urban grids were modified to cropland, UHI did not develop and the lake breeze was not as strong as with urbanized grids due to lower temperature (Figure 9(b)). Interestingly, although the lake breeze was weaker with cropland in place, it still penetrated inland as much as in the previous case. One possible explanation is the reduced surface friction of croplands in comparison with urban areas, which results in less reduction in wind velocity with distance in the case of cropland. Therefore, the inland penetration of lake breeze was dependent not only on the temperature difference between lake and land, but also on land-use roughness. As expected, the simulated night temperatures were lower when urban areas were modified to cropland (Figure 9(d)). For both cases, the circulation pattern was observed in the lower atmosphere with positive vertical velocities over urban areas and negative vertical velocities over Lake Michigan ( Figure S4).

Sensitivity of different urban parameterizations
For 10-m winds, in general, the uWRF model showed a mild land breeze (eastward) during the morning on 16 August 2013, and a strong lake breeze (westward) during the afternoon and evenings for all 3 days (Figure 10(a)). This diurnal variation was observed at all distances from the lake. Over land, closest to the lake, the simulated winds were almost as strong as over the lake (Figure 10(b)). However, observations of 10-m wind speeds close to the lake were of poor quality due to disruption of the flow by tall buildings in downtown Chicago. Further from the lake, but still in urban areas, the lake breeze wind speed in simulations and observations decreased due to the surface resistance (Figure 10(c) and (d)). Farthest from the lake ( Figure 10(d)), both simulations and observations began to show a systematic reduction in wind speed, and the local winds also more closely followed the synoptic flow from the large-scale forcing (e.g. Figure 7(d)). For all combinations of land-use and urban parameterizations shown in Table 2, simulated 10-m wind speed directions closely matched observations as shown in Figure 10. The statistics of 10-m wind speeds magnitude for rural and urban stations were all comparable and are shown in Tables 3 and  4, respectively. The sensitivity of 2-m temperature to different SLUCM parameterizations was minimal for rural stations, except for those close to the CMA that are affected by advection of lake breeze and urban heating. Nevertheless, simulations for 2-m temperature over rural stations showed reduced root mean square error (RMSE) and mean bias (MB), and higher hit rate (HR) with HRLDAS in comparison  to the case when HRLDAS was not employed, since soil moisture and soil temperature were well represented after land data assimilation; note that default NARR lacked a robust representation of land surface forcing data (Table 3 and Figure 11(d)). Inclusion of land-use heterogeneity using HRLDAS + SLUCM + mosaic also improved peak day rural 2-m temperatures (Table 3). In comparison to rural stations, urban stations also showed significant sensitivity to different SLUCM parameterizations. Within different SLUCM parameterizations, the performance of SLUCMs at night improved in cases when AH was included in the model land-use forcing. The night RMSE was lowest with HRL-DAS + SLUCMAH (Figure 11 and Table 4). Simulations for SLUCM with no AH showed cold biases (MB) for both during day and night for all urban classes (Table 4  and Tables S2-S4). Simulations with no HRLDAS also showed cold bias during night for all urban classes. HRLDAS + SLUCM + mosaic performed well in comparison with other SLUCM schemes, especially over low-intensity urban areas as it accounted for the aggregated effects of multiple land-use classes within the suburban areas ( Figure 6), which was missing in other SLUCM parameterizations.
Comparing all UCM schemes over urban areas (Figure 12), BEP + BEM with HRLDAS showed lowest MB and RMSE for both during day and night (Table 4) due to explicit accounting of AH from urban structures compared to SLUCMs diurnal climatological AH profile ( Figure S1). However, BULK showed least RMSE for high-intensity urban area, even though the parameterization is relatively simplistic. For low-intensity urban areas, BEP showed the least RMSE (a little lower than BEP + BEM). For all urban land-use classes, night MB and RMSE decreased from BEP to BEP + BEM because the BEM within BEP + BEM accounts for latent and sensible AH separately by prescribed A/C usage (Section 2.2; Salamanca and Martilli, 2010) and would lead to higher sensible heat released to the atmosphere to maintain prescribed indoor temperatures within comfort level limits. For all parameterizations, MB and RMSE were typically lower at night and higher at day for both rural and urban stations with different urban classes.

Discussions and conclusions
This article evaluates simulations of summertime high temperatures in CMA due to UHI effects, and interaction of UHI with lake breeze from the Lake Michigan. Sensitivity of these simulated phenomena to the urban parameterization schemes employed in the uWRF model was evaluated. All available urban parameterizations in uWRF were evaluated in the experiments: the BULK urban scheme, SLUCM with and without AH, BEP, and BEP + BEM scheme. The study also examined the use of a land surface data assimilation technique (HRLDAS) for potential reduction of uncertainties resulting from initial   conditions, along with mosaic approach for LSM which accounts for sub-grid variability of land-cover. In addition to normal UHI effects related to land-cover, the study evaluated simulations of more complicated urban phenomena where UHI effects due to land-cover interact with the lake breeze from Lake Michigan. The simulations showed that the lake breeze likely penetrates about 15-30 km inland over the CMA and then fades away upon interacting with the synoptic flow. The influence of UHI effects during the daytime was reduced by lake breeze along the highly urbanized CMA coastline. Simulations of urban land-use modified to cropland showed  Figure 9. Vertical cross-section of the CMA (as shown in Figure 2(a) as a black dashed line) with contoured temperature ( ∘ C) and horizontal wind speed vectors on 17 August 2013 showing a schematic of lake breeze phenomenon for SLUCMAH with HRLDAS (a) during day (1600 LST) for default urbanization, (b) with hypothetical situation when urban areas are modified to croplands. Panels (c) and (d) are same as (a) and (b) respectively except for night (0400 LST). Values in the plots are the water vapour mixing ratios (kg kg −1 ).
reduced temperatures but an unabated lake breeze with strength almost similar to the case with urban land-use.
In the cropland case, the temperature difference was low between urban areas modified to cropland and surrounding cropland/rural areas, so was surface friction on land, with one counteracting the other in determining the lake breeze. Initial conditions were found to be an important source of model bias. The inclusion of land surface data assimilation using HRLDAS significantly changed both soil moisture and soil temperature initialization, which in turn modified the simulation of 2-m air temperature and 10-m wind speed. Note, there were no quantitative comparisons against observations of soil moisture and soil temperature, only qualitative discussion about anticipated outcomes on near-surface temperatures and wind speeds with different land-covers in contrast to NARR initialization were performed. Urban schemes that used the mosaic approach better captured the spatial variability of temperature, particularly at night. These effects were particularly pronounced in the Chicago suburbs where noticeable surface LULC variability within respective grid cells was common. In absence of strong synoptic flow, with improved initial conditions and improved land surface modeling techniques, the more complicated urban schemes performed better. Urban schemes that used the mosaic approach did not necessarily achieve the best performance in comparison with other schemes in terms of reproducing point observations, but the ability to resolve sub-grid spatial variability in suburban areas was an important advantage of this approach.
With inclusion of appropriate and realistic parameter values for the CMA for all available urban schemes, we   Figure 11. 2-m temperatures ( ∘ C) for different SLUCM parameterizations and for different urban (high intensity, medium intensity, and low intensity) and rural LULC classes. The stations with latitude-longitude and LULC information are shown in Table S1.  Table S1.
have attempted to reduce the bias in the uWRF simulations as much as possible without artificially tuning parameters for each urban scheme to match the observations. We hope the methodology can be of use in guiding future urban studies, especially in view of the fact that the approaches used here can readily be applied to other urban areas, including coastal cities whose UHI effects are influenced by the marine environment. We also discussed the strengths of a mosaic approach to account for sub-grid land-use variability, which will be valuable in future studies if high-resolution land-use maps are available for other parts of the world. Overall, this article is a step forward in understanding and modeling the physics of urban environments using different urban simulation schemes combined with urban parameters customized for individual cities. It is our hope that improvements in modelling UHI effects will help predict the impacts of climate variability and climate change on cities and help develop appropriate adaptation/mitigation strategies to heat-related impacts.

Supporting Information
The following supporting information is available as part of the online article: Figure S1. Anthropogenic heat (AH) profiles for the Chicago metropolitan area (CMA): (a) average winter and summer profiles. Winter profile is higher than summer due to exacerbated building heating in extreme cold weather; and (b) AH profiles for summers for different urban land-use classifications. Figure S2. Comparison of CMA-averaged energy budget fluxes for different combinations of SLUCM with and without AH, and with and without HRLDAS. Figure S3. Comparison of domain averaged energy budget fluxes for dominant and mosaic parameterizations for rural and urban areas. Urban component of mosaic/dominant refers to the sub-grid relative contribution of the urban fraction within each grid cell for respective approaches. Figure S4. Vertical cross-section of the CMA on 17 August 2013 at 1600 LST for SLUCM during daytime showing contours of (a) horizontal wind component and (b) vertical wind component. Table S1. Detailed meteorological station information and their type of MODIS LULC classification used in the study. Table S2. Statistical comparison of simulated and observed 2-m temperatures ( ∘ C) and 10-m wind speeds (m s −1 ) for the high-intensity urban stations for 16-18 August 2013. The criterion for hit rate calculation is 1 ∘ C for temperature and 1 m s −1 for wind speed. Table S3. Statistical comparison of simulated and observed 2-m temperatures ( ∘ C) and 10-m wind speeds (m s −1 ) for the medium-intensity urban stations for 16-18 August 2013. The criterion for hit rate calculation is 1 ∘ C for temperature and 1 m s −1 for wind speed. Table S4. Statistical comparison of simulated and observed 2-m temperatures ( ∘ C) and 10-m wind speeds (m s −1 ) for the low-intensity urban stations for 16-18 August 2013. The criterion for hit rate calculation is 1 ∘ C for temperature and 1 m s −1 for wind speed.