Assessment of the Urban Impact on Surface and Screen-Level Temperature in the ALADIN-Climate Driven SURFEX Land Surface Model for Budapest

Land surface models with detailed urban parameterization schemes provide adequate tools to estimate the impact of climate change in cities, because they rely on the results of the regional climate model, while operating on km scale at low cost. In this paper, the SURFEX land surface model driven by the evaluation and control runs of ALADIN-Climate regional climate model is validated over Budapest from the aspect of urban impact on temperature. First, surface temperature of SURFEX with forcings from ERA-Interim driven ALADIN-Climate was compared against the MODIS land surface temperature for a 3-year period. Second, the impact of the ARPEGE global climate model driven ALADIN-Climate was assessed on the 2 m temperature of SURFEX and was validated against measurements of a suburban station for 30 years. The spatial extent of surface urban heat island (SUHI) is exaggerated in SURFEX from spring to autumn, because the urbanized gridcells are generally warmer than their rural vicinity, while the observed SUHI extent is more variable. The model reasonably simulates the seasonal means and diurnal cycle of the 2 m temperature in the suburban gridpoint, except summer when strong positive bias occurs. However, comparing the two experiments from the aspect of nocturnal UHI, only minor differences arose. The thorough validation underpins the applicability of SURFEX driven by ALADIN-Climate for future urban climate projections.


Introduction
Urbanized surfaces are characterized by highly different surface properties (e.g., imperviousness, large roughness, large heat capacity, and heat admittance of materials) from natural surfaces, that leads to altered energy budget components and boundary layer properties. One of the most studied and well-known manifestations of these differences is the higher surface, subsurface and air temperature in the city, forming the urban heat island (UHI) [1]. More than half of the world's population lives now in urban settlements, although the impervious surfaces occupy less than 0.5% of the Earth [2]. Such high population density can be observed in Hungary as well; for example, the Budapest urban agglomeration occupies 2.7% of the territory of Hungary, while its 2.5 million inhabitants represent one quarter of the whole country's population [3]. As a consequence of the high population density and elevated temperatures, cities are exposed to more severe heat related risks compared to rural areas. With climate change, this exposure may be further exacerbated, negatively affecting for example human health, environment and energy demand [4].
By the increasing resolution and complexity of global and regional climate models (GCMs and RCMs), urban processes tend to be parameterized, giving a first approximation about the impact of climate change on cities [5]. For example, it was shown that are validated against Moderate Resolution Imaging Spectroradiometer Sensor (MODIS) LST data on a 3 year period and against the synoptic station measurement on a 30 year period. Section 2 presents the SURFEX model, the evaluation and control simulation of the ALADIN-Climate providing forcings for SURFEX, the considered observations and the evaluation methods and measures. Section 3 is dedicated to presenting the spatial characteristics of LST and surface UHI (SUHI), and the temporal characteristics of urban climate in one suburban gridpoint. The results are discussed in Section 4 and finally, Section 5 summarizes our conclusions and presents future perspectives of our research.

The SURFEX Land Surface Model
SURFEX [26] describes land surface processes in the lower 10% (a few 10 m) of the planetary boundary layer, called surface layer. The surface heterogeneity is handled by the tiling method, i.e., in each grid cell, the fractions of sea, inland water, town and natural land surface are given, and for each surface type (i.e., tile) a dedicated scheme computes the prognostic variables and turbulent fluxes. The total flux of the grid box is resulted by addition of the individual fluxes over the composing tiles. The land cover information is provided for the model by the first version of the ECOCLIMAP database [27]. From urban climate perspective, the continental cities and their vicinity can be described dominantly by the town and natural land surface types, therefore the schemes applied for them are presented hereinafter.
Over natural land surfaces the 3-layer Interaction Soil Biosphere Atmosphere scheme (ISBA-3L) [28] is used, that computes the surface and soil temperature and moisture with the force-restore method. This scheme divides the soil into three layers (a thin layer representing the surface, and two thicker layers for middle and deep soil), and the prognostic evolution of the temperature and moisture is described with a relatively simple formula.
The urban physical properties are calculated with the TEB scheme [29] that approximates the complexity of urban morphology with street canyons. Prognostic equations for the roof, wall and road specify the surface temperature and (except for wall) water content. The surfaces are divided into three layers in order to consider heat conductivity. Only domestic heating was considered as the source of anthropogenic heat flux by preventing the indoor temperature to fall below 19 • C. Note that in SURFEX there is a simple building energy model that can simulate the interaction of internal building temperature and the outdoor temperature. The near surface variables (e.g., 2 m temperature, humidity, 10 m wind speed) are calculated with the surface boundary layer (SBL) scheme [30,31] that divides the surface layer into six sub-layers and prognostically computes the temperature, humidity, wind and turbulence, taking into account the canopy and canyon effect and drag forces.

The Driving ALADIN-Climate Model
The atmospheric forcings of SURFEX are temperature, humidity, wind speed and wind direction at a few 10 m above ground level, downward shortwave and longwave radiation, surface pressure, snow and rain. In our case the forcings are provided by the ALADIN-Climate version 5.2 [32] hydrostatic spectral RCM. The physical parameterization package of ALADIN-Climate is derived from the ARPEGE-Climat version 5 [33] atmospheric GCM. The longwave radiation transfer is described by the Rapid Radiation Transfer Model (RRTM) scheme [34], while the shortwave radiation transfer is parameterized according to Fouquart and Bonnel [35]. The large-scale precipitation is determined by the Smith scheme [36], and the convective cloud and precipitation formation are described according to Bougeault [37]. The surface scheme of ALADIN is SURFEX version 5, in which ISBA-3L was applied over natural land surfaces. The vertical profile of temperature, humidity and windspeed in the surface layer is parameterized according to Geleyn [38]. Urbanized areas are substituted with rocks and the physical processes are described by the ISBA scheme.

Experimental Design
In this study, two urban climate simulations are performed with the 5.1 version of SURFEX for Budapest driven by the evaluation and control run of ALADIN-Climate 5.2 RCM. The ALADIN-Climate simulations were achieved at 10 km horizontal resolution on a domain covering Central and South-Eastern Europe ( Figure 1). First, a 10-year long simulation was performed for 1996-2005 with SURFEX, where the lateral boundary conditions for ALADIN-Climate were derived from ERA-Interim [39]. This experiment allows to examine the RCM model's influences on the behavior of SURFEX (hereinafter the RCM and LSM simulations are referred to ALADIN-EI and SURFEX-EI, respectively). Then, SURFEX was driven by the control simulation of ALADIN-Climate for 1960-2005 to evaluate SURFEX from a climate perspective. In this case, the lateral boundary conditions of ALADIN-Climate were the CNRM-CM5 GCM (hereinafter the RCM and LSM simulations are referred to ALADIN-ARP and SURFEX-ARP, respectively).
by the Smith scheme [36], and the convective cloud and precipitation formation are described according to Bougeault [37]. The surface scheme of ALADIN is SURFEX version 5, in which ISBA-3L was applied over natural land surfaces. The vertical profile of temperature, humidity and windspeed in the surface layer is parameterized according to Geleyn [38]. Urbanized areas are substituted with rocks and the physical processes are described by the ISBA scheme.

Experimental Design
In this study, two urban climate simulations are performed with the 5.1 version of SURFEX for Budapest driven by the evaluation and control run of ALADIN-Climate 5.2 RCM. The ALADIN-Climate simulations were achieved at 10 km horizontal resolution on a domain covering Central and South-Eastern Europe ( Figure 1). First, a 10-year long simulation was performed for 1996-2005 with SURFEX, where the lateral boundary conditions for ALADIN-Climate were derived from ERA-Interim [39]. This experiment allows to examine the RCM model's influences on the behavior of SURFEX (hereinafter the RCM and LSM simulations are referred to ALADIN-EI and SURFEX-EI, respectively). Then, SURFEX was driven by the control simulation of ALADIN-Climate for 1960-2005 to evaluate SURFEX from a climate perspective. In this case, the lateral boundary conditions of ALADIN-Climate were the CNRM-CM5 GCM (hereinafter the RCM and LSM simulations are referred to ALADIN-ARP and SURFEX-ARP, respectively).  The SURFEX domain covers Budapest and its vicinity ( Figure 1). It consists of 61 × 61 gridpoints with 1 km horizontal resolution. The main urban cover types over the domain are dense urban, temperate suburban, industrial and commercial areas, mineral extractions and construction sites, sport facilities, urban parks and airport according to ECOCLIMAP. The dominant natural cover types are Central European crops and temperate woodlands. 3 hourly outputs of ALADIN-Climate simulations derived at 30 m above ground level are considered as forcings. The ALADIN simulations are interpolated from 10 km to 1 km resolution using the 927 configuration of ALADIN-Climate (in general this configuration is used to prepare lateral boundary conditions for the RCM). The advantage of this method-in contrast to simply multiplying the gridpoints in each 10 km × 10 km gridcell-is that it takes into account the orography effect during interpolation, and it is crucial for a hilly domain. The integration timestep of SURFEX is 300 s, to which the 3 h forcings are linearly interpolated.

MODIS LST Product
Simulated surface temperature and SUHI over Budapest is studied against the MODIS Terra (from 1999) and Aqua (from 2000) LST product (version 6 of MOD11A1 and MYD11A1, respectively) [40,41]. The two polar orbit satellites acquire data in 36 spectral bands from 0.4 to 14.4 µm wavelengths [42]. The LST products are produced by the split-window method based on 7 spectral bands 4 times a day [43]. Table 1 shows the averaged satellite overpass time in UTC for Aqua and Terra over Budapest. The spatial resolution of the retrieved LST images are 1 km [19], which coincides with the spatial resolution of SURFEX LST fields. The satellite observation data was first cleaned from the cloud contaminated grid cells in the following way: those times were excluded where the cloudy cells coverage exceeded 25 per cent of the whole domain. This filtering was applied for the SURFEX data as well using the mean total cloud cover of ALADIN-EI over Budapest, in order to keep the cloud-free cases only.
Finally, we applied the MODIS Terra + Aqua Combined Land Cover product [44,45] that includes the IGBP (International Geosphere-Biosphere Programme) land-use classification system which enabled us to select urban and non-urban grid cells in the MODIS LST analysis.

Gridded Observational Dataset and Station Measurements of 2 m Temperature
The 2 m temperature of SURFEX and ALADIN is evaluated against gridded and synoptic station measurements. The CarpatClim-Hu [17] gridded observation dataset was constructed based on homogenized and interpolated station measurements, and widely used for regional climate model evaluation for Hungary. Note that due to its 10 km resolution and the missing urban stations, urbanized areas are not emerging with higher temperatures in this dataset.
One station was chosen for pointwise validation based on the criteria of being continuously operated since 1970 and located in urbanized area. This station is situated in the Southeastern agglomeration area of Budapest (in Pestszentlőrinc) at 19 • 10 56 E, 47 • 25 45 N and at 139 m above sea level. The vicinity of the synoptic station is covered by grass, but the observatory is surrounded by houses with gardens except from south that is covered by forest ( Figure 1). According to ECOCLIMAP, the closest gridpoint to the station is classified as temperate suburban, which consists of 60% urban area and 40% natural area.
In order to eliminate temporal inhomogeneities (originated from e.g., station relocation or change of instruments or measurement methods), the daily observation data was homogenized with the Multiple Analysis of Series for Homogenization (MASH) method [46].

Surface Temperature and SUHI in SURFEX-EI
Since four measurements per day are available from the end of 2002 in MODIS (Table 2), we considered only the SURFEX-EI experiment on the period of 2003-2005 on a domain containing mainly urbanized and cropland areas ( Figure 2). SURFEX-ARP cannot be evaluated for such a short period, because it does not reflect the observed characteristics of the individual years (since it is driven by a GCM, in which only the observed greenhouse gas concentration describes the past reality, see Figure 8).
To compare the simulated value with the satellite data, the gridcell average surface temperature had to be computed beforehand, because it is not given by SURFEX v5. The mean surface temperature (LST SURFEX ) was derived based on the Stefan-Boltzmann law. Therefore, the emitted longwave radiations (LW) given by the TEB and ISBA schemes were taken (roof, wall and road values were averaged in TEB) using the same emissivity values (ε) as the model. Emissivity is constant in time and identical (0.97) for all natural land cover types, and for roof, wall and road as well (0.90, 0.85 and 0.94, respectively). Radiative surface temperature was produced for TEB and ISBA separately and the results were linearly combined according to the fraction of urban and nature land cover (x urban and 1 − x urban , respectively; Equation (1): First, SURFEX simulations at 0 and 12 UTC-representing nighttime and daytime caseswere considered and compared with measurements of the Aqua satellite. Note that the measurements happen 30 min earlier around noon and 43 min later at night. We analyzed the seasonal mean daytime and nighttime LST and its variability to investigate how SURFEX is able to describe the spatial and temporal distribution of surface temperature over the urban and rural areas.
Afterwards SUHI affected area was determined based on the method of Zhang and Whang [47], which statistically defines the size of the area, which significantly differs in temperature from its environment. Following the method, we calculated the spatial mean surface temperature (T in • C) and standard deviation across gridpoints (σ in • C) in each timestep. Then, LST (T) in each gridpoint and timestep was categorized based on Equation (2): Therefore, we identified those gridpoints as affected by SUHI where the temperature was higher than T + σ. The sum of the gridpoints, which fulfill this criterium were taken and divided by the number of all gridpoints for each timestep. This gave the relative SUHI affected area. In this analysis we considered a smaller domain bounded by the administrative border of Budapest (blue rectangle on Figure 2), because otherwise the relative SUHI affected area values (i.e., the proportion of SUHI affected gridcells and the all gridpoints of the domain) would be too low. The observed and modelled daily values are compared on histograms.
Finally, the seasonal mean diurnal cycle of SUHI and its variability was studied. SURFEX results were compared to all four measurements of Aqua and Terra, and SUHI was computed in the following way: in every timestep and in every gridpoint the LST was subtracted from the mean rural temperature. The mean rural temperature was determined as the average temperature of pure rural grid cells according to ECOCLIMAP in SURFEX and IGBP in MODIS.

2 m Temperature in SURFEX-ARP
Besides short-term validation focusing on the urban surface temperature, the ability of SURFEX to reproduce long-term climate characteristics is assessed based on the SURFEX-ARP experiment. First, the temperature bias of ALADIN, and the interaction between ALADIN and SURFEX is briefly validated on 10 km resolution considering CarpatClim-Hu as reference (Table 2). Moreover, the performance of SURFEX is studied in one gridpoint, closest to the synoptic meteorological station in Pestszentlőrinc ( Table 2). Measurements of 3-hourly and daily mean temperature were used to validate SURFEX in the period of 1971-2000. Annual and seasonal yearly mean values, distribution of daily mean temperatures and the mean diurnal cycle of 3 h temperature were assessed.

UHI in SURFEX-EI and SURFEX-ARP
Finally, SURFEX-EI and SURFEX-ARP was jointly evaluated and compared against each other, considering spatial distribution and seasonal mean diurnal cycle of UHI, and seasonal mean diurnal cycle of surface energy fluxes on a common 10-year period (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005). The UHI were computed following the same method as was presented for SUHI in Section 2.5.1.  Note that the surface temperature over the river is prescribed and kept constant throughout the simulation, therefore these gridpoints were masked out. Considering the nighttime results, the core of the city is emerging with the highest temperature in all seasons in SURFEX. This pattern can be detected in the observation too; however, it has different magnitude and shape. Generally, the isotherm lines portray roughly concentric circles around the downtown in the MODIS, while in SURFEX, except the warmest downtown area, the rest of the city is approximately equally warm. The spatial pattern of the observed daily SUHI largely differs from the nocturnal results, and the differences between the model and observation are larger as well. In MODIS, the hilly area of the domain (west from the Danube, mainly forests) is much colder than the plain areas from spring to autumn. The influence of topography and partly the cooling effect of trees on the daytime surface temperature can be seen in SURFEX as well, however over a smaller area, not propagating into the western edge of the city. Due to this orographic and land cover effect, the warmest area is observed over the Pest side of Budapest (east from the Danube), while in SURFEX almost the overall urban area is significantly warmer than its environment.

Surface Temperature and SUHI
The daytime and nighttime land surface temperatures averaged over the urban and rural areas are seasonally shown on boxplots for SURFEX and MODIS ( Figure 4) and temporally averaged values are shown in Table 3. The urban LST is overestimated almost in all seasons and in daytime and nighttime as well (as it was suggested earlier); while in Atmosphere 2021, 12, 709 9 of 20 winter, the model performs the best over the whole domain. The rural temperatures are colder in spring and autumn at night and in spring at noon in SURFEX than in MODIS, which-in line with the exaggerated urban temperatures-leads to stronger simulated SUHI. It can be also mentioned that variability is well described by the model with slightly overestimated variability of the mean urban temperature.

environment.
The daytime and nighttime land surface temperatures averaged over the urban rural areas are seasonally shown on boxplots for SURFEX and MODIS ( Figure 4) and t porally averaged values are shown in Table 3. The urban LST is overestimated almos all seasons and in daytime and nighttime as well (as it was suggested earlier); whil winter, the model performs the best over the whole domain. The rural temperatures colder in spring and autumn at night and in spring at noon in SURFEX than in MOD which-in line with the exaggerated urban temperatures-leads to stronger simulated SU It can be also mentioned that variability is well described by the model with slightly o estimated variability of the mean urban temperature.  The SUHI-affected area was investigated and validated following the method Zhang and Wang [47]. Figure 5 shows the seasonal distribution of the relative are SUHI affected gridpoints (the sum of the gridpoints where > + ) over the city of dapest. In MODIS at night the histograms are skewed left in each season except wi (when it is rather symmetric) with the most frequent values of 25-30%. The general o estimation of SUHI extent in SURFEX is obvious, because except winter the histogram  The SUHI-affected area was investigated and validated following the method of Zhang and Wang [47]. Figure 5 shows the seasonal distribution of the relative area of SUHI affected gridpoints (the sum of the gridpoints where T > T + σ) over the city of Budapest. In MODIS at night the histograms are skewed left in each season except winter (when it is rather symmetric) with the most frequent values of 25-30%. The general overestimation of SUHI extent in SURFEX is obvious, because except winter the histogram is shifted towards the larger values. The most frequent SUHI area is between 35-40%, which occurs in the great majority of time in spring and summer. However, in winter, the histogram of SURFEX has a bimodal shape and most often the SUHI area is below 5%, therefore in this season MODIS shows a more developed SUHI area. At daytime, SURFEX shows a similar distribution to nighttime results both in the shape and most frequent SUHI relative area. While the histogram of observed value portrays different shape than at nighttime, because they are rather symmetric except summer (when it is skewed left). The most frequent areas are 10-30% that is lower than of SURFEX again. In winter the same can be observed, SURFEX underestimates the SUHI affected area. Based on this analysis we can conclude that there is much smaller variability in the SUHI affected area in SURFEX both at day and at night in all seasons compared to MODIS measurement.
Atmosphere 2021, 12, x FOR PEER REVIEW 10 of 20 shifted towards the larger values. The most frequent SUHI area is between 35-40%, which occurs in the great majority of time in spring and summer. However, in winter, the histogram of SURFEX has a bimodal shape and most often the SUHI area is below 5%, therefore in this season MODIS shows a more developed SUHI area. At daytime, SURFEX shows a similar distribution to nighttime results both in the shape and most frequent SUHI relative area. While the histogram of observed value portrays different shape than at nighttime, because they are rather symmetric except summer (when it is skewed left). The most frequent areas are 10-30% that is lower than of SURFEX again. In winter the same can be observed, SURFEX underestimates the SUHI affected area. Based on this analysis we can conclude that there is much smaller variability in the SUHI affected area in SURFEX both at day and at night in all seasons compared to MODIS measurement. Finally, the seasonal mean daily cycle of SUHI is shown in Figure 6. It is obvious that besides the general overestimation in SURFEX, there is a much larger intraday variability in LST compared to MODIS, especially in spring and summer. SUHI varies between 1 °C and 4 °C in spring and summer, 0.5 °C and 2 °C in autumn and 0.5 °C and 1 °C in winter in SURFEX, while in MODIS it does not exceed 1.5 °C in any seasons. The best agreement is found in winter at 12 and 0 UTC compared to the Aqua results. Finally, the seasonal mean daily cycle of SUHI is shown in Figure 6. It is obvious that besides the general overestimation in SURFEX, there is a much larger intraday variability in LST compared to MODIS, especially in spring and summer. SUHI varies between 1 • C and 4 • C in spring and summer, 0.5 • C and 2 • C in autumn and 0.5 • C and 1 • C in winter in SURFEX, while in MODIS it does not exceed 1.5 • C in any seasons. The best agreement is found in winter at 12 and 0 UTC compared to the Aqua results. Atmosphere 2021, 12, x FOR PEER REVIEW 11 of 20

2-m Temperature and UHI Climatology
In the following, 2 m temperature simulated by SURFEX-ARP is studied and compared in one suburban gridpoint with station measurements in the period of 1971-2000. Before the SURFEX results are assessed, it can be mentioned that the ALADIN (providing forcings to SURFEX) largely (with 2.9 °C) overestimates the 2 m temperature in summer, while underestimates it in the rest of the year over the SURFEX domain with respect to the 10 km resolution CarpatClim-Hu gridpoint dataset (Table 4). Comparison of the 2 m temperature in ALADIN and SURFEX reveals that SURFEX heats not only the urban gridpoints (due to the detailed land cover and parameterization), but in autumn and in winter the lower elevated rural gridpoints are warmer as well (Figure 7). In the gridpoint nearest to the station located in the suburban area, SURFEX overestimates the measured value in summer and autumn, with 3.5 °C and 1.1 °C respectively (Figure 8). In spring and in winter smaller underestimation can be observed. The variance of the mean values are similar in SURFEX than the observation in the 30 year period, apart from summer, when SURFEX overestimates the interannual variability of the mean temperature (Table 5).

2-m Temperature and UHI Climatology
In the following, 2 m temperature simulated by SURFEX-ARP is studied and compared in one suburban gridpoint with station measurements in the period of 1971-2000. Before the SURFEX results are assessed, it can be mentioned that the ALADIN (providing forcings to SURFEX) largely (with 2.9 • C) overestimates the 2 m temperature in summer, while underestimates it in the rest of the year over the SURFEX domain with respect to the 10 km resolution CarpatClim-Hu gridpoint dataset (Table 4).
Comparison of the 2 m temperature in ALADIN and SURFEX reveals that SURFEX heats not only the urban gridpoints (due to the detailed land cover and parameterization), but in autumn and in winter the lower elevated rural gridpoints are warmer as well (Figure 7). In the gridpoint nearest to the station located in the suburban area, SURFEX overestimates the measured value in summer and autumn, with 3.5 • C and 1.1 • C respectively ( Figure 8). In spring and in winter smaller underestimation can be observed. The variance of the mean values are similar in SURFEX than the observation in the 30 year period, apart from summer, when SURFEX overestimates the interannual variability of the mean temperature (Table 5).    The seasonal distribution of daily temperatures is presented on histograms (Figure 9). In spring and winter the distributions of the simulated and observed temperature are in good agreement. In summer, the systematic overestimation is clear: the histogram based on SURFEX outputs has the same shape as the one based on the observations, but it is shifted towards higher values. The largest inconsistency is in autumn, when the model values show bimodal distribution, which (i.e., the lower occurrence of the values between the most frequent values) is not visible in the observation. The underestimation of the mean values can be explained with the higher frequency of higher temperature values. The seasonal distribution of daily temperatures is presented on histograms 9). In spring and winter the distributions of the simulated and observed tempera in good agreement. In summer, the systematic overestimation is clear: the hi based on SURFEX outputs has the same shape as the one based on the observati it is shifted towards higher values. The largest inconsistency is in autumn, when th values show bimodal distribution, which (i.e., the lower occurrence of the values the most frequent values) is not visible in the observation. The underestimatio mean values can be explained with the higher frequency of higher temperature v The diurnal evolution of 2 m temperature in the suburban area is best simu spring and winter (Figure 10), especially at 0 and 12 UTC. However, in the mornin the temperature is slightly underestimated, while in winter in the afternoon hou estimation occurs too. In summer in addition to the general overestimation, the t ture after 15 UTC does not drop as fast as in the observation. Moreover, the low perature is simulated to occur later at 6 UTC instead of 3 UTC (note, that data w able in 3 h frequency for both the observation and the simulation) and it increas in the morning compared to observations. In autumn a smaller overestimation is d from noon to 6 UTC. The diurnal evolution of 2 m temperature in the suburban area is best simulated in spring and winter (Figure 10), especially at 0 and 12 UTC. However, in the morning hours the temperature is slightly underestimated, while in winter in the afternoon hours overestimation occurs too. In summer in addition to the general overestimation, the temperature after 15 UTC does not drop as fast as in the observation. Moreover, the lowest temperature is simulated to occur later at 6 UTC instead of 3 UTC (note, that data was available in 3 h frequency for both the observation and the simulation) and it increases faster in the morning compared to observations. In autumn a smaller overestimation is detected from noon to 6 UTC.

Comparison of SURFEX-EI and SURFEX-ARP for Simulating the UHI
Finally, the two experiments are compared on a common period (1996-2005) from the aspect of simulating UHI. Figure 11 presents the seasonal mean UHI at 0 UTC i SURFEX-EI and SURFEX-ARP. The differences are not large between the two simulation however the GCM driven model gives stronger temperature surplus (with 0.1-0.2 °C) ove the city than its reanalysis driven counterpart. Note that the differences for the 2-m tem perature are much higher (e.g., the 2 m temperature bias is 1.1 °C and 2.5 °C for SURFEX EI and SURFEX-ARP, respectively, compared to CarpatClim-HU over Budapest; no shown), but the subtraction in UHI eliminates the great majority of the systematic bias.
The reasons behind the stronger UHI in SURFEX-ARP can be explained by compa ing the surface energy budget components ( Figure 12). The ground heat flux over the u ban surfaces is larger in SURFEX-ARP in the morning hours, while it is smaller during th night than in SURFEX-EI in all seasons except winter. Therefore, in the GCM driven ex periment the city absorbs more heat during the day and emits more during the night tha explains the stronger UHI. Moreover, less sensible heat and more latent heat are emitte in SURFEX-ARP than SURFEX-EI.

Comparison of SURFEX-EI and SURFEX-ARP for Simulating the UHI
Finally, the two experiments are compared on a common period (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005) from the aspect of simulating UHI. Figure 11 presents the seasonal mean UHI at 0 UTC in SURFEX-EI and SURFEX-ARP. The differences are not large between the two simulations, however the GCM driven model gives stronger temperature surplus (with 0.1-0.2 • C) over the city than its reanalysis driven counterpart. Note that the differences for the 2-m temperature are much higher (e.g., the 2 m temperature bias is 1.1 • C and 2.5 • C for SURFEX-EI and SURFEX-ARP, respectively, compared to CarpatClim-HU over Budapest; not shown), but the subtraction in UHI eliminates the great majority of the systematic bias.
The reasons behind the stronger UHI in SURFEX-ARP can be explained by comparing the surface energy budget components ( Figure 12). The ground heat flux over the urban surfaces is larger in SURFEX-ARP in the morning hours, while it is smaller during the night than in SURFEX-EI in all seasons except winter. Therefore, in the GCM driven experiment the city absorbs more heat during the day and emits more during the night that explains the stronger UHI. Moreover, less sensible heat and more latent heat are emitted in SURFEX-ARP than SURFEX-EI.

Discussion
The first part of this study focuses on the spatial and temporal characteristics of surface temperature and SUHI simulated by SURFEX-EI, considering the MODIS LST product as reference. The seasonal mean surface temperature is substantially warmer over the whole city than the ambient rural areas from spring to autumn, resulting in large SUHI extent and magnitude both at day and night. In contrast, MODIS shows that the highest temperature surplus of the city concentrates in the core of the city at night and in the Pest side (to the east of the Danube) by day. The less realistic variability and extent of SUHI in SURFEX can be explained partly by the lack of modelling the full boundary layer and its dynamics. Moreover, in the suburbs the canyon concept is less representative for the family houses than in the downtown area of the city. To solve this problem, the Local Climate Zone approach has been developed that classifies the urban areas based on compactness and height of buildings [48].
On the Buda side (west of Danube), the city is hilly and very green. Studies demonstrate that implementing trees in the urban canopy scheme can reduce surface temperature overestimation with several degrees [49,50]. These suggests that the strong summer and autumn surface temperature overestimation over the city (8.3 • C and 5.4 • C, respectively over the urbanized gridcells) and especially over Buda and the suburbs may be significantly lowered with a more explicit modelling of street trees' effect.
The best agreement between the model and observation was found in winter, when surface temperature bias remains below 1 • C for urban and rural areas both at day and night. The small biases may be explained by that we only considered cloudy-free cases and in winter there is no leafy vegetation.
Upon SURFEX limitations, it is worth noting that satellite LST products are burdened with some considerable uncertainties [19,51] as well. For example, LST observations are valid only during clear-sky conditions, that limits the meteorological conditions to be potentially studied. In addition, satellite cloud-screening algorithms are often imperfect, causing biases in LST estimation [19]. Moreover, large viewing angles enhance atmospheric radiation extinction that can result in biases in observed LST [19,51,52]. Lastly, the anisotropy of LST (caused by the high level of surface heterogeneity particularly in urban areas) modifies the viewing angle of the satellite sensor above certain objects, hence leads to the uncertainty of observation as well [19]. Finally, it must be also mentioned, that since SURFEX results are saved 3-hourly, there is a 30-45 min difference between the compared simulation and the Aqua satellite's observation time, that may contribute to a certain proportion of bias, especially during the day. Despite these limitations, the satellite observations provide important reference to assess the model's ability to simulate spatial characteristics of urban temperature.
After the surface temperature and SUHI analysis in SURFEX-EI on shorter timescale, the 2 m temperature was studied in SURFEX-ARP for 30 years. The behavior of the driving ALADIN model strongly affects the results of SURFEX, i.e., the strong positive summer bias is dominantly caused by the 2.9 • C overestimation of ALADIN-Climate over the model domain. According to [49] the more explicit parameterization of street trees does not improve significantly the 2 m temperature overestimation of the LSM.
It was also found, that SURFEX heats not only the urbanized gridpoints, but in autumn and in winter the lower elevated rural areas as well. The surface scheme of ALADIN is also SURFEX version 5, but TEB and the SBL scheme was not activated in the climate change runs. Instead, the screen level values were computed by the Geleyn diagnostic formula [38]. One advantage of the SBL scheme is that it does not let the 2 m temperature cool too much in very stable situations, in order to prevent the decoupling of the surface model and the atmospheric component [31]. It was found in short term sensitivity studies that the minimum temperatures computed with the SBL scheme are higher compared to the diagnostic formula (not shown).
Apart from summer, the ALADIN-Climate is too cold compared to CarpatClim-HU. The majority of this cold bias was eliminated by SURFEX resulting smaller underestimation in spring and winter (less than 1.5 • C) and small positive bias in autumn (less than 1 • C) over the domain. The validation in the suburban gridpoint suggests that SURFEX is able to sufficiently describe the seasonal distribution of the daily mean temperature and the seasonal mean diurnal cycle. However, it must be mentioned that these investigations were achieved only in one gridpoint and more long term station measurement is needed to justify our results.
The limitations of ECOCLIMAP-i.e., it is a static database based on land cover data from the mid-90s-can also explain parts of the mean 2-m temperature bias of SURFEX-ARP. The importance of built-up fraction was studied in short sensitivity tests and was found that the bias originated from the driving model is more important. e.g., if the temperature is overestimated, reducing the urban fraction in the gridcell could moderate the bias; however, in other periods when the model is too cold, a smaller built-up ratio caused larger underestimation.
Comparing SURFEX-EI and SURFEX-ARP from the aspect of seasonal mean diurnal cycle of UHI, it can be stated that the systematic bias of the land surface model (and therefore the influence of the driving model) largely diminishes in the UHI.

Conclusions
The aim of this study was a thorough validation of the SURFEX LSM from the aspect of simulating surface and 2 m temperature and the corresponding UHI over Budapest. SURFEX was first forced by the ERA-Interim re-analysis driven ALADIN-Climate regional climate model, in order to reveal how the ALADIN-Climate model itself influences the LSM. Due to the strong correlation with observations, only a shorter 10-year past period was modelled, and the results were compared to the MODIS satellite LST observations for 2003-2005. It was found that the spatial extent of warmer area and even the location of heat island much more varies seasonally and between day and night in MODIS than in SURFEX. The surface temperature of SURFEX is much warmer over all the urban gridpoints than the measurement, resulting in larger SUHI extent from spring to autumn both at day and night. It was statistically underpinned that the daily variability of SUHI affected area is smaller than observed. SURFEX was also forced by the GCM driven ALADIN-Climate model in the period of 1960-2005. The RCM was evaluated against the CarpatClim-Hu gridded observational dataset over the SUREX domain and 2 m temperature of SURFEX in one suburban gridpoint was validated with respect to station measurement for the period of 1971-2000. The main goal of this research was to assess the capability of SURFEX to simulate urban climate on decadal timescale. It was found that over the Budapest domain the ARPEGE driven ALADIN-Climate largely overestimates the temperature in summer, while underestimates it in the other seasons. This summer positive bias is inherited in SURFEX as well, leading to 3.5 • C warmer summer temperature compared to station measurement. In other seasons the difference between the land surface model and observation does not exceed 1.5 • C. The daily cycle of 2 m temperature is reasonably simulated and only smaller differences were found in each season except the systematic bias of summer. The long term gridpoint validation results of SURFEX are very promising, although to get more robust information a dense measurement network would be important.
Finally, the two experiments were compared on a common 10-year period from the aspect of simulating UHI. The spatial distribution and magnitude of UHI does not differ too much, because the majority of the systematic model bias is eliminated in the UHI calculation. However, small differences were found in the magnitude of the ground heat flux, that explains the small differences in the temperature excess over the city.  Data Availability Statement: The ALADIN and SURFEX simulation data used in this study are stored locally at the Hungarian Meteorological Service and can be provided upon request. The CarpatClim-Hu dataset and the station measurements are available at http://odp.met.hu. MODIS Aqua, Terra and IGBP products were retrieved from NASA's Land Processes Distributed Active Archive Center (https://lpdaac.usgs.gov/).

Acknowledgments:
The ALADIN-Climate simulations were performed within the framework of the RCMGiS (EEA-C13-10) project. We acknowledge the use of MODIS Aqua and Terra land surface temperature imagery from NASA's Land Processes Distributed Active Archive Center. We wish to say thanks to the Climate Division of the Hungarian Meteorological Service for providing the CarpatClim-Hu and station measurements. Great thanks are also dedicated for Gabriella Szépszó, who improved this paper with useful recommendations. Last but not least, the authors would like to express their gratitude to the three anonym reviewers for their detailed and valuable reviews.