Characteristics of LDAPS-Predicted Surface Wind Speed and Temperature at Automated Weather Stations with Di ﬀ erent Surrounding Land Cover and Topography in Korea

: We investigated the characteristics of surface wind speeds and temperatures predicted by the local data assimilation and prediction system (LDAPS) operated by the Korean Meteorological Administration. First, we classiﬁed automated weather stations (AWSs) into four categories (urban ﬂat (Uf), rural ﬂat (Rf), rural mountainous (Rm), and rural coastal (Rc) terrains) based on the surrounding land cover and topography, and selected 25 AWSs representing each category. Then we calculated the mean bias error of wind speed (WE) and temperature (TE) using AWS observations and LDAPS predictions for the 25 AWSs in each category for a period of 1 year (January–December 2015). We found that LDAPS overestimated wind speed (average WE = 1.26 m s − 1 ) and underestimated temperature (average TE = − 0.63 ◦ C) at Uf AWSs located on ﬂat terrain in urban areas because it failed to reﬂect the drag and local heating caused by buildings. At Rf, located on ﬂat terrain in rural areas, LDAPS showed the best performance in predicting surface wind speed and temperature (average WE = 0.42 m s − 1 , average TE = 0.12 ◦ C). In mountainous rural terrain (Rm), WE and TE were strongly correlated with di ﬀ erences between LDAPS and actual altitude. LDAPS underestimated (overestimated) wind speed (temperature) for LDAPS altitudes that were lower than actual altitude, and vice versa. In rural coastal terrain (Rc), LDAPS temperature predictions depended on whether the grid was on land or sea, whereas wind speed did not depend on grid location. LDAPS underestimated temperature at grid points on the sea, with smaller TE obtained for grid points on sea than on land.


Introduction
Weather directly and indirectly influences daily life and economic activity. Severe weather can cause disasters that lead to loss of human life and property [1][2][3][4]. Accurate and precise weather prediction can help mitigate such disasters and provide useful information for socioeconomic and cultural fields including agriculture, industry, transportation, tourism, and leisure [5][6][7]. In response to demands from the industrial sector [8,9], the Korean Meteorological Administration (KMA) has been working to supply customized weather information, such as real-time, on-site weather forecasting to enhance the success of the 2018 Pyeong-Chang Winter Olympics.
Previous studies in Korea have improved the accuracy of numerical weather prediction models. Jeong and Lee [10] contributed to enhancing the medium-range prediction of the weather research and forecasting (WRF) model by correcting bias errors. Lee et al. [11] decreased the negative bias of surface wind speeds of WRF by parameterizing a subgrid-scale orographic effect over mountainous terrain. Kwun et al. [12] analyzed the difference in the planetary boundary layer (PBL) scheme between the weather research and forecasting (WRF) model and the fifth-generation Pennsylvania State University-National Center for Atmospheric Research Mesoscale Model (PSU-NCAR MM5) to the optimal prediction of coastal and sea surface wind speeds. KMA conducts weather forecasting using various numerical weather prediction models to provide rapid, high-quality weather information. KMA developed the local data assimilation and prediction system (LDAPS) based on the unified model (UM) designed by the UK Met Office. To prepare for weather disasters caused by local severe weather events over the entire Korean peninsula, LDAPS employs a high-resolution grid system with 1.5 km horizontal resolution and 70 vertical layers [13]. However, the spatial and temporal resolution of LDAPS is insufficient to resolve small obstacles such as buildings and hilly terrain, which cause external forcing in urban-or smaller-scale flows [14]. Recently, many studies are trying to predict detailed flows and dispersion within urban canopy by using computational fluid dynamics models coupled to numerical weather prediction models (e.g., LDAPS) [15][16][17]. LDAPS enables smaller-scale numerical simulations within the atmospheric boundary layer by providing realistic initial and boundary conditions for numerical models with computational fluid dynamics models. To apply LDAPS to multi-scale numerical simulations, we need to grasp the characteristics of LDAPS-predicted surface winds and temperatures for the application studies of LDAPS predictions.
Some studies have analyzed the characteristics of LDAPS prediction results and have improved its prediction performance. Kang et al. [18] compared the air temperature, wind speed, and relative humidity observed at the Daegu and Gumi meteorological stations for 7 days with those predicted by LDAPS. Their building-scale resolved air temperature model improved the performance of LDAPS air temperature prediction by reflecting the heating effect in urban areas [19]. Other previous studies have contributed to our understanding of LDAPS prediction characteristics [20][21][22][23][24]. However, the extensive numerical experiments including sensitivity tests to spatial and temporal resolutions and physical parameterization schemes have not distinctly explained the reasons for the spatial variability in LDAPS biases for surface wind speeds and temperatures over the Korean Peninsula yet. One possible thing is that a land cover type and terrain undulation are significantly different region by region in Korea.
We investigated the characteristics of surface wind speeds and temperatures predicted by LDAPS at automatic weather stations (AWSs) with different surrounding land cover and topography. First, we classified 100 AWSs into four categories based on land cover (urban and rural areas) and topography (mountainous, coastal, and flat terrain). For each category, we identified the characteristics of wind speeds and air temperatures predicted by LDAPS, by comparing them with data collected during a 1-year period (January-December 2015). For comparison, we interpolated LDAPS-predicted wind speeds and temperatures to the same observation heights above the ground surface.

AWS Classification
First, we followed previously described classification methods [25][26][27] to characterize the region (1.5 × 1.5 km 2 , to match the horizontal LDAPS grid) surrounding each of 100 AWSs using land cover and topography data obtained from 1:25,000 land cover maps provided by the environmental geographic information service (EGIS) of the Ministry of Environment, Korea ( Figure 1). For example, if the region surrounding an AWS had a ratio of urban (U) to rural (R) land cover of >50%, then that AWS was classified as U; otherwise, it was classified as rural R [25]. Topography was classified based on digital maps created by the National Geographic Information Institute of Korea in 2015. An AWS with altitude of >200 m in the surrounding area was classified as mountainous (m) [26]. An AWS located within Atmosphere 2020, 11, 1224 3 of 16 horizontal distance of 500 m of a coastline was classified as coastal (c) [27]. Any AWS not classified as m or c was considered on flat terrain (f). The combination of both classification schemes yielded six categories: Um, Uc, Uf, Rm, Rc, and Rf. However, preliminary classification showed that AWSs in the Um and Uc categories were very rare in Korea. Therefore, in this study, we selected 25 AWSs ( Figure 2) for each of the four remaining categories (Uf, Rm, Rc, and Rf) to identify the characteristics of LDAPS-predicted wind speeds and temperatures. Dynamical and biophysical effects of land covers are not considered in the analysis.
Atmosphere 2020, 11, x FOR PEER REVIEW 3 of 16 with altitude of >200 m in the surrounding area was classified as mountainous (m) [26]. An AWS located within horizontal distance of 500 m of a coastline was classified as coastal (c) [27]. Any AWS not classified as m or c was considered on flat terrain (f). The combination of both classification schemes yielded six categories: Um, Uc, Uf, Rm, Rc, and Rf. However, preliminary classification showed that AWSs in the Um and Uc categories were very rare in Korea. Therefore, in this study, we selected 25 AWSs ( Figure 2) for each of the four remaining categories (Uf, Rm, Rc, and Rf) to identify the characteristics of LDAPS-predicted wind speeds and temperatures. Dynamical and biophysical effects of land covers are not considered in the analysis.    with altitude of >200 m in the surrounding area was classified as mountainous (m) [26]. An AWS located within horizontal distance of 500 m of a coastline was classified as coastal (c) [27]. Any AWS not classified as m or c was considered on flat terrain (f). The combination of both classification schemes yielded six categories: Um, Uc, Uf, Rm, Rc, and Rf. However, preliminary classification showed that AWSs in the Um and Uc categories were very rare in Korea. Therefore, in this study, we selected 25 AWSs ( Figure 2) for each of the four remaining categories (Uf, Rm, Rc, and Rf) to identify the characteristics of LDAPS-predicted wind speeds and temperatures. Dynamical and biophysical effects of land covers are not considered in the analysis.

LDAPS Data
We used the surface wind speeds and temperatures predicted by LDAPS from January 1 to December 31, 2015. LDAPS employs a self-analysis and prediction circulation system based on a three-dimensional variational data assimilation technique [19]. LDAPS performs 36 h predictions at 00, 06, 12, and 18 UTC for the prediction and data assimilation purposes and 3 h predictions at 03, 09, 15, and 21 UTC for the data assimilation only. LDPAS assimilates various surface and upper air observation datasets after the quality control process of LDAPS automatically selects datasets that will be included and excluded in the assimilation process. The AWS observations selected in the assimilation process may hinder the data independence in the analysis. However, most of the AWS observations are known to be excluded in the assimilation process. To consider LDAPS data expected to have minimal errors, we used the initial data of LDAPS at each operation time (that is, every 3 h from 00 UTC to 21UTC) and 1 and 2 h prediction data at the rest times (that is, 01, 02, 04, 05, . . . , 22, and 23 UTC). Table 1 summarizes the LDAPS configuration; further details are provided elsewhere [28,29]. The horizontal and vertical grid systems are based on the Arakawa C-grid and Charney-Phillips staggered grid, respectively [30]. LDAPS employs New Dynamics [31] as a dynamics core, JULES (the Joint UK Land Environment Simulator) as a land surface process, mixed-phase scheme [32] for microphysics, and Edward-Slingo general 2-stream scheme [33] for the radiative process. For comparison to the observations, we used LDAPS-predicted wind speeds and temperatures extracted to the same observation heights above the ground surface, even taking the building heights into account in cases of the AWSs installed above the building roofs. LDAPS does not have any grid cell including two or more AWSs. Meteorological variables predicted at a grid point by LDAPS represent those averaged over the 1.5 × 1.5 km 2 area around the grid point. On the other hand, the observations at an AWS give meteorological information at one point, not reflecting the effects of the observation environment in the area as wide as the grid cell of LDAPS. Despite the difference in data representativeness, we conducted a point-based comparison of LDAPS predictions to the AWS observations due to the evaluation data availability. Hence, our analysis has the limitation of the nonequivalent representativeness of data used in the comparison.

Results and Discussion
We calculated the mean bias error (MBE) of LDAPS prediction results for the 25 AWSs in each category during the study period and compared their averages. For details of the AWSs considered in this study, see Table A1 in Appendix A. Figure 3 shows boxplots of the MBE of wind speed (WE) and temperature (TE) for the four categories. In all categories, LDAPS generally overestimated the measured wind speed. Average WE was highest in Uf (1.26 m s −1 ) and lowest in Rf (0.42 m s −1 ). Rm showed the widest variation in WE (−2.38 to 2.68 m s −1 ), but a lower average (0.57 m s −1 ) than Rc, which had an average WE of 1.06 m s −1 (standard deviation (SD) = 0.52 m s −1 ). LDAPS overestimated surface wind speeds at all AWSs ( Figure 3a).

Results and Discussion
We calculated the mean bias error (MBE) of LDAPS prediction results for the 25 AWSs in each category during the study period and compared their averages. For details of the AWSs considered in this study, see Table A1 in Appendix A. Figure 3 shows boxplots of the MBE of wind speed (WE) and temperature (TE) for the four categories. In all categories, LDAPS generally overestimated the measured wind speed. Average WE was highest in Uf (1.26 m s −1 ) and lowest in Rf (0.42 m s −1 ). Rm showed the widest variation in WE (−2.38 to 2.68 m s −1 ), but a lower average (0.57 m s −1 ) than Rc, which had an average WE of 1.06 m s −1 (standard deviation [SD] = 0.52 m s −1 ). LDAPS overestimated surface wind speeds at all AWSs (Figure 3a).
At Uf AWSs, TE showed the opposite pattern to WE (Figure 3b). LDAPS underestimated temperatures at most Uf AWSs, with an average TE of −0.63 °C, which was the most significant difference among the four categories. At Rf AWSs, there was no distinct pattern in TE; LDAPS overestimated the surface temperature at some AWSs and underestimated it at others. However, as observed for average WE values, the average TE (0.12 °C) and its SD (0.46 °C) were lower than those of the other categories. At Rm AWSs, the variation in TE was highest (1.56 °C), but its average was not very high (0.41 °C). The lowest average TE (−0.12 °C) and SD (0.31 °C) were observed at Rc AWSs. Next, we analyzed the LDAPS prediction characteristics in detail, using the relationships between MBE and the land cover and topography surrounding the AWSs in each category.  At Uf AWSs, TE showed the opposite pattern to WE (Figure 3b). LDAPS underestimated temperatures at most Uf AWSs, with an average TE of −0.63 • C, which was the most significant difference among the four categories. At Rf AWSs, there was no distinct pattern in TE; LDAPS overestimated the surface temperature at some AWSs and underestimated it at others. However, as observed for average WE values, the average TE (0.12 • C) and its SD (0.46 • C) were lower than those of the other categories. At Rm AWSs, the variation in TE was highest (1.56 • C), but its average was not very high (0.41 • C). The lowest average TE (−0.12 • C) and SD (0.31 • C) were observed at Rc AWSs.
Next, we analyzed the LDAPS prediction characteristics in detail, using the relationships between MBE and the land cover and topography surrounding the AWSs in each category. Figure 4 shows WE and TE for Uf AWSs. Numerical weather prediction models are well known to overestimate the roughness lengths and surface wind speeds [34][35][36] in urban areas because most of them did not sufficiently reflect the effects of roughness such as buildings. LDAPS overestimated wind speeds measured at a height of 10 m and underestimated temperatures measured at a height of 1.5 m at most Uf AWSs ( Figure 4). Notably, the TE averages at AWSs 627 and 417 were slightly greater than zero. LDAPS employs the Joint UK Land Environmental Simulator (JULES) as a land surface model; using the mosaic method within a single grid, JULES classifies land cover into nine types and applies an urban parametrization weighted by the average of each type [37]. However, the limited resolution of LDAPS prevents it from accurately reflecting flow changes caused by buildings and differential heating among land use types in urban areas [38]. As most Uf AWSs were located on building roofs, LDAPS overestimated wind speeds and underestimated temperatures (Figure 4). Correlations between AWS measurements and LDAPS predictions were slightly better in summer (June, July, and August) than winter months (December, January, and February) ( Figure 5).
Atmosphere 2020, 11, x FOR PEER REVIEW 6 of 16 shows WE and TE for Uf AWSs. Numerical weather prediction models are well known to overestimate the roughness lengths and surface wind speeds [34][35][36] in urban areas because most of them did not sufficiently reflect the effects of roughness such as buildings. LDAPS overestimated wind speeds measured at a height of 10 m and underestimated temperatures measured at a height of 1.5 m at most Uf AWSs ( Figure 4). Notably, the TE averages at AWSs 627 and 417 were slightly greater than zero. LDAPS employs the Joint UK Land Environmental Simulator (JULES) as a land surface model; using the mosaic method within a single grid, JULES classifies land cover into nine types and applies an urban parametrization weighted by the average of each type [37]. However, the limited resolution of LDAPS prevents it from accurately reflecting flow changes caused by buildings and differential heating among land use types in urban areas [38]. As most Uf AWSs were located on building roofs, LDAPS overestimated wind speeds and underestimated temperatures ( Figure 4). Correlations between AWS measurements and LDAPS predictions were slightly better in summer (June, July, and August) than winter months (December, January, and February) ( Figure 5). LDAPS overestimated wind speeds at Rf AWSs except for AWS 706, and underestimated temperatures at nine AWSs ( Figure 6). The WE (<1 m s −1 ) and TE (<1 °C) of Rf AWSs were consistently lower than those of Uf AWSs, with averages of 0.43 m s −1 and 0.12 °C, respectively ( Figure 3). This result was expected because there were few buildings to create drag and act as heat sources; thus, LDAPS performed well in predicting wind speed and temperature at Rf AWSs ( Figure 3). Note that LDAPS slightly overestimated wind speed except at AWS 706, and negative TE was observed only when LDAPS altitude was higher than the actual altitude, at AWSs 321, 623, 708, 825, 829, 887, 900, 932, and 946. At AWSs 321, 900, and 946, LDAPS altitude was much higher (>170 m) than the actual altitude, and WE and TE showed similar characteristics to those of the Rm AWSs described below, i.e., wind speed overestimation and temperature underestimation. The ratio of urban areas was close to 50% at AWSs 708 (38%), 829 (46%), and 932 (48%), resulting in negative TE. AWS-measured and LDAPS-predicted temperatures were very strongly correlated, with R values of 0.96 in summer and 0.97 in winter, at AWS 615, which was located in a flat rural area, and less strongly correlated, with R values of 0.92 in summer and 0.81 in winter, at AWS 900, which was located in a basin area with a range of altitudes (Figure 7). The measured and predicted wind speeds were weakly correlated at both AWS 615 (R = 0.69 in summer and 0.71 in winter) and AWS 900 (R = 0.76 in summer and 0.69 in winter), indicating that LDAPS performed better in predicting wind speeds at Rf AWSs than at Uf AWSs (Figures 5b,d and 7b,d). Further investigation showed that the LDAPS altitude was higher than the actual altitude at AWS 900, resulting in wind speed overestimation and more frequent temperature underestimation.  LDAPS overestimated wind speeds at Rf AWSs except for AWS 706, and underestimated temperatures at nine AWSs ( Figure 6). The WE (<1 m s −1 ) and TE (<1 • C) of Rf AWSs were consistently lower than those of Uf AWSs, with averages of 0.43 m s −1 and 0.12 • C, respectively (Figure 3). This result was expected because there were few buildings to create drag and act as heat sources; thus, LDAPS performed well in predicting wind speed and temperature at Rf AWSs (Figure 3). Note that LDAPS slightly overestimated wind speed except at AWS 706, and negative TE was observed only when LDAPS altitude was higher than the actual altitude, at AWSs 321, 623, 708, 825, 829, 887, 900, 932, and 946. At AWSs 321, 900, and 946, LDAPS altitude was much higher (>170 m) than the actual altitude, and WE and TE showed similar characteristics to those of the Rm AWSs described below, i.e., wind speed overestimation and temperature underestimation. The ratio of urban areas was close to 50% at AWSs 708 (38%), 829 (46%), and 932 (48%), resulting in negative TE. AWS-measured and LDAPS-predicted temperatures were very strongly correlated, with R values of 0.96 in summer and 0.97 in winter, at AWS 615, which was located in a flat rural area, and less strongly correlated, with R values of 0.92 in summer and 0.81 in winter, at AWS 900, which was located in a basin area with a range of altitudes (Figure 7). The measured and predicted wind speeds were weakly correlated at both AWS 615 (R = 0.69 in summer and 0.71 in winter) and AWS 900 (R = 0.76 in summer and 0.69 in winter), indicating that LDAPS performed better in predicting wind speeds at Rf AWSs than at Uf AWSs (Figure 5b,d and Figure 7b,d). Further investigation showed that the LDAPS altitude was higher than the actual altitude at AWS 900, resulting in wind speed overestimation and more frequent temperature underestimation.
The variation in WE and TE was greatest at Rm AWSs among the four categories (Figure 3). At Rm AWSs, WE generally had the opposite sign to TE. WE (TE) increased (decreased) as the difference between the LDAPS and actual altitude increased (Figure 8) because LDAPS inevitably smoothed the steep terrain to avoid model blowup. Thus, for AWSs on mountain peaks or ridges, the LDAPS altitude was generally lower than the actual altitude. LDAPS resultantly underestimated (overestimated) wind speed (temperature), which is considered as a fundamental problem of numerical models for mountainous terrain [39]. By contrast, for AWSs in valleys or basins, the LDAPS altitude was generally higher than actual altitude, leading to overestimation (underestimation) of wind speed (temperature). Thus, MBE distributions distinctly reflected systematic LDAPS errors caused by smoothing ( Figure 8). TE had a monotonic relationship with the altitude difference, whereas WE had a more complicated relationship, due to local circulation patterns in mountains and valleys around AWSs that were not resolved by LDAPS [40,41]. Further studies should perform more detailed analyses of these relationships. Figure 9 shows scatterplots of wind speed and temperature at AWSs 316 and AWS 872. Temperature (wind speed) was distinctly overestimated (underestimated) at AWS 316, where the LDAPS altitude was lower than the AWS altitude, and underestimated (overestimated) at AWS 872, where the LDAPS altitude was higher than the AWS altitude. At AWS 316, predicted and observed temperatures were strongly correlated (R = 0.90 in summer and 0.94 in winter). However, predicted and observed wind speeds were very poorly correlated (≤0.32), showing apparent underestimation. The correlation between measured and predicted temperatures (wind speeds) at AWS 872 was weaker (stronger) than that at AWS 316, showing apparent underestimation (overestimation).
The average WE of Rc AWSs was comparable to that of Uf AWSs (Figure 3). At Rc AWSs, wind speed was overestimated, regardless of whether the LDAPS grid points were on sea or land ( Figure 10). The highest WE occurred at AWS 301, which was located on the sea. The average WE for AWSs near the West Sea (AWSs 300, 301, 606, 607, 631, 657, 662, 663, 697, 700, and 881) and East Sea (AWSs 310, 524, 661, 671, 800, 852, 901, 923, 924, 949, and 954) were 1.28 and 0.83 m s -1 , respectively. The coastline of the Korean Peninsula is more complicated near the West Sea than the East Sea. However, LDAPS simplified the coastline to a greater extent near the West Sea than the East Sea, resulting in greater overestimation of wind speed. LDAPS tended to underestimate temperature when the LDAPS grid points of the AWS were on the sea (right panel, Figure 10). TE and its variation were lower for AWSs on the sea than for those on land (left panel, Figure 10). We further analyzed Atmosphere 2020, 11, 1224 8 of 16 monthly averages of WE and TE. The average monthly WE and SD tended to increase in winter and decrease in summer ( Figure 11). In spring and summer, LDAPS underestimated temperatures at AWSs with LDAPS grid points on the sea (Figure 11a), where lower sea temperatures induced negative TE. Higher sea temperatures resulted in positive TE in the fall and winter. However, at AWSs with LDAPS grid points on land, LDAPS generally slightly underestimated temperature, regardless of the season (Figure 11b). The AWS-measured and LDAPS-predicted wind speeds and temperatures were more strongly correlated in winter than in summer ( Figure 12). Wind speed correlations at Rc AWSs were highest among the four categories.   The variation in WE and TE was greatest at Rm AWSs among the four categories (Figure 3). At Rm AWSs, WE generally had the opposite sign to TE. WE (TE) increased (decreased) as the difference between the LDAPS and actual altitude increased (Figure 8) because LDAPS inevitably smoothed the steep terrain to avoid model blowup. Thus, for AWSs on mountain peaks or ridges, the LDAPS altitude was generally lower than the actual altitude. LDAPS resultantly underestimated (overestimated) wind speed (temperature), which is considered as a fundamental problem of  The average WE of Rc AWSs was comparable to that of Uf AWSs (Figure 3). At Rc AWSs, wind speed was overestimated, regardless of whether the LDAPS grid points were on sea or land ( Figure 10). The highest WE occurred at AWS 301, which was located on the sea. The average WE for AWSs near the West Sea (AWSs 300, 301, 606, 607, 631, 657, 662, 663, 697, 700, and 881) and East Sea he average WE of Rc AWSs was comparable to that of Uf AWSs (Figure 3). At Rc AWSs, w AWSs with LDAPS grid points on land, LDAPS generally slightly underestimated temperature, regardless of the season (Figure 11b). The AWS-measured and LDAPS-predicted wind speeds and temperatures were more strongly correlated in winter than in summer ( Figure 12). Wind speed correlations at Rc AWSs were highest among the four categories.  AWSs with LDAPS grid points on land, LDAPS generally slightly underestimated temperature, regardless of the season (Figure 11b). The AWS-measured and LDAPS-predicted wind speeds and temperatures were more strongly correlated in winter than in summer ( Figure 12). Wind speed correlations at Rc AWSs were highest among the four categories.   Next, we analyzed WE and TE at 25 AWSs within in each category. Based on this analysis, we summarized the characteristics of wind speeds and temperatures predicted by LDAPS for Uf, Rf, Rm, and Rc (Table 2). At Uf AWSs, the average WE was >0.63 m s −1 , and the average TE was <0.04 °C. At Rf AWSs, the average WE ranged from −0.12 to 0.72 m s −1 . The average TE was dependent on the altitude difference between LDAPS and the actual terrain, with values > 0.24 °C when LDAPS altitude was lower than actual altitude and values ranging from −0.68 °C to 0.89 °C when LDAPS altitude was higher than actual altitude. At Rm AWSs, WE and TE was also dependent on the difference between LDAPS altitude and actual altitude, because LDAPS smoothed steep terrains in mountains and valleys. Jiménez and Dudhia [42] reported that the smoothed terrain in a numerical weather prediction model causes errors in predicting wind speeds by failing to capture momentum fluxes from unresolved topography. When the altitude difference was less (more) than 400 m, the average WE was lower (higher) than −0.92 (0.32) m s −1 except at one AWS (554). The average TE was higher (lower) than 0.14(−0.15) °C at altitude differences of less (more) than 150 m except at one AWS (831). At Rc AWSs, average WE and TE values were influenced by whether the LDAPS grid was on sea or land. At AWSs corresponding to LDAPS grid points on the sea, the average WE was >0.04 m s −1 , and the average TE varied seasonally, with values >0.19 °C in fall and winter and <0.02 °C in spring and summer. At AWSs corresponding to LDAPS grid points on land, the average WE was >0.14 m s −1 , and the average TE ranged from −0.56 to 0.46 °C, similar to those of Rf AWSs. Next, we analyzed WE and TE at 25 AWSs within in each category. Based on this analysis, we summarized the characteristics of wind speeds and temperatures predicted by LDAPS for Uf, Rf, Rm, and Rc (Table 2). At Uf AWSs, the average WE was >0.63 m s −1 , and the average TE was <0.04 • C. At Rf AWSs, the average WE ranged from −0.12 to 0.72 m s −1 . The average TE was dependent on the altitude difference between LDAPS and the actual terrain, with values > 0.24 • C when LDAPS altitude was lower than actual altitude and values ranging from −0.68 • C to 0.89 • C when LDAPS altitude was higher than actual altitude. At Rm AWSs, WE and TE was also dependent on the difference between LDAPS altitude and actual altitude, because LDAPS smoothed steep terrains in mountains and valleys. Jiménez and Dudhia [42] reported that the smoothed terrain in a numerical weather prediction model causes errors in predicting wind speeds by failing to capture momentum fluxes from unresolved topography. When the altitude difference was less (more) than 400 m, the average WE was lower (higher) than −0.92 (0.32) m s −1 except at one AWS (554). The average TE was higher (lower) than 0.14(−0.15) • C at altitude differences of less (more) than 150 m except at one AWS (831). At Rc AWSs, average WE and TE values were influenced by whether the LDAPS grid was on sea or land. At AWSs corresponding to LDAPS grid points on the sea, the average WE was >0.04 m s −1 , and the average TE varied seasonally, with values >0.19 • C in fall and winter and <0.02 • C in spring and summer. At AWSs corresponding to LDAPS grid points on land, the average WE was >0.14 m s −1 , and the average TE ranged from −0.56 to 0.46 • C, similar to those of Rf AWSs.

Summary and Conclusions
We analyzed the characteristics of surface wind speeds and temperatures predicted by LDAPS for 100 AWSs classified into four categories based on surrounding land covers and topography.
At Uf AWSs, LDAPS did not sufficiently reflect the influence of building drag and surface heating due to limited grid resolution and the applied coordinate system. At Uf AWSs, it generally overestimated wind speed (average WE = 1.26 m s −1 ) and underestimated temperature (average TE = −0.63 • C). At Rf AWSs, it performed best in predicting wind speed and temperature (average WE = 0.42 m s −1 , average TE = 0.12 • C), because there were few buildings to provide friction and heat around the AWSs. The average WE and TE of LDAPS predictions were very small in mountainous terrain (Rm) despite their extensive variation. At Rm AWSs, WE and TE were strongly correlated with the difference between LDAPS altitude and actual altitude. LDAPS underestimated wind speed and overestimated temperature when the LDAPS terrain was lower than the actual terrain, and vice versa. In coastal terrain (Rc), it overestimated wind speed, regardless of whether the LDAPS grid points corresponded to land or sea; WE was generally slightly higher for grid points on land than for those on the sea. TE and its variation were smaller for grid points on the sea than for those on land.
Our results demonstrate that LDAPS performance in predicting wind speed and temperature at AWSs depends on the surrounding land cover (urban or rural areas) and topography (mountainous, coastal, or flat terrain). We expect that the results would help researchers (including LDAPS improvers and users) to understand the characteristics of LDAPS-predicted surface wind speed and temperature around any concerning region. Additionally, our findings will contribute to future applications of LDAPS for forecasting local weather in Korea.