The metropolitan region of Seoul, the capital city of Korea, was considered in the analysis performed in this city. The city government of Seoul has developed several systems to facilitate the collection and analysis of data pertaining to urban phenomena across neighbourhoods. Among these, the Smart Seoul Data of Things (S-DoT) was deployed under a pilot program in 2019, the data obtained as part of which was publicly released in April 2020. This study was performed over the period between July 1, 2020 and September 30, 2020, i.e., the summer months. Fig. 1 presents a flowchart of the tasks performed in this study. First, the data required for analysis were first collected and pre-processed. Next, the temperatures obtained using S-DoT and AWS were compared via spatial statistical analyses, and their corresponding relationships with the solar radiation (SR), green-area ratio (GAR), and altitude were evaluated. Subsequently, the interpolation results obtained using AWS and S-DoT were comparatively examined. Each variable was selected to reflect land use, altitude, and urban structural factors that have been discussed as possible causes of temperature differences between the two observation networks (Park, 2021). For land use, GAR was selected as a variable to figure out differences in green space, and SR was used to reflect urban structural factors. Altitude has a significant impact on air temperature (Choi, 2011; Yim & Lee, 2017), so we used it to reflect factors related to differences in altitude within location.

Fig. 1. 
Flowchart of proposed study

The elevation, climate, land use, and satellite image data were collected and pre-processed. Elevation data were used to build a digital surface model (DSM), and the daily solar radiation (SR), which was corrected based on cloud cover, was calculated. The weather data were obtained from the S-DoT, a high-resolution temperature-monitoring network of Seoul, and the AWS system operated by the government was used as a reference temperature-monitoring network. Those stations were installed throughout Seoul, and each installation location is shown in Fig. 3. The interpolated data of the AWS were generated by using the daily temperature data of the AWS. Subsequently, the land-cover ratios around the weather stations were calculated, and the LSTs were calculated by using the Landsat 8 satellite images.

2.2.1. High-resolution air temperature data for Seoul (S-DoT)

S-DoT is a group of monitoring sensors that collect urban data generated in the city of Seoul. In total, 850 units have been installed across Seoul (as of 2020. S-DoT sensors are mainly installed on closed-circuit TV supporting plates, telephone poles, and exterior walls of buildings). With the S-DoT, hourly measurement information on 10 types of data (air temperature, humidity, illuminance, noise level, particulate matter including PM10 and PM2.5, ultraviolet (UV) rays, vibrations, number of visitors, wind direction, and wind speed) is collected. In this study, only the data from 8:00 to 20:00 on the observation date were used, and S-DoT points were excluded when errors were found in the data (Fig. 2).

Fig. 2. 
S-dot data with observational errors removed

2.2.2. National temperature data from the Automatic Weather System

The AWS is a terrestrial observation system that has been in operation since 1997 by the Korea Meteorological Administration (KMA). Monitoring stations are installed at 504 locations nationwide, and these automatically monitor the air temperature, wind, precipitation, humidity, and air pressure. As of 2020, there were 27 monitoring stations in Seoul. If nearby stations were included, there were 33 monitoring stations with data available. The AWS air temperature sensors are installed at a height of 1.2–2.0 m above the ground and in places where a ventilation wind speed of 2.5–10 m/s is maintained. The AWS constructs 1-min data by averaging six data points sampled at 10-s intervals at the measurement point. The hourly air temperature dataset was constructed by averaging 1-min data in units of an hour, and the air temperature data were recorded in units of 0.1°C (KMA, 2019).

Fig. 3. 
Locations of S-DoT and AWS installations in Seoul

2.2.3. Land cover around the monitoring stations (S-DoT and AWS)

A spatial statistical analysis was performed to investigate the relationship between the air temperature difference and land-use characteristics of the two types of monitoring stations (S-DoT and AWS). For the data used for land use, land-use maps created by the Ministry of Environment were used, and the land-use data were divided into seven categories (urban, agriculture, forest, grass, wet land, and water). Among the land uses, forest, grass, wet land, and water were defined as green areas and analysed accordingly. Agricultural land in Seoul was excluded from green areas because most of the land refers to greenhouses. For land-use characteristics around the S-DoT sites, the land-use ratio was calculated by using buffer information within a 300 m circle around the centre of the S-DoT site (Fig. 4).

Fig. 4. 
Method for determining land-use characteristics around S-DoT sites

First, the daily mean temperatures were compared between the S-DoT and AWS. The statistical difference in mean temperatures between the monitoring stations was validated using a t-test. We then checked whether there was a difference in temperature depending on the spatial distribution. The AWS temperature and mean temperature of the S-DoT distributed within 300 m of the AWS site were statistically compared to check if the spatial distributions were different. In addition, we compared the temperature differences between the monitoring stations located in forests and the city centre.

Next, the correlations between the temperature and three factors, namely, the SR, GAR, and altitude were analysed in two steps. In the first step, it was assumed that the rainy and sunny days would show different correlations, and all of the data were grouped into either rainy or sunny days, after which their correlations were analysed. In the second step, by using daily data from sunny days, we investigated the explanatory power of each model derived through a multilinear regression analysis of the AWA–S-DoT data. Here, the input data of the S-DoT were sorted and broken into 150 equal-sized quantiles based on the dependent variable for 850 monitoring stations to conduct the analysis (Pan et al., 2018).

Finally, the statistical difference between the AWS kriging and S-DoT temperatures at the installation locations of the S-DoT (n = 850) was analysed to determine the difference between the kriging data and the city monitoring data. Then, correlations between the temperature and the three factors (SR, GAR, and altitude) were analysed for each day (sunny days) to analyse which factors were correlated with the difference between the two datasets. The correlation coefficients were comparatively analysed using Fisher’s Z transformation (Fisher, 1921) to compare the daily correlations. As depicted in Fig. 5, the analysis was performed by generating separate installation locations for S-DoT and daily data sets.

Fig. 5. 
Comparison of correlations between air temperature (AWS kriging, difference between AWS kriging and S-DoT) and factors of interest (SR, GAR, and altitude)

In the results of the daily distribution comparison between the two datasets from the S-DoT and AWS, both monitoring station types showed similar patterns for the daily mean temperature at all monitoring stations (Fig. 6). The mean temperature obtained using S-DoT equalled 26.51°C in summer, whereas that obtained using AWS was 23.93°C. Accordingly, the mean S-DoT temperature equalled 2.58°C higher compared to its AWS counterpart. Furthermore, the standard deviation of the S-DoT data was greater than that of the AWS data, whereby the mean variation of the S-DoT data (7.3%) was higher than that of the AWS data (5.1%). In addition, the t-test results for the difference between the two monitoring station types showed that a statistically significant difference existed between the mean temperatures of the S-DoT and AWS. On a day-to-day basis, the difference between temperatures measured using S-DoT and AWS revealed the S-DoT temperature to exceed the corresponding AWS value by 2.54°C on average. The corresponding difference in the standard deviations equalled 1.02.

Fig. 6. 
Daily S-DoT air temperatures and AWS data

In the comparison of the mean temperatures of the S-DoT and AWS according to the monitoring point, the mean was statistically different between the two datasets (t-statistic = -12.489, p-value = 0.000) (Table 1). Furthermore, the AWS tended to show lower temperatures than the S-DoT within 300 m. The air temperatures measured by the AWS located near the forests (n = 19, mean = 24.5°C) were generally lower than those of the AWS located in the urban areas (n = 12, mean = 25.2°C), and the temperature distribution range of the monitoring stations was smaller for the AWS than for the S-DoT (Fig. 7). Specifically, the temperature range of AWS equalled 23.35–25.36°C (i.e., 2.01°C) while that of S-DoT equalled 24.01–28.37°C (i.e., 4.36°C).

Fig. 7. 
Spatial distribution of mean air temperatures (S-DoT versus AWS)

The data obtained from the AWS by kriging (AWS kriging) were compared to the temperature distribution of the S-DoT. The mean value of AWS kriging equalled 24.80°C, which differed the corresponding mean obtained using S-DoT by 1.72°C. In other words, AWS kriging revealed AWS temperatures to be were generally compared to their S-DoT counterparts. Furthermore, the standard deviation of AWS kriging was 0.20, whereas that of the S-DoT was 0.51, which was indicative of a narrower distribution. We compared the spatial temperature difference between the two monitoring station types and confirmed that the temperature difference decreased when the monitoring point was near (200 m) the forest area (Fig. 9). The mean temperature equalled 1.43°C for the monitoring points located near the forest area; in other regions, this mean temperature equalled 1.82°C.

Fig. 9. 
Spatial distributions of mean air temperatures

By analysing the changes in the AWS interpolation (kriging) temperatures of S-DoT points and SR, GAR, as well as altitude, this study confirms the absence of a radiation correlation with the above-mentioned parameters. Although the GAR revealed a weak negative correlation (-0.155), the same was statistically insignificant on 44.1% of the days. The altitude demonstrated a strong negative correlation (-0.657), which was statistically significant on all days. In contrast, the temperature difference between the S-DoT and AWS kriging was statistically significant on all days, as revealed by the results of the correlation analysis for SR (83.7%), GAR (100%), and altitude (98.9%). The mean Fisher’s Z value equalled 0.409 for SR, which indicates a positive correlation, whereas equalled -0.805 for GAR—i.e., a negative correlation. The altitude, too, demonstrated a negative correlation of -0.619.

To understand the formation of urban microclimates, this study performed an analysis based on the SR, GAR, and altitude. When a correlation analysis was performed for these three factors, the correlation with the temperature of the S-DoT and AWS varied depending on the rainfall. A positive correlation was observed for SR, and this correlation was particularly strong on sunny days. Thus, it appears that the rainfall had a direct impact on the temperature in the microclimate; furthermore, the microclimate changed as the urban evapotranspiration rate changed due to rainfall (Kubilay et al., 2019; Narayanan, 2017). When a multilinear regression analysis was performed for the daily temperature of sunny days, excluding the changes caused by rainfall, the results were similar to those of the correlation analysis conducted for all data in the case of the S-DoT but were different in the case of the AWS. Even when the regression analysis was performed using daily data, the S-DoT showed a positive correlation with the SR, but the AWS showed a negative correlation. This difference implies that the radiation difference on the boundary layer is not significant at the city scale, but the radiation difference on the canopy layer is significant. When data were examined in terms of the boundary-layer scale, most of the investigated areas (excluding the southern and northern slopes of mountains) had similar radiation levels, but at the canopy-layer scale, the radiation difference caused by the urban structure was reflected because the radiation was calculated to reflect the structure. Furthermore, in terms of the method involving the installation of monitoring stations, the S-DoT possesses a structure that can reflect the radiation according to the urban structure, whereas the AWS measures the temperature by removing the factors that can affect the surrounding temperatures while maintaining the ventilation. As a result, it was determined that the AWS did not reflect radiation, and the data thereby showed a negative correlation in the regression analysis results. It was also determined that the explanatory power of the model that excluded the radiation from variables (AWS Model 2) in the AWS regression model was high because this phenomenon was reflected.

The GAR showed a negative correlation with both the S-DoT and AWS data. This finding was interpreted as follows: the temperature decreased as the GAR increased because of the heat-island reduction effect of green areas (Jin et al., 2011; Li et al., 2011; Yuan and Bauer, 2007). The altitude showed a large difference in its influence, which seemed to be due to the difference in the installation locations of S-DoT and AWS sites. Unless the temperature is measured on mountains or building rooftops in urban areas, the altitude difference should not be large. Thus, it proved difficult to determine the temperature difference caused by the altitude. In contrast, there exist several locations where the AWS performs temperature monitoring at mountain sites. In such cases, spatial datasets are likely to demonstrate a greater impact of the altitude compared to S-DoT data within the city owing to a large altitude difference between monitoring stations.

A noticeable difference in the explanatory power was observed when comparing the S-DoT Model 1 and AWS Model 2. This difference can be considered a consequence of the data not reflecting the temperature increase caused by anthropogenic factors. The climate of the canopy layer in the city centre is affected by various factors, including the building envelopes and anthropogenic factors. Furthermore, anthropogenic factors (e.g., outdoor air-conditioning units, automobiles) emit substantial amounts of thermal energy, which has a direct impact on the temperature. This indicates that the S-DoT model considered in this study is characterised by a low explanatory power owing to inadequate consideration of anthropogenic factors. In contrast, the AWS measures the temperature in an aggregated state of urban microclimates (Armoogum and Bassoo, 2019; Petralli et al., 2014), and as a result, the effect of factors considered on the canopy layer is small, which led to the difference in the explanatory power.

The measuring method and characteristics of heat islands vary depending on the measurement method used for air temperature. The BLUHI approach measures the heat-island phenomenon while focusing on the city centre at a mesoscale. This measurement method characteristically assumes the atmosphere in the upper part of the city to represent a heat island, and some studies have noted that it is desirable to measure the temperature at the top of the highest buildings and to install the monitoring devices at locations that are not affected by the surrounding buildings (Branea et al., 2016). The CLUHI approach focuses on heat islands at a more local scale and requires more detailed high-resolution data than the BLUHI approach. On such a local scale, there is a high possibility of reflecting radiation (Nakamura and Mahrt, 2005). Park (2021), who compared S-DoT with AWS and ASOS, discussed that the difference in temperature between the two stations may be due to the height of the station installation. In addition, the S-DoT is installed on closed-circuit TV camera mounting plates, building exterior walls, and so forth, and as a result, it likely reflects the radiation depending on the installation location, which may be a result of the influence of artificial heat generated in cities such as buildings, roads, outdoor air conditioner units, etc (Park, 2021). In other words, the AWS is a mesoscale-monitoring network that measures a BLUHI and the S-DoT is a microscale-monitoring network that measures a CLUHI. Therefore, even if the same temperature is observed, S-DoT is an urban environment that reflects artificial heat and is measured at a lower height, while AWS is measured at a higher location, resulting in different characteristics.

Even if the same heat island is investigated, the results will be different depending on which data are used (i.e., air temperature or surface temperature data) (Fig. 10, which compares results obtained at three data points). This confirms that the heat-island hot spots in the investigated area appear differently depending on which data—AWS kriging (mesoscale), S-DoT (microscale), and LST (surface temperature)—are used. The results obtained considering the AWS kriging and S-DoT data revealed different hot-spot-distribution ranges. The S-DoT showed local hot spots, whereas the AWS kriging showed a global distribution. The LST results also showed local hot spots similar to those of the S-DoT, but the spatial distribution was different because the characteristics of surface temperature and air temperature are different from each other (Branea et al., 2016; Sheng et al., 2017). In brief, the type of heat island discussed will differ depending on the scale of the spatial solution, and the characteristics of the data required would differ as well. So, not only are the heat island regions different depending on the data used, but the features that need to be discussed are also different. Therefore, when discussing spatial solutions, it is important to consider the data used and the scale.

Fig. 10. 
Spatial air-temperature hotspots identified using AWS kriging, S-DoT, and LST

When the results of the interpolation at the mesoscale (AWS kriging) and the results of the microscale (S-DoT) were compared, a temperature difference was found. This difference seems to be a result of the aggregation of urban microclimates due to the low density of temperature-monitoring stations, whereby the heat and artificial heat were reflected according to the structure of urban buildings in the process of aggregation (Petralli et al., 2014). The canopy-layer temperature cannot be accurately determined via numerical interpolation owing to the difficulty involved in incorporating complex urban structures (Asdrubali and Desideri, 2019). In fact, the difference between the canopy- and boundary-layer temperatures increases with increase in the urbanisation ratio around monitoring stations and decrease in GAR. In urban areas, the difference between the said temperatures increases owing to the heat generated by the buildings and roads (Sailor 2013), and it decreases with increase in GAR owing to the heat-island reduction effect of green areas that reduce anthropogenic heat (Jin et al., 2011; Li et al., 2011; Yuan and Bauer, 2007). The radiation appears differently depending on the city structure, and this difference seems to affect the S-DoT results, wherein the microclimate is determined at the street level (Zhu et al., 2020). In contrast, as these differences are aggregated in AWS kriging, the difference between the S-DoT- and AWS-kriging-based temperatures increases as the radiation increases. In the case of altitude, because the target study areas were at high altitudes, the GAR increased and the temperature decreased as the altitude increased. As a result, the observed temperature difference decreases with increase in altitude owing to a reduced UHI effect.

To mitigate the UHI phenomenon, targeted spatial solutions need to be developed, and to this end, a temperature map of urban spaces must be created. In developing a reliable climate map and temperature map, it is essential to measure the air temperature in short intervals. However, because measured temperature data are point data, many studies have spatialised such data by using interpolation methods (Viggiano et al., 2019, Shtiliyanova et al., 2017), as described in this study. In the AWS kriging results of this study, it was difficult to reflect the temperature difference of the city caused by radiation because the interpolation was performed by using a small number of point data. While the station locations can reflect the effects of SR, their number is too small to reflect such effects when interpolated. In contrast, the S-DoT results reflect SR, and because the S-DoT data are constructed at a relatively higher resolution compared to the AWS data, the difference in the effect between various urban structures can be reflected. Furthermore, the observed variation can be investigated in the canopy layer. In other words, the S-DoT and AWS data should be used by dividing them in consideration of the scale under discussion. Additionally, policies and projects aimed at mitigating heat islands should evaluate their effectiveness using appropriate scales for discussion. For instance, when considering green space-based projects for heat island mitigation, the success of efforts to reduce urban heat islands by incorporating green spaces in cities has been measured by analysing temperature at a micro-scale near the green areas. However, in order to fully assess the effectiveness of heat island mitigation, it is crucial to examine data at various scales to comprehend its overall usefulness and impact. Monitoring plans or projects related to urban heat islands necessitates evaluating their effectiveness across multiple scales. This means assessing the project's impact not only in the immediate vicinity of the implementation area but also at the targeted scale or across multiple scales.