Mapping Local Climate Zones Using ArcGIS-Based Method and Exploring Land Surface Temperature Characteristics in Chenzhou, China

The local climate zone (LCZ) has become a new tool for urban heat island research. Taking Chenzhou as the research object, eight urban spatial form elements and land cover elements are calculated respectively through ArcGIS, Skyhelios and ENVI software. The calculation results are then rasterized and clustered in ArcGIS to obtain the LCZ map at a resolution of 200 m. Afterwards, the land surface temperature (LST) of different local climate zones in the four seasons from 2017 to 2018 is further analyzed using one-way ANOVA F-test and Student’s t-test. The results suggest that: (1) by adding localized LCZ classes and applying the semi-automatic algorithm on the Arc-GIS platform, the final overall accuracy reaches 69.54%, with a kappa value of 0.67, (2) the compact middle-rise buildings (LCZ-2′) and open low-rise buildings (LCZ-6) heavily contribute to the high LST of the downtown area, while the large low-rise buildings (LCZ-8) cause the high LST regions in the eastern part of the town, (3) obvious land surface temperature differences can be detected in four seasons among different LCZ classes, with high LST in summer and autumn. Built-up LCZ classes generally revealed higher LSTs than land cover LCZs in all seasons. The findings of this study provide better understandings of the relationship between LCZ and LST, as well as important insights for urban planners on urban heat mitigation.


Introduction
Consistent urbanization leads to the conversion of natural land surface covers using impervious surfaces, which in turn causes an increase of artificial heat emissions. As a result of urbanization, urban areas exhibit significantly higher surface and air temperatures than nearby suburban and rural areas, resulting in the urban heat island (UHI) phenomenon [1], which affects people's health and the sustainable development of cities [2,3]. A number of studies have attempted to achieve an understanding of the interactions between urbanization and UHI [4][5][6][7][8][9]. One of the most frequent topics is to evaluate the relationship between the urban physical properties and the UHI intensity. Rizwan pointed out that huge quantities of solar radiations are mainly stored and re-radiated in urban areas due to the massive construction materials used and the decreased sky view factor [10]. Morini found that an increase of albedo leads to a decrease of urban temperatures during daytime and in general also at night time, and the temperature may be increased by 1 • C in a few points of the domain when the average albedo of urban areas has been increased from 0.2 to 0.65, 0.6 and 0.45 for roofs, walls and roads, respectively [11]. Lima pointed out that buildings' geometry can affect their energy consumption, while reductions in the thermal load of 16-18% were observed when the urban environment was considered in energy simulations [12]. Chen concluded that both the spatial

Data
The workflow of LCZ mapping is illustrated in Figure 2. Landsat 8 remote sensing images, building data and land use data are used as the input data. The output data are LCZ indicators such as the building height (BH), building surface fraction (BSF), sky view factor (SVF), street aspect ratio (SAR), terrain roughness (TR), vegetation coverage ratio (VCR), impervious surface fraction (ISF) and pervious surface fraction (PSF). Other urban planning information was collected from the planning department of Chenzhou, which included the building data, street data and land use data. The building footprint and building height were collected from the building data, and street cover data in polygon were received from the street data. The urban and rural land use information were collected from the land use data.

Data
The workflow of LCZ mapping is illustrated in Figure 2. Landsat 8 remote sensing images, building data and land use data are used as the input data. The output data are LCZ indicators such as the building height (BH), building surface fraction (BSF), sky view factor (SVF), street aspect ratio (SAR), terrain roughness (TR), vegetation coverage ratio (VCR), impervious surface fraction (ISF) and pervious surface fraction (PSF). Other urban planning information was collected from the planning department of Chenzhou, which included the building data, street data and land use data. The building footprint and building height were collected from the building data, and street cover data in polygon were received from the street data. The urban and rural land use information were collected from the land use data.

Local Climatic Zone Mapping
Local climate zone has been commonly defined and classified using blocks [37] or fixed size grids [38]. The former is related to the shape, scale and land use of the block, while the latter divides the urban texture evenly, avoiding the errors caused by the different sizes and uses of land. For the latter, the grid size should first be determined in order to ensure the classification accuracy, which is already discussed in the previous studies [18,35]. The more even the building height is within the grid, the more appropriate the grid size is. By building a semi variogram model in ArcGIS, the height range of the strong autocorrelation of buildings was obtained. As can be seen from Figure 3, the building in the main urban area of Chenzhou indicates a stable building height at a distance of 167.22 m (dash line), thus the grid size was set at 200 × 200 m. Then, the LCZ indicators were calculated using different methods (Table 1). Finally, after visualizing eight indicators in ArcGIS (Figure 4), the clustering rules were made according to the indicator range to generate the LCZ map.

Spatial form elements
Average building height (BH) and building surface fraction (BSF) were calculated directly in ArcGIS based on the building data (Figure 4a,b). However, indirect spatial form elements needed to be calculated by using multiple steps: 1) The sky view factor values were extracted at every 200 m using Skyhelios software and then rasterized in ArcGIS (Figure 4c,d). 2) Calculating the SAR (street aspect ratio) of streets required the following steps: Firstly, we found out the total covered area of streets, then, we divided the total covered area by the total length of the same streets to achieve the average width. Finally, we worked out the SAR by dividing the average height of buildings along the streets by the average width (Figure 4e,f) [29].
3) The calculation of terrain roughness (TR) was based primarily on the Davenport classification (Table 2) [39]. According to the expression of roughness grades 4, 5 and 6, the rough and very rough grid cells were categorized according to the building height and building surface fraction index in ArcGIS. Then other roughness grades were manually selected according to the representative morphology ( Figure 4g).

Local Climatic Zone Mapping
Local climate zone has been commonly defined and classified using blocks [37] or fixed size grids [38]. The former is related to the shape, scale and land use of the block, while the latter divides the urban texture evenly, avoiding the errors caused by the different sizes and uses of land. For the latter, the grid size should first be determined in order to ensure the classification accuracy, which is already discussed in the previous studies [18,35]. The more even the building height is within the grid, the more appropriate the grid size is. By building a semi variogram model in ArcGIS, the height range of the strong autocorrelation of buildings was obtained. As can be seen from Figure 3, the building in the main urban area of Chenzhou indicates a stable building height at a distance of 167.22 m (dash line), thus the grid size was set at 200 × 200 m. Then, the LCZ indicators were calculated using different methods (Table 1). Finally, after visualizing eight indicators in ArcGIS (Figure 4), the clustering rules were made according to the indicator range to generate the LCZ map.

Spatial form elements
Average building height (BH) and building surface fraction (BSF) were calculated directly in ArcGIS based on the building data (Figure 4a,b). However, indirect spatial form elements needed to be calculated by using multiple steps: (1) The sky view factor values were extracted at every 200 m using Skyhelios software and then rasterized in ArcGIS (Figure 4c,d). (2) Calculating the SAR (street aspect ratio) of streets required the following steps: Firstly, we found out the total covered area of streets, then, we divided the total covered area by the total length of the same streets to achieve the average width. Finally, we worked out the SAR by dividing the average height of buildings along the streets by the average width (Figure 4e,f) [29]. (3) The calculation of terrain roughness (TR) was based primarily on the Davenport classification (Table 2) [39]. According to the expression of roughness grades 4, 5 and 6, the rough and very rough grid cells were categorized according to the building height and building surface fraction index in ArcGIS. Then other roughness grades were manually selected according to the representative morphology ( Figure 4g).

2.
Land cover elements Land cover elements of LCZs include the vegetation coverage ratio (VCR), impervious surface fraction (ISF) and pervious surface fraction (PSF). The vegetation coverage ratio is obtained by obtaining remote sensing images through ENVI software (Figure 4h,i). Due to the large number of unbuilt areas which are presented as bare soil (LCZ-F), the ISF cannot be simply calculated using the formula "ISF = 1 − (BSF + VCR)" as in a high-density city [29]; instead, the ISF needs to be obtained using linear spectral mixing model in ENVI. In that case, the radiometric correction and principal component analysis (MNF) of the cropped image were firstly used to select the end element of the two-dimensional scatter map, which was used as the region of interest for the mixed linear spectral decomposition. Then, the generated high albedo percentage image and the low albedo percentage image were added to obtain the preliminary impervious layer. After the vegetation and the water body were masked, the ISF vector and raster maps (Figure 4j,k) were obtained with the threshold (0, 1]. Finally, PSF was calculated using the formula "PSF = 1 − ISF".

Clustering Rules
Afterwards, output data above were overlaid and clustered in the ArcGIS software platform. In the first step, the built-up LCZs and the land cover LCZs were classified based on their BSF. When the Sustainability 2020, 12, 2974 6 of 18 BSF was greater than 10%, it was classified as a built-up LCZ. If the building surface fraction was less than 10%, it belonged to the land cover LCZs. In the second step, the built-up classes LCZ-1, LCZ-2, LCZ-2 , LCZ-3, LCZ-4 and LCZ-5 were further distinguished by their BSF, ISF and BH, as can be seen in Equation (1). Then the SVF, SAR and VCR were used to distinguish the built-up LCZ-6, LCZ-7, LCZ-8, LCZ-8B and LCZ-9, as can be seen in Equation (2). The third step was to further differentiate the land cover LCZ-E and LCZ-G using ISF, as can be seen in Equation (3), and then to use VCR and TR to divide the LCZ-A, LCZ-B, LCZ-C, LCZ-D and LCZ-F, as can be seen in Equation (4).
Here, w, x and y refer to the corresponding indicator value, while z represents the LCZ type distinguished in the step.
Sustainability 2019, 11, x FOR PEER REVIEW 6 of 20 formula "ISF = 1-(BSF + VCR)" as in a high-density city [29]; instead, the ISF needs to be obtained using linear spectral mixing model in ENVI. In that case, the radiometric correction and principal component analysis (MNF) of the cropped image were firstly used to select the end element of the two-dimensional scatter map, which was used as the region of interest for the mixed linear spectral decomposition. Then, the generated high albedo percentage image and the low albedo percentage image were added to obtain the preliminary impervious layer. After the vegetation and the water body were masked, the ISF vector and raster maps (Figure 4j,k) were obtained with the threshold (0, 1]. Finally, PSF was calculated using the formula "PSF = 1-ISF".

Clustering Rules
Afterwards, output data above were overlaid and clustered in the ArcGIS software pla the first step, the built-up LCZs and the land cover LCZs were classified based on their BSF. W BSF was greater than 10%, it was classified as a built-up LCZ. If the building surface fraction than 10%, it belonged to the land cover LCZs. In the second step, the built-up classes LCZ-1 LCZ-2', LCZ-3, LCZ-4 and LCZ-5 were further distinguished by their BSF, ISF and BH, as can in Equation (1). Then the SVF, SAR and VCR were used to distinguish the built-up LCZ-6, LCZ 8, LCZ-8B and LCZ-9, as can be seen in Equation (2). The third step was to further differentiate cover LCZ-E and LCZ-G using ISF, as can be seen in Equation (3), and then to use VCR and TR the LCZ-A, LCZ-B, LCZ-C, LCZ-D and LCZ-F, as can be seen in Equation (4). Here, w, x and y refer to the corresponding indicator value, while z represents the L distinguished in the step.

Seasonal Variation Characteristics of Land Surface Temperature in LCZ
The LSTs for all four seasons were obtained to further investigate the relationships betw climate zone and land surface temperature. First of all, remote sensing images with less clo from 2017 to 2018 were selected (Table 3) in order to retrieve the land surface temperature. S the seasonal differences of paired LCZ classes were analyzed using a one-way ANOV (Tukey), with "p < 0.05" indicating that the difference was significant. Finally, the tem differences between standard LCZs and sub-class LCZs were analyzed by using the Studen

Seasonal Variation Characteristics of Land Surface Temperature in LCZ
The LSTs for all four seasons were obtained to further investigate the relationships between local climate zone and land surface temperature. First of all, remote sensing images with less cloud cover from 2017 to 2018 were selected (Table 3) in order to retrieve the land surface temperature. Secondly, the seasonal differences of paired LCZ classes were analyzed using a one-way ANOVA F-test (Tukey), with "p < 0.05" indicating that the difference was significant. Finally, the temperature differences between standard LCZs and sub-class LCZs were analyzed by using the Student's t-test.

Localize the Local Climate Zone Class
According to the building data of Chenzhou, 76.11% of the total buildings are low-rises (1-3 floors), while multi-floor (4-6 floors) and middle-rise (7-11 floors) buildings account for 8.06% and 4.62% of total buildings, respectively, which are dominant in the city core. As such, some localized changes to the LCZ should be made in accordance with Chenzhou's current situation. Firstly, it is necessary to add LCZ-2 in addition to LCZ-2. On the one hand, the building height of standard LCZ-2 (12-18 m) cannot properly describe newly built buildings in Chenzhou because most of them are over 18 m. On the other hand, it is generally known that the building height has influence over air temperature and wind velocity [39], and therefore any misclassification may also affect future climate analysis. Secondly, even though low rise buildings are very common in Chenzhou, it is noticeable that the city is currently in the process of urban renewal, where some of the low-rise buildings are being replaced by high-rise buildings, especially in the early developed downtown area. Therefore, the mixing class LCZ-34 (compact low-rise and high-rise mixed) was added. Since most of the industrial regions in Chenzhou are covered with greenery, the large and low-rise buildings mixed with vegetation (LCZ-8B) has also been added (Table 4). Finally, LCZ-10 (heavy industry) was deleted because there is no heavy industry within the study boundary.

Accuracy Assessment
A Confusion matrix is an effective tool to evaluate the LCZ classification model [27,40] where each column represents the predicted results, and each row represents the real situation of the sample. Standard accuracy measures were calculated using the confusion matrix including: the overall accuracy (OA), which is the percentage of correct classifications; producer accuracy (PA), which is the percentage of correct classifications of the reference class; user accuracy (UA), which refers to the percentage of the pixels classified as class X that actually belongs to this class; and the κ-estimate, which summarizes different accuracies in a single measure. For this study, 8% (877 grids) of the sample were randomly selected to test the classification results in Chenzhou (Table 5).   It is found in Table 5 that the accurately detected classes are open building areas such as LCZ 4 (66.67%), LCZ-5 (71.43%) and LCZ-6 (81.93%), which are the main LCZ classes in Chenzhou. Also, the newly add class LCZ-8B, as well as the LCZ-E and LCZ-F have high accuracies of 74.60%, 80.00% and 84.38%, respectively. However, there are still some misclassifications with low UA values in LCZ-1, LCZ-34, LCZ-A and LCZ-B. We determined by checking the results of LCZ classes' spatial distribution patterns with the actual urban morphology of Chenzhou that the compact high-rise buildings (LCZ-1) have significant errors (50%) due to some LCZ-1 regions having been misclassified into compact mid-rise building class (LCZ-2 ). However, since LCZ-1 only takes up 1.03% of Chenzhou, its influence on OA is insignificant. Additionally, the land cover classes LCZ-A (50.00%), LCZ-B (54.35%) and LCZ-C (66.67%) tend to be confused with each other, which is because the vegetation cover index of remote sensing images better reflect the vegetation density rather than the vegetation height. It was also noted that the accuracy of the mixing class LCZ-34 is the lowest (45.83%), which was attributed to the fact that average building height within the 200 × 200 m range is easily confused with the parameter value of LCZ-5. In order to minimize the errors caused by the cluster process and have better accuracy results, one more step was added to distinguish LCZ-34 (buildings of 3-33 m) with LCZ-5 (buildings of 12-15 m), which was to examine the standard deviation of building height in each 200 × 200 m grid in ArcGIS. In that case, grids with average BH value more than one standard deviation above the average BH mean were defined as LCZ-34, and areas with BH value less than one standard deviation below the BH mean were defined as LCZ-5. The final LCZ map was generated and is shown in Figure 5b, with the changes mainly located in three areas of Chenzhou. The UA value of LCZ-34 reached 67.12%, and the overall accuracies increased to 69.54%, with the kappa growing to 0.67 (Table 6).

Analysis of Land Surface Temperature Characteristics in Local Climate Zones
As can be observed in Figure 6, even though land surface temperature varies across the four seasons, it revealed similar spatial distributions of LST in Chenzhou. Except for winter, the warmest temperatures tend to be experienced in the central area of western downtown, where LCZ-2′, LCZ 34, LCZ-5 and LCZ-6 have large proportions, as well as the industrial areas (LCZ-8 and LCZ-8B) to the east of Chenzhou in all seasons. In particular, the east town industrial areas experience a large proportion of extreme high LST up to 40.75 ℃ in the summer; however, the high LST areas of the east town shrink in the autumn. In contrast, LCZ-A, LCZ-B and LCZ-C have comparatively lower LSTs during the four seasons, which are generally 2-4 ℃ cooler than the built-up LCZs and can be up to 10 ℃ cooler than the hottest regions in summer.

LCZ Pattern in Chenzhou
According to the final LCZ map in Figure 5 even though land cover LCZs share an area coverage of 53.37%, which is higher than that of the built-up LCZ (44.63%), LCZ-6 (6.33%), LCZ-2 (3.48%), LCZ-5 (3.40%) and LCZ-34 (3.12%) are the dominant classes in the center of the city. Another LCZ class that covers a large proportion of land (12.96%) is LCZ-9 (sparsely built area), which is mostly found in suburban areas. Furthermore, the proportion of open building areas (33.64%) is higher than that of compact building areas (10.98%) in terms of the building surface faction parameter. Regarding the building height, high-rise buildings (LCZ-1 and LCZ-4) account for only 2.03% of Chenzhou's land, while the middle-rise buildings (LCZ-2, LCZ-2 and LCZ-4) and low-rise buildings (LCZ-8B and LCZ-6) are more common in Chenzhou. It is also observed in Figure 5b that the sub-classes LCZ-2 and LCZ-34 are concentrated in the early developed downtown, which mostly covers commercial areas and residential areas. Meanwhile, the LCZ map detects LCZ-8B (4.81%) and LCZ-4 (1.00%) in the industrial district in the east part of Chenzhou. As for land cover LCZs, LCZ-C (shrub) and LCZ-A (dense forest) constitute 17.42% and 14.05%, respectively of the inner urban area, followed by LCZ-B (9.74%) and LCZ-G (4.94%). LCZ-A-C are distributed in two parallel lines in the downtown area, which is well distinguished by built-up LCZs in the city core.

Analysis of Land Surface Temperature Characteristics in Local Climate Zones
As can be observed in Figure 6, even though land surface temperature varies across the four seasons, it revealed similar spatial distributions of LST in Chenzhou. Except for winter, the warmest temperatures tend to be experienced in the central area of western downtown, where LCZ-2 , LCZ 34, LCZ-5 and LCZ-6 have large proportions, as well as the industrial areas (LCZ-8 and LCZ-8B) to the east of Chenzhou in all seasons. In particular, the east town industrial areas experience a large proportion of extreme high LST up to 40.75 • C in the summer; however, the high LST areas of the east town shrink in the autumn. In contrast, LCZ-A, LCZ-B and LCZ-C have comparatively lower LSTs during the four seasons, which are generally 2-4 • C cooler than the built-up LCZs and can be up to 10 • C cooler than the hottest regions in summer.
A boxplot of seasonal LST (Figure 7) in different LCZs was then developed to further investigate their relations. As seen in Figure 7, the mean LSTs of all LCZs are usually the highest in the summer, especially in LCZ-8 (39.2 • C), LCZ-2 (38.3 • C) and LCZ-6 (38.1 • C). LCZ-2 and LCZ-E also contributed largely to the high LST regions. In contrast, the lowest temperatures occur in the land cover classes (LCZ-A, LCZ-B and LCZ-G) and some built-up classes (LCZ-7 and LCZ-9). In the spring, a small LST difference between built-up classes and land cover classes was observed. High LST is only found in LCZ-8 and LCZ-8B with mean values of 29.53 • C and 26.16 • C, respectively. Regarding the autumn season, the quartile ranges and outliers of most LCZs are higher than those in the spring, with high LST found in LCZ-8 (35.06 • C), LCZ-2 (31.73 • C) and LCZ-3 (31.45 • C). Even though the LST in winter is the coolest in all LCZ classes, a similar trend in LCZ has been anticipated, that is, values in built-up LCZ are higher than that of land cover LCZs except for LCZ-E.
The results of one-way ANOVA F-test satisfied the normality and homogeneity of variance. Afterwards, the Tukey analysis was applied and results revealed that most of the paired LCZ types (p < 0.05) have statistically significant differences in LST (Figure 8). The normality and homogeneity of variance of each LCZ group was tested by the Kolmogorov-Smirnov test and QQ diagram, with a test level of 0.1. The results of paired LCZ are listed in Figure 8, where the blue grids indicate that there is no significant difference in LST of paired LCZs (p < 0.05), while the white grids indicate that differences existed. Overall, the differences between LCZs are statistically significant in most of the LCZ pairs during all the seasons, with the LST values being more distinguishable in the spring and summer. Additionally, the difference between the new LCZs and the original LCZ types (LCZ-8 and LCZ-8B, LCZ-3 and LCZ-34, LCZ-2 and LCZ-2 ) were further detected using the Student's t-test, since it is an important assumption that newly-added LCZs make the classification more accurate and comprehensive. According to the Student's t-test results, there are significant LST differences between the paired LCZs. From Table 7, the results confirm that LCZ-8B features a cooler LST than LCZ-8 in all seasons, while the highest difference appears in the summer (3.92 • C), followed by spring and autumn.
Similarly, the LST of LCZ-34 is relatively lower than LCZ-3 seasonally. However, the LST difference between LCZ-2 and LCZ-2 is not significant in the summer and winter.
temperatures tend to be experienced in the central area of western downtown, where LCZ-2′, LCZ 34, LCZ-5 and LCZ-6 have large proportions, as well as the industrial areas (LCZ-8 and LCZ-8B) to the east of Chenzhou in all seasons. In particular, the east town industrial areas experience a large proportion of extreme high LST up to 40.75 ℃ in the summer; however, the high LST areas of the east town shrink in the autumn. In contrast, LCZ-A, LCZ-B and LCZ-C have comparatively lower LSTs during the four seasons, which are generally 2-4 ℃ cooler than the built-up LCZs and can be up to 10 ℃ cooler than the hottest regions in summer. A boxplot of seasonal LST (Figure 7) in different LCZs was then developed to further investigate their relations. As seen in Figure 7, the mean LSTs of all LCZs are usually the highest in the summer, especially in LCZ-8 (39.2 ℃), LCZ-2′ (38.3 ℃) and LCZ-6 (38.1 ℃). LCZ-2 and LCZ-E also contributed largely to the high LST regions. In contrast, the lowest temperatures occur in the land cover classes (LCZ-A, LCZ-B and LCZ-G) and some built-up classes (LCZ-7 and LCZ-9). In the spring, a small LST difference between built-up classes and land cover classes was observed. High LST is only found in LCZ-8 and LCZ-8B with mean values of 29.53 ℃ and 26.16 ℃, respectively. Regarding the autumn season, the quartile ranges and outliers of most LCZs are higher than those in the spring, with high LST found in LCZ-8 (35.06 ℃), LCZ-2 (31.73 ℃) and LCZ-3 (31.45 ℃). Even though the LST in winter is the coolest in all LCZ classes, a similar trend in LCZ has been anticipated, that is, values in builtup LCZ are higher than that of land cover LCZs except for LCZ-E.  The results of one-way ANOVA F-test satisfied the normality and homogeneity of variance. Afterwards, the Tukey analysis was applied and results revealed that most of the paired LCZ types (p < 0.05) have statistically significant differences in LST (Figure 8). The normality and homogeneity of variance of each LCZ group was tested by the Kolmogorov-Smirnov test and QQ diagram, with a

The LCZ Classification and LST Characteristics in Chenzhou
Some localized classification steps were adapted to refine the mapping by understanding the LCZ properties in Chenzhou. Firstly, as the optimal resolution of LCZ maps might vary among cities [34], the scale of the grid should be determined first. Unlike the earlier LCZ mapping research focusing on high-density cities [29,33,34], which were usually mapped at a resolution of 300 m, GIS-method classification in a developing city like Chenzhou applied grids of 200 × 200 m to ensure there was classification accuracy. Secondly, some modifications have been applied to the standard LCZ classes according to the dominant building forms of Chenzhou, which some previous studies already demonstrated in other countries. For example, a study in Nagpur developed twenty-one LCZ classes, including fourteen standard classes and seven distinct LCZ sub-classes in order to successfully deal with the problems associated with the intermixing of LCZs [25]. Similarly, a study in Colombo, Sri Lanka pointed out that the LCZ approach was not able to capture all of the nuances of urban growth everywhere and therefore a LCZ sub-classification system was needed based on urban spatial particularities [38]. In the current study, the LCZ map of Chenzhou is classified using twelve built-up LCZs and seven land cover LCZs, which includes three new localized LCZ types (LCZ-2 , LCZ-34 and LCZ-8B) according to the character of land surface properties. Lastly, even though the Arc-GIS based classification in this study shows high overall accuracy like many other previous research [28,30,37], it detects relatively low accuracies in some LCZ classes, which are mainly in the mixed building types (LCZ-34) and the land cover classes (LCZ-A-C). Although the sub-classification adds complexity to the LCZ map, the misclassifications which occurred in built-up LCZs can be improved by conducting statistical analysis of urban physical structure parameters in the Arc-GIS platform. By examining the standard deviation of building height after classification workflow (Figure 2) in this study, the accuracy of LCZ-34 has been improved from 45.83% to 67.12% and the final LCZ map (Version 2) was generated with higher overall accuracy (69.54%). However, further studies are needed to apply higher resolution satellite images to improve the accuracy of LCZ-A-C.
It was found that some relations existed between the distributions of LCZ and LST. On the whole, for a developing city, the proportion of open building area is larger than that of compact building area, with the open middle-rise regions (LCZ-5) and the open low-rise regions (LCZ-6) appearing with high frequencies. However, the areas with high LST in the city center were mostly compact urban areas (LCZ-2 and LCZ-2 ) where ventilation was likely to be blocked. Also, high LST regions appear in the eastern industrial zones (LCZ-8 and LCZ-8B) where impervious surfaces have a large proportion. It is generally known that impervious surfaces absorb more solar radiation and emit more long-wave radiation, and thereby leading to higher LST [21,41,42]. Even though the LSTs of the most built-up LCZs exceed that of land cover LCZs, simple low-rise buildings (LCZ-7) and very open lower-rise buildings (LCZ-9), which are common in urban suburbs, exhibit low LST values as well. Similar conclusions were reached in studies of the Yangtze River Delta and Pearl River Delta in China [20,43], and Brno and Prague in Europe [44]. It is also noted that the LCZ-F (bare land) takes up 5.47% of Chenzhou's land, indicating that the unbuilt area of Chenzhou accounts for a significant proportion of its land, which might be attributed to the process of transforming natural environment into an urban environment. Therefore, there is a possibility that the proportion of built-up LCZs in the environment will be larger than that of land cover LCZs in the future. In that case, attention will also need to be paid to LSTs during urban construction, because the LST of LCZ-F reaches 35 • C in the summer.
Generally, the characteristics of land surface temperature in each local climatic zone vary with the seasons, with a high LST in summer, followed by autumn and spring, and the lowest in winter. The LST differences also exist in most LCZ pairs in each season according to the Tukey tests. According to Figures 6 and 7, the highest differences in LSTs are between the built-up LCZ classes during the summer and autumn periods, which means the LSTs might be strongly influenced by LCZ classes during the hot seasons. Moreover, the interpretations of the Student's t test suggest that significant LST differences occur between the standard LCZs and the localized LCZ, except for LCZ-2 and LCZ-2 . The mean LST of LCZ-8B is lower than that of LCZ-8, which may due to the fact that the shadows of large low buildings in LCZ-8 are not sufficient to lower the LST, while the vegetation of LCZ-8B (large low-rise with vegetation) produces more shadows and brings soil moisture during the day. Meanwhile, similar results are observed in LCZ-34 when compared to LCZ-3, which might be explained by the shadow of high-rise buildings reducing the radiation temperature [44][45][46].

Implementation for Urban Planning
The study of LCZ and LST patterns is beneficial for understanding the land use planning as well as the urban thermal environment. This analysis showed that the key planning areas for mitigating the UHI phenomenon of Chenzhou are LCZ-8 (large low-rise), LCZ-2 (compact middle-rise) and LCZ-6 (open low-rise), which are mainly found in core downtown areas and industrial zones in Chenzhou. Particularly, LCZ-8 and LCZ-2 contribute largely to the hot spots of Chenzhou. Having taken the adverse impacts of these LCZ classes into account, some planning suggestions are made. Firstly, proper greenery should be applied in LCZ-8 areas, which will bring more shadows and surface moisture, as well as cut down the solar radiation in summer. Secondly, it is encouraged to convert compact middle-rise buildings (LCZ-2 ) to open middle-rise buildings  or open high-rise buildings (LCZ-4) in central Chenzhou, in order to provide sufficient public space for green areas and to increase urban ventilation. Also, the proper and reasonable distributions of LCZs with lower land surface temperatures such as LCZ-A-C (dense forest, sparse forest and shrub) and LCZ-G (water body) are advisable to mitigate the influence of UHI during the process of urban development

Limitations
This study aims to map local climate zones in Chenzhou and evaluate its relationship with land surface temperature patterns. However, there are some limitations. The calculation of the surface albedo, surface admittance and anthropogenic heat flux has not been realized. Therefore, it is unknown whether their calculation would have led to a more accurate result. Additionally, even though the accuracy of built-up LCZ classes is improved in this study, satellite data with higher resolutions, as well as other sensors and auxiliary sources such as Open Street Map are still needed in future studies, because the errors of land cover LCZs are possibly due to the spatial resolution of the Landsat data.

Conclusions
This paper aims to deepen the knowledge about GIS methods in the LCZ mapping of developing city and further evaluate the four seasons' land surface temperatures in LCZs. In this study, the LCZ scheme has been applied in Chenzhou and a LCZ map has been built using multiple types of software. Meanwhile, the semi-automatic algorithm on the Arc-GIS platform was found to improve mapping efficiency. By providing the sub-classification in Chenzhou, a better understanding of the urban structure, as well as the relations between LCZ map and LST patterns was achieved, which provided important insights for urban planners on urban heat mitigation.