Modeling of Daytime and Nighttime Surface Urban Heat Island Distribution Combined with LCZ in Beijing, China

Abstract

Along with urbanization, surface urban heat island (SUHI) has attracted more attention. Due to the lack of perspective of spatial heterogeneity in relevant studies, it is difficult to propose specific strategies to alleviate the SUHI. This study discusses the impact of spatial heterogeneity on the day and night SUHI by taking one day and night in Beijing as an example, and uses it to improve the efficiency of SUHI simulation for related planning. This study, based on the local climate zone (LCZ), deeply discusses the relationship between urban morphology and the SUHI. Then, an artificial neural network (ANN) model with the LCZ is developed to predict the distribution of the SUHI. The results show that: (1) In summer, the general SUHI intensity distribution patterns are compact zone > large low-rise zone > open zone and medium floor zone > low floor zone > high floor zone. (2) Building density and albedo in dense areas are higher correlated with the SUHI than open areas. The building height has a significant negative correlation with the SUHI in high-rise zone, but has a positive correlation in middle and low floors. (3) The LCZ improves the overall accuracy of the ANN model, especially the simulation accuracy in the daytime. In terms of regions, LCZ2, LCZ8, and LCZ10 are improved to a higher degree. This study is helpful to formulate the SUHI mitigation strategies of “adapting to the conditions of the LCZ” and provide reference for improving the sustainable development of the urban thermal environment.

1. Introduction

Urban heat island (UHI) is used to describe a phenomenon where the temperature in urban areas is higher than that in surrounding suburbs [1]. The surface UHI (SUHI) caused by urbanization has become a major problem affecting the development of the ecological environment [2,3,4,5]. It has aggravated air pollution [6] and endangered human health [7,8]. How to alleviate it has attracted much attention.

It has been considered that the UHI is the result of the combined effect of anthropogenic heat and solar radiation on urban space [6]. Obviously, controlling and optimizing urban morphology is an important way to alleviate the impact of the urban thermal environment and build a climate-friendly city [7]. Therefore, many researchers take this as a focus to explore the relationship between urban morphology and the UHI in order to provide theoretical support for related planning and management. For example, Guo et al. [8] pointed out that building height and building density can significantly affect the UHI, and the influence degree of building density is higher. Wang et al. [9] studied the quantitative relationship between the spatial distribution of a water body and surface temperature to help us understand the cooling effect of a water body on the surrounding thermal environment. Yin et al. [10] established the relationship between six morphological indexes, such as building density, impervious surface ratio, vegetation coverage, and surface temperature, by using spatial regression model.

The above research is of great significance to explore the formation mechanism of the UHI effect, but the conclusions are difficult to be transformed into specific and spatially targeted urban morphology regulation strategies, and they ignoredthe effect of spatial heterogeneity on the correlation between urban form and UHI intensity. The current research on the spatial distribution and variation characteristics of urban heat islands has achieved great results, but these studies on urban heat islands do not take into account the spatial variability of surface temperature. However, the linkage between elements at different scales in the urban area is significant, and the interdependence forms a complex system. It is particularly important to further understand the comprehensive action laws of different regional elements within the complex system. Therefore, some scholars try to divide the city into functional areas, such as commercial area, residential area, and industrial area, and further analyze the relationship between the urban morphology index and surface temperature on the basis of comparing the surface temperature of each functional area [11]. Some scholars have also found that changes in land-use types have led to significant changes in the spatiotemporal pattern of the SUHI [12]. Such research will help urban planners to assess and improve the heat island status of each functional area more specifically. However, criteria for urban functional zoning and land use are not strongly correlated with climate and urban form. Therefore, the proposed urban morphological regulation strategy, guided by mitigating the heat island effect, has strong uncertainty.

The LCZ (local climate zone) provides an opportunity to solve the above problems. It is a classification system proposed according to the difference of response ability of different urban form types to the thermal environment; cities can be divided into 10 types of building (LCZ1–LCZ10) and 7 types of natural coverage (LCZA–LCZG) [13]. Similar LCZs have similar morphological characteristics and temperature attributes. Bechtel et al. [14] used remote sensing image classification technology to develop the world urban database and access portal tool (WUDAPT) [15], which promoted the dissemination and application of local climate areas. At present, research on the UHI based on the LCZ has been gradually carried out in cities around the world. Saliba et al. [16] presented a methodology to assess the impact of Beirut’s urban structure on the UHI. Malley et al. [17] identified urban forms that are prone to heat storage in Japan based on the LCZ. Researchers analyzed the distribution of the UHI based on the LCZ and found that different LCZs had different impacts on land surface temperature [18]. In addition, different seasons will also lead to the difference of the UHI intensity of different LCZs [19]. Through the study of urban agglomeration, it is found that the LCZ types of areas with high LST are consistent, indicating that the research results are universal [20]. Simanjuntak et al. [21] found that not only the type of LCZ but also the composition and configuration pattern of the LCZ significantly affected the surface temperature.

The above research reveals the close relationship between the SUHI and urban morphology from the temperature difference of each LCZ, but research on LCZ-SUHI is not deep and detailed enough, ignoring the spatial heterogeneity of the relationship between relevant driving factors and the SUHI, and there are relatively few studies on the influence of different urban morphology indices on the SUHI in different climate regions, which leads to the lack of guidance of the research results to the planning in the specific implementation. In addition, the time period selected in the above study to explore the influence of urban morphology on the heat island effect is daytime, so the conclusions obtained are only applicable to the improvement of the daytime thermal environment. Studies have shown that the intensity of the night UHI in some areas is stronger than that of daytime [22], and the mitigation strategies for the night UHI need to be studied separately from those during the day. Therefore, in view of the shortcomings of the above research, this study adds spatial heterogeneity to the traditional heat island research, and further studies the differences between day and night urban heat islands. This study takes summer in Beijing as an example, explores differences in the driving indices of the SUHI in different LCZs, and develops an artificial neural network for SUHI distribution prediction through the combination of the LCZ and other indices. The improved model of the LCZ can more effectively predict and simulate the SUHI caused by urban planning and layout so that people can carry out reasonable urban planning and design, effectively reduce the SUHI in the region, and improve the livability and comfort of the city. The main contents of this study are as follows: (1) LCZ zoning based on the WUDAPT process and analyze the distribution of the SUHI in the study area during the day and night. (2) Build a parametric model to explore the influence of surface heat island intensity and parameters in different LCZs during the day and night. (3) Use city morphology indices to train an artificial neural network to predict the SUHI, and join the LCZ to improve the accuracy of the model and then analyze the prediction results. The research results of this study can provide a valuable reference for improving the urban thermal environment, community construction, and human environmental quality in future research.

2. Materials and Methods

2.1. Study Area

The research area of this study is located in the area within the Fifth Ring Road of Beijing, China (Figure 1). The built-up area of Beijing is in a concentric expansion mode. There are seven ring roads in Beijing, of which the area within the Fifth Ring Road is the main area for urban construction and development in Beijing, covering an area of about 667 km². Beijing is located in North China. It is hot and rainy in summer and cold and dry in winter. With the continuous increase in the built-up area in Beijing, the permanent resident population of the city has reached 21.893 million by the end of 2020. With the rapid development of the city, the problem of the thermal environment has become increasingly prominent [23]. How to alleviate the heat island effect is one of the key issues in Beijing.

2.2. Research Data

The data used in this study include AST_08 surface temperature products and a landsat 8 image on 17 August 2019, Google Earth image and impervious surface data, and building base data. The resolution, purpose, and source of each data are shown in Table 1.

Data processing mainly includes two aspects: remote sensing image preprocessing (radiation calibration, atmospheric correction) and research unit division. This study uses the semivariogram model as a reference for the division of research units [24]. Figure 2 shows that the semivariance of building height increases as the distance between two buildings increases. Red dots represent the binned value and are generated by grouping empirical semivariogram points. Average values are represented by blue crosses and are generated by binning empirical semivariogram points that fall within angle sectors. At 200 m, the model appears horizontal, and the distance at which the model first appears horizontal is called the range. Sample positions separated by a distance shorter than the variable range are autocorrelated with space, while sample positions separated by a distance farther than the variable range are not autocorrelated with space. As can be seen from the figure, the height of the building within 200 m has a strong autocorrelation, and the height of the building is relatively uniform. Therefore, the resolution of the grating-based LCZ classification map is determined to be 200 m [25,26].

2.3. Research Methods

This study determined the study unit through the semivariogram, then used ArcGIS Pro and ENVI to calculate the urban morphology indices and the SUHI of the unit, and used SAGA [27] and Google Earth Pro to divide the LCZs. The correlation between the SUHI and urban morphology indices was analyzed, the artificial neural network model was constructed for the distribution simulation of the SUHI, and the models with LCZ/without LCZ were evaluated and compared, relying on this to quantify the effect of adding spatial heterogeneity to the model. The research process flow chart of this study is shown in Figure 3.

2.3.1. Local Climate Zones (LCZs)

This study is based on the standard LCZ classification method proposed by Bechtel et al. in 2015 to perform LCZ division. The specific classification includes four stages: (1) Calculate the morphological indices of each grid, and extract the water body by using the improved normalized difference water body index (MNDWI) [28] to help identify LCZG. (2) Manually select the samples through the Google Earth platform, and combine the calculated morphological indices and Baidu Street View map to improve the accuracy of the sample selection. Sample types include LCZ1-LCZ6, LCZ8-LCZ10, LCZT (this study mainly studies the heat island effect of the built environment, so the vegetation is merged into a type of LCZ, namely, LCZT), and LCZG (Figure 4). (3) Based on the provided training area and the spectral characteristics of the satellite data, the random forest classifier is used to divide the entire study area into different LCZ categories. Finally, the obtained verification samples are used to verify the LCZ mapping results.

2.3.2. Land Surface Temperature Retrieval

Using a Landsat 8 image, the surface temperature is retrieved based on the radiative transfer equation method [29]. The main formula of this method is as follows:

(1)$LST=K_2/\ln \left(K_1/B+1\right),$

with

(2)$B=\left(L_{\lambda }-L_u-\tau \times \left(1-\epsilon \right)\times L_d\right)/\left(\epsilon \times \tau \right)$

where B is the radiant brightness value of a blackbody in a thermal infrared band; L_λ is the sensor spectral radiance; the three parameters L_u, L_d, and τ represent the atmospheric upward radiation, the atmospheric downward radiation, and the atmospheric transmittance, respectively, which can be found on the NASA official website (http://atmcorr.gsfc.nasa.gov/ (accessed on 18 July 2021)) according to the imaging time and central latitude and longitude of the image; K₁ and K₂ can be obtained from the image header file; and ε is the surface emissivity, which is an important parameter for temperature inversion. Due to the various types of surface coverage in the study area, this study adopts the algorithm proposed by Qin et al. [30].

To obtain nighttime LST, we used AST_08 LST data for 31 July 2019 based on five thermal infrared band inversion. The absolute accuracy of this product is 1–4 K, and the relative accuracy is 0.3 K [31]. Convert the temperature unit to degrees Celsius using Formula (3), where A is the gray value of the pixel.

(3)$C=A\times 0.1-273.15$

2.3.3. Definition and Classification of SUHI Intensity

The definition of the heat island in this study is expressed as the land surface temperature (LST_K) of each research unit minus the average land surface temperature (LST_T) of all research units belonging to the LCZ_T type in the research area, as shown in Formula (4) [32].

(4)$UH{I}_K=LS{T}_K-LS{T}_T$

The unit with UHI ≤ 0 is divided into cold island area, and the unit with UHI > 0 is divided into four levels: weak heat island area, medium heat island area, strong heat island area, and super heat island area, according to the mean standard deviation method. The specific division rules are shown in Table 2.

2.3.4. Parameter Value Extraction and Statistical Analysis

Based on the LCZ classification grid, the urban form index calculation in each grid is carried out. This study selects building height (BH), building density (BD), impervious surface abundance (ISF), normalized vegetation index (NDVI), albedo, and normalized water index (MNDWI) as indices. These six indices can describe the urban form from three aspects: building form, surface coverage, and surface material. The definition and description of each index are shown in Table 3.

Superimpose the LCZ classification map with the heat island intensity map; calculate and compare the heat island intensities of various types of LCZ; and select the Pearson correlation coefficient r (Formula (5)) to compare and analyze the correlation between various morphological indices and the heat island intensity from the global and different LCZs, where X_ik and X_uk are the values of the shape index and heat island intensity in the nth unit, respectively; ${\overline{X}}_i$ and ${\overline{X}}_u$ are the average values of the shape index and heat island strength, respectively; and n is the number of units.

(5)$r=\frac{{\sum }_{k=1}^n\left(X_{ik}-{\overline{X}}_i\right)\left(X_{uk}-{\overline{X}}_u\right)}{\sqrt{{\sum }_{k=1}^n{\left(X_{ik}-{\overline{X}}_i\right)}^2{\sum }_{k=1}^n{\left(X_{uk}-{\overline{X}}_u\right)}^2}}$

2.3.5. Artificial Neural Network Model

As a general nonlinear function approximation algorithm, an artificial neural network (ANN) is used for machine learning by evaluating and interpreting data relationships [33]. In this study, a three-layer MLP is developed by using a MATLAB neural network toolbox to predict the distribution of the SUHI, and the error reverse transmission algorithm is used [34]. The activation function we used is tanh function, the approximate range of the number of nodes is calculated based on Formulas (6)–(8), and then we obtain the best number of nodes through continuous experiments. Before adding the LCZ, enter six variables. Through experiments, we use four hidden layer nodes to achieve the best results. After adding LCZ, we use one-hot encoded discrete data and input 17 variables, and through repeated experiments, we use 13 nodes to achieve the best performance. This algorithm is called back propagation algorithm for short (i.e., BP algorithm). The multilayer perceptron using the back propagation algorithm is also called BP neural network [35]. Its basic idea include: (1) Calculate the state and activation value of each layer until the last layer (that is, the signal propagates forward). (2) The error of each layer is calculated, and the calculation process is pushed forward from the last layer. (3) Update the parameters (the goal is to reduce the error). The first two steps are iterated until the stop criterion is met. For example, the difference between the errors of the two adjacent iterations is very small. However, the BP neural network has its own limitations; that is, it is easy to fall into the local optimum. This study considers the genetic algorithm [36] to directly use the fitness function as the search information. The search process is not constrained by the continuity of the function. It has good global search capabilities, can overcome the problem of the BP neural network easily falling into local minima, and find the optimal value of the BP neural network quickly and accurately; therefore, the genetic algorithm is used to first optimize the BP neural network and strive to improve the accuracy of the prediction model. The network process is shown in Figure 5:

(6)$l<n-1$

(7)$l<\sqrt{\left(m+n\right)}+a$

(8)$l<{\log }_{2}n$

where $n$ is the number of nodes in the input layer, $l$ is the number of hidden layer nodes, $m$ is the number of output layer nodes, and $a$ is a constant between 0 and 10.

The six indices calculated above are used as the input data of the network, and the SUHI calculated in the selected area is set as the network target. In addition, the LCZ is one-hot encoded and input into the model to compare the improvement of the model. The input data set is randomly divided into three parts for training, verification, and testing. There are 11,853 pixels for training and 5078 pixels for verification and testing.

2.3.6. Evaluation Metrics

The following four metrics are used to evaluate the model [37,38]:

(9)$R^2={\sum }_{i=1}^n{\left(\widehat{y_i}-\overline{y}\right)}^2/{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2,$

(10)$RMSE=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left|{\widehat{y}}_i-y_i\right|}^2} ,$

(11)$MAE=\frac{1}{n}{\sum }_{i=1}^n\left|{\widehat{y}}_i-y_i\right| ,$

(12)$MAPE=\frac{1}{n}{\sum }_{i=1}^n\left|\frac{{\widehat{y}}_i-y_i}{y_i}\right|\times 100\%$

where $y_i$ is the value of the dependent variable, $\widehat{y_i}$ is the predicted value, $\overline{y}$ is the mean of the dependent variable, and n is the number of samples. The determination coefficient (R²) represents the extent to which the regression equation explains the change of dependent variables or how well the equation fits the observed values. The root mean square error (RMSE) is used to measure the deviation between the observed value and the true value. MAPE is more penalized for small values, while MAE is more susceptible to large values.

3. Results and Discussion

3.1. LCZ Mapping Results

According to the LCZ classification results (Figure 6), the built-up blocks in the Fifth Ring Road of Beijing can be mainly divided into 11 categories, of which the highest proportion is dense high-rise buildings (LCZ1), which are mostly concentrated in the Third Ring Road, accounting for 41.6% of the total area. Dense middle-level buildings (LCZ2) are concentrated in the Second Ring Road, with dense low-rise buildings and open high-rise buildings accounting for the same proportion, accounting for 7% each, mostly scattered between the Fourth Ring Road and the Fifth Ring Road. The proportion of open low layer (LCZ6) and large low layer (LCZ8) is the same, accounting for about 10%. The proportion of heavy industrial zone (LCZ10) and very open low floor (LCZ9) is very small. Among the natural cover types, vegetation areas (LCZT) are mostly distributed between the Fourth Ring Road and the Fifth Ring Road, accounting for about 7.8%, and water areas (LCZG) are mostly lakes, accounting for about 1.1%.

It can be seen from Table 4 that the total accuracy of this classification is 94.49%, the kappa coefficient is 0.93, and the producer accuracy and user accuracy of each LCZ category are higher than 85%, indicating that the classification results in this study are highly reliable and meet the application requirements of LCZ.

3.2. Distribution of SUHI

The classification of temperature inversion results by the mean standard deviation method is shown in Figure 7. It can be seen that during the day, the strong heat island areas are relatively concentrated, and 54% of the areas are located in the strong heat island and super heat island areas, and are concentrated in the central and southwest regions. At night, the strong heat island areas are relatively scattered, and 49% of the areas are located in the strong heat island and super heat island areas. Combined with the LCZ classification diagram (Figure 6), the super heat island areas during the day and night are mostly compact high-rise area (LCZ1), compact low-rise area (LCZ3), and large low-rise building area (LCZ8). LCZ10 and LCZG increased significantly at night, which may be due to the effect of night heat rejection in the LCZ10 industrial zone and the specific heat capacity of the water in the LCZG.

It can be seen from Figure 8 that the SUHI of each LCZ is obviously different, which is expressed as compact building area > large low-rise building area > open building area. For the built environment, the SUHI of the compact building area is significantly higher than that of the open type when the building height is equivalent, such as compact high-rise (LCZ1) > open high-rise (LCZ4), compact middle-rise (LCZ2) > open middle-rise (LCZ5). It is mainly due to the fact that the compact building area has a larger proportion of hard ground, more heat absorption, and poor ventilation [10], while the open building area has good air circulation and relatively high vegetation coverage. Both transpiration and daytime shading can help alleviate the heat island effect. In addition, it is found that the corresponding relationship between the SUHI and the height of the building is more complicated. In the case of the same density, the rule of middle floors > low floors > high floors appears. The influence of building height on the SUHI needs to be discussed separately. For high-rise building areas, the SUHI is relatively low due to mutual shielding of buildings, while for low-rise building areas, the SUHI increases with the increase in floors.

3.3. Analysis of the Correlation between Each Index and SUHI

By comparing the SUHIs of various LCZs, the significant differences of SUHIs of different LCZs can be qualitatively reflected. The correlation between selected morphological indices and the SUHI is further discussed based on the Pearson correlation coefficient r. Figure 9 and Figure 10 show the correlation between urban morphology indices and SUHI intensity during the day and at night, respectively (** indicates significance at p < 0.01, * indicates significance at p < 0.05, and unmarked * indicates no significance). From the perspective of the total correlation coefficient, the SUHI has a relatively low correlation with the height of the building and the surface albedo. Among other indices, the SUHI is positively correlated with building density, impervious surface ratio, and MNDWI, indicating that high-intensity land development can significantly enhance the heat island effect; there is a significant negative correlation with the NDVI, indicating that increasing vegetation cover can effectively alleviate the heat island effect [39].

Further study the correlation between the SUHI and various indices in different LCZs. The NDVI has a strong negative correlation in LCZ1–LCZ6 and LCZ8. The correlation between the area with more green space coverage (LCZG) and the SUHI is much smaller than that in other building areas. This shows that in the urban environment, the vegetation space is relatively scattered, and the influence of the surrounding environment (such as impervious surface) has become an important factor affecting the spatial differentiation of the SUHI in areas with large vegetation coverage [40]. Due to the higher heat capacity of the impervious surface, it absorbs more heat after receiving solar radiation, so the proportion of the impervious surface has a significant positive correlation in each LCZ area, and the highest daytime LCZ5 area reaches 0.64. The building density also showed a significant positive correlation in each area, and the dense area (LCZ1–LCZ3) was more correlated than the open area (LCZ4–LCZ6).

The height of buildings in high-rise building areas (LCZ1, LCZ4, etc.) has a significant negative correlation with the SUHI, and the correlation is greater during the day compared with night. For example, the Pearson coefficient is −0.2 during the day and −0.16 at night in LCZ1. The building area presents a positive correlation, even in some areas where the correlation is not significant, such as LCZ5 and LCZ6 in the daytime. It shows that high-rise buildings absorb less heat due to shadow occlusion, while the correlation is relatively small in the middle and low-rise building areas. Previous studies have also proved this conclusion [41].

The high MNDWI value reflects the distribution of surface water bodies and wetlands. The LCZG, MNDWI, and SUHI showed a significant negative correlation in the day and night due to the good cooling effect of water and wetland. The correlation coefficient at night is −0.51, and that during the day is −0.72. The MNDWI and SUHI have a large negative correlation, which is consistent with the conclusions obtained from relevant studies [42]. However, this study found that the cooling effect at night is not as obvious as that in the daytime, especially in other areas where there is no water distribution. The MNDWI values of artificial surfaces and buildings are relatively high, so this index is positively correlated with the SUHI in the construction area. Generally, in water areas, the MNDWI and SUHI are negatively correlated, while in nonwater areas, due to the surface structure, some artificial surfaces are positively correlated between the MNDWI and urban heat, which also proves that some indicators mentioned in previous studies may have different relationships with the correlation of the thermal environment in different ranges [14]. During the day, albedo shows a negative correlation in the building areas. Studies have shown that the decrease in urban albedo by 0.3 will increase the SUHI by about 0.8 k [43]. The correlation is large in the dense building areas (LCZ1–LCZ3); it indicates that the use of high-albedo building materials in these areas can effectively reduce the SUHI. At night, the correlation coefficients are smaller than that in the daytime and are not significant in LCZ4–LCZ6. Building albedo affects daytime surface energy balance in a way that it changes the ratio of reflected to incidence sunlight. Therefore, at night, albedo does not play a role in affecting surface energy balance and surface temperature. Therefore, the correlation between indices and the SUHI during the day and night in different LCZs is very different. When formulating specific thermal environment strategies, they should be formulated according to the characteristics of different LCZs.

3.4. Performance Evaluation of the Development Model

The neural network is trained by the calculated indices to predict the distribution of the SUHI and compare the model with the LCZ and without the LCZ. Figure 11 is the residual diagram of model training, validation, and testing during the day. It can be seen that the residuals of the two models are very close to the best fit line of 45°. The regression R value represents the correlation between the output and the target. An R value of 1 indicates a close relationship, and 0 indicates a random relationship. Both models have a higher R value, which means that both models have better predictive capabilities. After adding the LCZ, the total regression R value increased from 0.87 to 0.91. Similarly, it can be seen from Figure 12 that after the daytime model is added to the LCZ, the RMSE of the training, verification, and testing processes are reduced from 1.66, 1.70, and 1.76 to 1.05, 1.11, and 1.14, respectively. This shows that the addition of the LCZ has expanded the learning range of the model and improved model performance.

Figure 13 shows the residual diagram of the model’s prediction of SUHI distribution at night. It can be seen that the R value is relatively lower than the R value trained in the daytime study area. In addition, Figure 12 can clearly see that the RMSE of the night model is also greater than that of the day model, which shows that the accuracy of the night model is significantly lower than that of the day model. The indices selected in this study have a better ability to explain the distribution of the SUHI during the day and a relatively low ability to explain the distribution of the SUHI at night. The main reason may be that the indices mainly describe the characteristics of buildings and surface coverage in an urban construction area, which is the main influencing factor of SUHI differentiation in the daytime. In summer, with a large solar height angle, long sunshine time, and vigorous vegetation growth, the surface temperature with high heat capacity such as impervious surface and bare land increases rapidly, while the SUHI is significantly differentiated in space in places such as tall buildings and parks due to shadow shading and slow photosynthetic temperature rise of vegetation. At night, due to the weakening of vegetation photosynthesis, the SUHI differentiation mainly comes from the heat released by the artificial surface to the atmosphere and the thermal insulation effect of the atmosphere on the heat during the day, which is mainly related to human activities and the heat storage of suspended particles, such as atmospheric aerosols and building materials. The selected indices do not well reflect this difference, so the prediction accuracy at night is lower than that during the day. After the LCZ is added, it can be seen from Figure 12 that the RMSE decreases and the prediction accuracy is slightly improved, but the change range is not as obvious as that of the daytime model, indicating that the LCZ also has limitations in affecting the differentiation of the nocturnal heat island.

3.5. Model Validation

In order to evaluate the accuracy of the SUHI prediction model and the impact of adding the LCZ to the training process, the developed model was used to predict the SUHI in areas not included in the training process in the study area.

Table 5 shows various evaluation parameters obtained according to the validation area. The determination coefficient R² of the model represents the ability of the model to explain the distribution of the SUHI. It can be seen that R² of the daytime model is significantly increased from 0.74 to 0.81 after adding the LCZ. It means that the new model can explain the change of the SUHI by 81%. The significant increase in R² indicates that the LCZ type has a significant impact on the prediction of SUHI distribution. RMSE, MAPE, and MAE measure the prediction error from different calculation methods. It can be seen that after adding the LCZ, the three parameters are significantly reduced, indicating that the prediction error of the model is reduced and the accuracy of the model for SUHI distribution is improved. In addition, Table 6 shows that the night model has low R² value and relatively high error. After adding the LCZ, the model can only explain 59.5% of the change of SUHI distribution in this area, and other parameters are reduced after adding the LCZ, indicating that the accuracy has been improved.

In Figure 14b–d and Figure 15b–d, the predicted results are basically consistent with the distribution of the SUHI obtained from satellite image inversion. The distribution prediction effect is the best after adding the LCZ in the daytime. The degree of coincidence of distribution at night is not as good as that in the daytime.

According to Figure 14a and Figure 15a, through error analysis for different climate regions (LCZ9 and LCZG are not included in the validation region), it can be seen that there are great differences in prediction errors in different LCZ regions, especially the model without LCZ training. In the prediction of the daytime SUHI, the average errors of LCZ2, LCZ8, and LCZ10 are large, which are 1.51, 1.38, and 1.78, respectively. After LCZ training, the average errors of these areas are significantly reduced, reaching 0.99, 0.79, and 1.13, respectively. It shows that the impact of indices in these regions on the SUHI is more complex than that in other regions. Except for LCZT, the error in other areas decreases after adding the LCZ. It shows that adding LCZ partition can better fit in a complex artificial surface area, which may be caused by two reasons. One reason is that the LCZ adds factors affecting the temperature outside the indices to a certain extent, such as different ventilation conditions in open areas and dense areas, and human activities are included in the medium and heavy industrial areas of the LCZ. Some factors that may affect the intensity of the SUHI are included in the LCZ classification model, which improves the prediction accuracy. The other reason is that the addition of the LCZ makes the fitting effect between the existing indices in each climate region and the SUHI better. Figure 9 and Figure 10 show that the influence mechanism between the indices in different LCZs and SUHIs is different. The addition of zoning results can improve the nonlinear fitting accuracy of the network.

Combined with Figure 15a, the error of each climate region increases in the nighttime prediction results compared with the daytime, and even after the improvement, the error reduction is not obvious compared with the daytime. It shows that more factors need to be considered in night prediction, and the LCZ can slightly reduce the error. In Figure 14a and Figure 15a, the error of LCZT does not decrease after adding the type of climate area, indicating that the LCZ is not significantly helpful to SUHI prediction in a vegetation area, which is a single surface coverage area, and the prediction accuracy of existing indices for the SUHI of LCZT has been high. In the artificial zone, the zone with a relatively better prediction effect of the improved model after adding the LCZ is LCZ6 and LCZ8 in the daytime, and the zone with a better prediction effect at night is LCZ1 and LCZ5.

4. Conclusions

Based on remote sensing images and building data, this study analyzed the spatial distribution characteristics of the day and night SUHI in Beijing and the factors affecting the intensity of the SUHI in each LCZ, and trained a neural network model to compare the prediction accuracy of the model with and without the LCZ. The conclusions are as follows:

(1) The super strong heat island areas in Beijing are concentrated in the central and western regions during the day. The night heat islands are scattered, and the SUHI is generally higher. SUHI_N in Beijing is greater than SUHI_D. During the day and night, the super heat island zone mainly occurs in the compact high-rise zone (LCZ1), compact low-rise zone (LCZ3), and large low-rise building zone (LCZ8). There are big differences in the SUHI between different LCZs. The general rule of the day and night SUHI is compact building zone > large low-rise zone > open zone. Under the same density, the SUHI intensity shows the law of medium floor zone > low floor zone > high floor zone. At night, the SUHI of the open low-rise zone (LCZ6) decreases significantly, LCZ10 and LCZG increase significantly, and the SUHI of other climate zones increases.

(2) In terms of the overall correlation coefficient, the correlation between day and night is consistent, which shows that the correlation between the SUHI and building height and surface albedo is relatively low. There is a significant positive correlation with building density and impervious surface ratio, indicating that high-intensity land development can significantly enhance the heat island effect. There is a significant negative correlation with the NDVI, indicating that increasing vegetation cover in cities can effectively alleviate the heat island effect. It is found that the correlation between the SUHI and various indices is quite different in various LCZs, and the correlation of individual indices is different in the day and night. When formulating specific thermal environment optimization strategies, it should be adjusted according to the type of LCZ. As an example, the building density in the dense area (LCZ1–LCZ3) is more correlated than that in the open area (LCZ4–LCZ6); the building height is significantly negatively correlated with the SUHI in the high-rise building area, and it is positively correlated in the middle- and low-rise building area. Albedo has a greater correlation in dense areas during the day (LCZ1–LCZ3), while at night, the correlation between albedo and the SUHI decreases. In LCZ4–LCZ6, the correlation is not even significant. It shows that the use of high-albedo building materials can effectively reduce the surface heat island effect during the day, but the effect on the night heat island is limited.

(3) After adding the LCZ to the daytime model, the improved model’s ability to explain the SUHI distribution (R²) was significantly increased from 0.74 to 0.81, and the RMSE was reduced from 1.66, 1.70, and 1.76 to 1.05, 1.11, and 1.14 during training, verification, and testing, respectively, which shows that the addition of the LCZ improves the prediction accuracy of the model. The night model has a low R² and a large error. After adding the LCZ, there is only a slight improvement, indicating that the selected city morphology indices are less explanatory for the night SUHI distribution than during the day. After adding the LCZ, the artificial areas with the best prediction effect during the day of the improved model are LCZ6 and LCZ8, and the areas with the best prediction effect at night are LCZ1 and LCZ5. In the daytime SUHI forecast, the average error of each area (except LCZT) increases after adding the LCZ. Among them, LCZ2, LCZ8, and LCZ10 are the most obvious improvement compared with other LCZ regions, indicating that the consideration of regional heterogeneity has significantly promoted the fitting and prediction of urban morphology and the SUHI in these regions. Compared with the daytime prediction results, the errors in all LCZs at night have increased. Even after the improvement, the error reduction is not obvious compared with that in the daytime. This shows that night prediction needs to consider more factors, and the LCZ can only slightly reduce the error. The model for predicting the distribution of the SUHI constructed by this research provides a reference for the planning of mitigating the intensity of the SUHI. It also illustrates the advantages and necessity of adding the concept of LCZs to consider spatial heterogeneity when analyzing and predicting the intensity of the SUHI.

Based on the above conclusions, it can be concluded that spatial heterogeneity plays an important role in the study of the heat island effect. Since the traditional correlation between SUHI intensity and urban form mainly focuses on the global linear regression model, less attention is paid to the spatial heterogeneity of the role of thermal environment factors in different regions. This study has made a systematic analysis of the relationship between spatial heterogeneity and the urban heat island. The correlation of the SUHI and urban form in different LCZs is analyzed, and the comparison between day and night is added. The test is carried out by inputting the regional classification into the neural network model. Through the analysis, it is found that considering the spatial heterogeneity, we can better understand the relationship between SUHI intensity and different urban morphological parameters in different morphological regions. According to the simulation results of the constructed model, the model with the LCZ can more accurately simulate the thermal environment, especially in the daytime. The above conclusions can provide some reference value for the related research of the SUHI.

This study also has some limitations. Because the research focuses on the SUHI in summer, the available remote sensing images are limited. The study of the SUHI in the day and night is only based on one image, which is lack of universality to a certain extent. Additionally, in the selection of urban form indicators, the types of indicators can continue to expand. Further research should screen and increase the indicators of the nighttime heat island effect, and study more influencing factors of the nighttime heat island. A series of studies will be conducted on more regions and seasons to explore universal laws and patterns. In view of the difference in the prediction effect of this model on the SUHI of different LCZs, it is necessary to further analyze the reasons.

Author Contributions

Methodology, Y.X.; software, Y.X.; validation, Y.X. and C.Z.; writing—original draft preparation, Y.X.; writing—review and editing, W.H. and C.Z.; visualization, Y.X.; supervision, C.Z.; project administration, C.Z.; funding acquisition, W.H. and C.Z. All authors have read and agreed to the published version of the manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Figures and Tables

Figure 1. The location of the study area.

Figure 2. Semivariogram modelling of building height.

Figure 3. Research methods.

Figure 4. Samples of the LCZ from Google Earth: (a) LCZ1: compact high-rise; (b) LCZ2: compact midrise; (c) LCZ3: compact low-rise; (d) LCZ4: open high-rise; (e) LCZ5: open midrise; (f) LCZ6: open low-rise; (g) LCZ8: large low-rise; (h) LCZ9: sparsely built; (i) LCZ10: heavy industry; (j) LCZT: vegetation; (k) LCZG: water.

Figure 5. Flowchart of the BP neural network optimized by the genetic algorithm.

Figure 6. Spatial distribution of the LCZ in the study area (within the fifth ring of Beijing).

Figure 7. Distribution diagram of SUHI grade.

Figure 8. Box plots with SUHIs in LCZs: The line within the box indicates the median. The rectangular indicates the average. The bottom of the box is the first quartile and the top is the third quartile. The connecting line is connected to each average value.

Figure 9. Pearson correlation coefficients between the urban morphology indices and SUHID. (** The significance level is 0.01, * the significance level is 0.05).

Figure 10. Pearson correlation coefficients between the urban morphology indices and SUHIN. (** The significance level is 0.01, * the significance level is 0.05).

Figure 11. Daytime model performance results.

Figure 12. Comparison of model RMSE values.

Figure 13. Night model performance results.

Figure 14. Comparison and distribution of daytime model errors.

Figure 15. Comparison and distribution of nighttime model errors.

Table 1

Data sources and descriptions.

Types	Resolution	Purpose	Data Source
Landsat 8 image	30 m	Calculate NDVI, daytime surface temperature, and other indices	earth.explorer.usgs.gov
Building data	vector data	Calculate building height and building density	Baidumap
Impervious surface data	30 m	Calculate impervious rate	data.ess.tsinghua.edu.cn
Google Earth image	0.92 m	Select training samples	Google Earth Pro
AST_08 surface temperature products	90 m	Calculate night surface temperature	search.earthdata.nasa.gov

Table 2

Classification of SUHI intensity based on the mean standard deviation method.

Heat Island Strength Grade	Partition Interval
Cold island region	$UHI\le 0$
Weak heat island region	$0\le UHI\le u-std$
Medium heat island region	$u-std<UHI\le u$
Strong heat island region	$u<UHI\le u+std$
Super heat island region	$UHI>u+std$

μ and std are the average and variance of the heat island intensities of all units, respectively.

Table 3

Definition and description of urban form indices.

Parameters	Basic Data	Definition	Meaning of Thermal Environment
H	Building data	Average value of building height per unit surface	Reflect local airflow and heat dissipation
BD	Building data	Ratio of building base area per unit surface to unit surface area	Reflect surface radiation heat gain, local air flow, and heat dissipation
ISF	Remote sensing data	Proportion of impervious surface coverage in unit surface	Reflect surface radiant heat gain and surface runoff
Albedo	Remote sensing data	Ratio of the emitted energy to the incident energy of the target	Reflectivity of the earth’ s surface to solar radiation
NDVI	Remote sensing data	Proportion of vegetation coverage per unit surface	Reflect heat from surface radiation and conversion of material and energy
MNDWI	Remote sensing data	Proportion of water body coverage per unit surface	Reflect surface water and evaporation cooling

Table 4

Confusion matrix of LCZ classification results.

LCZ	1	2	3	4	5	6	8	9	10	T	G	Total	UA (%)
1	199	2	1	2	1	0	3	0	0	1	1	210	94. 8
2	0	20	0	0	0	0	0	0	0	0	0	20	100.0
3	0	0	38	0	0	0	2	0	0	0	0	40	95.0
4	0	0	0	37	1	0	3	0	1	0	1	43	86.1
5	0	0	0	0	20	0	0	0	0	0	0	20	100.0
6	0	0	1	0	1	18	0	0	0	0	0	20	90.0
8	0	0	2	1	0	1	60	0	0	0	0	64	93.8
9	0	0	0	0	0	0	0	10	0	0	0	10	100.0
10	0	0	0	0	0	0	0	0	9	0	0	9	100.0
T	0	0	0	0	0	2	0	0	0	40	1	43	93.0
G	0	0	0	0	0	0	0	0	0	0	29	29	100.0
Total	199	22	42	40	23	21	68	10	10	41	32	508
PA (%)	100.0	91.0	90.5	92.5	87.0	85.7	88.2	100.0	90.0	97.6	90.6
Total accuracy = 94.49%, kappa coefficient = 0.93

Table 5

Comparison of SUHI_D prediction error parameters with and without the LCZ model.

	R 2	RMSE	MAE	MAPE
WO-LCZ	0.745	1.318	1.003	2.61%
W-LCZ	0.817	1.076	0.821	2.15%

Table 6

Comparison of SUHI_N prediction error parameters with and without the LCZ model.

	R 2	RMSE	MAE	MAPE
WO-LCZ	0.563	2.149	1.608	3.75%
W = LCZ	0.595	2.062	1.525	3.58%

References

1. Howard, L. The Climate of London, Deduced from Meteorological Observations; Cambridge University Press: Cambridge, UK, 2012.

2. Georgescu, M.; Moustaoui, M.; Mahalov, A.; Dudhia, J. Summer-time climate impacts of projected megapolitan expansion in Arizona. Nat. Clim. Chang.; 2013; 3, pp. 37-41. [DOI: https://dx.doi.org/10.1038/nclimate1656]

3. Zheng, H.; Chen, Y.; Ding, W.; Pan, W.; Cai, Y. Characteristics of Land Surface Temperature Disturbance and Its Relationship to Built-up Land Expansion. J. Geo-Inf. Sci.; 2018; 20, pp. 1529-1540.

4. McCarthy, M.P.; Best, M.J.; Betts, R.A. Climate change in cities due to global warming and urban effects. Geophys. Res. Lett.; 2010; 37, L09705. [DOI: https://dx.doi.org/10.1029/2010GL042845]

5. Arshad, S.; Ahmad, S.R.; Abbas, S.; Asharf, A.; Siddiqui, N.A.; Islam, Z.U. Quantifying the contribution of diminishing green spaces and urban sprawl to urban heat island effect in a rapidly urbanizing metropolitan city of Pakistan. Land Use Policy; 2021; 113, 105874. [DOI: https://dx.doi.org/10.1016/j.landusepol.2021.105874]

6. Radfar, M. Urban Microclimate, Designing the Spaces Between Buildings. Hous. Stud.; 2012; 27, pp. 293-294. [DOI: https://dx.doi.org/10.1080/02673037.2011.615987]

7. Zhang, J.; Li, Z.; Hu, D. Effects of urban morphology on thermal comfort at the micro-scale. Sustain. Cities Soc.; 2022; 86, 104150. [DOI: https://dx.doi.org/10.1016/j.scs.2022.104150]

8. Guo, G.; Zhou, X.; Wu, Z.; Xiao, R.; Chen, Y. Characterizing the impact of urban morphology heterogeneity on land surface temperature in Guangzhou, China. Environ. Model. Softw.; 2016; 84, pp. 427-439. [DOI: https://dx.doi.org/10.1016/j.envsoft.2016.06.021]

9. Wang, Y.; Zhan, Q.; Ouyang, W. How to quantify the relationship between spatial distribution of urban waterbodies and land surface temperature?. Sci. Total. Environ.; 2019; 671, pp. 1-9. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2019.03.377]

10. Yin, C.; Yuan, M.; Lu, Y.; Huang, Y.; Liu, Y. Effects of urban form on the urban heat island effect based on spatial regression model. Sci. Total Environ.; 2018; 634, pp. 696-704. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2018.03.350]

11. Feng, Y.; Du, S.; Myint, S.W.; Shu, M. Do Urban Functional Zones Affect Land Surface Temperature Differently? A Case Study of Beijing, China. Remote Sens.; 2019; 11, 1802.

12. Zhang, H.; Qi, Z.-F.; Ye, X.-Y.; Cai, Y.-B.; Ma, W.-C.; Chen, M.-N. Analysis of land use/land cover change, population shift, and their effects on spatiotemporal patterns of urban heat islands in metropolitan Shanghai, China. Appl. Geogr.; 2013; 44, pp. 121-133. [DOI: https://dx.doi.org/10.1016/j.apgeog.2013.07.021]

13. Stewart, D.; Orke, T.R. Local climate zones for urban temperature studies. Bull. Am. Meteorol. Soc.; 2012; 93, pp. 1879-1900. [DOI: https://dx.doi.org/10.1175/BAMS-D-11-00019.1]

14. Bechtel, B.; Alexander, P.J.; Böhner, J.; Ching, J.; Conrad, O.; Feddema, J.; Mills, G.; See, L.; Stewart, I. Mapping Local Climate Zones for a Worldwide Database of the Form and Function of Cities. ISPRS Int. J. Geo-Inf.; 2015; 4, pp. 199-219. [DOI: https://dx.doi.org/10.3390/ijgi4010199]

15. Ching, J.; Mills, G.; Bechtel, B.; See, L.; Feddema, J.; Wang, X.; Ren, C.; Brousse, O.; Martilli, A.; Neophytou, M. et al. WUDAPT: An Urban Weather, Climate, and Environmental Modeling Infrastructure for the Anthropocene. Bull. Am. Meteorol. Soc.; 2018; 99, pp. 1907-1924. [DOI: https://dx.doi.org/10.1175/BAMS-D-16-0236.1]

16. Badaro-Saliba, N.; Adjizian-Gerard, J.; Zaarour, R.; Najjar, G. LCZ scheme for assessing Urban Heat Island intensity in a complex urban area (Beirut, Lebanon). Urban Clim.; 2021; 37, 100846. [DOI: https://dx.doi.org/10.1016/j.uclim.2021.100846]

17. O’Malley, C.; Kikumoto, H. An investigation into heat storage by adopting Local Climate Zones and nocturnal-diurnal urban heat island differences in the Tokyo Prefecture. Sustain. Cities Soc.; 2022; 83, 103959. [DOI: https://dx.doi.org/10.1016/j.scs.2022.103959]

18. Yang, J.; Ren, J.; Sun, D.; Xiao, X.; Xia, J.C.; Jin, C.; Li, X. Understanding land surface temperature impact factors based on local climate zones. Sustain. Cities Soc.; 2021; 69, 102818. [DOI: https://dx.doi.org/10.1016/j.scs.2021.102818]

19. Hu, J.; Yang, Y.; Pan, X.; Zhu, Q.; Zhan, W.; Wang, Y.; Ma, W.; Su, W. Analysis of the Spatial and Temporal Variations of Land Surface Temperature Based on Local Climate Zones: A Case Study in Nanjing, China. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2019; 12, pp. 1-11. [DOI: https://dx.doi.org/10.1109/JSTARS.2019.2926502]

20. Cai, M.; Ren, C.; Xu, Y. Investigating the relationship between local climate zone and land surface temperature. Proceedings of the 2017 Joint Urban Remote Sensing Event (JURSE); Dubai, United Arab Emirates, 6–8 March 2017.

21. Simanjuntak, R.M.; Kuffer, M.; Reckien, D. Object-based image analysis to map local climate zones: The case of Bandung, Indonesia. Appl. Geogr.; 2019; 106, pp. 108-121. [DOI: https://dx.doi.org/10.1016/j.apgeog.2019.04.001]

22. Peng, S.; Piao, S.; Ciais, P.; Friedlingstein, P.; Ottle, C.; Bréon, F.-M.; Nan, H.; Zhou, L.; Myneni, R.B. Response to Comment on “Surface Urban Heat Island Across 419 Global Big Cities”. Environ. Sci. Technol.; 2012; 46, pp. 6889-6890. [DOI: https://dx.doi.org/10.1021/es301811b]

23. Zhang, H.; Hanqiu, X.U.; Le, L.I.; Fan, Y.P. Analysis of the relationship between urban heat island effect and urban expansion in Chengdu, China. J. Geo-Inf. Sci.; 2014; 16, pp. 70-78.

24. Chen, Y.; Peng, R.C.; Zhang, L.H.; Dong, J.; Ning, L.I. Application of semi-variogram in analyzing spatial construction of sounding field. Hydrogr. Surv. Charting; 2011; 06, pp. 63-65.

25. Lau, K.; Chao, R.; Yuan, S.; Zheng, V.; Lai, D. Determining the optimal size of local climate zones for spatial mapping in high-density cities. Proceedings of the 9th International Conference on Urban Climate (ICUC9); Toulouse, France, 20–24 July 2015.

26. Zheng, Y.; Ren, C.; Xu, Y.; Wang, R.; Ho, J.; Lau, K.; Ng, E. GIS-based mapping of Local Climate Zone in the high-density city of Hong Kong. Urban Clim.; 2017; 24, pp. 419-448. [DOI: https://dx.doi.org/10.1016/j.uclim.2017.05.008]

27. Conrad, O.; Bechtel, B.; Bock, M.; Dietrich, H.; Fischer, E.; Gerlitz, L.; Wehberg, J.; Wichmann, V.; Böhner, J. System for Automated Geoscientific Analyses (SAGA) v. 2.1.4. Geosci. Model Dev.; 2015; 8, pp. 2271-2312. [DOI: https://dx.doi.org/10.5194/gmd-8-1991-2015]

28. Han-Qiu, X.U. A study on information extraction of water body with the modified normalized difference water index (MNDWI). J. Remote Sens.; 2005; 9, 595.

29. Gillespie, A.; Rokugawa, S.; Matsunaga, T.; Cothern, J.S.; Hook, S.J.; Kahle, A.B. A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images. IEEE Trans. Geosci. Remote Sens.; 1998; 36, pp. 1113-1126. [DOI: https://dx.doi.org/10.1109/36.700995]

30. Qin, Z.; Li, W.; Gao, M.; Zhang, H. Estimation of land surface emissivity for Landsat TM6 and its application to Lingxian region in North China. Proceedings of the Remote Sensing for Environmental Monitoring, GIS Applications, and Geology VI; Stockholm, Sweden, 3 October 2006; 6366.

31. Song, J.; Wang, Z.-H.; Myint, S.W.; Wang, C. The hysteresis effect on surface-air temperature relationship and its implications to urban planning: An examination in Phoenix, Arizona, USA. Landsc. Urban Plan.; 2017; 167, pp. 198-211. [DOI: https://dx.doi.org/10.1016/j.landurbplan.2017.06.024]

32. García, D.H.; Díaz, J.A. Modeling of the Urban Heat Island on local climatic zones of a city using Sentinel 3 images: Urban determining factors. Urban Clim.; 2021; 37, 100840. [DOI: https://dx.doi.org/10.1016/j.uclim.2021.100840]

33. Li, B.; Wang, N.; Chen, J. GIS-Based Landslide Susceptibility Mapping Using Information, Frequency Ratio, and Artificial Neural Network Methods in Qinghai Province, Northwestern China. Adv. Civ. Eng.; 2021; 2021, pp. 1-14. [DOI: https://dx.doi.org/10.1155/2021/4758062]

34. Equere, V.; Mirzaei, P.A.; Riffat, S.; Wang, Y. Integration of topological aspect of city terrains to predict the spatial distribution of urban heat island using GIS and ANN. Sustain. Cities Soc.; 2021; 69, 102825. [DOI: https://dx.doi.org/10.1016/j.scs.2021.102825]

35. Liu, S.Q. Research and application on matlab BP neural network. Comput. Eng. Des.; 2003; 11, pp. 81-83.

36. Sun, Y.; Xue, B.; Zhang, M.; Yen, G.G.; Lv, J. Automatically Designing CNN Architectures Using the Genetic Algorithm for Image Classification. IEEE Trans. Cybern.; 2020; 50, pp. 1-15. [DOI: https://dx.doi.org/10.1109/TCYB.2020.2983860] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32324588]

37. Aras, S.; Kocakoç, I.D.; Polat, C. Comparative study on retail sales forecasting between single and combination methods. J. Bus. Econ. Manag.; 2017; 18, pp. 803-832. [DOI: https://dx.doi.org/10.3846/16111699.2017.1367324]

38. Davydenko, A.; Fildes, R. Measuring forecasting accuracy: The case of judgmental adjustments to SKU-level demand forecasts. Int. J. Forecast.; 2013; 29, pp. 510-522. [DOI: https://dx.doi.org/10.1016/j.ijforecast.2012.09.002]

39. Estoque, R.C.; Murayama, Y. Classification and change detection of built-up lands from Landsat-7 ETM+ and Landsat-8 OLI/TIRS imageries: A comparative assessment of various spectral indices. Ecol. Indic.; 2015; 56, pp. 205-217. [DOI: https://dx.doi.org/10.1016/j.ecolind.2015.03.037]

40. Zhang, X.; Zhong, T.; Feng, X.; Wang, K. Estimation of the relationship between vegetation patches and urban land surface temperature with remote sensing. Int. J. Remote Sens.; 2009; 30, pp. 2105-2118. [DOI: https://dx.doi.org/10.1080/01431160802549252]

41. Xi, C.; Ren, C.; Wang, J.; Feng, Z.; Cao, S.-J. Impacts of urban-scale building height diversity on urban climates: A case study of Nanjing, China. Energy Build.; 2021; 251, 111350. [DOI: https://dx.doi.org/10.1016/j.enbuild.2021.111350]

42. Ramaiah, M.; Avtar, R.; Rahman, M. Land Cover Influences on LST in Two Proposed Smart Cities of India: Comparative Analysis Using Spectral Indices. Land; 2020; 9, 292. [DOI: https://dx.doi.org/10.3390/land9090292]

43. Liang, S.L.; Shuey, C.J.; Russ, A.L.; Fang, H.L.; Chen, M.Z.; Walthall, C.L.; Daughtry, C.S.T.; Hunt, R. Narrowband to broadband conversions of land surface albedo: II. Validation. Remote Sens. Environ.; 2003; 84, pp. 25-41. [DOI: https://dx.doi.org/10.1016/S0034-4257(02)00068-8]