4.2. Model Fitting and Performances
Due to the heterogeneity of UFZs, the results of model fitting vary hugely. For each model, the model parameters (Alpha and L1_Ratio) are set by K-fold cross-validation to achieve the best model fitting (
Table 7). After modeling, an evaluation of the models’ results is conducted to check if these models are reliable and worth investigating in the future. In this study, the models are evaluated by the goodness of fit and spatial autocorrelation of the model residual.
The goodness of fit of the seven models is listed in
Table 8. The values of
of all the seven models are above 0.4, while the goodness of fit is low in some models for two reasons. One is caused by the heterogeneity among UFZs, and the other is that the penalty constraint term of the regression shrinkage methods that reduce the goodness of fit to keep the model stable. Model-A and Model-CZ have a low goodness of fit, since they are based on highly heterogeneous units. In contrast to parks with a unified style and residential zones with regular building patterns, commercial zones lack both unified styles and spatial patterns, not to mention the more complex UFZs in Model-A, which includes different kinds of UFZs. As for the relatively homogeneous models, like Model-IZ and Model-ST, their performances are much better, as they have large values of
(0.94 and 0.75, respectively). This validates the necessity of modeling the LST–landscape relationships for each UFZ category.
Before modeling, a spatial autocorrelation analysis of the LST is first conducted. For LST, the Z-value of Moran’s I is 5.66 and the P-value is 0.00, indicating that the LST has a strong spatial autocorrelation and has a significantly clustered distribution. If the model can accurately explain the LST–landscape relationship, the model residuals should no longer show spatial autocorrelation but random distribution. The autocorrelation indices of the seven models are shown in
Table 9.
The Z-value (3.43) of Model-A indicates that the residuals are still significantly auto-correlated. As a result, the mixed model (Model-A) is weak and inappropriate to explain the LST–landscape relationship. In contrast, the spatial autocorrelations of the residuals are weakened to varying degrees in six single models. The model residuals of industrial zones, campuses, parks, and shanty towns are not auto-correlated, indicating that these models perform well. The model residuals for commercial zones and residential districts show a tendency toward aggregation, but their aggregation degree is much weaker than the LST itself, owing to the good model fittings. In conclusion, in terms of the spatial autocorrelation of the residuals, the six single models for each UFZ type explain the LST–landscape relationship to varying degrees, and the coefficients of explanatory variables can be trusted. The performance of the mixed model (Model-A) is poor, but it still partly reflects the general laws in the urban areas and reveals that different UFZ kinds may have different LST effects and need to be modelled and analyzed separately.
4.4. Result Analysis of the UFZ Level
The variables of UFZs describe the basic attributes of the UFZs, landscape composition, and configuration within UFZs, as well as the interrelationships between neighboring UFZs. First, as the basic attributes of UFZs, area, perimeter, and SI are only included in Model-A and Model-P, but not in the other five single models. In Model-A, the coefficient of area (−0.04) suggests that the areas of UFZs are negatively related to the LST, while perimeter and SI are regarded to have minimal effects on LST, as their coefficients are set as 0.0. In Model-P, the coefficients of area (−0.16), perimeter (−0.02), and SI (−0.02) indicate that a large area, large perimeter, and complex shape could reduce the LST in parks.
Second, CONTAG embodies the distribution of patches within UFZs. A greater CONTAG value, which signifies more clustered patches, will lead to a lower LST in residential districts and parks, with the negative coefficients (−0.10 and −0.12) in Model-RD and Model-P respectively. However, in shanty towns, a greater CONTAG value indicates a stronger heating effect, with the positive coefficient (0.07) in Model-ST. In addition, the LSTs in commercial zones, industrial zones, and campuses are not affected by CONTAG.
Third, the diversity of the patches in the UFZs is measured mainly by SIHI and SDHI. The coefficients of SIHI in Model-CZ, Model-IZ, and Model-ST are −0.04, −0.05, and −0.03, respectively. The coefficients of SDHI in Model-CZ and Model-ST are −0.06 and −0.06, respectively. Moreover, these two variables are not significant in Model-A, Model-RD, Model-C, and Model-P. The negative coefficients show the patch diversity in the three UFZ types (i.e., commercial zones, industrial zones, and shanty towns) is negatively related to the LST, which can alleviate the increasing of LST.
Lastly, the results indeed show that the LST of a UFZ is greatly affected by its neighbors. UFZs might have heating or cooling effects on the neighboring UFZs, depending on the adjoined UFZ types. Like parks, the coefficient of Adjacency-P in Model-RD (−0.09) shows that the LST will decrease in residential districts adjoined to parks, while the coefficient of Adjacency-P in Model-C (0.03) indicates that the LST will increase in campuses adjoined to parks. Among these UFZ types, shanty towns have the most significant impact on the LST of adjoined UFZs.
Table 11 shows that adjoining to shanty towns is not a good thing for almost all UFZ types. The positive coefficients of Adjacency-ST in the models, except for Model-C, reflect the positive relationships between the adjacency and the LST. However, neighboring campuses and shanty towns can reduce the LST for each other, embodied by the negative coefficients of Adjacency-ST in Model-P (−0.04) and Adjacency-P in Model-ST (−0.08). Moreover, in the interrelationship analysis (adjacent to its own type of UFZs) no longer reflects the adjacency relationship but the effect of size. Two adjacent UFZs can be regarded as a larger one. For residential districts, campuses, and parks, the adjacency variables of their own type are all negative (−0.03, −0.01, and −0.09), showing the cooling effects in different degrees. On the other hand, two adjoining shanty towns will both increase their LSTs, as revealed by the coefficient of Adjacency-ST in Model-ST (0.08).
4.5. Result Analysis of Land-Cover Level
The land-cover types contained in each UFZ are similar, but due to the differences in landscape composition, configuration, and human activities, the same land-cover type may have different LST effects and intensities on different kinds of UFZs. Next, we will show the different roles of six land-cover types in different models.
1. The variables of buildings. Four basic variables (i.e., BLD-ED, BLD-IJI, BLD-PD, and BLD-PLAND) depict the attributes of buildings the same as the attributes of other ground objects. Five more variables are added to characterize their spatial patterns, including the area standard deviation of buildings (BLD-AREA_SD), the number of buildings (NB), the number of building chains (NC), the length sum of all the chains (LC), and the ratio of buildings in chains to the total (RB).
From the results of each model, the LST effects of buildings are complex and diverse. The effects of buildings are most significant in the residential zones. There are five building variables significantly related to the LST, while commercial areas and parks have only one related variable.
In Model-A, the BLD-AREA_SD and BLD-PD tend to increase LST. For commercial zones, the LST is only increased by BLD-AREA_SD. For industrial zones, BLD-ED, BLD-PLAND, and NB are all positively related with the LST. Conversely, variables about buildings in Model-C and Model-P play a part in cooling the LST. The result of Model-C shows that BLD-AREA_SD and BLD-PLAND have a negative association with the LST in campuses, and the result of Model-P shows that NB is the only significant variable of the LST that shows parks with more buildings to be cooler.
In Model-ST, the impact of buildings on LST mainly emerges from three aspects: BLD-AREA_SD, BLD-PLAND, and BLD-IJI, with coefficients 0.05, 0.01, and −0.01, respectively. The former two are positively related with the LST, and the last one is negatively related with the LST. On the other hand, the absolute value of the coefficients of the last two (0.01 for BLD-PLAND and −0.01 for BLD-IJI) are very small, indicating the weakness of their influence.
The results of Model-RD show that the LST of residential districts is greatly influenced by the variables of buildings. The coefficients of BLD-AREA_SD (−0.04) and BLD-PLAND (−0.04) reflect their cooling effects while the coefficients of NB (0.13), NC (0.04), and RB (0.01) show that these variables have heating effects on LST.
2. The variables of vegetation. In most models, VEG-IJI has a negative relationship with LST and embodies the dispersion of vegetation. In Model-A, Model-CZ, Model-RD, Model-IZ, and Model-ST, cooling effects occur in different degrees. VEG-ED and VEG-PD also have cooling effects on shanty towns. On the other hand, VEG-ED in parks has heating effects on the LST.
3. The variables of roads. For all the models, Road-PLAND is positively related to the LST, suggesting that increasing road area could cause the LST to rise. In Model-A and Model-P, a greater Road-PD value could reduce the LST. In Model-IZ, the increasing Road-ED value could result in a large LST. As for Road-IJI, it affects LST differently in different kinds of UFZs. A greater Road-IJI value means a higher LST in residential districts, campuses and parks, and a lower LST in industrial zones and shanty towns, but nothing in commercial zones or the whole urban area.
4. The variables of shadows. The modeling results show that shadows are the most effective among all land cover types in terms of lowering the LST by the negative coefficients of Shadow-PLAND (−0.26, −0.20, −0.22, 0, −0.14, −0.13, and −0.03). For all the models, most variables of shadows are negatively related with the LST, indicating the strong cooling effects of shadows, except for Shadow-IJI in Model-RD, with a positive coefficient (0.16). This phenomenon may result from the mixed effects of the building warming LST and shadow lowering LST, which will be further discussed in
Section 5.3.
5. The variables of bare soil. For all the models, most variables of bare soil are positively related with the LST, especially Bare Soil-PLAND, showing significant heating effects on the LST in every type of UFZ. However, Bare Soil-IJI in Model-ST and Bare Soil-ED in Model-RD have negative relationships with the LST.
6. The variables of the waterbody. Model-A, Water-ED, Water-IJI, Water-PD, and Water-PLAND show cooling effects on the LST. However, a waterbody is not a common ground object type in UFZs except parks. Many UFZs do not have a waterbody, not to mention the influence of waterbodies on the LST. Therefore, the variables related with the waterbody are only discussed for parks. For parks, Water-ED and Water-PLAND are negatively related with the LST. Notably, Water-PLAND has a coefficient with a great absolute value (−0.18), showing that the Water-PLAND in parks has a strong cooling effect on LST.