Diurnal–Seasonal Contrast of Spatiotemporal Dynamic and the Key Determinants of Surface Urban Heat Islands Across China’s Humid and Arid Regions

Abstract

Regional management of the urban thermal environment is essential for sustainable development. However, both the surface urban heat island (SUHI) spatiotemporal patterns and driving mechanisms across humid–arid regions remain uncertain. Therefore, 329 cities from various humid–arid regions were selected to investigate the interannual, seasonal, and diurnal distribution characteristics of SUHIs across regions. By constructing six-dimensional influencing factors and using CatBoost-SHAP and SEM methods, the contributions and action pathways of these factors to SUHIs were analyzed across humid–arid regions. The influence mechanisms, differences in feature importance, and similarities and discrepancies in action pathways were thoroughly examined. The findings are as follows: 1. During the day, higher SUHII values occur in humid and semihumid regions, exceeding those in arid and semiarid regions by 1.521 and 0.921, respectively. At night, arid and semiarid regions exhibit UHI effects (SUHII > 0). The SUHI distribution across humid–arid regions demonstrates seasonal variations. 2. ΔSA and ΔNDVI are stable dominant influencing factors across all regions. The contribution rank varies along the humid–arid region: Pollution factors are more important in arid and semiarid regions, whereas surface features and 2D/3D dominate in humid and semihumid regions at night. 3. SUHI regulation by influencing factors across humid–arid regions follows both similar paths and regional variations. This study reveals the SUHI distribution across humid–arid regions and provides reference data for regional thermal environment management.

1. Introduction

Rapid urban development has significantly altered local climates [1]. The expansion of impervious surfaces that accompanies rapid urbanization has many negative ecological impacts [2]. One of the most prominent consequences is the urban heat island effect, in which temperatures in urban areas are generally higher than in the surrounding countryside [3]. This localized urban heat effect can cause multiple adverse effects, including increased mortality among vulnerable groups [4], increased energy demand, and peak load stress on power infrastructure [5], all of which can ultimately impede urban sustainability. The UHI shows heterogeneity in different climate zones, urban agglomerations, or city size. Hence, understanding the spatiotemporal distribution of urban heat islands in different zones, as well as identifying key drivers of UHIs, has theoretical and practical importance.

SUHIs and their driving mechanisms have been extensively studied. Researchers have explored SUHIs across different climatic zones [6] and have examined SUHI spatial patterns for different gross domestic product (GDP) levels, landforms and population sizes [7]. More importantly, the spatial distribution of SUHI across humid–arid regions is complicated by inconsistent findings. Refs. [8,9] reported that SUHIs are more notable in humid regions than in arid regions, whereas Peng reported the opposite conclusion. The mechanisms underlying SUHI variations across humid–arid regions have not been examined. Moreover, a systematic investigation into the spatial heterogeneity of SUHI driving mechanisms through the insight of humid–arid climate zoning is still lacking. Therefore, the exploration of the distribution and mechanisms underlying SUHIs in different humid–arid regions is crucial.

SUHIs are influenced by multiple factors, among which background climate [10] and surface features [11] are particularly critical at the macroscale. Regarding background climate, increased wind speeds can reduce temperature differences [12]; Yang reported that the effect of precipitation on the heat island intensity is nonlinear [13]. However, when surface features are considered, the proportions of impervious surfaces and water areas are closely related to SUHIs [14]; water can help mitigate SUHIs in megacity clusters [15]; surface albedo affects SUHIs; and a decrease in albedo increases solar radiation absorption [16]. The cooling effect of vegetation cover is particularly notable during the summer season [17].

From a socioeconomic development perspective, the population density and nighttime light distribution influence the heat island intensity [18]. Furthermore, urban air pollution and UHIs exhibit spatial consistency [19]. Notably, this study specifically focuses on the influence of urban morphology on SUHII. Urban morphology significantly affects the thermal environment [20,21]. Both two-dimensional (2D) factors (such as patch density, landscape fragmentation, and connectivity) and 3D factors (such as building height and volume) interact with urban land surface temperature [22,23,24]. These factors yield substantial contributions at both the city [25] and regional [26,27] levels. With respect to the importance of 2D and 3D factors, different scholars hold different views. Ref. [28] proposed that 3D factors exhibit superior performance, whereas [29] suggested that the importance of 2D factors is modulated by seasonal variations. More importantly, precipitation, vegetation, and development differ systematically across humid and arid regions. These differences change the baseline conditions of key factors. As a result, their pathways and relative contributions to SUHIs also change. Identifying region-specific pathways helps reveal regional heterogeneity in SUHI formation. It also provides a scientific basis for targeted thermal environment management in each climatic zone.

In recent UHI attribution studies, machine learning methods and artificial intelligence technologies have been adopted [30,31]. The SHapley Additive exPlanations (SHAP) technique provides a framework for interpreting complex machine learning models in which feature contributions are quantified [32,33]. SHAP values represent the positive or negative contribution of each feature to the prediction outcome, increasing model interpretability. Structural equation modeling (SEM) can simultaneously resolve the complex relationships among multiple dependent and independent variables.

Existing studies have increased the understanding of SUHI drivers and attribution methods. However, the distribution of SUHIs across diurnal, nocturnal, and seasonal patterns in different humid–arid regions and the relative importance and contribution of influencing factors remain uncertain. To address the gaps, a systematic research framework is established. Machine learning, SHAP, and SEM methods are applied to analyze 329 cities across four humid–arid regions of China. Within this framework, the importance, contribution direction, and impact pathways of various influencing factors are analyzed from both diurnal and seasonal perspectives, aiming to investigate the spatial heterogeneity of SUHI driving mechanisms across humid–arid regions. The objectives of this study are as follows:

• How do the diurnal and seasonal distributions of SUHIs vary across humid–arid regions?
• What are the differences in the relative importance and contribution directions of various influencing factors across humid–arid regions?
• What are the pathways through which these factors affect SUHIs across humid–arid regions?

The research framework (Figure 1) comprises the following stages:

(1)

Data collection and processing. Based on previous research, we selected several important influencing factors that have a relatively significant impact on SUHIs and established a six-dimensional system of influencing factors. This part includes the classification of humid–arid region, SUHII calculations, and the construction of six-dimensional influencing factors.

(2)

SUHI Seasonal and Diurnal Distribution. Based on the humid–arid region classification, 329 cities from different regions were selected. The SUHII of these cities was calculated to explore the distribution characteristics of SUHII across different seasons and day/night periods. The analysis of SUHI changes across humid–arid regions revealed regional differences in seasonal variations and diurnal distribution.

(3)

Correction Between Influencing Factors and SUHI. Based on the six-dimensional influencing factors and calculated SUHII, the correlations between these factors and SUHII across different seasons and day/night periods in different humid–arid regions were analyzed. This correlation analysis provides an initial, region-specific overview of the relationships between individual factors and SUHII and offers a preliminary basis for subsequent multivariate interpretation.

(4)

The Contribution of Influencing Factors to the SUHI and the Pathway Analysis. First, Optuna V4.2 is used for automated hyperparameter optimization based on Bayesian optimization with a Tree-structured Parzen Estimator (TPE), with the mean error from five-fold cross-validation defined as the objective for maximizing model performance. Using this model and the SHAP method, we analyzed the contributions and feature importance of seasonal and day-night influencing factors across humid–arid regions. Finally, the key factors identified by SHAP are incorporated into an SEM to examine their direct and indirect pathways associated with SUHII across humid–arid zones. In this framework, the analysis proceeds from preliminary association screening (correlation) to model-based explainability and variable prioritization (SHAP), and then to structured pathway assessment using SEM, with each stage building on and reinforcing the previous one.

2. Materials and Methods

2.1. Study Area

The regional differences in SUHIs across China are largely attributed to its extensive latitudinal and longitudinal spans (73°33′–135°05′ E, 18°05′–53°33′ N). The analysis was confined to mainland China, which was divided into four climate regions, namely, humid, semihumid, arid, and semiarid regions. A total of 188, 80, 37, and 24 cities, respectively, fall into these regions (Figure 2). These regions are distinguished not only by environmental diversity but also by dynamic differences in socioeconomic orientation, urbanization trajectories, and climate change patterns.

2.2. Research Framework

A systematic analysis was conducted after the humid–arid region was defined. This involved data processing, SUHI calculation, and six-dimensional influencing factor system development. Through the application of different models, the diurnal, seasonal, and annual distribution characteristics of SUHIs were revealed, the correlations with influencing factors were identified, and their contributions and regulatory pathways were explored. The detailed framework is shown in Figure 1.

2.3. Data and Process

The research data sources are listed in

Table 1. Details of the data used in this study.

Data Name	Spatial Resolution	Source	Use
LST	1 km, 2018–2020	Google Earth Engine (GEE)	SUHI calculation
GUB	1 km, 2018–2020	Zhao et al. [34]
DEM	1 km	Resource and Environmental Science Data Platform
GAIA	1 km, 2018–2020	[35]
Wind Speed	10 m, 2018–2020	[36]	Influencing factors
Precipitation	1 km, 1991–2020	[37]
NTL	1 km, 2018–2020	[38]
PD	1 km, 2018–2020	[39]
PM2.5	1 km, 2018–2020	[40]
PM10	1 km, 2018–2020	[41]
AOD	1 km, 2018–2020	GEE
CNLUCC	1 km, 2018–2020	Xu [42]
3D-GloBFP		[43]
NDVI	1 km, 2018–2020	GEE
WSA\BSA	500 m, 2018–2020	GEE

. Specific processing details are provided in Supplementary File Text S1. Data from 2018 to 2020 are processed to derive daytime and nighttime SUHII at annual and seasonal scales, along with the corresponding influencing factors for the same periods. At the annual scale, SUHII is calculated as the three-year mean (2018–2020); at the seasonal scale, SUHII is calculated as the mean of the corresponding season across the three years.

2.4. Methods

2.4.1. Humid–Arid Region Classification

Following the World Health Organization (WHO) standard climate reference period (1991–2020), a 30-year average dataset was employed to classify climate zones on the basis of the annual precipitation: >800 mm (humid), 400–800 mm (semihumid), 200–400 mm (semiarid), and <200 mm (arid). The study area was defined as the entirety of mainland China and was divided into four climate zones for systematic SUHI analysis. After the exclusion of cities with significant data gaps or low representativeness, a total of 329 cities were retained, namely, 188 in humid regions, 80 in semihumid regions, 37 in arid regions, and 24 in semiarid regions. The detailed procedures are provided in Text S2.

2.4.2. Construction of Six-Dimensional Influencing Factors

To investigate the driving mechanisms of SUHIs across humid–arid regions, a process-oriented framework is adopted in which the potential drivers are organized into (a) regional boundary conditions, (b) surface energy and moisture processes, and (c) human-built environmental structure and intensity. Following this logic, a six-domain indicator system is constructed, comprising background climate, surface features, pollution, development level, and two-dimensional (2D) and three-dimensional (3D) urban morphology. The corresponding factor names and detailed indicator explanations are provided in

Table 2. Details of the influencing factors.

Dimension	Factor	Description
Background climate	PRE	Precipitation
	WIND	Wind speed
Surface features	ΔSA	Difference in albedo between urban and rural areas
	ΔNDVI	Difference in the normalized difference vegetation index (NDVI) between urban and rural areas
	ISF	Urban impervious surface area
	WATER	Urban water body area
Pollution	ΔPM2.5	Difference in fine particulate matter (PM2.5) levels between urban and rural areas
	ΔPM10	Difference in inhalable particulate matter (PM10) levels between urban and rural areas
	ΔAOD	Difference in the aerosol optical depth (AOD) between urban and rural areas
Development	NTL	Nighttime light
	PD	Population density
2D	FRAC	Fractal dimension index
	CONTIG	Contiguity index
	SHEI	Shannon’s evenness index
	AI	Aggregation index
	LPI	Largest patch index
3D	BHmean	Mean building height
	BVmean	Mean building volume

and Text S3, respectively.

To mitigate the effects of multicollinearity on model stability and the interpretability of SHAP-based analyses, multicollinearity is assessed among influencing factors using the Variance Inflation Factor (VIF). Factors with VIF values greater than 10 are considered to indicate severe multicollinearity. Only factors with VIF ≤ 10 are retained for subsequent interpretability analysis and structural equation model establishment to enhance the robustness of the results. Detailed VIF values for each factor are provided in the Supplementary File Table S7.

2.4.3. Calculation of the Surface Urban Heat Island Intensity

The warming trend in the core areas of cities was 29% greater than that in surrounding suburban areas [44], indicating that major developed areas exhibited higher sensitivity in reflecting UHIs. The global annual urban extent dataset developed by [34] was employed to define urban areas. Rural areas were defined as the buffer zones created outside these urban areas, with an area equal to that of the corresponding urban area [45]. On the basis of the digital elevation model (DEM), China Land Use/Land Cover Change (CNLUCC), and Global Artificial Impervious Area (GAIA) datasets, pixels within an elevation range of ±100 m from the rural area elevation were first retained. Pixels classified as impervious surfaces, water bodies, snow and ice, or forests were subsequently excluded [46]. The SUHI intensity (SUHII) can be calculated as follows:

$${\mathrm{S}\mathrm{U}\mathrm{H}\mathrm{I}\mathrm{I}}_{\mathrm{i}}={\mathrm{L}\mathrm{S}\mathrm{T}}_{\mathrm{U}\mathrm{r}\mathrm{b}\mathrm{a}\mathrm{n}}-{\mathrm{L}\mathrm{S}\mathrm{T}}_{\mathrm{R}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{l}}$$

(1)

where i denotes the city, and LST_Urban and LST_Rural denote the average LST values in the urban and rural areas, respectively.

To strengthen credibility and evaluate robustness to rural-definition choices, SUHII was recalculated for 329 cities across ten periods using three methods (EA, IEA, and MEA). Differences among the three methods result were evaluated using the Friedman test (Table S8). The test indicated no statistically significant differences in SUHII among the three delineation methods (p > 0.05), suggesting that the results are robust to the choice of rural reference definition.

2.4.4. Pearson Correlation Coefficient

The Pearson correlation coefficient is adopted to evaluate the correlation between the SUHI and the variable. The Pearson correlation coefficient is a statistical measure of the strength and direction of linear relationship between two continuous variables. It assesses the degree to which one variable changes linearly as the other variable changes. The value of r ranges between [−1, +1]. r > 0: Indicates a positive correlation, meaning that as one variable increases, the other variable also tends to increase. r < 0: Indicates a negative correlation, meaning that as one variable increases, the other variable tends to decrease. The formula is

$${\mathrm{r}}_{\mathrm{X}\mathrm{Y}}={\displaystyle \frac{\mathrm{C}\mathrm{O}\mathrm{V}(\mathrm{X},\mathrm{Y})}{{\mathrm{\sigma }}_{\mathrm{X}}{\mathrm{\sigma }}_{\mathrm{Y}}}}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-\overline{\mathrm{y}})}^2}}}$$

(2)

x and y are the sample means of variables X and Y, respectively; COV(X,Y) is the sample covariance; σ_X and σ_Y are the sample standard deviations.

2.4.5. Selection of the Optimal Model

Due to the imbalance in the number of cities across arid and humid regions, oversampling is applied to arid and semiarid regions during model training to improve their representativeness. In addition, a weighted loss function is incorporated to further strengthen the contribution of underrepresented regions. By tuning the loss weights, regions with smaller sample sizes receive greater emphasis in the objective function, thereby improving the model’s predictive performance for these regions.

To avoid bias introduced by manual parameter tuning [47], Optuna, based on a Bayesian optimization framework, is adopted for automated hyperparameter search to maximize model performance. The search is configured to minimize the mean cross-validation error under five-fold cross-validation, and the Tree-structured Parzen Estimator (TPE) sampling algorithm is used to guide the exploration of the hyperparameter space [48]. Model selection was guided by the generalization performance metrics R² and root mean square error (RMSE) (Tables S1–S5). The dataset was split into training and testing sets at a 70/30 ratio. With the independent test set, the categorical boosting (CatBoost) model demonstrated superior performance (R² = 0.68; RMSE = 0.47) and was therefore selected for subsequent analysis. Finally, feature importance and model interpretability were assessed via the SHAP method.

2.4.6. SHAP Analysis

Machine learning models often face black-box problems. The SHAP method addresses these problems by quantifying the marginal contribution of each feature to predictions via Shapley values from game theory. In this framework, features are considered players contributing to a payout (prediction), with SHAP scores assigning each feature an importance value [49]. This method reveals the individual contribution of each feature to the prediction target, increasing model interpretability [50]. The SHAP method can be expressed as follows:

$$\mathrm{S}\mathrm{H}\mathrm{A}\mathrm{P}(\mathrm{f},\mathrm{x},\mathrm{S})={\sum }_{\mathrm{T}\subseteq \mathrm{S}}{\displaystyle \frac{\left|\mathrm{T}\right|!(\mathrm{n}-|\mathrm{T}\left|\right)!}{\mathrm{n}!}}[\mathrm{f}(\mathrm{x})-\mathrm{f}({\mathrm{x}}_{\mathrm{S}\mathrm{T}})]$$

(3)

where f denotes the model prediction function, x denotes the observed feature vector, S denotes the set of all features, n denotes the total number of features, and x_ST denotes the observed values of all the features, excluding those in subset T.

2.4.7. SEM

In SEM, confirmatory factor and path analyses are integrated to verify multivariate causal hypotheses [51]. SEM can be used to model multiple causal pathways and explore direct, indirect, mediation, and moderation effects between latent and observed variables [52]. In this study, a path model was established grounded in existing theoretical knowledge. Factors exhibiting high multicollinearity are removed based on a VIF threshold of 10. A CatBoost model is then trained using the remaining factors, and SHAP is applied to quantify feature importance as the mean absolute SHAP value. The eight highest-ranked factors are subsequently incorporated into the SEM. The priori assumption has been established: background climate, surface characteristics, 2D and 3D, and development factors are specified to be directly associated with SUHIs. Background climate and 2D/3D are also specified to be indirectly associated with SUHIs via surface characteristics. In arid and semiarid regions, direct associations between pollution-related factors and SUHIs are additionally included. Seasonal and annual SEMs are fitted for each climatic zone to test these pathways and compare regional differences. All the SEM fit indices were satisfactory. Detailed evaluation results are listed in Table S6.

3. Results

3.1. Day/Night Seasonal and Annual Variations in the SUHII Across Humid–Arid Regions

The diurnal and seasonal variations in the SUHII across humid–arid regions are shown in Figure 3. During the day, the average annual SUHII in most arid/semiarid regions (−0.607, −0.515) was lower than that in humid/semihumid regions (0.914, 0.406). At night, the average annual SUHII in semihumid regions was 1.208, which is significantly greater than that in other regions. In most arid/semiarid regions, compared with daytime levels, the SUHII increased and exhibited a heat island effect (SUHII > 0).

In humid regions, the daytime SUHII peaked (2.085) in summer, reaching a minimum (0.211) in winter, while nighttime levels remained stable annually. In semihumid areas, where the daytime SUHI also peaked in summer, the nighttime SUHII was highest in spring (1.810) and lowest in summer (1.348), with a heat island effect (SUHII > 0). In arid regions, the daytime SUHII was mostly negative except in summer (0.103), when limited warming occurs. This region is characterized by sparse vegetation and low soil moisture, which facilitates faster solar energy absorption [53]. At night, arid regions exhibited seasonal variations in the SUHII (1.547). Semiarid regions resemble arid regions by day (cooling except in summer). Nocturnally, the SUHII peaked in spring (1.169), decreased in summer, and increased through autumn and winter, remaining positive throughout the year.

3.2. Correlations Between the SUHI and Various Factors Across Humid–Arid Regions

National-scale correlations between influencing factors and SUHII are presented in Figure 4. Regional-scale correlation are presented in Figures S1–S4. For clarity, only the factors showing the strongest correlations across seasons are reported.

3.2.1. National-Scale Correlation

For background climate, PRE exhibits a pronounced diurnal contrast, showing positive correlations during the daytime and negative correlations at night (p < 0.001). Among surface characteristics, ΔSA shows the strongest association during summer daytime (r = 0.77, p < 0.001), whereas ΔNDVI displays consistently negative correlations throughout the year. For air pollution, ΔPM10 and ΔPM2.5 are positively correlated with daytime SUHII (r > 0, p < 0.05). Moreover, all pollution indicators show significant positive correlations at night in summer (r > 0, p < 0.05). Development-related factors, including nighttime lights (NTL) and population density (PD), are positively correlated with daytime SUHII, although PD is not significant at night (p > 0.05). For 2D, FRAC and CONTIG are negatively correlated with SUHII across all seasons. LPI shows significant positive correlations at night (r > 0, p < 0.05). Regarding 3, BHmean is positively correlated with daytime SUHII (r > 0, p < 0.05) but negatively correlated at night (r < 0, p < 0.05). BVmean shows a weak but significant positive correlation only during summer daytime (r = 0.13, p < 0.05), with no significant correlations in other seasons.

3.2.2. Regional-Scale Correlation

Across humid and arid regions, significant differences are observed in the correlations between influencing factors and SUHII. For background climate, PRE shows significant correlations mainly in humid and semihumid regions, whereas its significance weakens or disappears toward arid and semiarid regions, particularly at night. For surface characteristics, ΔNDVI consistently exhibits significant negative correlations across all regions, but the magnitude is markedly stronger in arid and semiarid regions, especially in summer. For development, including NTL and PD, more stable positive correlations are observed in humid and semihumid regions, particularly at night. For 2D, FRAC, CONTIG, and SHEI display clear regional differentiation, with shifts in correlation direction between humid and semiarid regions.

3.3. Contributions of Influencing Factors to SUHIs Across Humid and Arid Regions

The SHAP contribution and feature importance of each influencing factor to the annual daytime and nighttime SUHIs at the national and regional scales are shown in Figure 5 and Figure 6. The corresponding SHAP contributions and feature importance for each season at the national and regional scales are shown in Figures S5–S8.

3.3.1. National-Scale Contribution

On an annual scale, the daytime SUHI was driven primarily by PRE, ΔSA, and ΔNDVI. At night, PRE, ΔNDVI, and BVmean dominated, with the importance of BVmean significantly increasing at night. Precipitation amplifies soil moisture differences, causing rural areas to cool faster than urban areas do, thereby exacerbating daytime SUHI while reducing nighttime SUHI [54]. With decreasing ΔNDVI, urban vegetation cover decreases, reducing the cooling capacity in urban areas and thereby increasing the SUHII. ΔSA is much more important during the day. A higher albedo reduces solar radiation absorption and thus mitigates the SUHI [55]. However, in several cases, higher albedo corresponds to higher land surface temperatures, indicating that elevated ΔSA can increase the SUHII. An increased building volume typically enhances the heat island effect by increasing the thermal capacity and suppressing heat dissipation, yielding intensified nighttime heat accumulation and delayed release [56]. However, once BVmean exceeds a critical threshold, its shading-induced cooling effect outweighs the warming effect, thereby decreasing the SUHII.

3.3.2. Regional-Scale Contribution

The contributions of the influencing factors across humid–arid regions are shown in Figure 6. Substantial commonalities are observed in the dominant determinants of SUHII across humid and arid regions. In all regions, SUHII is consistently characterized by a pattern centered on ΔSA and ΔNDVI, indicating that the influence of these two factors remains stable across climatic zones. In contrast, clear regional differences are also evident. In arid and semiarid regions, pollution-related variables generally exhibit higher importance, suggesting pronounced regional heterogeneity in the drivers of SUHII across humid–arid gradients. Seasonal variations are also apparent. Building on these shared patterns and regional/seasonal differences, the dominant daytime and nighttime drivers within each climatic zone at the annual scale are presented below, and the corresponding seasonal variations are briefly summarized.

In humid regions, ΔSA, ΔNDVI, and ISF are the dominant daytime drivers of SUHIs, whereas PRE, ΔNDVI, and BVmean dominate at night, which is consistent with the national-scale results. A higher ISF enhances urban warming and thus amplifies the UHI effect. Seasonally, in spring, summer, and autumn, at night, the importance of 2D urban morphology, such as LPI, increases compared with that during the day. Moreover, higher LPI values correspond to increased SUHIIs, which agrees with previous findings [57].

During the day in semihumid regions, the SHAP values of AI increase compared with those in other areas. AI partially represents the extent of urban expansion, with a higher expansion intensity yielding a higher SUHII. At night, the feature importance of PD and NTL increases. Areas with higher PD and NTL values exhibit higher anthropogenic heat emissions, which leads to an increase in the SUHII [58]. Seasonally, across spring, summer, and autumn, the feature importance of development factors at night is consistently greater than that during the day.

During the day in arid regions, the dominant factors are ΔNDVI, WATER, and WIND. At night, ΔNDVI, ΔSA, and ΔAOD prevail. Water bodies serve as significant urban heat sinks and play a crucial role in mitigating the UHI effect; expanding water areas can enhance urban cooling effects [59]. Increased wind speeds facilitate faster heat dissipation, thereby reducing the SUHII [60]. In spring and summer, surface features remain the dominant factors, whereas 3D urban morphological factors prevail in autumn and winter. At night, the feature importance of pollution factors increases in spring, summer, and autumn.

The SHAP values of ΔPM₁₀ and ΔPM_2.5 are greater in semiarid regions than in humid and semihumid regions. When pollutant concentrations are higher in urban areas than in rural areas, the urban radiation dissipation rate is lower than that in rural areas, thereby amplifying the UHI effect. Surface features dominate across all four seasons during the day, whereas the importance of 2D features (SHEI and AI) increases at night.

3.4. Structural Pathways of the Factors Influencing SUHIs Across Humid and Arid Regions

Given the significant regional differences in SUHIs and the varying importance of influencing factors across China’s humid–arid regions, structural equation models were developed, including a model for the entire region at the annual scale and separate models for different periods across humid–arid regions (Figures S9–S12).

3.4.1. National-Scale SEM

At the national scale, PRE (β = 0.51) and ΔNDVI (β = −0.44) exhibited the strongest standardized direct associations with the SUHI (Figure 7). During the daytime, the direct effect of ΔNDVI was regulated by FRAC and PRE, with additional positive contributions from pollution factors and no significant paths of 2D factors. At night, the standardized direct association between ΔSA and the SUHI was stronger within the SEM in relation to AI and ΔNDVI. The nighttime SUHII increased due to 2D and development factors but decreased due to 3D factors.

3.4.2. Regional-Scale SEM

Figure 8 and Figure 9 present the daytime and nighttime SEM results across humid–arid regions. Both shared and region-specific paths are observed among regions. The potential drivers of these commonalities and heterogeneities are discussed in Section 4.2.

Humid regions (Figure 8a and Figure 9a) exhibited the strongest standardized direct associations with SUHIs during the day and at night, respectively. During the day, surface features exhibited dual regulatory effects. ISF and ΔSA had positive direct associations with the SUHI (cumulative direct effect of 0.24), whereas the negative direct association between ΔNDVI and the SUHI was primarily reflected through paths involving LPI and NTL (cumulative direct effect of −0.20). Notably, 2D, pollution, and development factors contributed to the SUHI (β > 0) through direct paths. At night, the negative direct association between ΔSA and the SUHI was primarily linked to FRAC, PRE, and ΔNDVI (β = −0.42). Background climate and 2D and 3D factors were negatively associated with the SUHI, whereas positive effects were associated with pollution. Seasonally, background climate showed more notably associations with the SUHI in spring, 2D in summer, and surface features in autumn and winter. At night, most factors exhibited negatively associations with SUHIs across all seasons.

In semihumid regions (Figure 8b and Figure 9b), PRE and ΔNDVI exhibited the strongest standardized direct associations with day- and nighttime SUHIs, respectively. During the day, background climate, 2D, pollution, and development factors were positively associated with SUHIs, with PRE (β = 0.28) and AI (β = 0.20) revealing stronger direct paths. Similarly to the findings for humid regions, surface features exerted a dual effect on SUHI. However, the direct effect of ΔSA (β = 0.27) was greater, whereas the direct effect of ΔNDVI remained similar. At night, surface features and background climate were generally negatively associated with SUHIs annually, whereas development factors yielded positive effects. Two-dimensional factors exerted a dual effect on the SUHI, primarily driven by indirect effects, resulting in an overall negative effect. Among these, the positive pathway was caused mainly by the indirect effect of AI, whereas the negative pathway was attributed primarily to LPI (β < 0). Seasonally, daytime SUHIs showed stronger associations with the background climate in summer and winter, by surface features in spring, and by 2D factors in autumn. At night, surface features played a dual role during all four seasons, and 2D factors exerted greater positive effects on SUHIs (β > 0) in summer, autumn, and winter.

In arid regions (Figure 8c and Figure 9c), ΔNDVI and ΔSA exhibited the strongest standardized direct associations with the day- and nighttime SUHIs. During the day, background climate, surface features, and pollution were negatively associated with the SUHI, whereas 3D factors yielded a positive effect. Among these factors, ΔNDVI showed the strongest direct association with the SUHI (β = −0.48). Two-dimensional factors exerted a dual regulatory effect on the SUHI. Specifically, LPI and FRAC contributed positively to SUHIs (β = 0.32 and 0.26, respectively), whereas CONTIG imposed a negative direct association (β = −0.36). At night, 3D factors remained positively associated with SUHIs. The negative association between the surface features and SUHI was stronger at night than during the daytime and was mainly captured by direct paths (β = −0.94). In contrast, 2D and pollution factors exerted relatively weak negative associations with the SUHI. Seasonally, surface features imposed a dual effect on daytime SUHI in spring, summer, and autumn. At night, 3D, 2D, and pollution factors also demonstrated a dual regulatory effect on the SUHI.

In semiarid regions (Figure 8d and Figure 9d), PRE and SHEI exhibited the strongest standardized direct associations with the day- and nighttime SUHIs. During the day, background climate, 3D, and pollution factors positively were positively associated with SUHIs, and the associations for 3D and pollution factors were relatively weak. Surface features exerted a dual effect on the SUHI. Specifically, ΔNDVI and WATER showed negative direct associations with the SUHI (β = −0.59), whereas ΔSA showed a positive direct association to the SUHI (β = 0.22). At night, background climate and surface features were negatively associated with SUHIs, although to a lesser extent than they did during the day, whereas 2D, 3D, and pollution factors were positively associated with SUHIs. Seasonally, the daytime behavior of surface features in semiarid regions was similar to that in arid regions, whereas at night, most factors exhibited significant positive associations with SUHIs.

4. Discussion

4.1. Differences in SUHI Distributions and Contributing Factors Across Humid–Arid Regions

A comparison with the literature revealed a lack of consensus regarding the spatial patterns of SUHIs across humid–arid regions. In contrast to the present findings, Peng [7] reported greater SUHI effects in arid and semiarid regions than in humid and semihumid regions. There are two reasons for this phenomenon. First, the number of cities selected in the two studies differs. Second, SUHIs in arid and semiarid regions are more significantly influenced by urban–rural definitions [61].

According to the results in Section 4.2, ΔSA and ΔNDVI are the main influencing factors across humid–arid regions. However, certain factors exhibit differences in the threshold (SHAP = 0) across regions. ΔSA yields a dual effect in humid regions, where a high ΔSA value can either increase or decrease the SUHII (Figure 5a). These findings may be attributed to land cover differences because bare soil or cropland can reflect sunlight but can retain heat [62]. In humid regions, the importance of 3D factors is greater at night than during the day (Figure 6e), which is consistent with the findings of previous studies [63]. Solar radiation is absorbed and later released by buildings with high heat capacity and thermal conductivity, leading to an increase in the nighttime SUHII. Large, high-rise structures increase heat storage and impede air circulation via the valley effect [64], thus limiting heat diffusion and dissipation.

Compared with arid and semiarid regions, PRE is more important in humid and semihumid regions, whereas WATER is more important in arid and semiarid regions than in their humid counterparts. This pattern highlights the role of water bodies as stable heat sinks under the conditions of sparse vegetation and rapid surface warming. The relatively dense vegetation and relatively high precipitation in humid regions attenuate the thermal regulation function of water bodies [65,66], thereby enhancing the relative cooling efficiency of WATER and explaining the difference in feature importance across humid–arid regions.

4.2. Commonality and Heterogeneity in the SEM Paths and Effect Direction Across Humid–Arid Regions

We observed two primary patterns in the pathway relationships. First, similar pathways occurred across multiple regions. The results of annual (Figure 8 and Figure 9) and seasonal (Figure S9) analyses indicated that PRE positively influences ΔSA. Precipitation regulates the heat island effect by adjusting the albedo [67]. The negative regulation of the SUHI by ΔSA occurs mainly in winter. This phenomenon may be attributed to the enhanced anthropogenic heat flux from heating and other anthropogenic activities in winter, along with larger ΔSA values due to seasonal leaf fall and/or snow cover [68]. Previous studies suggested air pollution as a mediating factor in the relationship between urban morphology and UHIs [69], and the results of our pathway analysis indirectly corroborated these findings. Our research focused on the differences in pollutant concentrations between urban and rural areas. Two-dimensional factors such as FRAC predominantly negatively influence urban–rural pollutant differences (Figure S9g). An increase in FRAC can lead to more contact opportunities between urban and suburban areas, thereby promoting air flow [70]. We also observed influences of 2D factors (LPI, AI, and SHEI) on ΔNDVI. LPI and AI reflect the expansion of impervious surfaces, in which higher values indicate greater urban expansion accompanied by vegetation reduction, thereby indirectly regulating the SUHI. This aligns with previous research [71].

Second, the pathway effects also exhibited regional differentiation. At night in autumn (Figures S9h and S10h), BHmean negatively influenced SUHIs in humid areas, while it yielded a positive effect in semihumid regions. SHAP scatter plots (Figure 10a) were created to clarify these results. In humid regions, the SUHI increased when BHmean ranged from 5 to 12. In semihumid regions, the SUHI demonstrated a decreasing trend, specifically when BHmean varied between 7 and 10. Previous studies have revealed varying impacts of the building height on temperature across different climatic zones [28,72], indicating that the effect of the building height on the thermal environment is modulated by background climatic conditions [8]. During the day in winter, WIND was positively correlated with SUHIs in humid regions, whereas it negatively influenced SUHIs in all the other regions. While conventional understanding suggests that increased wind speed generally reduces the SUHI effect [60], the SHAP dependency plots (as shown in Figure 10b) reveal a more complex relationship in humid regions: notably, when the wind speed ranges from 1.5 to 3.0, it can simultaneously increase and decrease the SUHI effect. This phenomenon may be due to population differences, as cities with higher population densities exhibit higher wind speed thresholds than those in cities with lower population densities [73]. The urban population density in humid regions is significantly greater than that in other areas, but it fails to reach the threshold needed to reduce the SUHII.

4.3. Robustness of SUHI Mechanisms Across Humid–Arid Regions

The limitations of the dataset and the unusual environment in 2020 introduce uncertainties. The results indicated that 3D factors generate greater effects at night. Therefore, incorporating 3D factors is reasonable. To investigate whether the conclusions for different years are generalizable, we selected 2018 and 2019 as non-pandemic reference years for comparison with 2020. The rationale for selecting these two years is as follows:

• The employed global urban built-up area dataset [34] covers only the period from 1992 to 2020, making it difficult to extend the analysis to longer time series.
• Three-dimensional factors remain stable over short periods, especially structural metrics such as building height and volume, which do not significantly change.

A comparative analysis with non-pandemic years was conducted. The results demonstrated remarkable consistency and stability of the key influencing factors across the different years. The ranking based on feature importance for most major variables exhibited no significant shifts. This consistency demonstrates the high robustness of the SUHI influencing mechanisms identified in this study. While individual variables such as PD exhibited minor fluctuations during the pandemic period because of altered urban activities, these changes primarily reflected short-term anthropogenic disturbances resulting from urban lockdowns [74,75] rather than fundamental changes in urban thermal environmental mechanisms. Overall, the conclusions derived from the 2020 data remain representative and robust when applied to other years. Our understanding of the SUHI effect has been enhanced by the inclusion of 3D factors (Figure 11).

4.4. Suggestions for Mitigating the Urban Heat Island Effect Across Humid–Arid Regions

Regional differences are observed in how influencing factors affect SUHIs across humid–arid regions. Therefore, mitigation strategies should be tailored to local conditions. First, for key drivers that are consistently important across all regions (e.g., ΔNDVI), expanding and improving urban greening remains an effective approach for reducing SUHII [76]. However, in arid and semiarid regions, these measures depend more heavily on irrigation and long-term operation and maintenance, which may increase water demand and maintenance costs. In humid and semihumid regions, reducing ISF and increasing FRAC can further alleviate SUHIs. Practically, this can be achieved by designing newly developed areas with more complex boundaries to increase FRAC, thereby promoting airflow and heat exchange and weakening heat accumulation. Arid regions also exhibit substantial cooling potential [77]. In arid and semiarid regions, irrigation and expanding blue–green spaces can yield pronounced cooling benefits [18,78]. These measures support vegetation growth by improving water availability and enhance cooling through evapotranspiration, which absorbs heat from the surrounding air and lowers near-surface temperatures [79], while it may cause heat stress in humid/semihumid regions [80]. In addition, establishing green corridors that connect fragmented patches and enhance local wind circulation can further strengthen cooling effects in arid and semiarid environments [81].

4.5. Uncertainty

In this study, the factors influencing SUHIs across humid–arid regions and their diurnal variations were quantified, but this research exhibits several limitations. First, several morphological indicators, such as the sky view factor, could not be included because of data constraints and the extensive spatial scale of this research. Second, the unequal number of cities across climate zones may influence feature importance, despite weighted adjustments and cross-validation. Third, the analysis was confined to 2018–2020 because of data availability. Future long-term urban footprint data could reveal the temporal evolution of SUHI mechanisms. Fourth, although a comparative analysis with non-pandemic years revealed general stability, short-term mechanistic shifts in some regions or variables remain possible. Despite these uncertainties, the results consistently identify SUHI drivers across humid–arid regions and provide region-specific sustainability insights.

5. Conclusions

In this study, 329 cities across humid–arid regions were selected to investigate the seasonal and diurnal distributions of SUHIs. The findings indicated that during the day, the SUHII was greater in humid and semihumid regions than in arid and semiarid regions. At night, however, the SUHII increased in arid and semiarid regions. Seasonally, all the regions exhibited the highest daytime SUHIIs in summer. At night, the SUHII remained relatively stable in humid and semihumid regions across all seasons, whereas in arid and semiarid regions, the nighttime SUHII was consistently higher than the daytime SUHII across all seasons.

The contributions and feature importance rankings of the various factors influencing SUHIs were subsequently analyzed across humid–arid regions. The analysis was first conducted at the national scale, and surface features, 3D, and background climate factors were identified as the primary drivers. At the regional scale, the results for humid and semihumid regions were similar to the global pattern. In contrast, pollution factors were found to be more critical in arid and semiarid regions. Additionally, seasonal variations in factor importance were observed across the different regions.

Finally, the top eight influencing factors based on feature importance were selected to construct structural equation models. Two types of action pathways were identified in each region. The first type indicated similar paths across regions. The second type revealed regional path differences. Additionally, within the same dimension of the influencing factors, the effects on SUHIs may be both positive and negative, which reduces the total effect value of that dimension.