Introduction

The urban heat island (UHI) effect refers to the phenomenon that the rising motion of prevailing airflow over urban areas causes suburban near-surface air to flow into the city, resulting in urban regions experiencing noticeably higher temperature than suburban regions1. In the context of accelerated urbanization, global climate warming, and frequent extreme weather events, the adverse UHI effects on human health, daily life, and urban eco-environment are also increasing severe2. In summer, rising temperatures could exacerbate the risk of heat stroke to large extent, leading to a significant increase in heat-related illnesses3,4. Additionally, high temperatures may put pressure on urban water and power supply, disrupting normal living and production activities5. Due to the sluggish air circulation in winter, the UHI effect often aggravates the accumulation of air pollutants, thus causing a series of urban ecological problems such as air pollution6,7,8. Therefore, exploring the spatial–temporal evolution differences of UHI in different seasons, and understanding the influence of different driving factors are crucial for improving urban living and eco-environment.

Earlier methods for monitoring UHI primarily involve meteorological data analysis9, stationary observations10, and mobile surveys11, which rely on data gathered from meteorological stations or mobile monitoring devices to examine and track long-term temperature patterns12,13. Although these methods offer relatively accurate discrete temperature data, they fall short of providing comprehensive surface temperature investigation and analyzing UHI’s spatial patterns14,15. Due to extensive coverage, strong comparability, and available data acquisition, remote sensing technology has increasingly become the primary tool for UHI studies16,17. In 1972, Krishna et al.18 pioneered the use of remote sensing technology in UHI research, becoming the first to utilize thermal infrared remote sensing data to investigate UHI. Major data sources for UHI remote sensing research include MODIS and Landsat data. Xu et al.19 utilized MODIS data to map the UHI (2001–2020) of Guangdong-Hong Kong-Macao Greater Bay Area of China, revealing an overall upward trend in UHI. However, the spatial resolution of the MODIS data is relatively low, making it unsuitable for high-precision spatial analysis20. In contrast, Landsat data have higher spatial resolution, making them ideal for monitoring surface heat island intensity changes in individual cities or localized areas21. Wang et al.22 employed Landsat 8 imagery to map LST for Shenzhen, revealing that the surface temperature decreases gradually from northwest to southeast, while impervious urban surfaces exhibit significantly higher temperatures than vegetated areas. Furthermore, seasonal features of UHI reveal that the UHI effect is most noticeable in summer as primarily temperatures and stronger solar radiation increase, which result in heightened heat accumulation in urban environments23. In contrast, winter conditions, characterized by lower solar radiation and different thermal demands, exhibit distinct UHI features and trends24. Therefore, a systematic analysis of the summer and winter UHI could provide a thorough comprehension of the spatial–temporal patterns and intensity variations of UHI25. Moreover, to elucidate the driving mechanisms of UHI, many previous studies have employed simple correlation and regression models to analyze the relationships between various influencing variables and UHI26,27. While these methods help understand the effect of individual factors on UHI, there are still limitations in determining the complex mechanisms of different driving factors and their interactions, causing uncertainties in understanding the UHI driving forces28.

The Shenzhen city, designated as China’s first special economic zone since the 1980s, demonstrates globally leading urbanization speed and scale, with its urban planning and environmental policies serving as an exemplary model29. In-depth research on Shenzhen’s UHI would provide a foundation for scientific evidence in urban planning and policy-making, while simultaneously offering valuable insights into the relationship between urbanization and climate change30. Considering the significance of UHI research and the limitations of prior studies, this study concentrates on the Shenzhen city. The purposes of this study are to: (1) examine the spatial–temporal pattern differences of UHI in summer and winter; (2) conduct a quantitative analysis of various factors influencing on UHI; (3) explore the underlying mechanisms driving UHI in summer and winter. This study aims to provide technical insights and data to inform the development of effective, scientifically grounded strategies for mitigating UHI effects in summer and winter.

Methods

Study area

The Shenzhen city is one of the most important megacities in China, situated in the southern region of Guangdong province, spanning from 113° 43′ E to 114° 38′ E and from 22° 24′ N to 22° 52′ N31. The city comprises ten districts, including Guangming, Baoan, Nanshan, Longhua, Futian, Luohu, Longgang, Pingshan, Yantian, and Dapeng, which are interconnected by three major urban ring roads (Fig. 1). The region is characterized by a subtropical monsoon climate, notable for abundant sunshine and high temperatures, with an average annual sunshine duration of 2120 h and an average temperature of 22.4 °C32. Since 2004, as the urbanization rate achieving 100% urbanization, its surface thermal environment has significantly been altered with the rapid economic development over the past 2 decades33.

Fig. 1
figure 1

Location of the Shenzhen city. Map generated by the authors using ArcGIS 10.8 (Environmental Systems Research Institute, USA. https://www.esri.com).

Full size image

Data source and preprocessing

To perform land cover mapping and land surface temperature (LST) retrieval for 2013, 2018, and 2023, high-quality images from the Landsat series were selected for their spectral discrimination and time sequencing of remote sensing. The images collected included multispectral and thermal infrared bands, and cloud cover in each image was less than 10%. To obtain calibrated images covering the study area from 2013 to 2023, the images were subjected to radiometric calibration, atmospheric correction and image cropping. 4670 validation sample points were collected via field surveys and high-resolution image interpretation using Google Earth to assess the accuracy of the land cover classification. Moreover, to ensure comparability, remote sensing data for the same season are acquired within the same month to minimize temporal discrepancies. Details of the selected images are provided in Table 1.

Table 1 The specific information of the selected images in the study.
Full size table

Methodological framework for investigating UHI

Firstly, through the radiation transmission equation and object-oriented random forest method, LST data and land cover data were obtained from the thermal infrared and multispectral data in Landsat imagery. Secondly, the spatial–temporal evolution characteristics of UHI in the study area were analyzed from a seasonal perspective (summer and winter) using the standard deviation ellipse (SDE) model. Finally, the driving mechanisms of summer and winter UHI in the study area were comprehensively analyzed using the GeoDetector model, profile analysis and Pearson correlation coefficient. The specific technical process is depicted in Fig. 2.

Fig. 2
figure 2

Flowchart of the study.

Full size image

LST retrieval

LST was retrieved using an atmospheric correction approach based on Landsat thermal infrared data. The specific calculation formulas are as follows34,35.

$$L_{\lambda } = [\varepsilon B(LST) + (1 - \varepsilon ) \times L_{ \downarrow } ]\tau + L_{ \uparrow }$$
(1)
$$B(LST) = \frac{{L_{\lambda - } L_{ \uparrow } - \tau \times (1 - \varepsilon ) \times L_{ \downarrow } }}{\tau \times \varepsilon }$$
(2)
$$LST = K_{2} /{\text{ln}}\left( {\frac{{K_{1} }}{B(LST)} + 1} \right)$$
(3)

where \(L_{ \uparrow }\) represents the rising atmospheric radiation, \(L_{ \downarrow }\) represents the downwelling atmospheric radiation, \(\varepsilon\) is the surface emissivity, B(LST) denotes the blackbody radiance and \(\tau\) is the atmospheric transmittance. K1 is the calibration coefficients of the sensor, K2 is the calibration coefficients of the sensor. For Landsat ETM + and Landsat OLI imagery, the K1 values are 606.76 W/(m2·sr·μm) and 774.89 W/(m2·sr·μm), respectively, while K2 is 1260.56 K.

Following LST retrieval, the mean-standard deviation method was applied to classify LST into six categories: extreme high, high, strong medium, weak medium, low, and extreme low temperatures. Heat island regions involve in the areas with extreme high, high, and strong medium temperatures, which are further classified as strong, medium, and weak heat islands, respectively. Weak medium, low, and extreme low temperatures are categorized as non-heat island regions (Fig. 3).

Fig. 3
figure 3

Comparison of LST grades with UHI types. T represents the mean value of the LST. σ represents the standard deviation of the LST.

Full size image

Land cover mapping

In light of the unique circumstances of the land cover and study purpose in the study area, the land cover classification system was built, which consists of seven land cover types (i.e., artificial surface, constructed wetland, natural open water, natural wetland, cropland, green land, and unvegetated land) divided into five land cover classes (i.e., artificial surface, wetland, cropland, green land, and unvegetated land). A comprehensive overview of this classification system is provided in Table 2.

Table 2 Land cover categorization system.
Full size table

Land cover maps for the study area in 2013, 2018, and 2023 were generated based on the object-oriented random forest method. This classification method reduces the risk of overfitting by constructing multiple decision-making trees and integrating their predictions, while improving model generalization and robustness through diverse tree structures, making it suitable for handling high-dimensional and large-scale datasets36. The accuracy of land cover classification was comprehensively evaluated using metrics including the confusion matrix, Kappa coefficient, producer accuracy, user accuracy, and overall accuracy.

SDE model

To conduct a quantitative analysis of the spatial–temporal evolution patterns of summer and winter LST in the study area, the spatial arrangement and orientation change of LST can be comprehensively reflected by calculating the area, major and minor axes, centroid, and azimuth of the SDE model. The specific calculation formulas for SDE are as follows37,38.

$$\sigma_{x} = \sqrt {\frac{{\sum\nolimits_{i = 1}^{n} {\left( {\overline{x}_{i} - \overline{X}} \right)^{2} } }}{n}}$$
(4)
$$\sigma_{y} = \sqrt {\frac{{\sum\nolimits_{i = 1}^{n} {\left( {\overline{y}_{i} - \overline{Y}} \right)^{2} } }}{n}}$$
(5)
$$\left( {\overline{X}_{w} ,\overline{Y}_{w} } \right) = \left( {\frac{{\sum\nolimits_{i = 1}^{n} {w_{i} x_{i} } }}{{\sum\nolimits_{i = 1}^{n} {w_{i} } }},\frac{{\sum\nolimits_{i = 1}^{n} {w_{i} y_{i} } }}{{\sum\nolimits_{i = 1}^{n} {w_{i} } }}} \right)$$
(6)
$${\text{tan}}\theta = \frac{{\left( {\sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{x}_{i}^{2} } - \sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{y}_{i}^{2} } } \right) + \sqrt {\left( {\sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{x}_{i}^{2} } - \sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{y}_{i}^{2} } } \right)^{2} + 4\sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{x}_{i}^{2} \overline{y}_{i}^{2} } } }}{{2\sum\nolimits_{i = 1}^{n} {w_{i}^{2} \overline{x}_{i} \overline{y}_{i} } }}$$
(7)

where σx is the long axes of the SDE, σy is the short axes. xi denotes the longitude coordinates, yi denotes the latitude coordinates. \(\overline{x}_{i}\) and \(\overline{y}_{i}\) denotes the coordinate offsets of latitude and longitude from the ith region to the SDE’s centroid, correspondingly. \(\overline{X}\) and \(\overline{Y}\) denotes the mean center of each area’s latitude and longitude, correspondingly. \(\left( {\overline{X}_{w} ,\overline{Y}_{w} } \right)\) stands for the centroid of the SDE coordinate. wi signifies the ith area’s weight. \(\theta\) denotes the clockwise angle between the SDE’s long axis and the north direction.

Driving factor selection of UHI effect

Analyzing the influencing factors is beneficial for investigating the UHI effect mechanism. To fulfill this, 8 indicators were chosen from natural and socio-economic perspectives in the study. The specific indicator information is shown in Table 3.

Table 3 Driving factors of UHI effect selected in the study.
Full size table

GeoDetector model

By identifying regional stratified heterogeneity, the GeoDetector model can quantitatively assess the ability of independent variables explaining dependent variables, making it become a useful tool for examining the driving impacts of intricate geographical features. Accordingly, in this study, factor analysis and interaction analysis of GeoDetector Model are used to assess the capacity of each driving factor to clarify the spatial disparity of urban LST and to examine the interaction patterns among the driving factors39,40.

$$w = 1 - \frac{{\sum\nolimits_{a = 1}^{L} {N_{a} \sigma_{a}^{2} } }}{{(L - 1)N\sigma^{2} }}$$
(8)

where a represents the count of dependent variables, Na denotes the number of samples in the ath layer, and N is the sum of the number of samples across the entire region. \(\sigma_{a}^{2}\) and \(\sigma^{2}\) are the ath layer variances and the overall region variances, respectively. w denotes the explanation ability of the driving factor of the dependent variable, ranging from 0 to 1. A greater explanatory capacity is indicated by a higher w value. The interaction detection method utilizes the factor detection computation of the w values to examine the combined effects of paired driving factors on LST. It determines their interaction (strengthens, weakens and stays independent). The interaction between two driving factors can be categorized into five types, as detailed below (Table 4).

Table 4 Interaction relationships of dual independent variables on dependent variables.
Full size table

Pearson correlation coefficient

Considering the factor analysis can only determine the correlation between various driving factors and the UHI effect without indicating the influence direction (positive or negative), the Pearson correlation coefficient analysis is employed to examine the relationship between the UHI effect and each driving factor. With a value range of [− 1, 1], the Pearson correlation coefficient can assess the degree of linear correlation between two variables41.

$$R = \frac{{\sum\nolimits_{i = 1}^{n} {(x_{i} - \overline{x})(y_{i} - \overline{y})} }}{{\sqrt {\sum\nolimits_{i = 1}^{n} {(x_{i} - \overline{x})^{2} \sum\nolimits_{i = 1}^{n} {(y_{i} - \overline{y})^{2} } } } }}$$
(9)

R represents the Pearson correlation coefficient, while x and y denote the two different variables, respectively.

Profile analysis

Conducting spatial profile analysis in the study could effectively depict the change characteristics of LST on a micro-scale. In this study, three profile baselines were selected, i.e., Shenzhen First Ring Road, Second Ring Road and Third Ring Road, to explore the LST variations at the micro spatial scale in summer and winter. The profile baselines are shown specifically in Fig. 1c.

Results

Spatial–temporal dynamics of land cover

The land cover maps of the Shenzhen city for the years 2013–2023 were created using the object-oriented random forest method, and validation sample points were utilized to assess the classification accuracy. These land cover maps allowed for the analysis of land cover dynamics, as evidenced by the accuracy assessment findings in 2013 (Table S1), 2018 (Table S2), and 2023 (Table S3). The overall accuracy of land cover in each year exceeded 0.91.

Based on the proportion and spatial–temporal distribution of land cover types from 2013 to 2023 in the Shenzhen city (Fig. 4), the artificial surface area increased from 715.68 to 783.44 km2. The area of green land decreased by 11.33 km2. The area of constructed wetlands was reduced by 6.12 km2. The decrease in cropland was the most significant, declining from 401.39 to 363.81 km2. The area of natural wetland, natural open water, and unvegetated land changed by 0.05%, − 0.2%, and − 0.02%, respectively.

Fig. 4
figure 4

Spatio-temporal dynamics of the land cover from 2013 to 2023 in the Shenzhen city. Map generated by the authors using ArcGIS 10.8 (Environmental Systems Research Institute, USA. https://www.esri.com).

Full size image

Spatial overlay analysis provided insights into the spatial–temporal distribution of land cover type transitions (Fig. 5 and Table 5). The land cover types experienced the most drastic conversion during 2018–2023, followed by that during 2013–2018. The conversion of cropland to artificial surface, cropland to green land, green land to artificial surface, and green land to cropland accounted for 75.56% and 69.26% in 2013–2018 and 2018–2023, respectively. These changes were evenly distributed throughout the study area, while conversions between other land cover types were relatively limited.

Fig. 5
figure 5

Spatial–temporal dynamics of land cover changes in the study area from 2013 to 2023. 1 is artificial surface, 2 is constructed wetland, 3 is cropland, 4 is green land, 5 is natural wetland, 6 is natural open water, 7 is unvegetated land. Map generated by the authors using ArcGIS 10.8 (Environmental Systems Research Institute, USA. https://www.esri.com).

Full size image
Table 5 Land cover change types from 2013 to 2023 in the Shenzhen city.
Full size table

UHI differences between summer and winter

Figure 6 illustrates the spatial–temporal features of various UHI types and LST levels in the study area. The areas of strong heat islands initially increased and then decreased. Non-heat island areas displayed an increasing trend followed by a decrease in summer, while in winter, they showed a consistent decrease. Weak and medium heat islands initially decreased and then increased in summer, exhibiting an increasing trend in winter. Notably, heat island areas were primarily located in the central and western regions of the study area, with strong heat islands primarily located in the western and southern parts of the study area. Non-heat island areas were largely distributed in the east of the study area. Furthermore, overlaying land cover data revealed that strong heat island areas were predominantly found on artificial surfaces and unvegetated land, whereas non-heat island areas were mainly present in green land, natural open water, natural wetlands, and constructed wetlands.

Fig. 6
figure 6

Spatial–temporal distribution of summer and winter UHI type in the Shenzhen city from 2013 to 2023.

Full size image

Based on the SDE model, the spatial–temporal evolution features of UHI in the Shenzhen city were analyzed (Fig. 7). Summer results showed that the heat island centroid generally migrated from Southwest to Northeast. The standard deviation ellipse area expanded from 1060.56 to 1131.96 km2, revealing that, as it moved northeast, the spatial clustering features of UHI areas progressively became more intense. Similarly, the winter results indicated that the heat island centroid exhibited an overall migration from Southwest to Northeast. However, the standard deviation ellipse area slightly decreased, from 1223.08 to 1219.28 km2, reflecting that the spatial agglomeration features of the UHI area progressively became stronger to the Northeast.

Fig. 7
figure 7

Spatial–temporal evolution patterns of UHI in the Shenzhen city from 2013 to 2023.

Full size image

Correlation between different factors and LST/UHI

By comparing the mean LST across different land cover types (Table 6), it was observed that from 2013 to 2023, the summer LST showed an overall trend of initially increasing and then decreasing, while the winter LST exhibited a consistently rising trend. Moreover, the average temperature in summer was higher than that in winter. Specifically, the mean LST for different land cover types in summer were as follows: unvegetated land > artificial surface > cropland > constructed wetland > natural open water > green land > natural wetland. In winter, the mean LST were: unvegetated land > artificial surfaces > cropland > natural wetland > green land > constructed wetland > natural open water. Additionally, factor analysis using the GeoDetector model (Table 7) revealed that NDVI, NDWI, and NDBI had significant effects on LST. Furthermore, through the profile analysis, a micro-level comparison was conducted on LST changes across different land cover types along the three major urban ring roads. As shown in Fig. 8, LST changes along the first ring road remained relatively stable in summer, while fluctuations were observed along the second and third ring roads. In winter, all three ring roads exhibited significant variations in LST. Besides, compared to other land cover types, artificial surfaces, cropland, and unvegetated lands demonstrated noticeably higher LST values.

Table 6 Mean LST for each land cover category.
Full size table
Table 7 Evaluation values of each driving factor used for factor detection.
Full size table
Fig. 8
figure 8

LST variations of three profile baselines.

Full size image

Based on Pearson correlation analysis, the relationship between LST and various driving factors in the Shenzhen city was analyzed (Table 8). The results showed that NDBI, GDP, PD, and PM2.5 were positively correlated with LST. Conversely, NDVI, NDWI, and DEM were negatively correlated with LST. Meanwhile, factor detection using the GeoDetector model (Table 7) revealed that land cover type, PD, NDVI, NDBI, and NDWI had a significant impact on UHI. Furthermore, the interaction patterns between driving factors were characterized as either nonlinear or exhibiting bivariate mutual enhancement, with their combined impact on LST surpassing that of individual factors (Fig. 9). In summer, the interplay of land cover type and NDBI had the most substantial effect on LST, whereas, in winter, the interaction between land cover type and PM2.5 had the most pronounced effect on LST.

Table 8 Correlativity between LST and various driving factors.
Full size table
Fig. 9
figure 9

Interaction strength between different driving factors of UHI.

Full size image

Discussion

UHI driving mechanisms associated by urbanization

The process of urbanization is closely associated with rapid population growth and swift economic development, leading to artificial surface expansion and increased energy consumption. On the one hand, at macro and micro scales, built-up areas consistently exhibit the highest temperatures. The artificial surface expansion, with an area increasing by 67.76 km2 from 2013 to 2023, has resulted in the conversion of natural land cover types, such as farmland, forests, and wetlands, into built-up areas, industrial zones, and transportation infrastructure42. Additionally, water and green land have cooling benefits. The reduction of natural surfaces like water, with an area decrease of 3.78 km2 from 2013 to 2023, has weakened cooling effects, while impervious surfaces (e.g. roads, concrete buildings) with high thermal capacity and low albedo absorb and store more solar radiation, significantly elevating urban temperatures43. Furthermore, the dense concentration of buildings hinders the effective heat dissipation, resulting in high-temperature areas being concentrated in high-density building districts44. This outcome manifests as a significant temperature difference of 6–9 °C between these areas and green spaces, which in turn intensifies the formation of localized heat islands (Figure S1 and Figure S2).

On the other hand, the high energy consumption associated with urbanization generates significant waste heat. In summer, extensive air conditioning systems transfer indoor heat outside, which further elevates environmental temperatures45. Industrial activities and vehicular emissions also produce large quantities of exhaust and dust, exacerbating the accumulation of heat and pollutants in the atmosphere. In winter, Shenzhen is frequently influenced by cold air masses from the north, which often lead to the formation of temperature inversion layers46. These inversion layers occur when cooler air near the ground is overlain by warmer air, causing temperature to increase with altitude47. In addition, the retention of pollutants under temperature inversion conditions intensifies their radiative effects, particularly the absorption of long-wave radiation, which further contributes to the amplification of the UHI during winter48. Therefore, the temperature inversion phenomenon significantly hinders the vertical dispersion of pollutants, including PM2.5, within the lower atmosphere. Meanwhile, the accumulation of PM2.5 and other atmospheric suspended particles enhances the absorption and scattering of solar radiation49. These processes alter the urban energy balance and reduce the heat dissipation effect, thereby exacerbating the UHI effect. In addition, by enhancing long-wave radiation, these particles impede nighttime cooling, thereby amplifying the UHI effect50.

Urban policies and planning are another key factor influencing the spatial–temporal dynamics of UHI. The development of the Shenzhen city has adopted a multi-center planning model, wherein different regions have evolved as relatively independent functional centers. For example, Futian and Nanshan serve as commercial and technology hubs, whereas Baoan and Longgang are primarily industrial areas51,52. Each of these functional zones tends to become a separate heat accumulation zone as a result of its high human activity density and specific architectural features. The intensive development and human activity in each zone independently contribute to the formation of localized heat islands, which, to some extent, explains the multi-center distribution pattern of the UHI effect in the Shenzhen city.

Measures for mitigating UHI

The driving factors of UHI effects exhibit notable differences between summer and winter. In summer, UHI is primarily driven by the combined effects of the expansion of built-up areas, where the density of urban buildings and their heat absorption features significantly elevate surrounding temperatures53. Moreover, the main driving factors for winter UHI are land cover changes and PM2.5. PM2.5 emissions not only exacerbate urban air pollution but also have a substantial impact on local temperatures.

To effectively mitigate UHI, Shenzhen promulgated the “Sponge City Construction” policy in 201654. Through systematic planning and interdepartmental collaboration, Shenzhen adopted the “Prioritize Green over Gray” principle (giving precedence to green infrastructure, such as rooftop gardens and permeable pavements) in sponge city construction. The measure effectively manages rainwater and reduces surface temperatures through evapotranspiration. Furthermore, Shenzhen integrated sponge city construction requirements into the entire process of construction projects, which aims to comprehensively mitigate the UHI. Likewise, by increasing the area of urban green land55 (lawns, parks, etc.) and water features56 (ponds, fountains, etc.), the evaporative cooling effects of plants and water also could be enhanced to lower surrounding temperatures57. Moreover, industrial and vehicular emissions have been reduced while clean transportation methods, such as electric vehicles and public transit, have been promoted. These actions have improved air quality and created more open spaces that facilitate natural cooling, particularly in areas with elevated PM2.5 concentrations.

The implementation of these measures could alleviate the UHI effect, improve the urban microclimate, and enhance the general quality of life for inhabitants. By mitigating the impact of these factors on summer and winter UHI, more sustainable and livable environments can be fostered in cities.

Advantages, uncertainties, and future research directions

To systematically analyze the driving mechanisms of the UHI, it is essential to clarify the features of UHI evolution across various spatial–temporal scales and understand the interactions among the factors influencing the UHI effect58. This study employs remote sensing data and spatial statistical methods to comprehensively analyze the LST spatial–temporal dynamics at macro and micro scales. Such a method is instrumental in understanding LST variations across different land cover types. Besides, the GeoDetector model is employed to quantitatively evaluate the impact of each factor and the interactions between factors on UHI. Previous studies that often focused on the influence of individual factors (e.g. urban expansion59, vegetation and atmospheric concentration60) on UHI, with neglecting the interactions between different factors. In comparison, the GeoDetector provides a more precise representation of the relationships among the driving factors and offers an in-depth analysis of their synergistic or antagonistic effects on UHI, which allows for effectively exploring the UHI driving mechanisms.

Urban ecosystems are complex and large-scale systems, with LST being significantly influenced by seasonal variations. In the analysis of seasonal UHI, the temporal span and coverage of the data must adequately represent the typical features of different seasons. However, this study only examined the UHI effects in summer and winter, excluding spring and autumn, which results in some incompleteness in the research findings. Moreover, the analysis of land cover changes in the Shenzhen city over the past decade is based on satellite remote sensing images, and the results are constrained by the temporal resolution of Landsat images, potentially affecting the accuracy of the study. In future research, spring and autumn should be included to provide a comprehensive understanding of seasonal UHI variations. Additionally, data sources would be further enhanced. By integrating climatic data and hourly temperature data14, etc., further comparisons of UHI between day and night across different seasons would be conducted, thereby investigating the variations in UHI under different climatic conditions61,62,63. Moreover, urban morphological parameters, such as building height, density, and street canyon effects, would be integrated into the analysis to better understand the spatial heterogeneity of micro-scale UHI. Furthermore, advanced technologies, including machine learning algorithms and urban 3D morphological modeling, would be leveraged to analyze large datasets and identify complex patterns in UHI dynamics.

Conclusions

Understanding the spatial–temporal evolution patterns and driving factors of UHI effect across different seasons is essential for assessing the influence of natural landscapes and human activities on the urban heat surroundings, and informing sustainable urban development planning. Based on the land cover classification and LST inversion, the SDE model, GeoDetector model and profile analysis were utilized to investigate the spatial–temporal variation features and driving factors of summer and winter UHI effect in the Shenzhen city from 2013 to 2023. The results revealed that the LST in the Shenzhen city exhibited an initial decline followed by an increase in summer, whereas, it displayed a continuously rising trend in winter. The summer and winter UHI effect demonstrated a migration from southwest to northeast, along with increased spatial clustering. Artificial surfaces, croplands, and unvegetated land had higher LST than other categories. The summer UHI was collectively driven by land cover changes and NDBI, and the winter UHI was jointly driven by land cover changes and PM2.5. These conclusions could not only provide valuable insights for other UHI studies but also act as a guide for creating workable and sensible urban development plans.