Monitoring Changes in Landsat Thermal Features in Urban and Non-Urban Interfaces from 1986 to 2023 in Two International Urban Centers: Implications for Climate and Global Issues

Highlights

What are the main findings?

• We developed a prototype approach combining long-term Landsat thermal data with dynamic land cover products to quantify SUHI intensities and hotspots at city scale.
• Urban growth significantly increased land surface temperatures, with Wuhan showing stronger warming trends (0.04 °C/year) compared to Brasília (0.01 °C/year).

What are the implications of the main findings?

• Provides a scalable framework for combining land cover and thermal satellite data to monitor surface urban heat dynamics.
• Supports sustainable urban planning and climate adaptation through improved SUHI tracking and analysis.

Abstract

Rapid urbanization is reshaping thermal environments worldwide, with the strongest impacts occurring at the interface between urban and non-urban areas. Impervious surfaces, as key indicators of urban expansion, are critical for monitoring urban growth and assessing surface urban heat island (SUHI) effects. Land use and land cover change (LULCC) provides an essential link between urban dynamics and their environmental and societal consequences. Here, we integrated the U.S. Geological Survey (USGS) Climate Global Issues (CGI) Land Cover Product with Landsat thermal time-series to investigate SUHI evolution in two contrasting metropolitan regions: Wuhan, China, and Brasília, Brazil. Using data spanning 1986–2023, we analyzed the relationships between land cover, Landsat-based land surface temperature (LST), and SUHI intensity, and identified persistent thermal hotspots. Results demonstrate that the land cover data utilized increases the accuracy of impervious surface mapping along urban–rural gradients. Average SUHI intensities were 3.4 °C in Wuhan and 3.3 °C in Brasília, with statistically significant warming trends of 0.04 °C/year and 0.01 °C/year, respectively. Maximum temperature proved to be a robust indicator of SUHI intensification, capturing long-term upward trends. Our findings highlight the important role of urban land cover dynamics in shaping temporal SUHI variability and hotspot emergence. This prototype framework demonstrates the scientific and policy value of combining long-term land cover monitoring information with satellite thermal monitoring to quantify and track SUHI at city scale, supporting sustainable urban planning and climate adaptation strategies.

1. Introduction

Urban areas have expanded dramatically over recent decades, driven by rapid population growth and economic development. Global urban populations are projected to increase by 2.5 billion people over the next 30 years, with nearly 68% of the world’s population expected to reside in urban areas by 2050 [1,2,3]. This transformation of natural landscapes into built environments alters surface energy balances, increases land surface temperatures (LST), and amplifies surface urban heat island (SUHI) effects, where urban LSTs are consistently higher than those of surrounding rural areas [4,5]. The intensification of SUHI is a biophysical process and a pressing societal concern, because SUHI interacts with climate change to exacerbate heat stress, biodiversity loss, fire risk, air quality degradation, urban hydrological challenges, and energy consumption, ultimately threatening public health and socio-economic resilience [6,7,8,9].

Many studies have worked to understand and quantify SUHI by either comparing in situ temperatures from weather/climate station records [5,10] or comparing wall-to-wall land surface temperature (LST) obtained from thermal remote sensing information [4,11,12]. Currently, the most widely used approach for studying SUHI is leveraging satellite remote sensing data to quantify its intensity and spatial distribution across large areas. These datasets are derived from multiple sensors, including Landsat [13,14,15], MODIS [16], VIIRS [17], and ECOSTRESS [12,18]. Studies utilizing thermal remote sensing data over extensive regions indicate that urban-induced warming is evident in many major cities worldwide, and addressing the global implications of SUHI will require innovative solutions [5,18,19].

The magnitude and variability of SUHI are shaped by complex factors, including city size, geographic setting, background climate, vegetation cover, and the intensity of human activities. These factors are expected to intensify under global warming and continued urbanization [2,6,20]. Understanding how urban land cover dynamics drive SUHI is therefore critical for advancing mitigation and adaptation strategies in support of sustainable urban development. Numerous studies have highlighted the value of accurately monitoring impervious surface expansion to understand SUHI trends [21,22,23,24]. Precise mapping of impervious surfaces remains central to linking land use and land cover change (LULCC) with thermal responses in urban and peri-urban environments [20,25].

Advances in remote sensing have enabled monitoring of SUHI at fine spatial and temporal scales. Landsat data, in particular, provide a unique 30-m resolution archive extending over five decades, enabling consistent long-term monitoring of urbanization and its thermal impacts [26]. The U.S. Geological Survey (USGS) has further advanced this potential through the provision of U.S. Landsat Analysis Ready Data (ARD), which reduces preprocessing requirements and supports systematic land change assessments [27]. These datasets underpin continental-scale efforts such as the Land Change Monitoring, Assessment, and Projection (LCMAP) Initiative and SUHI monitoring across the conterminous United States [5,28]. Moreover, Xian et al. [25] applied Landsat ARD towards an assessment of SUHI effects in U.S. cities, and demonstrated a novel approach based on Landsat time series of land cover and Landsat LST data that characterized changes in urban and surrounding non-urban lands to quantify the SUHI intensity, hotspots, and thermal conditions.

Urban development continues to convert non-urban land into urban landscapes worldwide. Despite this global trend, large-scale, systematic assessments of long-term surface urban heat island (SUHI) variations—particularly outside the conterminous United States (CONUS)—remain limited. This study addresses that gap by evaluating SUHI dynamics using Landsat thermal archives and available regional land cover data [25,29,30]. We focus on a universally applicable protocol for deriving LST from Landsat thermal data, integrating ancillary land cover datasets with global coverage. This approach overcomes key challenges in computation, storage, and data management, enabling comprehensive “wall-to-wall” assessments of SUHI effects and land cover change. Our objectives are to: (1) process and analyze multi-decadal Landsat-derived land surface temperature (LST) and land cover data to detect SUHI patterns and quantify their effects; (2) assess the influence of urban land cover dynamics on SUHI intensity trends over time and identify spatial hotspots of change; (3) examine the relationship between impervious surface expansion and SUHI variability across different urban contexts; and (4) demonstrate the scalability and applicability of the proposed workflow for global SUHI monitoring and its potential integration into climate adaptation and urban planning strategies.

To achieve these goals, we used Landsat time series data and regional, long-term land cover datasets developed by the USGS Earth Resources Observation and Science (EROS) Center. This enabled quantification of annual LST and land cover changes, including urban extent, persistence, and urban–rural interface dynamics. We applied this framework to Wuhan, China, and Brasília, Brazil, analyzing SUHI patterns from 1986 to 2023. The findings offer critical insights into long-term SUHI trends and support the development of scalable mitigation strategies. Ultimately, this methodology enhances our ability to monitor and characterize SUHI intensity across diverse geographic regions worldwide.

2. Materials and Methods

We completed this research using several steps. We first selected two urban centers as study areas (Section 2.1) and prepared annual land cover dynamics (Section 2.2) from existing land cover data published by the USGS. We developed the models and retrieved annual LST from Landsat’s Collection 2 thermal product (Section 2.3) following protocols of consistent estimation required for a statistically rigorous analysis with collected reference datasets from various existing sources with multiple spatial resolutions and temporal frequencies (Section 2.4). These reference datasets provided the basis for validation and comparisons. We analyzed SUHI intensity and trends (Section 2.5). Finally, we identified SUHI hotspots based on probability of annual hotspots analysis during the study period (1986–2023) (Section 2.6). Figure S1 shows the workflow chart of assessing SUHI effects by using Landsat time series LST and land cover datasets for better understanding the key steps for the proposed method. The method used in the study comprises several elements including Landsat research of interest (ROI) LST processing, land cover preparation, urban center selection, determination of urban and rural buffer zones, SUHI intensity and hotspots, and temporal trend quantifications. The general approach is summarized by a flowchart in Figure S1.

2.1. Study Areas

We applied the new land change monitoring and SUHI characterization approach to two international urban centers: Wuhan, China, and Brasília, Brazil (Figure 1). Each region of interest (ROI) spans 300 × 300 km² and was selected to capture contrasting continents, climates, and urban development patterns, while also benefiting from relatively rich Landsat archives dating back to the 1980s. Existing land cover data were derived from the Global Land Cover and Estimation (GLanCE) project, which uses continental Lambert Azimuthal Equal Area projections to minimize distortion [31]. GLanCE tiles measure 150 × 150 km (5000 × 5000 pixels at 30 m resolution), and this geometry guided our preprocessing and study area boundaries (https://measures-glance.github.io/glance-grids/, accessed 5 January 2025). Wuhan and Brasília were chosen because they represent markedly different urbanization trajectories, climatic conditions, and socio-economic contexts, making them ideal for assessing how diverse drivers influence SUHI. Wuhan is a rapidly urbanizing megacity in a humid subtropical climate, whereas Brasília is a planned city in a tropical savanna climate with slower urban growth. This comparison provides insights into global urban heat dynamics under contrasting environmental and developmental conditions.

Brasília (15°47′38″S, 47°52′58″W), the capital of Brazil (Figure 1a), has a tropical savanna climate with distinct wet (October–April) and dry (May–September) seasons. The metropolitan population exceeds 3 million, with rapid urban expansion since its inauguration in 1960. It has the highest GDP per capita among Brazilian cities, driven by public administration and services [33]. The climate’s strong seasonality, with dry-season relative humidity often <30%, makes it an important case for SUHI research under semi-arid conditions [34].

Wuhan (29°58′–31°22′N, 113°41′–115°05′E), located in Hubei Province along the Yangtze River (Figure 1b), has a humid subtropical climate with hot summers (>30 °C in July) and a mean annual temperature of 17.4 °C. With more than 11 million residents, it is a major economic hub of central China, experiencing rapid urbanization, industrialization, and associated thermal stresses in recent decades [35].

2.2. Annual Land Cover Dynamics

The USGS EROS Center has previously developed and published provisional land cover data for our two study regions [29]. Annual land cover data (1986–2023) for Wuhan and Brasília were produced to evaluate the potential of redeploying methods developed as part of the LCMAP Initiative to other regions of the world. LCMAP produced annual land cover for CONUS using a time-series approach based on the Continuous Change Detection and Classification (CCDC) method [36] which makes use of every cloud-free observation in the Landsat 30-m data record [5,28]. Production of land cover for our two study areas used all available 30-m Collection 2 Landsat imagery in the USGS global archive that was reprojected from various Universal Transverse Mercator (UTM) projections to appropriate continental Lambert Azimuthal Equal Area and tiled according to the GLanCE grid system. The CCDC method was applied on a per-pixel basis to create harmonic, time-series models across the entire period and the resulting coefficients used were as inputs to a gradient boosted classifier for the assignment of land cover labels. Land cover training data were initially sourced from the GLanCE project global training dataset [37] which was developed from a variety of preexisting sources along with new analyst-collected data.

The Stanimirova et al. [37] training dataset was inadequate for producing desired results in the Brasília and Wuhan regions for the seven targeted land cover classes; Developed, Cropland, Grass/Shrub, Tree Cover, Water, Wetlands, and Barren. An approach similar to that applied in Bratic et al. [38] was used to improve the training dataset by sampling from regions where preexisting, published maps were in agreement. For classes where multiple, preexisting maps were not available, the training data were augmented by random sampling from the global GLC_FCS30D dataset [39].

Table 1. Published land cover dataset from which additional land cover training data were sampled.

Land Cover Class	Dataset	Citation
Developed	GLC_FCS30D	[39]
	GLAD BuiltUp	[40]
Cropland	GLC_FCS30D	[39]
	GLAD Cropland	[40]
	LGRIP30	[41]
Grass/Shrub	GLC_FCS30D	[39]
Tree Cover	GLC_FCS30D	[39]
	GLAD Tree Cover	[40]
Water	GLC_FCS30D	[39]
	LGRIP30	[41]
	GLAD Water	[40]
Wetlands	GLC_FCS30D	[39]
Barren	GLC_FCS30D	[39]

lists the land cover classes mapped and the data sources used to augment training data. The resulting annual land cover maps for Wuhan and Brasília are openly available from USGS ScienceBase [29].

2.3. Annual Land Surface Temperature (LST) Retrieval

Initial LST data preparation required reprojection and tiling. As with Landsat spectral data used in the land cover component, LST data published by the USGS to the global archive are made available in Universal Transverse Mercator (UTM) projections. These data were prepared with the same protocols as the spectral data; first reprojecting imagery into the appropriate regional Lambert Azimuthal Equal Area projection followed by tiling into the grid system of each study area. The LST calculation, including empirical direct methods that incorporate remotely sensed thermal data with quality assessment (QA) into semi-empirical models, was divided into two parts: seasonal and annual. Seasonal LST values were derived from clear observations for each year. The annual mean LST represents the average of all clear observations within a given year. To estimate the annual maximum LST for each pixel, the seasonal maximum temperatures were averaged [25]. We implemented a data processing algorithm to analyze all Landsat LST records from 1985 to 2023, derived from the Landsat thermal band (

Table 2. Data used in the study.

Sensor	Data	Resolution	Time	Source
Landsat	Land cover	30 m	1986–2023	USGS [29]
	Thermal	30 m	1986–2023	USGS [13,14,15]
	QA 1	30 m	1986–2023	USGS [13,14,15]
MODIS	Surface temperature	1000 m	2000–2023	NASA [16]
VIIRS	Surface temperature	1000 m	2018–2023	NASA [17]
GHCN 2	Air temperature	Station (point)	1986–2023	NOAA [44]
Vector data	Global urban boundary	Polygon	2023	ESRI [32]

¹ Quality Assessment (QA) Bands; ² Global Historical Climate Network (GHCN).

) at a 30-m resolution for the selected cities (Wuhan and Brasília) [42]. The algorithm calculates annual means of LST and identifies temporal trends. Over 2,250 Landsat records in each ROI tile were processed to estimate these annual means, with temporal trends determined using a linear regression model. Validation indicated small residual errors of −0.27 K ± 0.89 K for cloud-free conditions and acceptable errors of −0.93 K ± 2.46 K in cloud-affected areas [27,43]. We acquired the Landsat LST products from the USGS global Landsat archive, filtering out pixels contaminated by clouds, shadows, and snow/ice using USGS-supplied QA data to ensure data quality. Clear observations were prioritized, and the analysis showed an increase in clear observations post-1998 due to additional sensors, while seasonal distributions remained consistent, with summer yielding the highest number of clear observations. To analyze trends, we used linear regression on the annual LST from 1985 to 2023, utilizing p-values and coefficient of determination (R²) values to assess statistical significance. In addition to the annual summary statistics, seasonal mean LST and maximum temperatures were also calculated and organized into four seasonal groups (December–February, March–May, June–August, September–November). This comprehensive approach facilitates a clearer understanding of LST variations and their implications for SUHI studies [5,42].

2.4. Reference Data, Extracting Strategy, and Validation

We selected three existing reference datasets for this study (Table 2) to evaluate and compare with our LST products. The first dataset is the NOAA GHCN data, which consists of an integrated database of daily climate summaries from land surface stations worldwide [44]. The GHCN includes daily climate records from various sources, all integrated and subjected to a comprehensive quality assurance process. The NOAA National Centers for Environmental Information (NCEI) offers numerous daily variables, including maximum and minimum temperatures. After applying the necessary filters for this study, only one station in Wuhan had complete records from 1985 to 2023, while Brasília had very limited records from 1990 to 1997. The second dataset is from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard NASA’s Terra and Aqua Earth Observing System satellites, which provides multiple daily LST products available from 2000 to 2023. The latest Collection 6.1 MODIS LST includes three refinements over bare soil surfaces compared to the earlier Collection 6 product [45]. The third dataset comprises VIIRS-derived products used to measure cloud and aerosol properties, ocean color, LST, ice movement, temperature, fires, and Earth’s albedo, available from 2018 to 2023 [17].

We utilized the GHCN station located within the ROI and converted the station points to match the Landsat resolution (30 m × 30 m) by extracting all pixels in the defined study area as a mask to spatially align with the other LST datasets. This extraction strategy was chosen to facilitate straightforward comparisons across multiple spatial resolution LST datasets and to match the locations of GHCN stations. Additionally, since GHCN represents field observations, it provides more accurate air temperature readings compared to other remote sensing-derived LST data. We evaluated the Landsat-derived annual LST against other remote-sensor-derived LST products. Thus, we utilized weekly composites of MODIS and VIIRS data, which enhance quality through composite algorithms that mitigate cloud impacts, to compare them with GHCN data and evaluate the GHCN comparability with remote sensing-derived LST data. Additionally, we used the GHCN station locations to create 900 × 900 m masks, equivalent to 3 × 3 30 m Landsat pixels, to obtain average values from Landsat LST and extract MODIS and VIIRS data (at 1000 m), ensuring comparability within the same land cover class. All these steps were implemented using geospatial analysis and geostatistical models (Figure S1). Specifically, satellite-derived LST was compared and validated against station-based air temperature data through spatial and temporal matching procedures. We applied buffer-based spatial alignment to ensure that LST pixels corresponded to the vicinity of meteorological stations, and temporal matching was performed by selecting satellite overpass times closest to station observation periods. To reconcile the inherent differences between LST and air temperature, we employed regression-based models and statistical correlation analysis, enabling us to quantify relationships and assess consistency across varying conditions [46].

2.5. Quantification of SUHI Intensities, Variations, and Trends

The SUHI intensity refers to the difference in temperature between urban areas and their surrounding rural areas, typically quantified as the temperature anomaly or excess heat observed in urban environments. In our study, the pixel-level SUHI intensity was determined by the magnitude of LST difference between an urban pixel and a mean of the surrounding rural buffer. We used persistent urban land cover extent during 1985–2023 to create a 5 km buffer zones for two study areas. We conducted a sensitivity analysis to determine the optimal buffer size for SUHI calculations. By comparing different buffer sizes around various urban centers of differing area and population, we found that a 5 km buffer provides the most reliable estimate of SUHI intensity [5]. The mean of temperature of persistent for non-urban class is calculated annually, which non-urban class includes tree cover, cropland, barren, grass/shrub, water, wetland, and snow/ice. Therefore, SUHI annual intensities associated with the persistent urban land cover type can be compared directly. To analyze trends, we employed linear regression on the annual LST intensity from 1985 to 2023, utilizing p-values and R² values to assess statistical significance [47]. Pixels affected by clouds and shadows were excluded from the regression analysis [25].

2.6. SUHI Hotspot Analysis

The SUHI hotspot refers to specific areas within an urban environment that experience significantly elevated temperatures compared to their surroundings, particularly during warmer months or at certain times of the day. These hotspots can adversely affect local microclimates, influence energy demand for cooling, compromise air quality, and impact public health. Identifying these hotspots is essential for urban planning and implementing mitigation strategies, such as increasing green spaces, enhancing urban design, and improving reflective surfaces. Persistent urban areas, identified as areas classified Developed across the entire study period, are derived from land cover datasets and are selected for identifying SUHI hotspots from 1986 to 2023 at the pixel level. The objective of analyzing persistent urban hotspots using annual mean LST is to pinpoint areas that are relatively warmer than their neighborhoods. The spatial extent, land cover types, and thermal conditions of these neighborhoods influence the location and LST intensity of persistently urban areas. For example, urban areas with high-intensity impervious surfaces typically have higher LSTs than those with low-intensity imperviousness. Hotspot analysis focuses on relatively warmer areas within persistent urban lands surrounded by similar landscape conditions. We calculate annual hotspots based on LST intensity, defining LST intensity of persistent urban pixels as greater than the mean of all persistent urban LST intensities plus one standard deviation during 1986 to 2023. Additionally, we assess the probability of SUHI hotspots by stacking annual hotspot layers and calculating the percentage of hotspots at the pixel level during the study period. The thresholds used to classify hotspot levels are: >50% as High, 20–49% as Medium, and < 20% as Low. Pixels with High levels are identified as SUHI hotspots for the selected study areas. Utilizing annual mean LST intensity over the long term to quantify SUHI hotspots in selected urban centers allows for rapid identification at the neighborhood scale within persistent urban areas, reducing potential biases caused by data quality and annual LST variations associated with different intensities of developed impervious surfaces. Annual land cover and change data are used to characterize SUHI hotspots as persistent over the recognized urban areas during the 1986–2023 period. This analysis provides a direct and systematic estimate of SUHI impacts in areas with substantial thermal effects [25].

3. Results

3.1. Land Cover Changes in Urban Areas and Urban-Rural Interfaces

We developed a dynamic land cover dataset for two study areas to support SUHI analysis. The distribution of land cover for the years 1987, 1990, 1995, 2000, 2005, 2010, 2015, 2020, and 2023 for these areas is displayed in Figure 2 (Wuhan) and Figure 3 (Brasília). Urban land cover changes in Wuhan from 1987 to 2023 are illustrated in Figure 2j. Cropland is the predominant non-urban land cover surrounding the urban center, as shown in Figure 2. Most new urban development occurs along existing urban areas, expanding in nearly all directions. The distribution of urban development from 1987 to 2023 is presented in Figure 3a–i, with intervals of three to five years. Changes in urban and non-urban land cover within 5-km buffers are depicted in Figure 3j,k. Notably, total urban land cover within this buffer zone increased from 33% to 45% between 1987 and 2023, while cropland experienced the largest decline. Total non-urban land cover decreased from 67% to 55% in the same period, indicating that urban expansion primarily occurred at the expense of cropland in Wuhan.

Similarly, urban land cover changes in Brasília between 1987 and 2023 are shown in Figure 3j. Grass/shrub and cropland are the major non-urban land covers surrounding the urban center (Figure 3). As with Wuhan, most new urban development follows existing urban areas and expands in nearly all directions. The distribution of urban development from 1987 to 2023 is displayed in Figure 3a–i with three- to five-year intervals. Changes in urban and non-urban land cover within 5-km buffers are illustrated in Figure 3j,k. For Brasília, total urban land cover increased from 10% to 16% of the total area within the buffer zone from 1987 to 2023. The most significant decreases were observed in grass/shrub and cropland, with total non-urban land cover declining from 90% to 85% in the same timeframe. Urban expansion in Brasília primarily involved the conversion of cropland and grass/shrub areas.

3.2. The Spatial Distributions of LST and Trends Analysis

Processing time-series annual LST datasets is critical for studying the SUHI effect. To illustrate the spatial distribution of SUHI in two selected study areas, we used LST and its changes within a 5-km buffer (Figure 4 and Figure 5). The annual average LST for selected years reveals numerous high-temperature patches within the urban areas of Wuhan (Figure 4a–i). For example, in 1987, relatively low LST values on the left side of the map can be attributed to limited availability of Landsat data prior to 1998, observed in both urban and rural areas.

LST differences in most years indicate that temperature magnitudes in larger urban centers and surrounding towns are generally higher than those in adjacent non-urban areas (Figure 4). The annual LST for regions converted to urban land after 1986 suggests elevated temperatures for these urban pixels. Significant surface temperature differences are observed in the central part of the buffer areas for Wuhan, notably around older downtown areas with business complexes (before 1990). New developments, post-1995, especially after 2000, exhibit substantial positive differences in LST, particularly in regions adjacent to older urban areas, newly developed towns, and major infrastructure such as the new international airport. In contrast, many older residential buildings in central urban areas close to rivers, lakes, and wetlands, or areas with dense vegetation cover, display comparatively smaller or even negative LST differences. The distribution of LST differences also reveals several areas with relatively low magnitudes, likely due to the limitations of Landsat thermal data availability. Generally, LSTs in developed, urban pixels exceed those in surrounding non-urban areas.

Furthermore, LST trends have been increasing in both persistent urban and non-urban areas from 1986 to 2023 (Figure 4j), but the growth magnitude in persistent urban areas is significantly stronger than in non-persistent regions. Both urban and non-urban areas exhibit positive trends of 0.32 °C/year and 0.24 °C/year in annual LST, though these trends have p-values greater than 0.10 (Figure 4j). The overall change (R²) in annual mean LST for urban and non-urban lands from 1986 to 2023 is displayed in Figure 4k.

Similarly, the annual average LST for selected years in Brasília also shows numerous high-temperature patches within urban areas (Figure 5a–i). LST differences in most years indicate that larger urban areas and surrounding satellite urban regions generally have higher magnitudes than those in adjacent non-urban areas (Figure 5). The annual LST for regions converted to urban areas after 2000 suggests elevated temperatures for these urban pixels. Significant surface temperature differences are observed in the central part of the buffer for Brasília, particularly around earlier urban areas with government administration buildings and business complexes. New developments after 2000, especially in both eastern and western directions, also show substantial positive LST differences, particularly around older urban areas, newly developed towns, and the new international airport on the south side of the city.

The distribution of LST differences reveals some years with lower magnitudes, likely due to the constraints of Landsat thermal data availability. Generally, LSTs in developed pixels exceed those in surrounding non-urban pixels. Additionally, trends in LST have risen in both persistent urban and non-urban areas from 1986 to 2023 (Figure 5j), with persistent areas experiencing significantly stronger growth compared to non-persistent areas. Urban and non-urban areas show positive annual LST trends of 0.34 °C/year and 0.29 °C/year, respectively, although both trends have p-values greater than 0.10 and are therefore not statistically significant (α = 0.05; Figure 5j). The overall change (R²) in annual LST for urban and non-urban lands from 1986 to 2023 is illustrated in Figure 5k. The average maximum LST magnitudes of 38 years are 36.3 °C for urban areas and 35.2 °C for non-urban areas, with urban regions exhibiting smaller standard deviations (2.7) than their non-urban counterparts (3.2) across all years. The urban area also has relatively smaller standard deviations than non-urban areas for all years, although this varies by individual year.

3.3. The Evaluation and Comparison of LST

We utilized available GHCN observation data and other existing remote sensing-derived LST datasets within the two study areas as evaluation (or comparison) datasets to assess the LST data prepared for this study [42]. Due to the limited availability of GHCN stations in these areas, we found only one station with temperature data for each study site. Figure 6 and Figure 7 illustrate the time series of field-observed GHCN data compared to Landsat, MODIS, and VIIRS LST data during study the period within the two study areas. Given the very limited number of GHCN stations (Figure 6c shows daily data) and the availability of MODIS (weekly), VIIRS (weekly), and Landsat (annual) LST data, we compared the 3 × 3 pixel average annual Landsat LST to these remote sensing datasets alongside weather station records. Our analysis revealed that MODIS and VIIRS weekly and monthly data align well in their temporal patterns (refer to Figure 6a,b for 2020, and Figure 6d,e for 2023) for Wuhan. Additionally, we found good agreement between these annual remote sensing-derived LST data and the GHCN data, with the GHCN air temperature serving as a reference for the annual temperature pattern in Wuhan (Figure 6f). We calculated the R² and root mean square error (RMSE) of annual LST derived from Landsat, MODIS, and VIIRS compared to GHCN temperature from 1986 to 2023. Temporally, Landsat exhibited a lower RMSE than both MODIS and VIIRS. Spatially, for the selected years, VIIRS demonstrated a better relationship with Landsat than MODIS, showcasing a higher R² value and a lower RMSE.

Given the very limited number of GHCN stations (Figure 7c shows daily data) and the availability of MODIS (weekly), VIIRS (weekly), and Landsat (annual) LST data, we compared the 3 × 3 pixel average annual Landsat LST to these remote sensing datasets alongside weather station records. Our analysis revealed that MODIS and VIIRS weekly and monthly data align well in their temporal patterns (see Figure 7a,b for 2020, and Figure 7d,e for 2023) for Brasilia. Additionally, we could not find any stations that have data in 2020 and 2023. We compared MODIS and VIIRS weekly and monthly data specifically for Brasília. Additionally, we analyzed annual remote sensing-derived LST datasets against GHCN air temperature data, which was used as a reference for annual temperature patterns in Brasília (Figure 7f). The blue circle indicates that only a few years of data were available from the 1990s. Furthermore, we calculated the R² and RMSE of annual LST derived from Landsat, MODIS, and VIIRS in relation to GHCN temperature records covering the period from 1986 to 2023 for Brasília. Due to the limited GHCN observations (only five years of records from the 1990s), we were unable to obtain comprehensive temporal results. Spatially, for the selected years, VIIRS exhibited a stronger correlation with Landsat than MODIS, showing a higher R² value and a lower RMSE.

A statistical regression analysis (the Theil-Sen estimator) returns positive values for both study areas indicating the spatial relationship of LST among Landsat, MODIS and VIIRS. Figure 8 shows the results in 2020 (top) and 2023 (bottom) for Wuhan, and Figure 9 shows the results for Brasília, depicting the relationship between Landsat and MODIS within the buffer zone in 2020 (top) and 2023 (bottom). Note that LST is measured at ground level and may differ from surface air temperature, which is measured at a height of 2 m. Generally, the LST is higher than air temperature. From Figure 8 and Figure 9, VIIRS has better relationship with Landsat LST than MODIS, and 2020 has better results than 2023 for both MODIS and VIIRS data.

3.4. Spatial Patterns and Temporal Variations of SUHI Intensities

We used a 5 km buffer with urban and non-urban interface to examine the impact of buffer width on the estimation of SUHI intensity in two study areas. The land cover proportions indicate that urban land increased from 32% to 45% in Wuhan and from 10% to 16% in Brasília within the 5-km buffer area between 1986 and 2023 (Figure 3 and Figure 4). The annual LST demonstrates a greater increase in urban areas compared to non-urban areas within the buffer zones in both study areas. However, the variance of LST is reduced. Additionally, the variance of LST differences in urban areas is larger than in non-urban areas, suggesting that the significant annual LST differences could be influenced by the variance of LST in urban regions.

The trends in SUHI intensity for both study areas are similar. Moreover, the temporal trends of SUHI intensities in urban lands exhibit consistent patterns regardless of urban size (Figure 10). However, average SUHI intensity is relatively higher in urban centers than in urban areas adjacent to non-urban land cover. The LST difference between the highest and lowest intensity urban areas can exceed 3 °C in some years. The temporal variation of annual SUHI intensity from 1986 to 2023 shows a positive trend for both study areas, with increases of 0.06–0.07 °C/year (p > 0.10) for Wuhan and 0.02–0.04 °C/year (p > 0.10) for Brasília, indicating an overall rising pattern in SUHI intensity for both locations (Figure 10a,b for Wuhan and Figure 10c,d for Brasília).

This intensity trends vary based on location. In low-density urban areas surrounded by water, wetlands, and forests, the SUHI intensity is relatively low. The intensity trend for areas that transitioned from low to high-intensity urbanization has a relatively small p-value of 0.07. Figure 10 compares annual maximum LST with annual mean LST for capturing SUHI trends. The results indicate that maximum land surface temperature is more We have moved the figure so that it now appears after its first citation in the text.effective We have moved the figure so that it now appears after its first citation in the text.than mean annual temperature for detecting SUHI intensification and identifying extreme heat conditions.

3.5. Spatial Distribution of SUHI Hotspot

The annual LST over persistent urban land and persistent urban hotspots for the two study areas reveal distinct differences (Figure 11). The maximum LST in these persistent urban areas indicates that the highest temperatures are concentrated in the central parts of Wuhan (Figure 6c) and Brasília (Figure 7c). Large LSTs associated with new growth extend in all directions from the urban edges in both the Wuhan and Brasília metropolitan areas. However, most hotspots are located in the town centers of Wuhan (Figure 11a) and Brasília (Figure 11b).

The multi-year average maximum LST suggests that temperatures in high, medium, and low probability areas are approximately 35.1 °C, 33.3 °C, and 30.8 °C in Wuhan, and 39.1 °C, 37.4 °C, and 35.9 °C in Brasília. The average mean LSTs are about 29.6 °C, 28.3 °C, and 26.5 °C in Wuhan, and 35.1 °C, 33.2 °C, and 32.3 °C in Brasília. Comparing these with land cover dynamics (Figure 3 and Figure 4), most temporal trends and associated coefficients of determination correlate with new urban growth in both areas. The trends of intensity in persistent built-up lands are similar for both study areas.

The trend of maximum LST over persistent urban land is statistically significant in Wuhan. While the trends of maximum LST in persistent urban land cover are also significant in Brasília, they are slightly weaker than those in Wuhan. In contrast, the trends of maximum LST in non-urban areas are significant in both study areas, but exhibit a lower slope compared to urban areas.

4. Discussion

4.1. Land Cover Change and LST Variations and Trends

Our analysis of land cover within 5-km buffer zones revealed substantial urban expansion in both Wuhan and Brasília over the past four decades. In Wuhan, urban pixels are frequently interspersed with wetlands, cropland, trees, and water bodies, with impervious surfaces now exceeding 60%. By contrast, Brasília’s urban pixels are mixed primarily with cropland, grass/shrub, and rangeland, with impervious cover below 40%. Most development occurred in the past three decades, a pattern consistent with prior studies documenting rapid peri-urbanization in China and Brazil [48,49,50].

When comparing the LSTs of these mixed pixels to those of surrounding dominant non-urban lands, noticeable differences emerge. In Wuhan, non-urban land is primarily composed of cropland, trees, water bodies, and wetlands, while new urban growth has mostly come from the conversion of cropland and wetlands. Consequently, the urban center has shifted continuously, moving from the northwest to the southeast within the buffer zones (Figure 2), with corresponding gradual increases in LSTs (Figure 4). Positive trends in LST are observed for all urban pixels and those with more than 20% impervious cover. The temporal trend of LST in persistent urban areas exceeds that of persistent non-urban areas. This upward trend may persist if more non-urban lands, particularly those with trees and wetlands, are transformed into urban spaces in the future.

In Brasília, cropland and grass/shrub (or rangeland) dominate the non-urban landscape (Figure 3), which reduces aerodynamic resistance. The urban land is rougher, allowing for more efficient heat dissipation from urban areas. Additionally, lower rates of evapotranspiration may elevate LST in cropland and grass/shrub land, further narrowing the LST difference between urban and non-urban areas (Figure 5). The increase in LST over persistent developments is relatively smaller compared to that observed in Wuhan. Temporal trends in LST differences between persistent urban and non-urban areas show that both are positive, but urban areas exhibit a more rapid increase. The growing vegetation cover within these buffer zones could enhance evaporation and cooling, thus reducing LST differences and leading to negative trends in SUHI intensity. The maximum LST is derived from seasonal LST peaks, with the magnitude representing the highest levels observed during clear conditions across different seasons for both study areas.

4.2. The Evaluation and Comparison of LST

This study also focuses on using available GHCN station observation and multi-datasets comparison for evaluating the Landsat derived LST data. Our results show that the Landsat LST annual product, which is derived from Landsat archive, has relatively high accuracy for estimating annual thermal conditions, and provides inputs for SUHI and trend analysis applications. Our results also demonstrate that the Landsat LST, with its high spatial resolution, can capture the detailed variation associated with Landsat derived land cover data in a timeseries. However, the accuracy of the Landsat LST data varied depending on the two study areas with time, availability of Landsat clear observations, and climate conditions. With limited GHCN station data availability within the 5 km buffer zones of two selected study areas (one station is available for each), the Landsat LST has similar temporal pattern with GHCN observations (Figure 6 and Figure 7). The comparison results show that in Wuhan, VIIRS LST demonstrated stronger temporal correlation (R² > 0.65 in both selected years) with Landsat LST than MODIS (R² is about 0.46). R² values in Brasília are lower compared to values for Wuhan, R² is about 0.57 for 2023 and 0.5 for 2021 with VIIRS, and about 0.45 with MODIS, but linear relationships were not significantly different between Wuhan and Brasília. The correlation coefficient was higher in the Wuhan area than in Brasília area likely because of the availability of clear Landsat observations. The availability of clear Landsat observations is a major factor affecting the accuracy of annual Landsat LST products, alongside other issues like modeling errors, QA band errors, and the quantity and quality of Landsat thermal data. Based on the comparison results (R² and RMSE), in Wuhan, Landsat exhibited higher R² values and lower RMSE than both MODIS and VIIRS over time. In both study areas, spatially, VIIRS demonstrated a stronger linear relationship with Landsat data than MODIS for the selected years, based on higher R² values and lower RMSE. This comparative analysis reveals that Landsat LSTs are more accurate in spatial estimates and possess significant potential for applications in SUHI studies. The long-term records and higher spatial resolution of Landsat LST enable more robust statistical analyses to be completed.

4.3. Land Cover and SUHI Intensity Trends

Using the global LST archive and existing USGS land cover information, this study assessed the spatial distribution and temporal trends of the SUHI intensity associated with land cover change. Substantial Landsat thermal data were processed to obtain the annual LST status for SUHI intensity analysis. The incorporation of land change information with time-series LST data revealed the differences between persistent urban and non-urban areas, with SUHI intensity being influenced by the physical and thermal characteristics of land cover in both urban and surrounding rural areas. The conversion of non-urban land to urban areas alters radiative forcing, convection, and surface heat fluxes by modifying surface physical and biophysical conditions. LST data in urban areas compared to dominant non-urban land cover in surrounding rural regions are crucial in determining the magnitude of SUHI intensity because the transition from non-urban to urban land can significantly affect the LST difference.

Several previous studies have indicated that vegetation conditions, water bodies, and wetlands can impact SUHI intensity [51,52,53,54]. Green spaces and impervious surfaces are the key urban characteristics associated with land surface temperature, while water bodies and wetlands within urban areas also influence the urban microclimate [55]. SUHI intensity varies across different climate zones and landform positions. On a global scale, the mean surface SUHI intensity can reach 0.85 °C during the day and 0.55 °C at night [3,19]. Our findings reveal that LSTs over cropland in Wuhan are more than 2–3 °C warmer than those over water bodies, while LSTs over grass/shrub land in Brasília are over 2 °C warmer than those over trees. This conclusion aligns with previous studies that used data from various sensors, although the magnitudes differ slightly [56,57,58].

Land cover components within a 5-km buffer zone and geographic location are major factors influencing SUHI intensity. Water bodies possess a higher heat capacity, and trees have greater aerodynamic roughness, allowing them to dissipate heat more effectively than croplands or grass/shrub land, which feature relatively less dense and shorter vegetation cover. This results in lower LSTs over water bodies and trees during the daytime. Additionally, the difference in heat capacity and aerodynamic resistance, combined with reduced evapotranspiration, contributes to warming effects in urban areas within the 5-km buffer. The SUHI intensities derived from the maximum LST difference have large magnitudes and temporal trends in both study areas, and most of these trends are statistically significant. However, the temporal trends derived from maximum LST have strong increasing trends and could be significantly impacted by new growth. The SUHI intensity derived from the maximum LST represents the maximum potential of the SUHI effect. The temporal trends of the maximum SUHI intensities are strong and positive for almost all urban land overs. The increase in the maximum SUHI intensity could have implications in human and ecosystem health under the trend of climate heat wave [4,18,56,57,58].

Urban growth rate plays a critical role, as rapid expansion often leads to increased impervious surfaces and reduced vegetation, amplifying heat retention. Population density further intensifies SUHI effects by driving energy consumption and anthropogenic heat emissions. Climate differences among regions also contribute significantly. Additionally, land cover characteristics, such as the proportion of green spaces, water bodies, and built-up areas, affect surface thermal properties and heat dissipation. By integrating these factors, our analysis provides a more comprehensive understanding of spatial and temporal variations in SUHI intensity and highlights the complexity of urban thermal environments.

4.4. SUHI Hotspot Probability in Persistent Urban Areas

Analysis of SUHI hotspot probabilities reveals that the distribution of hotspots in persistent urban areas, particularly those with high probabilities (greater than 50%), is concentrated in city centers characterized by intense impervious surfaces. These areas include older city centers, shopping districts, central business districts (CBDs), and industrial zones, where tall buildings, airports, and large warehouses are prevalent, and where green space is limited in both study areas. Additional significant hotspots are found in various locations across the study sites, such as the southeastern part of Wuhan, which is situated away from larger water bodies and surrounded by cropland. Most persistent hotspots are characterized by flat, wide-open landscapes with little green space. The SUHI hotspot probability analysis effectively captured both the magnitude and spatial distribution of these areas, indicating relatively high LST values during the daytime in both study regions.

The impact of SUHI on surface temperatures, surface water temperatures via runoff, and of human health and urban ecosystems can be substantial. Socioeconomic disparities exacerbate the unequal effects of heat, leading to increased energy costs, health issues, and mortality rates [6]. The social groups most affected by heat exposure are often the most vulnerable. As urbanization progresses, the implications of SUHI-induced warming and related scaling effects become increasingly critical for global warming and public health. Additionally, the identification of SUHI hotspots across various spatial scales underscores the need for urbanization strategies that account for different climatic contexts to effectively adapt to future global climate change.

4.5. Limitations and Future Directions

The goal of bridging urban and non-urban interfaces is to enhance our understanding of the interconnected patterns and processes associated with SUHI effects in relation to time-series LST and land cover dynamics. This is particularly relevant given advancements in remote sensing technologies, especially with Landsat, which offers a long historical record and medium-resolution data capable of capturing SUHI information, both spatially and temporally. Over the past several decades, urban remote sensing has progressed significantly, with sensors providing greater detail and algorithms becoming more adept at discerning patterns on a global scale.

Limitations of this study include the focus on only one Region of Interest (ROI) tile for each area. Data availability poses significant challenges for SUHI analysis, including limited clear observations, potential biases introduced by cloud cover and shadows, and variability driven by regional climate conditions. Furthermore, the scarcity of historical remote sensing data adds uncertainty to long-term SUHI trend estimation. Incomplete temporal coverage can obscure early stages of urbanization, potentially leading to underestimation of SUHI intensity or misinterpretation of its evolution over time, such as in Brasília. These limitations highlight the need for cautious interpretation of historical trends and underscore the importance of integrating complementary datasets or modeling approaches to improve temporal continuity and accuracy. Future research could explore multiple ROI tiles that represent a wider range of thermal conditions impacted by land cover changes, particularly in various suburban centers and their surrounding non-urban areas. The urban land cover classes in this study are defined quite broadly, which could be refined in future work to better understand how the resampling of input data affects specific SUHI patterns across different urban land cover categories (e.g., developed high intensity, developed medium intensity, developed low intensity, and developed open space). Without precise data on urban land cover transitions, the SUHI effects on some low-intensity urban areas may be underestimated [25].

Although the Landsat archive provides ample observations for time-series analyses to estimate annual means and long-term trends in LST, there are some gaps in the data, particularly in the earlier years across both study areas. Many pixels remain contaminated by cloud cover and shadows, which were not used in calculating the annual means. In certain years and specific locations, a considerable number of pixels exhibit a relatively low percentage of clear observations, which may skew the seasonal cycle and, consequently, the annual mean While this issue does not significantly bias the overall findings for Wuhan, it could potentially affect the annual mean LST in some regions of Brasília during the 1980s. Implementing additional procedures to fill in these cloud and shadow contaminations using clear observations could help reduce potential effects of these biases [46,59]. The heat-related health impacts stemming from heat wave events have emerged as a leading cause of climate change-related illnesses and fatalities worldwide [6]. Our findings underscore the importance of large-scale, systematic efforts to quantify the effects of SUHI in relation to land cover dynamics against the backdrop of climate change, utilizing time-series remote sensing data. The extensive geographic scope of these data enables the assessment of spatial patterns and temporal variations in impacts while identifying areas disproportionately affected by human-induced climate change. Furthermore, these findings support the scientific basis for developing more effective mitigation and adaptation strategies to minimize the public health impacts of climate change.

5. Conclusions

This study demonstrates how long-term Landsat thermal data can be integrated with dynamic land cover products to monitor and quantify SUHI intensities and hotspots. By examining two climatically and geographically distinct urban centers—Wuhan, China, and Brasília, Brazil—we provide several key insights:

• Urban expansion over the past four decades has significantly influenced land surface temperatures (LSTs). Wuhan exhibited stronger warming trends due to extensive conversion of cropland and wetlands, while Brasília showed more moderate increases linked to cropland and grass/shrub conversion.
• Landsat-derived LST products offer robust, spatially detailed records of SUHI dynamics, outperforming coarser MODIS and VIIRS datasets in capturing fine-scale temperature variations, despite limitations from cloud contamination and data gaps.
• SUHI intensity trends are more pronounced when assessed using maximum LST values, highlighting that urban growth amplifies extreme heat conditions more strongly than mean annual temperatures.
• Persistent SUHI hotspots were concentrated in dense urban cores with high impervious surface cover, underscoring the importance of green infrastructure and water bodies in mitigating urban heat stress.
• Land cover composition is a critical mediator of SUHI effects: forests and water bodies consistently provided cooling benefits, while cropland and grass/shrub contributed to elevated LSTs.

This research introduces several novel contributions to SUHI studies. We developed a comprehensive framework integrating multi-decadal Landsat-derived LST and land cover data for comparative analysis across diverse climatic and urban contexts. We implemented an automated modeling system for data preprocessing, LST calculation, SUHI mapping, intensity estimation, hotspot identification, and linkage analysis between land cover and human activities. Additionally, we introduced a dual-metric approach—annual mean LST and annual maximum LST—to assess SUHI trends and extreme heat events, a dimension rarely explored in previous studies. Finally, our comparative analysis of two cities in distinct hemispheres and socio-economic settings provides insights into global urban heat dynamics and their drivers.

Overall, this study underscores the importance of integrating land cover dynamics with medium-resolution, long-term Landsat thermal data to advance understanding of SUHI evolution in rapidly urbanizing regions. These findings offer a strong scientific foundation for developing urban heat mitigation strategies—such as targeted greening, climate-sensitive urban design, and heat-resilient infrastructure—particularly at the urban–rural interface. Future research could broaden the geographic scope, improve land cover classification accuracy with detailed change information, increase LST temporal frequency, and incorporate socio-economic and health datasets to comprehensively assess SUHI impacts on human well-being under a changing climate.