Global urban vegetation exhibits divergent thermal effects: From cooling to warming as aridity increases

Abstract

Urban vegetation, a key nature-based solution for mitigating heat stress, is critical as global warming, and urban heat islands amplify high temperatures in cities, affecting over half the global population. Yet, its potential warming effects remain unquantified globally, with mechanisms unclear. Using high-resolution satellite and climate data, we provide the first global assessment of vegetation’s temperature regulation across 761 megacities across 105 countries, uncovering a paradox: cooling weakens in arid environments; and in 22% of cities with <1000-millimeter annual precipitation, vegetation, particularly grasslands and croplands, causes net warming. This results from lower albedo and reduced heat storage outweighing limited evapotranspiration in arid regions. During extreme heat, trees fail to cool 25% of cities, while grasslands and croplands fail in 71 and 82%, respectively, due to reduced evapotranspiration under high vapor pressure deficits and impeded canopy conductance. Climate-adaptive greening and irrigation are critical, while high-albedo surfaces may better mitigate heat in water-scarce cities. Misguided greening risks are worsening urban warming.

City plants cool, yet with increasing aridity, they can heat urban areas.

INTRODUCTION

As global warming intensifies, extreme heat critically increases, particularly in urban areas where the urban heat island effect exacerbates these conditions (1, 2). The urgency of addressing this issue is underscored by the harmful environmental impacts of high temperatures in global cities, such as heightened energy consumption, urban ecosystem degradation, and disrupted precipitation patterns (2, 3). Amid these challenges, urban trees stand out as a nature-based solution to mitigate these impacts, offering shade to reduce direct sunlight and facilitating cooling through evapotranspiration (ET) (4, 5). Consequently, the United Nations advocates for protecting urban vegetation as a critical resource in reducing urban temperature and fostering sustainable urban environments. This directly contributes to achieving Sustainable Development Goal 11: Make cities inclusive, safe, resilient, and sustainable (4, 6).

Despite the recognized cooling benefits of urban trees, a growing body of evidence suggests that urban vegetation may unexpectedly cause higher surface temperatures than nonvegetated areas (i.e., warming effect) (7–10). Observations indicate a rise in daytime surface temperatures with increased urban tree coverage in some African and European cities. However, the global prevalence, seasonal dynamics, and dominant drivers of this warming effect remain unclear (7, 8). This counterintuitive finding suggests that the interaction between urban vegetation and temperature is more complex than simple cooling as previously understood. Overlooking the potential warming effect of increased urban tree cover could render heat mitigation efforts ineffective or even counterproductive.

Similar warming effect from vegetation is well-documented in natural ecosystems (11–13). Boreal forests, for instance, can have a net warming effect because the cooling from ET is outweighed by the radiation-forcing warming (12, 13). Significant differences between natural and urban areas in environmental conditions, surface properties, and vegetation composition (7, 14) preclude extrapolating these findings to cities. This critical knowledge gap underscores the urgent need to directly investigate urban vegetation’s biophysical effects. Understanding these climate-vegetation feedbacks is essential for securing urban sustainability, which we define here as the development of thermally resilient cities that mitigate extreme heat while responsibly managing vital resources like water and vegetation (15).

Before implementing urban vegetation as a heat mitigator, distinct characteristics of various vegetation types and associated biophysical effects must be carefully considered. Trees, for example, typically exhibit high ET rates due to large leaf surface areas, promoting cooling through increased moisture exchange with the atmosphere (7, 16). Trees also have low albedo, meaning they absorb more solar radiation and may contribute to localized warming (13, 17), rendering the net effect uncertain. In contrast, grasslands, compared to trees, generally display lower ET rates but higher albedo, thus reflecting more solar radiation (18). Croplands, increasingly common in expanding urban agriculture (19, 20), have their own distinctive climatic effects, shaped by species, phenology, and management practices. These differences mean that the net thermal impact of urban vegetation depends on vegetation type, local climate, and surrounding land cover—factors that must be explicitly considered in urban planning (21).

To optimize urban vegetation-driven heat mitigation and avoid unintended warming, we provide the first global, city-specific assessment of vegetation’s temperature regulation capability (TRC) across major climate zones. TRC is defined as the temperature difference between urban vegetated and built-up areas (i.e., ∆T = T_veg − T_built-up), where a negative value indicates cooling (22). We ask three key questions: (i) How do the direction and magnitude of TRC vary across climate zones and seasons? (ii) Does vegetation mitigate or exacerbate temperature increases during extreme heat? and (iii) What drives global variation in TRC? To address these questions, we analyze high-resolution datasets of land surface temperature (LST), near-surface air temperature (T_air), and land use and land cover (LULC) for 761 global megacities across 105 countries. We use both LST and T_air to ensure that our results are robust across different measures of thermal exposure. To enable a direct comparison among urban trees, grasslands, croplands, and built-up areas, we use a validated machine learning model that estimates the temperatures of purely vegetated and built-up surfaces, effectively accounting for heterogeneous land cover, topography, and mixed pixel effects (see Materials and Methods). This approach enables robust, globally comparable insights into the thermal effect of urban trees, grasslands, and croplands, providing critical evidence to guide sustainable, climate-resilient urban design.

RESULTS

Divergent temperature-regulating effect of urban vegetation: From cooling to warming

The TRC of urban vegetation, as indicated by the temperature difference between urban vegetated and built-up areas, exhibits distinct spatial patterns (Fig. 1). Across global cities, urban grasslands provided a cooling effect in 78% of cases, while urban trees did so in 98%, showing consistently lower surface temperatures than built-up areas (blue points in Fig. 1, A, B, E, and F; real examples in fig. S1A). However, in up to 22% of cities, predominantly in arid regions with annual precipitation below ~1000 mm (an empirical threshold), urban grassland (n = 164) and cropland (n = 145; fig. S2) exhibit higher surface temperatures than built-up areas (red points in Fig. 1, A to D), indicating a net warming effect (fig. S2). In 13 cities (2%) of arid regions, urban trees also exhibit a warming effect (Fig. 1, E to H). Among the vegetation types, urban trees exhibit significantly stronger TRC (∆T, −3.71° ± 1.73°C) compared to grassland (∆T, −1.44° ± 1.91°C) and cropland (∆T, −1.86° ± 2.04°C).

Fig. 1. Annual mean temperature difference (∆T) between areas covered 100% by urban vegetation and areas covered 100% by built-up surfaces.

(A) ∆T between urban grassland (UG) and built-up (BU) (i.e., T_UG − T_BU). (B) Distribution of ∆T across mean annual temperature (MAT) and precipitation (MAP) gradients. (C) ∆T between UG and BU from different temperature datasets, including Landsat LST, MODIS LST, and near-surface air temperatures (red circles, ∆T from MODIS LST versus that from Landsat LST; blue stars, ∆T from MODIS LST versus that from Tair). (D) A case for warmer UG than BU at Tabriz city, Iran [longitude (Lon) 46.334°E; and latitude (Lat), 38.064°N], including city red-green-blue (RGB) image, region of interest (ROI) RGB image, ROI Landcover, and ROI LST at February, June, and October. (E) ∆T between urban tree (UT) and BU (i.e., T_UT − T_BU). (F) Distribution of ∆T in (E) across temperature and precipitation gradients. (G) ∆T between UT and BU from different temperature datasets (red circles, ∆T from MODIS LST versus that from Landsat LST; blue stars, ∆T from MODIS LST versus that from Tair). (H) A case for warmer UT than BU at Benghazi city, Libya (Lon, 20.047°E; Lat, 32.076°N). (I) ∆T between urban vegetation and BU (i.e., T_Veg − T_BU) across different climate zones derived from Landsat LST. (J) T_Veg − T_BU across different climate zones derived from T_air. Error bar length refers to one standard error. Tropical, temperate, boreal, and arid regions are classified on the basis of Köppen-Geiger climate classification.

When considering different climatic zones (Fig. 1, I and J), all urban vegetation types exhibit stronger TRC in tropical cities (∆T_tree, −5.24° ± 1.90°C; ∆T_grass, −2.38° ± 2.17°C; and ∆T_crop, −3.18° ± 2.34°C), followed by temperate cities (∆T_tree, −4.18° ± 1.54°C; ∆T_grass, −1.75° ± 1.51°C; and ∆T_crop, −2.24° ± 1.82°C), boreal cities (∆T_tree, −3.69° ± 1.20°C; ∆T_grass, −2.04° ± 1.31°C; and ∆T_crop, −2.22° ± 1.44°C), and arid cities (∆T_tree, −1.93° ± 1.38°C; ∆T_grass, 0.60° ± 1.90°C; and ∆T_crop, 0.15° ± 1.86°C). Given the TRC similarity between urban grassland and cropland (Fig. 1, I and J, and fig. S2), we present cropland results in the supplementary material (fig. S2). Meanwhile, although TRCs derived from LST (tree, −3.71°C; Fig. 1I) are stronger than those derived from air temperatures (tree, −0.60°C; Fig. 1J), their spatial patterns are consistent (fig. S3). The correlations of TRCs derived from three datasets (Landsat LST, MODIS LST, and near-surface air temperature) are high [correlation coefficient (r > 0.68; Fig. 1, C and G]. Thus, we only display Landsat LST results in the subsequent analysis, as it offers the highest resolution to better capture the temperature heterogeneity of urban surfaces, enabling a more precise TRC estimation.

An XGBoost model was built and validated for each month to quantify the seasonal variability of TRC under varying background temperatures (Fig. 2). Cities were divided into two categories: those with net vegetating cooling or warming, and analyses were performed separately for trees (Fig. 2A) and grassland (Fig. 2C). Both vegetation types show similar seasonal patterns. Cities with net warming experienced it consistently throughout most months (see fig. S4 to S6 for the detailed cases of all vegetation types). Conversely, cities with net cooling experience it year-round, with the strongest effect in summer (Fig. 2, A and B). We then categorized global cities into four groups based on Köppen-Geiger climate classification. In extratropical cities, urban vegetation exhibits the strongest TRC in summer and the weakest in winter (Fig. 2, B and D). This can be attributed to the fact that summer coincides with plants’ peak growing season, when greenness and ET cooling are highest (fig. S7). Conversely, tropical cities show the strongest TRC during late summer and early autumn, corresponding to the wet season, with sufficient rainfall to support ET cooling (fig. S8), and weaker TRC during dry season (December to February).

Fig. 2. Monthly changes in mean temperature differences (∆T) between urban vegetation, including urban trees (T_UT) and urban grassland (T_UG), and built-up (T_BU) areas.

(A) ∆T between UT and BU for cities where UT exhibits a net warming effect (red line) and a cooling effect (blue line). (B) ∆T between UT and BU across different climatic zones. (C) ∆T between UG and BU for cities where UG exhibits a net warming effect (red line) and a cooling effect (blue line). (D) ∆T between UT and BU across different climatic zones. For consistency between the northern and southern hemispheres, months are ordered by season. Four seasons are defined differently in the northern and southern hemispheres. For the northern hemisphere, spring is March to May, summer is June to August, autumn is September to November, and winter is December to February. For the Southern Hemisphere, spring is September to November, summer is December to February, autumn is March to May, and winter is June to July.

Urban vegetation can either mitigate or aggravate extreme heat intensity

To examine how urban vegetation regulates local extreme heat, we compared the temperature difference between normal and extreme hot summers (T_Extreme − T_Normal) for different land cover types. Here, extreme hot summers are defined as months with LST exceeding the 85th percentile of historical summer values. Sensitivity analyses using different thresholds (80th and 90th percentiles; fig. S9) and temporal windows (late spring-early autumn; fig. S10) confirmed the robustness of our conclusions. Our results reveal that vegetation types differed markedly in their response. Specifically, during extreme hot summers, urban trees generally mitigate temperature increases compared to built-up areas, reducing the temperature increase magnitude (−0.49° ± 0.10°C) in ~75% of global cities. The T_Extreme − T_Normal for urban trees is negatively correlated with tree coverage (Fig. 3, A and B, and fig. S11A). In contrast, during extreme hot summer, urban grasslands and croplands tend to exacerbate temperature rises (0.48° ± 0.11°C and 1.08° ± 0.15°C) compared to built-up areas in around 71 and 82% of cities, respectively. The T_Extreme − T_Normal is positively correlated with grass and crop coverage (Fig. 3, C and D, and fig. S11). This discrepancy occurred because grassland/cropland surface temperatures increased more than built-up areas under extreme heat, opposite to trees (fig. S12).

Fig. 3. Patterns and drivers of urban vegetation regulating local extreme heat.

(A) The magnitude of temperature change between normal (T_Normal) and extreme (T_Extreme) hot summer months across different climatic zones and land cover types. Relationship between the magnitude of temperature change and the fraction of urban vegetation, including (B) urban tree, (C) urban grassland, and (D) urban cropland. These relationships are derived from calibrated XGBoost models (see Materials and Methods for more details). (E) ET rate changes between normal and extreme hot summer months across urban vegetation types. ET data are from gap-filled MODIS ET products (MOD16A2GF). Error bars refer to one SD. *P < 0.05 and ***P < 0.001, t test.

To explore the reasons why urban trees and grassland respond differently during extreme heat events, we analyzed changes in their transpiration cooling between extreme hot and normal summers (Fig. 3C). The net radiation changes were not analyzed because of the homogeneity of solar radiation within cities, resulting in similar net radiation across different vegetation types. Our findings revealed that, during extreme hot summers, the atmospheric drought [indicated by vapor pressure deficit (VPD)] significantly increases (+12.7%; fig. S13A). Under these conditions, grasslands and croplands significantly reduce their total canopy conductance (−15.5 and −15.2%, respectively; fig. S13B), which is greater than the VPD increase, leading to reduced ET rate (Fig. 3C). Whereas trees reduce their canopy conductance (−10.5%) less than the VPD increase (fig. S13B), maintaining slight increase in ET cooling (Fig. 3C).

Biophysical factors driving the variation in temperature regulation of urban vegetation

Using observational data and a process-based attribution model, namely, the two-resistance mechanistic (TRM) model, we attributed temperature difference between urban vegetated and built-up areas to radiative forcing (R_n), anthropogenic heat emission (Q), aerodynamic resistance (r_a), surface resistance (r_s), and heat storage (G). The TRM performed well, with moderate-to-strong agreement between simulated and observed TRC (r = 0.72, bias = 0.26°C; Fig. 4, A and C) and between remote-sensing-derived and flux tower-derived biophysical factors [coefficient of determination (R²) = 0.43 to 0.73; fig. S14]. Net radiation and ground heat flux contribute to warming (Fig. 4, B and D). Vegetation’s lower albedo, indicated by negative albedo differences between urban vegetated and built-up areas (fig. S15A, i and ii), leads to greater solar absorption and net radiation-induced warming (fig. S15A). Vegetation also stored less heat than built-up areas, releasing more energy to the surface and causing a warming effect (fig. S15E, i and ii).

Fig. 4. Attribution of annual temperature difference (∆T) between urban vegetation and built-up areas in cities from four climatic zones and all cities together (Global).

(A) Comparison between observed and predicted ∆T of urban trees and built-up areas. (B) Attribution of ∆T between urban trees and built-up areas. (C) Comparison between observed and predicted ∆T of urban grasslands and built-up areas. (D) Attribution of ∆T between urban grasslands and built-up areas. Variables in (B) and (D) are radiative forcing (R_n), anthropogenic heat emission (Q), aerodynamic resistance (r_a), surface resistance (r_s), and heat storage (G).

In contrast, aerodynamic resistance and surface resistance contribute to cooling effects (Fig. 4B). Vegetation increases surface roughness (i.e., reduces aerodynamic resistance) compared to built-up areas, thereby enhancing heat convection and leading to cooling. Surface resistance facilitates cooling through vegetation transpiration, which absorbs energy to convert latent heat (fig. S15D, i and ii). Vegetation also mitigates anthropogenic heat emissions to facilitate cooling (fig. S15B, i and ii). Among these three factors, surface resistance emerges as the predominant cooling element, indicated by its larger negative ∆T (Fig. 4, B and D), which is also supported by the strongest correlation between latent heat flux and TRC (fig. S15D). This underscores the crucial role of transpiration cooling.

Across climate zones, the cooling magnitude of r_s decreases in the order: tropical, temperate, boreal, and arid cities. Analysis by vegetation types indicates that trees exhibit a stronger cooling effect than grasslands due to their lower surface and aerodynamic resistance (r_s and r_a; Fig. 4, B and D). In arid zones, net warming occurs when transpiration cooling is exceeded by albedo-induced radiative warming and reduced heat storage (Fig. 4, B and D). This explains why some arid cities showed warmer vegetation than built-up areas (Fig. 1).

Even within arid regions, TRC outcomes varied (from net cooling to warming) under similar climates, suggesting a role for nonclimatic drivers like human management (e.g., irrigation). To preliminarily investigate this hypothesis in the absence of global irrigation data, we used national economic development level as a proxy for a city’s capacity to invest in vegetation management. Our analysis revealed a clear correlative pattern: TRC was consistently stronger in cities within developed economies (e.g., Phoenix, USA; fig. S1B; net cooling) compared to those in developing economies (e.g., Benghazi, Libya; fig. S1E; net warming). This pattern held across different vegetation types and climatic zones (fig. S16).

DISCUSSION

We used high-quality, high-resolution data to evaluate the TRC of urban vegetation in 761 megacities across 105 countries, providing the first global, urban-specific assessment of its spatial and temporal patterns. Our findings reveal that, in up to 22 and 2% of these cities, urban grassland/cropland and tree, respectively, exhibit a net warming effect (Fig. 1), predominantly in arid cities worldwide. Given their already high background temperatures (23, 24), the diminished cooling effect in these cities underscores the need for alternative heat-mitigation approaches, such as light-colored roofs or cool valleys. Additionally, the significant costs of expanding urban vegetation (25, 26) necessitate an adaptive greening strategy, selecting species well suited to local climates that provide sustained cooling, on a global scale (27).

Previous studies have reported warming effects from vegetation in boreal and arid natural ecosystems (17, 28, 29). Our research demonstrates that urban vegetation contributes to net warming exclusively in arid cities (Figs. 1 and 2). The difference in thermal impacts of vegetation between natural and urban environments in boreal regions is because of their different effects on snow cover. In natural regions, vegetation can promote snow melting, resulting in lower surface albedo and more absorbed energy (warming effect). Compared to natural regions, boreal cities inherently experience less snow cover due to their warmer climates, and urban vegetation generates a negligible effect on the melting of urban snow cover (30, 31).

We further elucidate the biophysical drivers governing the TRC of urban vegetation across climates. Consistent with local-scale studies, we confirm that TRC is driven by the balance of net radiation, latent heat, and sensible heat (16, 32, 33). Our global analysis advances this understanding by identifying surface resistance, a function of ET, as the dominant process explaining spatiotemporal TRC patterns. This is clearly demonstrated by the pronounced cooling in tropical cities, where water is abundant, versus the weak cooling or net warming in arid cities (Fig. 1), where water scarcity severely limits ET (23, 34). The seasonality of TRC further reinforces the primacy of water availability. Arid-warming and tropical cities both exhibit relatively flat seasonal TRC trends (red and dark green lines in Fig. 2), which can be attributed to perennial ET conditions. In arid cities, soil moisture and ET are perennially low, creating a persistent imbalance where warming forcings (e.g., albedo-induced net radiation) consistently outweigh the limited cooling capacity (fig. S17). Conversely, in tropical cities, ET is perennially high, sustaining a stable and strong cooling effect year-round.

Urban trees exhibit a stronger TRC than grasslands during both normal and extreme heat events (Figs. 1 and 3). This enhanced cooling stems from two key mechanisms: (i) higher transpiration rates sustained by deeper roots and a higher leaf area index (leading to lower surface resistance, r_s); and (ii) greater canopy roughness that reduces aerodynamic resistance (r_a; fig. S18), thereby enhancing turbulent heat exchange (18). During extreme heat, this physiological advantage is critical: Trees mitigate temperature increases in 75% of global cities, while grasslands and croplands exacerbate them in up to 82% of cities (Fig. 3). Trees’ deeper roots provide access to deeper soil moisture, allowing them to maintain transpiration cooling even under high VPD (18). In contrast, the shallow roots of grasses and crops make them vulnerable to water shortage (18, 35), leading to reduced canopy conductance and diminished transpiration during heat extremes (Fig. 3C), a finding supported by recent observations (36). Our results demonstrate that vegetation’s cooling benefits are highly dependent on vegetation type and local context, underscoring the critical importance of strategic species selection. Although other strategies may be more efficient for cooling in arid zones, urban vegetation remains essential for its co-benefits (e.g., biodiversity, stormwater management, and well-being) (26, 37). A comprehensive resilience strategy must, therefore, integrate climate-appropriate species with other heat mitigation measures.

While our study provides a robust, macroscale assessment of urban vegetation effects, several avenues exist for future refinement. First, our 100-m-resolution analysis captures neighborhood-scale aggregate effects but necessarily integrates over critical microclimatic variations caused by urban geometry (e.g., building shading and street canyons) and fine-scale greenspace configuration (e.g., patch size and connectivity) (38). Consequently, our findings on surface radiative temperature cannot directly translate to pedestrian-level thermal comfort (Fig. 1J) and do not negate the proven efficacy of well-placed shade trees. Second, by classifying vegetation into broad categories (trees and grassland), we inherently average over significant intraclass variability in biophysical properties, such as species-specific traits (e.g., albedo and ET rates of deciduous versus evergreen trees) and management practices (e.g., irrigation). This means that our results represent the macroscale performance of generalized classes rather than optimized, species-specific outcomes. Last, our models are based on land-cover fractions and do not incorporate three-dimensional urban or vegetation structure, which modulates shading and airflow (39). Notably, the use of national economic development as a proxy for vegetation management capacity is also a simplification. Future work that integrates microclimate measurements, species-specific traits, local management, and configuration data will be valuable for translating our macroscale insights into precise, context-specific design guidelines, further enhancing the practical application of this research.

In conclusion, urban vegetation does not always cool cities; in arid areas, it can cause net warming as transpiration cooling is outweighed by the warming effects of albedo and heat storage changes. During extreme heat, urban grasslands and croplands can exacerbate temperature increases. These findings underscore that greening initiatives are not one-size-fits-all solutions. To achieve sustainable urban cooling, cities must adopt a climate-adaptive approach. In humid regions, vegetation should be leveraged for its potent cooling and ecological benefits. In contrast, water-limited arid cities require a more strategic framework: Greening plans must prioritize water-efficient practices and integrate vegetation with complementary strategies like high-albedo materials and shading structures to effectively build urban resilience without straining limited water resources.

MATERIALS AND METHODS

Research design

We used a comprehensive four-step framework to investigate urban vegetation’s TRC across global cities (fig. S19). (i) Data aggregation: We aggregated 10-m LULC data to 100 m to calculate fractions of urban tree, grassland, cropland, and built-up areas (fig. S19A). (ii) Machine learning modeling: We trained a validated XGBoost model to predict temperature using land cover fractions and topography as inputs. This allowed us to disentangle the effects of different vegetation types from other urban surfaces (fig. S19B). (iii) TRC calculation: We defined TRC as the temperature difference (∆T = T_veg − T_built-up) between surfaces 100% covered by a given vegetation type and those 100% covered by built-up areas. A negative ∆T indicates a cooling effect. These model-based estimates were consistent with direct satellite observations (fig. S19C). And (iv) TRC attribution: We used the TRM model to attribute the calculated ∆T to five biophysical processes, explaining the global spatial variation in TRC (fig. S19D).

Data processing

Land surface/air temperature data

To calculate TRC of urban vegetation, we used daytime MODIS LST, near-surface air temperature, and Landsat LST data. We chose to focus on daytime data, excluding nighttime data, for two primary reasons: (i) the highest spatial resolution data (Landsat LST) is only available during the daytime; and (ii) daytime data are captured close to midday, roughly representing the highest daily temperature, which is when most city dwellers require cooling. The MODIS LST data were obtained from the gap-filled datasets with a resolution of 1 km in daily time steps (40). The 1-km near-surface air temperature dataset was generated through a seamless integration of MODIS LST, DEM dataset, and air temperature observations from weather stations using the spatially varying coefficient model with sign preservation algorithm (41). This air temperature dataset has been validated using observations from 103,156 weather stations across the world, showing high accuracy (R² = 0.92, bias =1.86°C) (41).

The Landsat 8–Level 2–Collection 2 LST data at a resolution of 30 m were derived from the ST_B10 band, with clouds and shadows removed using the QA band. All temperature data were resampled to a 100-m resolution for consistency. The 1-km data were resampled to 100-m resolution using bilinear interpolation (42), and the 30-m data were resampled to 100 m by averaging all values within each 100-m grid. TRCs from all three products are highly correlated (Fig. 1C). For seasonal analysis, we used monthly composites from 2014 to 2023 to minimize data gaps (Landsat 8–Level 2–Collection 2 product started in March 2013). The period from 2014 to 2023 did not account for potential climate change impacts over a longer period. Cities with more than 50% missing pixels (accounting for less than 1% of selected cities) were excluded from our analysis.

Global high-resolution land cover data

We used the ESA WorldCover 2021 product (10-m resolution) to calculate land cover fractions (22). This dataset can capture urban green cover details (e.g., small park and street trees) (fig. S20) and has the highest accuracy among comparable global products (43). To mitigate the impact of urban land cover change during the study period, we used annual 30-m land cover change data (44) to remove any pixels that transitioned between major classes (e.g., vegetation to built-up) between 2014 and 2023.

Fractions for each 100 m grid were calculated by dividing the number of pixels of a specific class by the total pixels. Dominant urban vegetation types include trees, grassland, and cropland, which on average make up 28.9, 13.8, and 7.7% of the urban land cover, respectively (fig. S21).

Topographic data

Elevation data were obtained from the digital elevation model Shuttle Radar Topography Mission (SRTM), with a spatial resolution of 30 m (45).

Global urban areas (GUB)

Our analysis focused on 761 global cities across 105 countries with an area larger than 100 km², which ensured that each city contained enough pixels for pattern and driver analysis. Urban areas were identified and extracted using urban boundary shapefiles from a publicly available dataset for the year 2018 (46), the most recent year available. The use of a static 2018 urban boundary is robust for analyzing stable urban land covers, as we excluded pixels with land cover change. Any potential residual bias would be minimal and likely result in a slight overestimation of built-up area temperatures in rapidly expanding cities.

Köppen climate map

Cities were classified using the Köppen-Geiger climate classification system, the global standard for defining climate types based on temperature and precipitation patterns. Each city was assigned to one of four primary climate groups: tropical climates (i), arid climates (ii), temperate climates (iii), and boreal climates (iv) (47). Notably, the “arid” group encompasses both arid (BWh and BWk) and semiarid (BSh and BSk) climates.

Economic data

Income data were obtained from the World Bank dataset. We reclassified cities into developed (>$12,375) and developing (<$12,375) categories, on the basis of the 2018 gross national income per capita, following the same threshold as prior studies (48). These data were used as a provisional proxy for human management capacity, acknowledging that it is a simplification that misses local policy variation.

Urban flux data

Urban flux data were used as benchmarks to validate the reliability of our remote-sensing based energy balance framework. Urban flux data are scarce due to the requirement for homogeneous underlying surfaces, which are typically heterogeneous in urban environments. Thus, only 20 urban flux sites across the world are openly available (49). These datasets offer gap-filled, harmonized measurements at 30-min intervals. We selected high-quality observations (quality control flag = 0 or 1) for analysis. We used two-step processing to match flux measurements with satellite observation: (i) select the observation at 10:30 a.m. to match the Landsat and MODIS Terra cross-over time; and (ii) aggregate the daily value to monthly average to align with monthly TRC evaluation.

Calibrating machine learning model to calculate temperature differences between urban vegetation and built-up areas

Model description and settings

We used the XGBoost model (v3.0.2 in Python) to estimate the temperature of purely vegetated and built-up surfaces from mixed urban pixels (50). XGBoost was chosen to capture complex interactions that linear unmixing models oversimplify. Linear models assume additive contributions, failing to account for nonlinear processes such as (i) advection, where heat transfer between adjacent land cover types introduces nonlinear thermal signals (51); (ii) shadowing effects from variations in solar angle and urban geometry (51); and (iii) ET variability, driven by nonlinear interactions between canopy conductance and microclimate (52). XGBoost is also robust to collinearity between predictors (e.g., negative correlations between built-up areas and vegetation/topography; fig. S22). The model also addresses spatial autocorrelation, performing comparably to spatial statistical models (53).

We built a separate model for each city and month, with temperature as the response variable and land cover fractions (i.e., tree, grass, crop, built-up, and water) and topography as predictors. We used the trained models to predict temperature under 100% coverage scenarios for each vegetation type and for built-up areas. TRC was calculated as ∆T = T_veg − T_built-up.

XGBoost model validation

We validated the model with three methods: (i) 10-fold cross validation; (ii) synthetic data validation; and (iii) observed pure-pixel validation. We first conducted 10-fold cross-validation, with 90% of all pixels for training, and 10% validation per fold. The modeled and observed LST show high consistency (mean coefficient of determination: R² = 0.69; fig. S23). Cities with R² < 0.5, indicating low explanatory power, were excluded from analysis (~1%, n = 11).

Second, to evaluate XGBoost’s capability in recovering theoretical endmember LST values from mixed pixels, we designed a controlled synthetic data experiment with known ground truth. The procedure consisted of three phases:

1) Synthetic dataset generation: Reference endmember LST values ($T_{\text{pure},i}$) were established for four land cover classes—urban tree, grass, cropland, and built-up—on the basis of characteristic thermal ranges observed in urban environments: [28.0, 30.0, 32.0, 40.0] °C, respectively (40). A total of 10,000 synthetic mixed pixels were generated with randomized fractional compositions (f_i) using Dirichlet distributions to ensure ∑f_i = 1. The composite LST for each mixed pixel ($T_{\text{mixed}}$) was calculated linearly (54): $T_{\text{mixed}}={\sum }_{i=1}^4(f_i\times T_{\text{pure},i})+\mathrm{ϵ}$, where ϵ ∼ N(0,1.0) represents Gaussian noise (σ = 1.0°C) simulating observational uncertainties.

2) Run the XGBoost model to estimate $T_{\text{pure},i}$: In the XGBoost model, we designated the $T_{\text{mixed}}$ as the output and land cover fractions [f_tree, f_grass, f_crop, f_built] as inputs. The model was trained on 80% of the dataset (8000 samples) with 20% used for testing to prevent overfitting. Subsequently, the model predicted theoretical LSTs for pure land cover scenarios (e.g., f_i: [1, 0, 0, 0] for 100% tree cover). To assess robustness, we varied the ground truth temperatures of pure endmembers within a ±5°C range and repeated the experiment 500 times, generating 500 pairs of predicted and true pure-pixel LSTs for evaluation.

3) Validation protocol: The XGBoost model demonstrated high performance in estimating pure-pixel LST across all land cover types (R² = 0.98, bias = 0.32°C, root mean square error = 0.49; fig. S24). This enabled robust TRC calculations in our analysis.

Third, we compared pure-pixel’s LST between modeling and observation from Landsat LST. We identified pure pixels (>98% cover) for trees, grassland, cropland, and built-up areas across 388 cities and calculated their mean LST as observed endmember values. XGBoost model outperformed the standard linear unmixing model (55) in estimating pure-pixel LST and TRC (mean R² = 0.82 versus 0.70, bias = 0.33°C versus 0.60°C; figs. S25 and S26). To further assess spatial autocorrelation, we calculated Moran’s Index for temperature residuals with values between −0.08 and 0.13 (fig. S27), which is far from the values of 1 and −1, indicating that our models are not affected by strong spatial autocorrelation.

Vegetation’s role on heat extremes regulation

We quantified how urban vegetation regulates extreme heat by analyzing the temperature rise (ΔT_Extreme = T_Extreme − T_Normal) of each land cover type from normal to extreme summer conditions. Extreme summers were defined as months where LST exceeded the 85th percentile of summer (three consecutive hottest months for each year) values during 2014 to 2023 (fig. S28A). For this, we calculated the T_Extreme as the mean LST for all extreme summer months and T_Normal as the mean LST for all nonextreme summer months, thus deriving ΔT_Extreme. We used an XGBoost model to investigate this regulatory capacity, using ΔT_Extreme as the response variable and land cover fractions and topography as predictors. Using the temperature rise over built-up areas as a baseline, we interpreted a smaller (or larger) rise over vegetated areas as mitigation (or exacerbation) of extreme heat. The robustness of our conclusions was confirmed through sensitivity analyses using different temperature thresholds (80th and 90th percentiles; fig. S9) and expanded temporal windows (late spring to early autumn; fig. S10). Extreme hot summer months were defined using LST to ensure direct physical relevance to surface-driven heat fluxes and methodological consistency with our primary LST-based TRC analysis. This approach was validated against ERA5-Land air temperature data, showing a high correlation (r = 0.97) and >94% agreement in event identification (fig. S29).

TRM model for $\mathbf{∆}$T attribution

We used the TRM to attribute the temperature difference between vegetated areas and built-up areas (56). The model requires inputs including the sensible heat flux, latent heat flux, LST, net radiation, and air temperature for each land cover type; aerodynamic and surface resistances are then derived from these flux variables (56). Previously, the TRM model has been applied primarily in natural systems with relatively uniform surfaces (56). To extend the TRM model to urban environments with heterogeneous surfaces, we assumed that energy balance components in mixed pixels are a linear combination of each land cover endmember (54); however, we acknowledge that shadows may introduce nonlinear effects (57). Then, we used an XGBoost model to unmix the energy flux components from mixed pixels into values for pure land cover endmembers, which were input into the TRM to attribute TRC variability.

TRM model description

In TRM model, temperature differences (positive or negative) caused by different land cover types can be quantitatively attributed to variations in different biophysical factors, including radiative forcing, aerodynamic resistance, surface resistance, and heat storage (58). The TRM model was derived from the energy balance theory, in which surface temperature variations across urban land types are determined by the balance of five energy components

$$R_{\mathrm{n}}+Q=\mathit{LE}+H+G$$ (1)

where $R_{\mathrm{n}}$, $Q$, $\mathit{LE}$, H, and $G$ refer to net radiation (watts per square meter), anthropogenic heat emission (watts per square meter), latent heat (watts per square meter), sensible heat (watts per square meter), and ground heat storage (watts per square meter), respectively.

Net radiation at the urban surface was calculated as

$$R_{\mathrm{n}}={\mathrm{\epsilon }R}_{\text{LW}\downarrow }+(1-\mathrm{\alpha })R_{\text{SW}\downarrow }-\mathrm{\epsilon }\mathrm{\sigma }{T}_{\mathrm{s}}^4$$ (2)

where $\mathrm{\epsilon }\ast R_{\text{LW}\downarrow }$ is absorbed longwave radiation by land surface; $\mathrm{\epsilon }$ and $R_{\text{LW}\downarrow }$ (watts per square meter) are emissivity (equals to longwave absorptance) and downward longwave radiation, respectively; $(1-\mathrm{\alpha })\ast R_{\text{SW}\downarrow }$ refers to the absorbed shortwave radiation, in which $\mathrm{\alpha }$ and $R_{\text{SW}\downarrow }$ (watts per square meter) refer to surface albedo and downward shortwave radiation, respectively. The albedo ($\mathrm{\alpha }$) is calculated from Landsat reflectance data based on the empirical equation (Eq. 3) (59); $\mathrm{\epsilon }\mathrm{\sigma }{T}_{\mathrm{s}}^4$ is reemitted longwave radiation, in which $\mathrm{\sigma }$ and $T_{\mathrm{s}}$ are Stefan-Boltzmann constant (5.6704 × 10⁻⁸ Wm⁻² K⁻⁴) and LST (kelvin), respectively. In our analysis, $R_{\text{SW}\downarrow }$ and $R_{\text{LW}\downarrow }$ were derived from ERA5_Land product, and $\mathrm{\epsilon }$ is calculated based on Eq. 4 (60)

$$\mathrm{\alpha }=\frac{0.356b_1+0.130b_3+0.373b_4+0.085b_5+0.072b_7-0.0018}{0.356+0.130+0.373+0.085+0.072}$$ (3)

where b represents Landsat bands 1,3,4,5, and 7

$$\mathrm{\epsilon }=0.2122{\mathrm{\epsilon }}_{29}+0.3859{\mathrm{\epsilon }}_{31}+0.4029{\mathrm{\epsilon }}_{32}$$ (4)

where ${\mathrm{\epsilon }}_{29}$, ${\mathrm{\epsilon }}_{31}$, and ${\mathrm{\epsilon }}_{32}$ are the values of bands 29, 31, and 32 of MYD21A1D product.

The anthropogenic heat emissions data were obtained from the AH4GUC product (61), which is generated on the basis of energy consumption statistics, population distribution data, satellite imagery, urban growth models, and climate models. This dataset has improved accuracy compared to other anthropogenic heat datasets (61).

Ground heat storage (G) was estimated on the basis of its relationship with net radiation and the normalized difference vegetation index (NDVI) (Eq. 5). Furthermore, we examined alternative empirical methods for estimating ground heat flux. For more details, refer to the “TRM Model validation and uncertainty” section. Ground heat storage was calculated as follows

$$G=R_{\mathrm{n}}\times 0.583\text{exp}(-2.13\text{NDVI})$$ (5)

Theoretically, latent and sensible heat can be parameterized using Eqs. 6 and 7. However, due to the lack of necessary variables, such as r_a and r_s, and the large uncertainties in the estimation of these variables, we obtained the LE directly from the MOD16A2GF product, which is an 8-day gap-filling composite dataset produced at a resolution of 500-m pixels and has been validated by ground-based flux observations with high accuracy (62). In MOD16A2GF, LE is calculated on the basis of the Penman-Monteith equation, which combines inputs from daily meteorological reanalysis data and MODIS remote sensing data products (e.g., vegetation attribute dynamics, albedo, and land cover). Here, we assumed that the energy balance is closed, as such, H can be calculated from the energy balance equation: $H=R_{\mathrm{n}}+Q-\mathit{LE}-G$

$$\mathit{LE}=\frac{\mathrm{\rho }{C}_{\mathrm{p}}}{\mathrm{\gamma }}\frac{e_{\mathrm{s}}^{\ast }-e_{\mathrm{a}}}{r_{\mathrm{a}}+r_{\mathrm{s}}}$$ (6)

$$H=\frac{\mathrm{\rho }{C}_{\mathrm{p}}}{r_{\mathrm{a}}}(T_{\mathrm{s}}-T_{\mathrm{a}})$$ (7)

where $\mathrm{\rho }$ and $C_{\mathrm{p}}$ are air density and air specific heat at standard pressure, respectively; $\mathrm{\gamma }$ is psychrometer constant, $e_{\mathrm{s}}^{\ast }$ and $e_{\mathrm{a}}$ are saturated and actual vapor pressure (pascal), respectively, (they are derived from ERA5-Land data); $r_{\mathrm{a}}$ is aerodynamic resistance; and $r_{\mathrm{s}}$ is surface resistance.

By combining Eqs. 1 and 7, we can calculate the aerodynamic resistance (Eq. 8)

$$r_{\mathrm{a}}=\frac{\mathrm{\rho }{C}_{\mathrm{p}}}{H}(T_{\mathrm{s}}-T_{\mathrm{a}})=\frac{\mathrm{\rho }{C}_{\mathrm{p}}}{R_{\mathrm{n}}-G-\mathit{LE}}(T_{\mathrm{s}}-T_{\mathrm{a}})$$ (8)

By combining Eqs. 1, 6, and 8, we can calculate the surface resistance (Eq. 9)

$$r_{\mathrm{s}}=\frac{\mathrm{\rho }{C}_{\mathrm{p}}}{\mathrm{\gamma }}\frac{e_{\mathrm{s}}^{\ast }-e_{\mathrm{a}}}{\mathit{LE}}-r_{\mathrm{a}}$$ (9)

By linearizing the surface energy balance equation using a first-order Taylor expansion on the outward longwave radiation and saturated vapor pressure terms, the temperature difference can be attributed to four biophysical factors, including net radiation, heat conversion efficiency, evaporation, and heat storage, respectively (Eq. 10)

$$\begin{array}{c}∆T={\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }{R}_{\mathrm{n}}}\right)}^{\text{TRM}}\hfill \\ \times ∆R_{\mathrm{n}}+{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }Q}\right)}^{\text{TRM}}\times ∆Q+{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }{r}_{\mathrm{a}}}\right)}^{\text{TRM}}\hfill \\ \times ∆r_{\mathrm{a}}+{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }\mathrm{\beta }}\right)}^{\text{TRM}}\times ∆\mathrm{\beta }+{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }G}\right)}^{\text{TRM}}\times ∆G\hfill \end{array}$$ (10)

where $\mathrm{\Delta }T$, $\mathrm{\Delta }{R}_{\mathrm{n}}$, $\mathrm{\Delta }{r}_{\mathrm{a}}$, $\mathrm{\Delta }{r}_{\mathrm{s}}$, and $\mathrm{\Delta }G$ are the contributions of net radiation, heat conversion, evaporation, heat storage, and anthropogenic heat to ∆T, respectively. The partial differential derivations in each term can be expressed as Eqs. 11 to 15

$${\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }{R}_n}\right)}^{\text{TRM}}={\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }Q}\right)}^{\text{TRM}}=-{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }G}\right)}^{\text{TRM}}=\frac{{\mathrm{\lambda }}_0}{1+f_{\text{TRM}}}$$ (11)

$$\begin{array}{c}{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }{r}_{\mathrm{a}}}\right)}^{\text{TRM}}=\frac{{\mathrm{\lambda }}_0\mathrm{\rho }{L}_{\mathrm{v}}[q_{\mathrm{a}}^{\ast }\left(T_{\mathrm{a}}\right)-q_{\mathrm{a}}]}{{(r_{\mathrm{a}}+r_{\mathrm{s}})}^2}\frac{1}{1+f_{\text{TRM}}}\hfill \\ -[R_{\mathrm{n}}+Q-G-\frac{\mathrm{\rho }{L}_{\mathrm{v}}[q_{\mathrm{a}}^{\ast }\left(T_{\mathrm{a}}\right)-q_{\mathrm{a}}]}{r_{\mathrm{a}}+r_{\mathrm{s}}}]\frac{{\mathrm{\lambda }}_0}{{(1+f_{\text{TRM}})}^2}\frac{\mathrm{\partial }{f}_{\text{TRM}}}{\mathrm{\partial }{r}_{\mathrm{a}}}\hfill \end{array}$$ (12)

$$\begin{array}{c}{\left(\frac{\mathrm{\partial }{T}_{\mathrm{s}}}{\mathrm{\partial }{r}_{\mathrm{a}}}\right)}^{\text{TRM}}=\frac{{\mathrm{\lambda }}_0\mathrm{\rho }{L}_{\mathrm{v}}[q_{\mathrm{a}}^{\ast }\left(T_{\mathrm{a}}\right)-q_{\mathrm{a}}]}{{(r_{\mathrm{a}}+r_{\mathrm{s}})}^2}\frac{1}{1+f_{\text{TRM}}}\hfill \\ -[R_{\mathrm{n}}+Q-G-\frac{\mathrm{\rho }{L}_{\mathrm{v}}[q_{\mathrm{a}}^{\ast }\left(T_{\mathrm{a}}\right)-q_{\mathrm{a}}]}{r_{\mathrm{a}}+r_{\mathrm{s}}}]\frac{{\mathrm{\lambda }}_0}{{(1+f_{\text{TRM}})}^2}\frac{\mathrm{\partial }{f}_{\text{TRM}}}{\mathrm{\partial }{r}_{\mathrm{s}}}\hfill \end{array}$$ (13)

$$\frac{\mathrm{\partial }{f}_{\text{TRM}}}{\mathrm{\partial }{r}_{\mathrm{a}}}=-\frac{r_0}{r_{\mathrm{a}}^2}[1+\frac{\mathrm{\delta }}{\mathrm{\gamma }}{\left(\frac{r_{\mathrm{a}}}{r_{\mathrm{a}}+r_{\mathrm{s}}}\right)}^2]$$ (14)

$$\frac{\mathrm{\partial }{f}_{\text{TRM}}}{\mathrm{\partial }{r}_{\mathrm{s}}}=-\frac{\mathrm{\delta }}{\mathrm{\gamma }}\frac{r_0}{{(r_{\mathrm{a}}+r_{\mathrm{s}})}^2}$$ (15)

where ${\mathrm{\lambda }}_0=1/\left(4\mathrm{\epsilon }\mathrm{\sigma }{T_{\mathrm{a}}}^3\right)$ is local climate sensitivity, $\mathrm{\delta }=\frac{\mathrm{\partial }{e}^{\ast }}{\mathrm{\partial }T}\mid T_{\mathrm{a}}$, $\mathrm{\gamma }=\frac{c_{\mathrm{p}}P}{0.622L_v}$ and $r_0=\mathrm{\rho }{c}_{\mathrm{p}}{\mathrm{\lambda }}_0$. $f_{\text{TRM}}=\frac{r_0}{r_{\mathrm{a}}}[1+\frac{\mathrm{\delta }}{\mathrm{\gamma }}\left(\frac{r_{\mathrm{a}}}{r_{\mathrm{a}}+r_{\mathrm{s}}}\right)]$ is energy redistribution factor, $e^{\ast }$ is saturation vapor pressure, $P$ is air pressure, $L_{\mathrm{v}}$ is latent heat of vaporization, $q_{\mathrm{a}}$ is atmosphere specific humidity, $q_{\mathrm{a}}^{\ast }$ is saturated specific humidity at temperature of $T_{\mathrm{a}}$, $c_{\mathrm{p}}$ is specific heat at constant pressure, ${\mathrm{\epsilon }}_{\mathrm{s}}$ is surface emissivity, $\mathrm{\rho }$ is air density, and $r_{\mathrm{a}}$ and $r_{\mathrm{s}}$ are aerodynamic resistance and surface resistance, respectively.

TRM model validation and uncertainty

To quantify uncertainties in the application of the TRM model in urban areas, we performed two evaluations: (i) assessing the effects of spatial resolution and mixed-pixel issues and (ii) comparing remote-sensing-derived biophysical factors with those derived from ground-based urban flux data.

The effect of spatial resolution and mixed pixels

We selected 30-m ET data from the OpenET (Ensemble version) (63) to quantify the impact of spatial resolution on our model for two reasons: (i) ET (or latent heat) flux is a critical input of the model; and (ii) it is the only energy-related variable available at high spatial resolution, serving as benchmark data. We selected all megacities with available 30-m OpenET data (n = 56, fig. S30A) for analysis. We aggregated the annual ET data (2020) from 30 to 90, 510, and 990 m (fig. S30, B to E). Corresponding land cover fractions were derived from 10-m ESA LULC data. For each resolution, we developed an XGBoost model (with default parameters) using land cover fractions as inputs and ET as the output. The XGBoost model was trained on 80% of samples, with 20% used for testing to prevent overfitting. The trained model then predicted theoretical ET for pure land cover endmembers. Comparing these predictions to the 30 m benchmark revealed that estimation bias increased with coarser resolution, reaching ~6% at 990 m (fig. S30, F and G). We acknowledge this uncertainty but consider it unlikely to significantly affect our attribution analysis.

Comparison between remote-sensing- and ground-based biophysical factors

We compared biophysical parameters derived from remote sensing with those measured at 20 urban flux sites, including net radiation, latent heat flux, sensible heat flux, air temperature, surface temperature, surface-to-air temperature difference, aerodynamic resistance, and surface resistance. Flux data provided direct measurements of net radiation, latent heat, sensible heat, and air temperature. Remote-sensing-derived net radiation was calculated using Eq. 2, latent heat was obtained from gap-filled MODIS data, and sensible heat was derived using Eq. 7. Aerodynamic resistance and surface resistance were calculated using Eqs. 8 and 9, respectively. Surface temperature (T_s) from flux sites was derived from longwave thermal radiation using Eq. 16, while remote-sensing T_s was obtained from Landsat LST data

$$T_{\mathrm{s}}={\left\{\frac{1}{\mathrm{\sigma }\mathrm{\epsilon }}[R_{\text{LW}\uparrow }-(1-\mathrm{\epsilon })R_{\text{LW}\downarrow }]\right\}}^{1/4}$$ (16)

Given the high land cover heterogeneity around urban flux sites, we generated representative remote-sensing values for each site. This was done by calculating land cover percentages (tree, grassland, cropland, and built-up areas) within a 100 m–by–100 m grid and using them as weights to compute an average of endmember flux values. Linear regression revealed moderate to strong agreement (R² = 0.43 to 0.92) between methods, confirming the reliability of our remote sensing approach (fig. S14).

The effect of heat storage estimations

We conducted sensitivity analyses comparing three methods (64–66) for calculating ground heat flux (G) in urban tree (G_UT), urban grassland (G_UG), and built-up areas (G_BU): (i) G = R_n × 0.583 × exp(−2.13 × NDVI), (ii) G = R_n × (0.32 to 0.21 × NDVI), and (iii) G = R_n × (LST − 273.15) × (0.0038 + 0.0074 × α) × (1 to 0.98 × NDVI⁴), where R_n, NDVI, LST, and α represent net radiation, NDVI, LST, and albedo, respectively. The heat storage from these methods showed high agreement (R² = 0.69 to 0.95; fig. S31). Crucially, the TRC attributions for urban trees and grasslands remained consistent regardless of the G estimation method used (fig. S32).

Uncertainty propagation in the TRM

A sensitivity analysis was performed to quantify the propagation of uncertainty in the TRM model and to assess the influence of potential errors in its input parameters on the estimated TRC (i.e., ΔT). We focused on the four primary energy balance components: R_n, Q, G, and LE. Following an established protocol (33), the baseline value of each flux was independently perturbed by ±10% for both vegetated and built-up areas, while all other parameters were held constant. The resultant change in the modeled ΔT was computed for each perturbation, revealing that the model output was most sensitive to R_n and LE (fig. S33). Although these uncertainties influence the absolute magnitude of the TRC estimates, the attribution patterns remain robust. Specifically, increased R_n consistently drives warming, while increased LE drives cooling, underscoring the reliability of our conclusions (Fig. 4 and fig. S33).

Supplementary Materials

This PDF file includes:

Figs. S1 to S33

References