Introduction

Building sustainable cities is a central objective of the United Nations’ Sustainable Development Goals (SDGs)1. A significant challenge in achieving this goal is the urban heat island (UHI) effect, where temperatures in urban areas are significantly higher than those in neighboring rural regions2, threatening the well-being of over half of the urbanites worldwide3. From 1983 to 2016, global urban exposure to extreme heat increased by nearly 200%4, and by 2050, 1.6 billion people living in over 970 cities will be regularly exposed to extreme high temperatures5. Meanwhile, since the 1980s, the diurnal temperature range (DTR) has also risen, driven by factors such as aerosols and cloudiness, and is anticipated to persist under unmitigated climate change6,7. DTR reflects the variability in the temperature difference between day and night, and sudden changes in short-term temperature (higher DTR) can increase the risk of illnesses8. UHI and augmented DTR have severe implications for human health, including an increased risk of mortality7,9. Tackling urban heat challenges, particularly those associated with UHI and DTR, has become an urgent priority in achieving sustainable urban development7,10.

Nature-based Solutions, including urban green and blue infrastructure (UGBI)11,12, have been recognized as sustainable and cost-effective strategies in mitigating urban heat13,14,15,16. Urban blue infrastructure refers to networks of water bodies such as wetlands, rivers, and ponds, which enhance evaporative flux and provide notable cooling effects11,17. For example, urban wetlands can reduce local temperature by approximately 5 °C18. Urban green infrastructure includes vegetative systems such as green walls, parks, and border trees13, which cool cities through evapotranspiration, shading, and modifying surface energy balance19. These green spaces have been estimated to provide substantial urban cooling capacity, reducing warm season temperatures by 2–4 °C globally20. However, disparities in the quality and quantity of UGBI between and within cities lead to inequalities in cooling benefits20,21,22,23. Additionally, the limited availability of land in high-density urban areas constrains opportunities for expanding conventional greening and blue infrastructure24.

While ground-level land is limited, the dense concentration of buildings in cities offers vast rooftop areas, estimated to account for 20–25% of the total urban surface area25. These rooftops offer an opportunity to implement green roofs on existing buildings without the need for additional land. This approach holds significant potential for cooling urban environments and protecting city populations from extreme heat26. Globally, several green-roof initiatives have been implemented. For instance, Basel in Switzerland has implemented a series of roof greening programs since the 1990s, and ~40% of its flat roofs were greened (Table S1). In North America, Toronto is the first city to adopt a green-roof policy that requires new buildings above a certain size to be topped with plants27, and the size of green roof required ranges from 20 to 60% of the available roof space. San Francisco has also mandated that new buildings include 15–30% green roofs, solar installations, or a combination of both (Table S1).

Quantifying the cooling effects of green roofs is critical for informing urban heat mitigation policy decisions. To examine the cooling effects of green roofs, previous studies have employed various methods28, including field experiments29,30,31 and modeling approaches such as ENVI-met32 and Weather Research and Forecasting models33,34. Field experiments were typically used to assess the cooling effects of existing green roofs at the building scale and to compare the cooling performance of different vegetation configurations29,30,31, while modeling approaches were generally applied to simulate the cooling potential of various green roof scenarios at the city scale32,33,34. These studies, despite variation in methods, consistently indicated that green roofs provide significant cooling effects/potentials, with temperature reductions of up to 3 °C28. However, these approaches are difficult to scale to national or global levels due to their labor-intensive nature or high data and parameter requirements. Therefore, the cooling potential of roof greening at a global scale remains unclear.

To address this knowledge gap, here we present the first global-scale assessment of the cooling potential of urban rooftop greening. Our assessment leverages the recently published ultra-high-resolution global building footprints (vector boundaries derived from submeter satellite imagery), generated using state-of-the-art remote sensing and deep learning-based image recognition techniques35,36,37. To simulate the cooling potential of roof greening, we relied on the well-documented linear relationship between Land Surface Temperature (LST) and urban vegetation cover (VC)34,38, incorporating the average summer LST and VC from 2016 to 2020 to develop city-specific models. This approach allowed us to predict the cooling potential of roof greening by treating urban rooftops as additional vegetated areas. We finally applied the model to the increased VC under three roof greening scenarios based on existing roof greening practices (see “Methods” and Table S1): low (20% of each roof area covered by vegetation), medium (40% of each roof area covered), and high (60% of each roof area covered).

Specifically, we focused on the following questions: (1) What is the magnitude and global pattern of cooling potential of rooftop greening under different greening scenarios? (2) What is the size and pattern of the population that will benefit from the cooling effects of rooftop greening? (3) How might green roofs influence the urban diurnal temperature range (DTR)? Our study provides a simple yet effective framework for estimating the large-scale cooling potential of green roofs by combining high-resolution global rooftop footprints with observed relationships between VC and LST. Using this framework, we identified substantial cooling potential for numerous global cities, along with an additional benefit of reduced DTR. This suggests that implementing rooftop greening, with meticulous consideration of engineering aspects, could serve as an effective nature-based climate solution for urban areas. As a forward-looking evaluation, our work could pave the way for the future large-scale implementation of green roofs (Fig. 1).

Fig. 1: Method framework for estimating the cooling potential of green roofs.
Fig. 1: Method framework for estimating the cooling potential of green roofs.
Full size image

a examples of green roof practices worldwide. Photograph by CHUTTERSNAP (Singapore), TonyTheTiger (Chicago, US), and Hao Bai (Beijing, CN). b illustration of the relationship between land surface temperature (LST) and vegetation cover (VC). c remote-sensing datasets used for modeling. Empirical models were developed based on the relationship between LST (day and night separately) and VC, incorporating building height, building height variation, building density, proportion of materials, and elevation as covariates. d scenarios of roof greening and corresponding cooling potential. Rooftop greening scenarios (low, medium, and high) were set based on existing practices, which generally have 60% as the maximum (“Methods”). Scenario graphics were created using Icograms (https://icograms.com/).

Results

Cooling potential of global urban roof greening

Our results reveal a substantial cooling potential of roof greening at the global scale. The average cooling potentials under low, medium, and high extent of greening scenarios were 0.57 ± 0.47 (mean ± SD), 1.12 ± 0.92, and 1.58 ± 1.31 °C, during the day, and 0.14 ± 0.14, 0.27 ± 0.27, and 0.39 ± 0.38 °C at night, respectively. Across three scenarios, daytime cooling potential consistently exceeded nighttime cooling potential. Spatially, cities with stronger cooling potential were primarily located near the equator (Fig. 2a–c). Cities with stronger nighttime cooling potential were predominantly concentrated in Central America, South America, and Southeast Asia (Fig. 2d–f). In terms of mitigating surface urban heat island intensity (SUHII, calculated as the difference between urban LST and neighboring rural LST) through rooftop greening, it could reduce SUHII by 16.1–44.9% during the day and 6.4–18.2% during the night on average (see Fig. S1 for spatial distribution of SUHII reduction) for cities with significant SUHI effect (SUHII ≥ 1 °C). The uncertainty of cooling potential (half of the 95% confidence interval relative to the estimated cooling potential) was low: for daytime, the uncertainty was 22 ± 17% (mean ± SD), with 57.7% cities having an uncertainty below 20%; for nighttime, the uncertainty was 29 ± 19%, with 40.2% cities having a uncertainty below 20% (Fig. 3).

Fig. 2: Spatial distribution of the cooling potential of urban rooftop greening.
Fig. 2: Spatial distribution of the cooling potential of urban rooftop greening.
Full size image

spatial distribution of daytime cooling potential under low (a), medium (b) and high (c) scenarios (n = 3913 cities). spatial distribution of nighttime cooling potential under low (d), medium (e) and high (f) scenario (n = 1687 cities). The inset bar chart in a-f displays the number of cities corresponding to the five levels of cooling potential.

Fig. 3: Spatial distribution of the cooling potential uncertainty of urban roof greening.
Fig. 3: Spatial distribution of the cooling potential uncertainty of urban roof greening.
Full size image

a cooling potential uncertainty for daytime (n = 3913). b cooling potential uncertainty for nighttime (n = 1687). The uncertainty represents half of the 95% confidence interval, expressed as a percentage of the estimated cooling potential. For example, for a cooling potential of 1 °C with uncertainty of 20%, the 95% confidence interval ranges from 0.8 °C to 1.2 °C.

To better illustrate the pattern of the cooling potential, we categorized cities into groups according to their background daytime LST and nighttime LST (2016–2020 summer average, aligned with sampled LST for modeling), VC (2016–2020 average), and roof cover in urban area (Fig. 4). We found higher daytime cooling potential for cities with higher background LST under all three greening scenarios (Fig. 4a), suggesting that green roofs could provide higher cooling potential for cities experiencing higher heat during the day. However, the trend was not observed during the night, as there was little difference in the cooling potential corresponding to the background temperatures of the first three groups (Fig. 4d). We also found higher cooling potential in both daytime and nighttime for cities with higher background urban VC (Fig. 4b, e) and higher urban roof cover (Fig. 4c, f), but the cooling potential plateaued under high background urban VC (around 60%). Moreover, cooling potential during nighttime increased more obviously with roof cover (Fig. 4f) than with VC (Fig. 4e).

Fig. 4: Average cooling potential by background LST, urban vegetation cover and urban roof cover.
Fig. 4: Average cooling potential by background LST, urban vegetation cover and urban roof cover.
Full size image

Daytime cooling potential grouped by urban daytime LST (a), urban vegetation cover (b) and urban roof cover (c). Nighttime cooling potential grouped by urban nighttime LST (d), urban vegetation cover (e), and urban roof cover (f). The number of cities is annotated for groups with fewer than 10 cities.

Cooling benefits extend to urbanites

Additionally, we evaluated the benefits of urban roof greening to urbanites, calculated as the product of populations living in urban areas39 and cooling potential, which we referred to as “Cooling benefit” (unit: °C * 105 people). Most cities exhibited modest cooling benefits of 0–0.25 °C * 105 people, accounting for 66.9–78.3% (2617–3065) of all the cities during the daytime and 78.4–86.1% (1323–1452) at night (inset bars in Fig. 5a–f). Nevertheless, the proportion of cities with high cooling benefit (>2 °C * 105 people) increased from low to high greening scenarios: it increased from 5.8% (228 cities) to 12.2% (478 cities) for daytime but from 3.4% (58 cities) to 7.2% (122 cities) for nighttime.

Fig. 5: Cooling benefit of urban roof greening.
Fig. 5: Cooling benefit of urban roof greening.
Full size image

spatial distribution of daytime cooling benefit under low (a), medium (b) and high (c) scenarios (n = 3913 cities). spatial distribution of nighttime cooling benefit under low (d), medium (e), and high (f) scenario (n = 1687 cities). cooling potential (g) and cooling benefit (h) across continents. The inset bar chart in a–f displays the number of cities corresponding to the five levels of cooling benefit. In (g, h), daytime cooling benefit and potential are represented by solid bars, while nighttime cooling benefit and potential represented by hollow bars. The solid line indicates the standard deviation of cooling benefits and cooling potential, showing the variability among cities.

We also found that, due to the uneven distribution of population across continents, the spatial pattern of cooling benefit for urbanites differed from that of cooling potential. While Oceania and South America exhibit the greatest average cooling potential (0.85–2.42 °C during the day and 0.18–0.61 °C during the night, Fig. 5g), their average benefit for urbanites was not the highest (0.17–1.48 °C * 105 people during the day and 0.16–0.57 °C * 105 people during the night, Fig. 5h). Continents characterized by dense urbanites, namely Africa, Asia and Europe, had higher cooling benefit (0.34–2.33 °C * 105 people during the day and 0.17–1.14 °C * 105 people during the night, Fig. 5h) though the cooling potential was lower than that in other continents (0.38–1.21 °C during the day and 0.07–0.33 °C during the night, Fig. 5g).

Influence of roof greening on urban diurnal temperature range

Finally, we analyzed the effects of roof greening on diurnal temperature range (DTR). This analysis was performed on 765 cities with high-confidence cooling potential estimates for both daytime and nighttime [model R2 > 0.7, RMSE < 2 K, VC is a significant variable (p < 0.01), and maximum VIF among predictor variables <5, see section “Model construction” in “Methods”]. Interestingly, we found consistent reductions in DTR at the global scale. For the low (Fig. 6a), medium (Fig. 6b), and high (Fig. 6c) roof greening scenarios, the DTR changes (mean ± SD) were estimated to be −0.39 ± 0.47, −0.77 ± 0.92, and −1.10 ± 1.33 °C, respectively. Only 16, 26, and 36 cities show increased DTR with a magnitude of >0.1 °C under low, medium, and high greening scenarios, respectively. Such reduction was mainly due to an asymmetry in LST reduction during nighttime and daytime, meaning that the cooling magnitude was overall larger in daytime than in nighttime at the global scale, as illustrated by the distance between the red dashed line (roof greening LST) and the black solid line (current LST) in each panel in Fig. 6d–i.

Fig. 6: Spatial pattern of DTR (Diurnal Temperature Range) change and average DTR across different scenarios.
Fig. 6: Spatial pattern of DTR (Diurnal Temperature Range) change and average DTR across different scenarios.
Full size image

spatial pattern of DTR change under low (a), medium (b) and high (c) roof greening scenarios (n = 765). The inset bar chart in (a–c) displays the number of cities (left) and population (right) corresponding to the five levels of DTR change. d–i Daytime and nighttime LST under different scenarios. The black solid line and red dashed line in (d–i) mark represent the mean value of LST under current and roof greening conditions for all the cities, respectively.

We also found that more cities would experience notable reductions in DTR as roof greening increases. Under the low roof greening scenario, 67 cities would experience a decrease in DTR with a magnitude of <−1.0 °C. This number rises to 206 cities under the medium scenario and 319 cities under the high scenario. Consistent with this trend, the populations exposed to a DTR decrease of < −1.0 °C also grow substantially, from about 3 million under the low scenario to 36 million and 52 million under the medium and high scenarios, respectively. Spatially, cities in Oceania, South America, and North America exhibited higher DTR changes, ranging from −0.72 ± 0.21 (mean ± SD) to −2.10 ± 0.59 °C, −0.65 ± 0.59 to −1.82 ± 1.64 °C, and −0.56 ± 0.56 to −1.58 ± 1.65 °C, respectively (Fig. S2). Asian cities showed the lowest DTR changes from −0.20 ± 0.25 to −0.56 ± 0.71 °C (Fig. S2).

Discussion

Using global building footprints and empirical modeling of LST–vegetation relationships, we provide the first global assessment of the cooling potential of green roofs. Our results indicate that rooftop greening consistently delivers substantial LST reductions during both daytime and nighttime, helping to mitigate urban heat risk. In addition, we observed that daytime cooling generally exceeds nighttime cooling. This day-night cooling asymmetry could reduce diurnal temperature range (DTR), thereby mitigating the health stress associated with abrupt diurnal temperature changes.

Our results revealed strong spatial variability in green roofs’ cooling potential. Cities located near the equator generally exhibited higher cooling potential, particularly during the day (Fig. 2), consistent with our finding that cities with higher background daytime LST tend to achieve greater cooling potential (Fig. 4a). This spatial pattern is largely driven by differences in evapotranspiration12. In humid, low-latitude cities, sufficient moisture enables strong evaporative cooling, whereas arid regions experience moisture constraints that limit vegetation cooling performance12,40.

Additionally, our results show that daytime cooling potential increased with higher background urban VC and a larger proportion of rooftop area available for greening, which reflects the inherent vegetative capacity of a city and the proportion of roofs that can be converted into green infrastructure, respectively. In contrast, nighttime cooling potential is primarily driven by the extent of rooftop cover, while its associations with background LST and background VC were relatively weak (Fig. 4d–f). These findings suggest that nighttime cooling is less influenced by vegetation and is instead likely driven by impervious surface heat storage. Impervious surfaces, such as buildings and roads, exhibit higher heat capacity, lower thermal conductivity, and reduced albedo compared to natural surfaces or vegetation, making them more prone to absorbing heat. At night, however, stored heat from the day is released, involving more intricate heat transfer processes and being less directly influenced by vegetation41,42.

These mechanisms can be further illustrated using example cities (Fig. S3). For instance, Nairobi (Kenya, Fig. S3a), located near the equator, exhibited a much stronger daytime cooling potential of 3.5 °C, compared to ~0.5 °C in higher-latitude cities such as Xi’an (China, Fig. S3b) and Baltimore (United States, Fig. S3c) under the high greening scenario. This is consistent with Nairobi’s higher background VC (71.8% vs. 58.1% in Xi’an) and greater rooftop area available for greening (29.2% vs. 18.0% in Xi’an). Although Baltimore has a comparable background VC to Nairobi (71.5%), its urban roof cover is substantially lower (16.0%), resulting in weaker cooling potential. In contrast, the differences in nighttime cooling potential among the three cities were much smaller than those observed during the daytime, leading to the DTR reduction patterns similar to daytime cooling potentials.

We note that roof greening is not the only approach being carried out to mitigate the impacts of climate change and to achieve the UN SDG of sustainable cities1. In fact, various roof modification approaches have been proposed, such as white roofs (high-albedo surfaces) and solar roofs43. As a practical example, Singapore has announced plans to apply cool paints to all public housing buildings by 2030 (Table S1). We focused on roof greening rather than other roof modification approaches because green roofs provide not only cooling potential, but also other benefits such as alleviation of urban air pollution44, reducing urban noise45,46, sequestering carbon content in the air47,48,49, enhancing urban biodiversity by providing complex vegetation structures50,51.

Despite substantial cooling potential and other environmental benefits of green roofs, their implementation is far more complex than in idealized scenarios. First, not all roofs are suitable for retrofitting—for example, historical or cultural buildings, functional roofs, and those with steep slopes may not be viable candidates52. The roof’s load-bearing capacity must also be considered52. Additionally, constructing green roofs requires water for irrigation, making the careful selection of appropriate substrates and plant configurations crucial in regions with different climate conditions53,54, especially in arid areas. In such areas, green roofs could be replaced or integrated with other sustainable roof solutions55,56. For instance, artificial turf with suitable NDVI/albedo values would be an alternative option, as NDVI/albedo also performed well in many cities when modeling the cooling potential (Figs. S4 and S5). The integration of green roofs with solar PV panels or rooftop agriculture could offer multiple co-benefits, such as providing electricity or food55,56,57, while simultaneously maintaining their cooling potential. It’s partly out of these reasons that the high rooftop greening scenario in this study was set at 60%, which also corresponds to the highest ratios in current green-roof policies (Table S1).

Implementing green roofs also calls for financial subsidies and obligations by law to support their adoption to mitigate urban heat risk58. For example, the city of Basel in Switzerland has implemented a series of green roof policies since the 1990s, requiring or incentivizing new and retrofitted buildings to include green roofs. As a result, approximately 100 new green roofs were added annually, with about 23% of Basel’s flat roofs covered by vegetation in 2006; this proportion is now estimated to have reached ~40% (Table S1). Green roofs also offer an equitable opportunity to expand urban green infrastructure and advance urban heat mitigation22,23. Previous studies have shown that cities in the Global South possess only about 70% of the green spaces’ cooling capacity of cities in the Global North20, while heat extremes pose particularly high risks for populations in low-latitude regions59,60. Our results indicate that green roofs in these areas have greater cooling potential, suggesting that promoting low-cost green roof solutions could help enhance heat-risk mitigation equity in the Global South. Compared with conventional green space development, green roofs offer greater flexibility in dense urban environments, helping expand access to cooling benefits and advance environmental justice61. At the same time, urban green roofs are often concentrated in wealthier neighborhoods, leaving low-income communities with limited access to cooling benefits62,63. Future policies should therefore integrate considerations of both heat risk and socioeconomic inequality to ensure more equitable urban climate adaptation62,63.

In this study, we used the MODIS LST product at a 1 km resolution. While this resolution may limit the ability to represent fine-scale intra-urban temperature variations64, MODIS provides consistent daytime and nighttime observations globally, which are essential for estimating diurnal cooling potentials and temperature changes. Although the Landsat LST product provides higher spatial detail (100 m), its use is constrained by cloud contamination65 and the lack of temporally paired daytime and nighttime retrievals at the global scale66. Additionally, our modeling was based on satellite-derived LST, which reflects surface thermal conditions and cannot be directly interpreted as air temperature or human thermal comfort4,64. Finally, we assumed uniform rooftop greening coverage (20–60%) across each city. Although these scenarios were informed by existing green roof policies in some cities (Table S1), their applicability elsewhere may be constrained by practical limits on vegetation and rooftop suitability.

Future advancements in 3D reality technology and deep learning could play a crucial role in addressing the challenge of selecting suitable rooftops for retrofitting with vegetation43,67,68. First, future studies should improve mechanistic models by incorporating high-resolution LST products, robust LST–air temperature conversions, and tighter coupling with urban climate models to enhance our understanding of how green roofs cool cities. Second, advancements in rooftop recognition, such as improved 3D reality data, deep learning methods, and consideration of socioeconomic constraints, are necessary to accurately identify suitable roofs for greening, thereby enabling optimized spatial planning and design53,69. With careful design, thoughtful planning, and appropriate government regulations48, green roofs could serve as a practical solution to address urban heat and other environmental challenges12.

Conclusion

In summary, by leveraging global building footprint data, we developed a simple yet robust framework to assess the cooling potential of rooftop greening at a global scale. We found that green roofs provide substantial land-surface temperature (LST) reductions across cities worldwide, with average daytime cooling potentials of 0.57, 1.12, and 1.58 °C, and nighttime cooling potentials of 0.14, 0.27, and 0.39 °C, under the low, medium, and high greening scenarios, respectively. Due to the generally stronger daytime cooling than nighttime cooling, rooftop greening also contributes to reductions in diurnal temperature range (DTR), with decreases of 0.39, 0.77, and 1.10 °C under the low, medium, and high greening scenarios, respectively. Our results underscore the importance of rooftop greening in mitigating intensifying urban heat under global climate change. With future advancements in 3D reality technologies, mechanistic modeling, and supportive rooftop greening policies, green roofs hold strong potential to be widely implemented as an effective strategy to reduce urban heat risk and deliver additional co-benefits simultaneously.

Methods

Overview

In this study, we established multiple linear models for each city based on the relationship between VC and both daytime and nighttime LST38, incorporating building height, building height variance, building density, building materials, and elevation as covariates to improve the accuracy of parameter estimation. Using high-resolution rooftop footprint data, we defined three greening scenarios (VC increased by 20, 40, and 60%) for urban rooftops, enabling the estimation of daytime and nighttime cooling potentials. We considered scenarios where urban rooftops are uniformly greened, which could significantly influence the urban thermal environment70,71. We also calculated changes in the surface urban heat island intensity (SUHII) and diurnal temperature range (DTR) under different roof greening scenarios. In the following sections, we first introduce the datasets employed in this study. We then describe the procedures for model construction, the simulation of cooling potentials and uncertainties, and the subsequent analytical methods (see Fig. S6 for Methodological flowchart).

Datasets

Land surface temperature

Temperature was represented by LST, which measures the urban skin temperature (pavements and buildings)42. Although it remains debatable whether LST is closely connected with air temperature72, LST is the only wall-to-wall temperature measurement over global cities, and has been widely used in recent urban heat studies73,74. The Moderate Resolution Imaging Spectroradiometer (MODIS) instruments aboard the Terra and Aqua satellites collect global LST data four times a day. Terra’s orbit crosses the equator at approximately 10:30 AM and 10:30 PM local solar time, while Aqua’s orbit does so at around 1:30 PM and 1:30 AM. Here, we utilized the MODIS Aqua LST Daily product at a 1 km resolution. We did not use Landsat LST products despite their higher spatial resolution (100 m), since separate LST products for daytime and nighttime are currently not available with Landsat, and the longer revisit cycle makes Landsat LST more susceptible to cloud contamination65,66.

We composited averaged daytime and nighttime LST during summer (June, July, and August for the Northern Hemisphere; December, January, and February for the Southern Hemisphere) between 2016 and 2020 from MODIS MYD11A1.061 Aqua LST and Emissivity Daily Global 1 km. We selected high-quality observations of LST by quality control (QC) band (Bits 0-1 equals 0 or 1, Bits 2-3 equals 0, and Bits 6-7 equals 1, see User’s Guide for detailed interpretation).

Selection of vegetation cover data and other covariates

Previous studies have found a widespread negative relationship between LST and VC38,75,76. Roof greening can involve a variety of configurations of short trees, shrubs, and lawns52. To fully represent the status of vegetation in roof greening, we used MODIS MOD44B.006 Terra Vegetation Continuous Fields Yearly Global 250 m data. This dataset offers sub-pixel-level estimates of surface VC at a global scale, providing a continuous representation of the terrestrial surface as proportions of three primary components: percent tree cover, percent non-tree VC, and percent bare ground. We defined VC as the sum of percent tree cover and percent non-tree VC to align with the fact that some roof greening involves grass and shrubs. VC from 2016 to 2020 were averaged to ensure temporal consistency with the composited LST data. We also tried higher-resolution tree cover data provided by Landsat at a 30 m resolution77. However, this dataset does not provide information on non-tree VC, making it impossible to accurately describe the actual vegetation used in green roof practices.

To improve the accuracy of modeling, variables closely related to urban heat effects were added into the model as covariates, including building height, building density, building materials, and elevation. Building height, building height variance, building density, and building materials have been proven to have significant influences on urban area temperature, as the three-dimensional landscape morphology could exert direct impacts on the process of surface energy exchange78,79,80,81,82. Building height data were derived from the Global Human Settlement Layer (GHSL) project, which aims to assess human presence across the planet. The GHSL building height dataset represents the global distribution of building heights at a 100-m resolution, based on data referenced to the year 2018 [ALOS Global Digital Surface Model (30 m), NASA’s Shuttle Radar Topographic Mission data (30 m), and a global Sentinel-2 image composite from L1C data collected during 2017–2018]. Building height variation was calculated as the standard deviation of building heights within a 1 km area. Building density was computed as the percentage of building pixels within a 1 km area. Building materials were obtained from the World Settlement Footprint 3D (WSF3D) building stock dataset83. The dataset provides consistent and spatially detailed data indicating the building material stock (biomass, fossil, metal, and mineral-based building materials) used in the construction of every settlement worldwide at a 90 m resolution. We calculated the proportion of each material and reprojected it to 1 km resolution. Additionally, variations in elevation can influence temperature patterns84. Elevation was obtained from NASA’s Shuttle Radar Topographic Mission (SRTM) digital elevation dataset, covering latitudes between 60°N and 56°S.

Building footprint/rooftop map

We obtained global roof areas from Google35 and Microsoft36. Since this dataset does not include China’s roof areas, we added China’s latest roof data in 2021, sourced from the China Building Rooftop Area (CBRA) dataset with a 2.5 m resolution37. Using an enhanced U-Net model, the Google Open Buildings dataset contains building outlines derived from ultra-high-resolution satellite imagery (50 cm). This dataset includes 1.8 billion building detections across a vast area of 58 million km², covering regions such as Africa, South Asia, Southeast Asia, Latin America, and the Caribbean. The Microsoft Building Footprints dataset, which includes 1.4 billion building detections, was created from Bing Maps imagery between 2014 and 2024, incorporating data from Maxar, Airbus, and IGN France. Using deep neural networks (DNNs), it detects building pixels in aerial images and converts them into polygons. However, both datasets exclude buildings in China. The CBRA dataset was specially designed for mapping China’s roof footprint, produced using Sentinel-2 imagery from 2016 to 2021 and a Spatio-Temporal aware Super-Resolution Segmentation framework. To ensure consistency, we rasterized the Google and Microsoft building footprints to a 2.5 m resolution, matching that of the China rooftop map.

City boundary and urban area

To construct city-specific models, we adopted the second-level administrative units from the FAO Global Administrative Unit Layers 2015 to define city boundaries. The corresponding urban area extent of each city was obtained from the GHSL Degree of Urbanization 1975–2030. This raster dataset provides a global classification of rural and urban areas, following the Stage I “Degree of Urbanization” methodology recommended by the UN Statistical Commission. The classification is derived from global gridded population data and built-up surface information produced by the GHSL project. We defined urban areas as semi-dense urban clusters, dense urban clusters, and urban centers in the GHSL Degree of Urbanization layer.

Urbanites

For the analysis of the population that would benefit from the cooling effect of roof greening, we defined “cooling benefit” as the product of urban population and cooling potential (°C * 105 people). The population was obtained for the year 2020 from the Gridded Population of the World Version 4.11 dataset at ~1 km resolution39. The urbanites correspond to people residing in urban areas (semi-dense urban clusters, dense urban clusters, and urban centers, as defined by the GHSL).

Model construction

We used city boundary data to obtain all available pixels within each city, enabling the modeling of LST changes in relation to VC. Multiple linear models were constructed for each city using the LST, VC, building height, building height variation, building density, proportions of building materials, and elevation. Notably, we did not use the fraction of mineral-based materials to reduce multicollinearity, as the sum of the proportions of these four materials is always equal to one. Also, the 1 km resolution used in our modeling has been shown in previous studies to be less affected by spatial autocorrelation38,85. The model took the following form (1):

$${LST}=a* {VC}+b* {BH}+c* {BHV}+d* {BD}+e* {p}_{{bio}}+\,f* {p}_{{fossil}}+\,g* {p}_{{metals}}+h* {Elevation}+\varepsilon \,$$
(1)

where LST is the daytime or nighttime LST, VC is the VC, BH is the building height, BHV is the building height variation, BD is the building density, \({p}_{{bio}}\), \({p}_{{fossil}}\) and \({p}_{{metals}}\) are the proportion (in %) of biomass, fossil and metal based building materials respectively, Elevation is the mean SRTM elevation value of the city, and \(a-i\) represent coefficients and \(\varepsilon\) is the error term. To ensure model accuracy, predictor variables were further automatically selected per city basis using stepwise regression based on the Akaike Information Criterion (AIC).

The model was applied to 25,500 global cities with urban areas, and 1448 cities were excluded for insufficient sample pixels. To further ensure the reliability of our modeling results, we focused on cities where model accuracy was high. The cities with qualified models were chosen according to the following criteria: (1) VC must be a statistically significant (p < 0.01) variable with a negative coefficient, indicating a cooling effect of green roofs; (2) The model R2 > 0.7; (3) The model RMSE < 2 K; (4) maximum of Variance Inflation Factors (VIF) among predictor variables <5, indicating no concerning multicollinearity among predictors. In this way, 4570 and 1810 cities were finally used for daytime and nighttime cooling potential estimation (867 paired). Additionally, not all cities have available building footprint data. This results in roof greening potential estimations for 4835 cities, with 3913 cities for daytime, 1687 for nighttime, and 765 cities having both daytime and nighttime cooling potentials.

Note that our model accounted for the cooling effect of water bodies86. Towards this purpose, we compared the VC coefficients in the models when pixels within a 2 km buffer area around water bodies were masked versus unmasked. The MOD44W.005 Land Water Mask, derived from MODIS and SRTM, was used for this analysis. We found minimal influence on the VC coefficients when sampling at the city-wide scale (Fig. S7). Therefore, roofs within the 2 km water buffer were also simulated under three greening scenarios.

Simulating cooling potentials and uncertainties

After model construction, we simulated the cooling potential of green roofs. This was done by changing VC in Eq. (1) by adding rooftops as additional vegetated areas. We defined three potential roof greening scenarios as follows: (1) Low roof greening effort (Low), that 20% of the surface area of rooftops in the city is covered by vegetation. (2) Medium roof greening effort (Medium), that 40% of the surface area of rooftops in the city is covered by vegetation. and (3) High roof greening effort (High), that 60% of the surface area of rooftops in the city is covered by vegetation. These three scenarios, especially the highest greening effort (60%), were decided by referring to existing cities, including Toronto, Portland, and San Francisco (Table S1). The 60% is also a practical, conservative consideration, as not all the roofs in a city can be vegetated.

Specifically, the cooling potential [Eq. (2)] was calculated by multiplying the increase in vegetation cover (i.e., ΔVC) with the coefficient a calculated in Eq. (1), and uncertainty [Eq. (3)] was represented as half of the 95% confidence interval relative to the estimated cooling potential:

$$\Delta {LST}=\,a\Delta {VC}$$
(2)
$$\Delta {{LST}}_{{uncertainty}}=\frac{\left({a}_{2.5 \% }-\,{a}_{97.5 \% }\right)\Delta {VC}}{2a\Delta {VC}}$$
(3)

where \(\Delta {LST}\) represents the cooling potential, \(a\) is the coefficient of VC obtained from the model in Eq. (1), ΔVC is the greening effort of pixels in 1 km resolution, \(\Delta {{LST}}_{{uncertainty}}\) denotes half of the 95% confidence interval relative to the estimated cooling potential, and \({a}_{2.5 \% }\) and \({a}_{97.5 \% }\) correspond to the lower and upper confidence limits respectively. Since we assume uniform roof greening scenarios in the urban core area70,71, the ΔVC for the scenario of low, medium, and high levels of roof greening were 20%, 40%, and 60%, respectively, according to existing mandates of green roofs (Table S1).

It should be noted that we assumed that the greening effect of green roofs cannot surpass the natural vegetation in VC. Therefore, we calculated the maximum original VC within a 10 km buffer area (matching the sampling area) to control the VC after roof greening. If the VC of a pixel after greening exceeded the maximum original VC, the pixel was assigned the maximum original VC, instead of adding the greening effort (20%, 40%, or 60% corresponding to the low, medium, or high scenarios) to the original VC. The resulting ΔVC was reprojected to a 1 km resolution to align with the original VC resolution. Also, urban heat island typically occurs in the city center, where dense buildings and impervious surfaces dominate, but green space is limited87. These urban areas are in the greatest need of roof greening to mitigate elevated temperatures. Consequently, it’s reasonable to restrict roof greening simulation to semi-dense urban clusters, dense urban clusters, and urban centers as defined by the GHSL.

Testing alternative models for simulating the cooling potential of green roofs

To ensure the accuracy of our simulations, we also tested three alternative models using tree canopy cover, NDVI, and albedo as the predictor variables, respectively. The Landsat Vegetation Continuous Fields tree cover layer provides estimates of the percentage of each 30 m pixel covered by woody vegetation taller than 5 m. The Normalized Difference Vegetation Index (NDVI) is a metric that quantifies vegetation greenness and measures the density and health of vegetation. The MODIS MOD13A3 V6.1 Vegetation Indices product provides monthly NDVI at 1 km resolution. We composited the averaged NDVI during summer (June, July, and August for the Northern Hemisphere; December, January, and February for the Southern Hemisphere) between 2016 and 2020 to align with the LST we processed above. Only pixels with good observation quality were used (band “SummaryQA” equals 0). Albedo represents the proportion of shortwave radiation reflected by a surface material and is one of the factors influencing LST. An albedo of 0 indicates a completely black surface with no reflective capability, while an albedo of 100% indicates a completely white surface that reflects all radiation88. A green roof has a much higher albedo (0.7–0.85) than conventional roof materials ( < 0.1)89. We used Landsat 8 Level 2 Collection 2 Tier 1 dataset which contains atmospherically corrected surface reflectance. We calculated albedo by Eq. (4)90:

$${albedo}=0.356* b2+0.130* b4+0.373* b5+0.085* b6+0.072* b7-0.0018$$
(4)

where \({bx}\) is the specific reflectance of each band (x is the band number).

We then composited the averaged albedo during summer (June, July, and August for the Northern Hemisphere; December, January, and February for the Southern Hemisphere) between 2016 and 2020. All layers with a resolution higher than 1 km were reprojected to 1 km when sampling.

The performance of these models is quite similar in terms of R², and the number of cities that meet the selection criterion of R² > 0.7 is also nearly the same. However, the model using NDVI performs significantly better during the daytime than the others, with 10,648 cities having R² > 0.7 (Fig. S4). The spatial distributions of R² for the models are also largely consistent (Fig. S5), suggesting the reasonability of our modeling results. We ultimately chose VC because, compared to tree cover, it also includes non-tree vegetation. We did not use NDVI or albedo because it remains unclear what NDVI or albedo values the green roofs have. In contrast, VC provides a simpler and more accurate estimate of cooling potential compared to these two factors.

SUHII reduction analysis

For the analysis of surface urban heat island intensity (SUHII) reduction due to the cooling potential of roof greening, we calculated the proportion of cooling potential relative to daytime and nighttime SUHII. The SUHII for each city was calculated as the difference of the averaged summer LST (See section “Land surface temperature” in “Methods”) between rural areas and urban areas. We defined rural as very low density rural, low density rural, and rural cluster, and urban areas as semi-dense urban clusters, dense urban clusters, and urban centers in the GHSL Degree of Urbanization layer. Only cities with a significant SUHI effect (SUHII ≥ 1 °C) were included in the analysis.

DTR change estimation

As we simulated the summer temperature changes under roof greening scenarios (June, July, and August for the Northern Hemisphere; December, January, and February for the Southern Hemisphere), we defined the average diurnal temperature range (DTR) as the difference between the daytime mean LST and the nighttime mean LST. The average DTR change was calculated by subtracting the nighttime LST from the daytime LST under different scenarios. This can be simplified as the difference between the nighttime and daytime cooling potential, as shown in Eq. (5):

$$\Delta {DTR}={{\Delta {LST}}_{{night}{time}}-\Delta {LST}}_{{daytime}}$$
(5)

where \(\triangle {DTR}\) represents the DTR change, \({\Delta {LST}}_{{night}{time}}\) and \({\Delta {LST}}_{{daytime}}\) correspond to the cooling potential during the night and day, respectively.