Worldwide rooftop photovoltaic electricity generation may mitigate global warming

Two-stage process for global rooftop area estimation

The global rooftop area estimation followed a two-stage process (Extended Data Fig. 1): (1) a top-down approach using deep learning to quantify the rooftop area in selected representative regions from high-resolution remote sensing imagery and (2) a bottom-up approach using random forest ensembles and various geospatial datasets to extrapolate rooftop area estimates at a global scale.

In the top-down stage, we used the SegFormer model, a state-of-the-art Vision Transformer-based deep learning model in computer vision47. The choice of this architecture was motivated by its capacity to process large datasets and its ability to capture global context while handling complex patterns and long-range dependencies—crucial for large-scale building identification48. For pretraining, we used publicly available benchmark datasets focused on building identification13,49 (Supplementary Table 1). These datasets, generated by various institutions between 2013 and 2020, cover ~2,500 km2 across diverse regions with varying spatial resolutions, ranging from very fine (0.1 m per pixel) to coarser (3 m per pixel). The selected datasets represent a wide range of global regions and building types, including urban, suburban and rural areas, as well as a mix of built-up and natural landscapes. They also cover various urban layouts and densities, offering high-resolution imagery that captures different architectural styles and building sizes. This broad representation ensures that the model is trained across diverse environments, enhancing its ability to generalize to different geographic contexts.

For fine-tuning the deep learning model, we used high-resolution Google Earth imagery with a consistent spatial resolution of 1.2 m. Google Earth imagery is obtained through the Google Static Map API, which applies rigorous sifting and preprocessing to produce cloud-free, high-quality images50. The imagery is sourced from a combination of airborne and satellite platforms, including WorldView, QuickBird, IKONOS, GeoEye-1, Pleiades, SuperView-1 and Kompsat-3A, resulting in harmonized imagery51. Further details on the SegFormer architecture, along with the pretraining and fine-tuning processes, are provided in Supplementary Note 3.

We applied the trained deep learning model to quantify rooftop area in 1,724 cities, where ‘city’ refers to administrative divisions sourced from the database of global administrative areas (Extended Data Fig. 4). These regions were strategically selected using a spatial sampling scheme, with cities treated as sample units. The objective was to maximize the distance between samples to ensure representativeness, optimized using a simulated annealing algorithm52. To further enhance the geographic and environmental diversity of the selected regions, we first stratified the global grids using K-means clustering, based on natural and human characteristics of each grid (Supplementary Fig. 6 and Supplementary Table 2). The optimal number of clusters for the K-means algorithm was determined using silhouette coefficients. After stratification, spatial sampling was performed within each stratum.

In the bottom-up stage, we collected multi-source geospatial datasets and aggregated them as statistical variables—such as built-up proportion, night-time light intensity, road length, population, tree-cover proportion, terrain elevation, terrain slope and geographical coordinates (Supplementary Fig. 2 and Supplementary Table 2)—into global discrete spatial grids. These variables are strong predictors of rooftop areas because they capture both human activity and the physical characteristics of the environment9,10. For instance, night-time light intensity is often correlated with urbanization and building density, while terrain elevation and slope can influence building placement and rooftop orientation. Each spatial grid covered an area of 1 km2 and was based on the world Mollweide equal area projection (EPSG: 6933). To model the nonlinear relationship between the rooftop area and these geospatial variables, we aggregated the rooftop area determined in the selected regions into the corresponding grids. Grids lacking high-resolution remote sensing imagery were excluded, resulting in 8.5 million grid samples. The rooftop area was treated as the dependent variable in these samples, while the other statistical variables were used as independent variables.

Considering the spatial heterogeneity of grid cell variables and addressing the long-tail regression modelling issue, we developed both regression- and classification-oriented random forest models at both continental and global scales. The trained random forest ensembles were then applied to estimate the global rooftop area using a set of global independent variables. Further details on the random forest ensemble can be found in Supplementary Note 4. In the postprocessing stage, a water map was used to refine the initial estimates by assigning a rooftop area of 0 m2 to grids that were 100% water-covered based on land use and land cover data.

Evaluation of estimated global rooftop area

The evaluation involved assessing the accuracy of the deep learning model in the top-down stage and the random forest ensemble in the bottom-up stage using independent test datasets. Additionally, the quality of the global rooftop area estimates was evaluated through qualitative and quantitative comparisons with reference datasets.

To evaluate the performance of the deep learning model, we created 386 plots, each covering an area of 1 km2 (Extended Data Fig. 2b). These plots were selected to ensure global representation, with two plots randomly distributed within the built-up areas of each country. High-resolution remote sensing images were obtained for all 386 plots and each was manually labelled to identify rooftop areas. The average rooftop area across the plots was 0.14 ± 0.10 km2 (mean ± s.d.), with a maximum of 0.53 km2. These plots span different countries and geographic regions, encompassing a wide range of building distributions and densities, providing a comprehensive evaluation. The plots were further divided into 2,951 image patches, each processed to match the size and colour bands of the training images used for the deep learning model, ensuring consistency in the validation process (examples in Supplementary Fig. 1). This test dataset was kept independent from the training data to prevent bias in the performance assessment of the model.

Against the labelled image patches, the deep learning model successfully identified 76% of the rooftops (true-positive rate), with a 2.7% false-positive rate where non-rooftop objects were mistakenly identified as rooftops. This level of performance is acceptable for building identification at a global scale: we compared the MBF and GBF—currently state-of-the-art building footprint datasets—against our test dataset. MBF had a true-positive identification rate of 61.6% and a false-positive rate of 4%, while GBF had a true-positive rate of 66.5% and a false-positive rate of 3.8%.

Additionally, we compared the rooftop area quantified by the deep learning model with the manually labelled rooftop area for each image patch. Overall, the predictions of the model showed a strong correlation with the actual data, with an r2 value of 0.93 and a slope of 1.04 across all test samples (Extended Data Fig. 3a). However, we observed variations in prediction accuracy across different macroregions (Extended Data Fig. 3b–u): the model performed with greater accuracy in economically developed regions, where r2 values exceeded 0.95. In contrast, accuracy decreased in less economically developed regions, with r2 values around 0.9. These discrepancies could be due to the varying ability of the model to recognize landscapes that differ notably across regions.

To evaluate the accuracy of the global rooftop area estimation by the random forest ensemble, we randomly selected 16,000 grids across various macroregions (800 grid samples per macroregion). These grids were not part of the training process of the random forest model. We used deep learning and high-resolution remote sensing imagery to quantify the rooftop area within these grids. The rooftop area determined from the images was then compared with the estimates generated by the random forest ensemble. Our results demonstrated a high level of accuracy, with an r2 value of 0.89 and a slope of 0.87 (Extended Data Fig. 5), indicating strong consistency between the estimated and observed rooftop area, although with a slight underestimation.

However, the results also indicated a drop in accuracy for the Pacific Islands (r2 = 0.61, bias error = 55%; Extended Data Fig. 5c) and Western Asia (r2 = 0.67, bias error = 24%; Extended Data Fig. 5k). This suggests potential limitations in capturing relevant factors in these regions, probably due to a lack of local data in the training set, especially in areas where high-resolution remote sensing imagery is scarce. The bias error here is defined as the relative bias, normalized by the absolute value of the sum of the observed rooftop areas and is calculated using the following equation53:

We also compared the estimated and observed rooftop area by identifying deviations (residuals) across the 16,000 test grids. The results revealed a correlation between rooftop area and residuals: grids with larger rooftop area tended to exhibit greater residuals. Specifically, for grids with rooftop area ranging from 0.01 to 1 km2, the variation in residuals was substantial, showing a slight underestimation (Extended Data Fig. 6a). Residuals demonstrated a bell-shaped normal distribution, with most clustering between ±5,000 m2 (Extended Data Fig. 6b). Further examination of the global distribution of estimated rooftop area revealed that most 1-km2 grids had rooftop area <0.4 km2 (Extended Data Fig. 7a). This consistent pattern of rooftop area distribution was evident across various scales (Extended Data Fig. 7b–u).

To assess the quality of our global rooftop area estimates, we compared them with several reference datasets, including GBF, MBF, GHSL and WSF3D. The qualitative characteristics of these reference datasets are described in Supplementary Table 10. While GBF and MBF represent advanced building footprint products, their coverage is incomplete, omitting extensive areas across multiple continents (Supplementary Fig. 7). For example, aggregated GBF lacks large portions of North America, Europe, East Asia and Australia, while MBF lacks coverage over China, Russia and parts of Europe. Compared with these datasets, our estimates provide globally seamless coverage, produced through the two-stage method. By incorporating natural and human-related datasets from 2020, our estimates capture current urban development and building areas, with the exception of the digital elevation model from SRTM v.4 (collected in the 2000s), which remains appropriate given the relatively stable nature of global terrain over time.

For the quantitative comparison, we conducted a detailed grid-level and continental-level analysis. Owing to the limited coverage and potential spatial completeness issues in the GBF and MBF datasets, we restricted our comparison to grids that contain building areas in both datasets. For comparisons with GHSL and WSF3D, we used global grids. The results of these grid comparisons are shown in Extended Data Fig. 8, where our global rooftop area estimates exhibit high correlations with the reference datasets. However, our estimates tend to be lower than those of the reference datasets, as reflected by the slopes of the linear fits, all of which are <1 (0.71 for GBF, 0.83 for MBF, 0.58 for GHSL and 0.72 for WSF3D). This underestimation trend is consistent with our test dataset results, which also produced a slope of 0.87. Among the reference datasets, the comparison with GHSL shows the largest discrepancies, with an r2 value of 0.76 and a bias error of 37%. We attribute this to the tendency of GHSL to overestimate rooftop areas by including impervious surfaces.

We also evaluated the estimates across different macroregions. The grid-level analysis shows high correlations with both GHSL and WSF3D (Supplementary Figs. 8 and 9). However, in less economically developed regions—such as the Pacific Islands, Central Asia and various regions of Africa (Eastern, Southern, Western and Middle Africa)—we observed relatively lower r2 values and higher biases. Furthermore, the comparison of total rooftop area across macroregions, summarized in Supplementary Table 11, shows that WSF3D and our estimates are generally consistent, while the total area of GHSL is substantially higher. This trend persists in less economically developed regions, where our estimates are typically lower. The potential underestimation could be due to geographic variations in building identification accuracy, as previously discussed. At the global level, our estimates closely align with those of WSF3D, although the totals of WSF3D are slightly higher. In contrast, the global area of GHSL is substantially larger, probably indicating overestimation. Compared with a previous study that estimated a total global rooftop area of 193,875 km2 (which was noted for systematic underestimation)9, our estimates (286,393 km2) demonstrate improved alignment with reference datasets. Although slight underestimation remains in our results, we have mitigated many of the issues identified in earlier studies.

Assessment of electricity generation and carbon mitigation potential

Unified assumptions were applied to the RPV systems in this study. On the basis of the current technical level of the PV industry54, the scale and performance parameters of PV applications were determined (Supplementary Tables 12 and 13), including a PV panel conversion efficiency of 20%, an overall PV system efficiency of 0.8 and a rated power of 200 W m−2. All the PV panels in the system were assumed to be fixed horizontally. The available installed capacity of RPVs, essential to determine their potential for electricity generation and carbon mitigation, was calculated by converting the total rooftop area into the available rooftop area using a scaling factor. The factor typically considers a range of geographical constraints for RPV system installation, such as the societal function, rooftop slope, orientation, shadows and obstacles of the building. On the basis of a literature review (Supplementary Table 3), the scaling factor used in the main results was assumed to be 30%.

The potential installed capacity, \({P}_{\rm{installed}}\), was calculated using equation (2) as follows:

where \({P}_{\rm{rated}}\) is the rated power of the PV panel, \({S}_{\rm{rooftop}}\) is the total rooftop area obtained from the global rooftop area estimation and \({C}_{\rm{scaling}}\) is the scaling factor for calculating the available rooftop area for RPV installation.

The potential electricity generation \({P}_{\rm{power}}\) of the RPVs was estimated using equation (3):

$${P}_{\rm{power}}={S}_{\rm{rooftop}}\times {C}_{\rm{scaling}}\times {\rm{GHI}}\times {C}_{\rm{PV}}\times K$$ (3)

where GHI is the global horizontal irradiance received by the surface. We took the average values from 2010 to 2018 to represent the general radiation conditions at different assessment times. C PV is the conversion efficiency of the PV panel and K is the overall efficiency of the RPV system.

In this study, our focus was on the electricity generation stage of RPV systems, excluding consideration of other life-cycle stages. This decision stems from the considerably lower life-cycle carbon emissions of RPV systems compared with the emissions mitigated during the operational stage55 (Supplementary Table 14). We followed the baseline methodology56 proposed by the United Nations Framework Convention on Climate Change (UNFCCC) to measure the environmental benefits of renewable energy projects. The baseline methodology provides a simplified analytical framework and facilitates a uniform comparison across different projects, thereby promoting the development of related policies. The carbon emissions mitigated by RPV power replacing grid power were calculated using the baseline emission factors of the national grid57,58. The baseline emission factors included the operating margin (OM) and build margin (BM). While OM represents the cohort of existing power plants most affected (reduced) by the project (generally high-emission or high-cost power plants), BM represents the cohort of prospective/future power plants whose construction and operation could be affected by renewable energy projects, based on an assessment of planned and expected new generation capacity. The combined margin (CM), obtained from the weighted average of the OM and BM, represents the overall impact of both aspects, and reflects the existing carbon intensity of the national grid. Because the electricity generated through RPVs is considered clean energy, the CM factor, \({\rm{EF}}_{\rm{grid,CM}}\), can be used as the carbon mitigation factor, \({\rm{EF}}_{\rm{PV,mitigation}}\), of the RPV system. The calculations were conducted using equation (4) as follows:

where \({\rm{EF}}_{\rm{grid,OM}}\) is the OM factor and \({\rm{EF}}_{\rm{grid,BM}}\) is the BM factor. W OM is the weight of the OM factor and W BM is the weight of the BM factor. The values for PV projects were set to 0.75 and 0.25, respectively, based on the UNFCCC methodologies58.

According to the carbon mitigation factors of the RPV system for different power grids, carbon mitigation, \({P}_{\rm{carbon},{mitigation}}\), was calculated according to equation (5):

The selected baseline emission factors reflect the average conditions of national grids in recent years and are updated at least every 2 years (ref. 58). For assessment purposes, we designated 2020 as the base year. Future baseline emission factors are subject to alterations in the power structure of each country. We assume that this change is consistent with the trend in grid carbon intensity projected by the International Energy Agency for 2020–205059. On the basis of this assumption, we obtained future changes in the \({\rm{EF}}_{\rm{PV,mitigate}}\) under three climate policy scenarios: STEPS, SDS and NZE (Supplementary Notes 1 and 2 and Supplementary Fig. 4).

In this study, the energy assessment relied on conversion factors, such as rooftop area availability and PV panel efficiency, using harmonized global reference values. We acknowledge the importance of examining these factors at the regional level. However, such a detailed analysis was beyond the scope of this study. Sensitivity analyses were included to mitigate this limitation and provide an initial understanding of how variations in these factors could affect our main results (Supplementary Tables 4–6).

Assessment of global and regional warming mitigation potential

The relationship between global warming and cumulative carbon emissions was established using the well-documented TCRE metric. TCRE is estimated as the change in global surface air temperature per cumulative CO 2 emissions60. It considers both physical climate processes and the dynamics of land and ocean carbon sinks, presumed to be constant over time and independent of emission pathways61. The relationship between regional climate warming and carbon emissions was established using the RTCRE, which describe changes on a regional scale. In this study, the TCRE was scaled to a regional scale to generate an RTCRE using simple pattern scaling62. According to a previous study27, RTCRE can be calculated using equation (6):

where \(\varDelta T(m,x,t)\) is the temperature change simulated by the mth model in the ensemble, for the spatial domain x at time t. Parameter \(E(m,t)\) is the diagnostic value of the cumulative CO 2 emissions (calculated as the sum of the annual changes in atmospheric CO 2 concentration and the CO 2 uptake by ocean and land carbon sinks) of the model.

We used the 1pctCO2 simulations from nine ESMs with dynamic carbon cycle from the CMIP6 archive. The ensemble comprised ACCESS-ESM1.5, BCC-CSM2-MR, CanESM5, CESM2, IPSL-CM6A-LR, MIROC-ES2L, MPI-ESM1.2-LR, NorESM2-LM and UKESM1-0-LL. We followed the standard method63 for calculating \(\varDelta T(m,x,t)\), averaging over a 20-yr window (years 60–79) centred on the year of the CO 2 doubling (year 70). We applied this calculation to the spatial temperature fields from each model and then calculated the ensemble mean of the nine models. We considered the ensemble mean as the best estimate of the RTCRE values and represented the uncertainty range using one standard deviation of the model responses.

After obtaining the RTCRE, the warming mitigation potential, \({P}_{\rm{warming},{mitigate}}\) of RPVs were calculated according to equation (7):

While the concept of TCRE simplifies the relationship between CO 2 emissions and global warming into a linear model, this relationship is fraught with uncertainties owing to numerous complex factors, such as the effects of permafrost melting and non-CO 2 GHGs64. Land surface changes caused by PV panels could also impact the regional climate to some extent by redistributing surface energy, water fluxes and even atmospheric circulation65; however, global climate change mitigation should be dominated by carbon emission reductions. In addition, future solar energy generation could be modulated by climate change to some extent66. Future climate change impacts should be considered for finer assessments.

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.