Estimating high-resolution albedo for urban applications

By 2050 the global urban population is expected to increase from 55% to 68%¹, a projected doubling of the global urban footprint^2,3. This projected urban land expansion is expected to cause warming almost as strong as that caused by global greenhouse gas emissions^2,4. Urbanization produces profound changes in the surface-energy balance through the conversion of vegetation to impervious surfaces, a change that typically reduces surface reflectivity-also known as albedo. At the global level, these changes increase radiative forcing, producing a small but significant global temperature increase⁵. At the local level, these land conversions cause a suite of climate modifications known as the urban heat island (UHI) effect, which causes urban air temperatures to elevate above background climate conditions⁶. These global and local effects of urbanization combined with overall global warming are increasing temperatures in cities at the same time that urban population growth is increasing exposure⁷. In fact, between 1950 and 2015 about 60% of urban residents experienced warming twice the global average⁸. High temperatures in cities are detrimental to human health and well-being, can damage infrastructure and disrupt services, and increase energy demand for air conditioning ⁹.

There are two primary city-scale pathways for lowering air temperatures in urban areas that do not rely on the use of air conditioning: the evaporation of water—primarily by vegetation—and the reflection of sunlight. Common strategies for these pathways include increasing tree canopy cover and other vegetation, and increasing the reflectivity of roofs and roads, respectively. Regional-scale simulations using a combination of these strategies suggest that air temperature reductions on the order of several degrees Celsius are feasible¹⁰ and could have substantial positive health impacts⁷.

Roofs occupy anywhere from 20–25% of urban land areas¹¹ and vary greatly in size. The majority of urban roofs are covered with dark materials, absorbing as much as 80–90% of incoming solar radiation¹². Cool roofs are capable of reflecting almost all incident sunlight¹³. Increasing the reflectivity of roofs reduces the surface heat gain and the subsequent transfer of heat from the roof surface to both the atmosphere and the interior of the building, thereby lowering air temperatures and reducing energy loads^13,14. While the average albedo of most cities ranges from 0.12–0.16¹⁵—meaning that they reflect only 12–16 percent of received solar radiation—cool roofing materials can have initial solar reflectivities of above 0.9¹⁶ presenting a substantial opportunity for increasing city-wide albedos. However, research by Ouyang et al.⁵ suggests that at the global scale highly reflective materials are not widely implemented in cities.

Cool roofs are a high-impact solution for mitigating urban heat that are relatively low cost compared to traditional roofing materials and produce substantial energy savings without considerable aesthetic conflict¹⁴. Additionally, most roofs require regular maintenance necessitating periodic resurfacing, yielding ample opportunities for conversion to cool roofs¹⁴. At the global scale, conservative estimates show that increasing urban albedo by 0.1 would result in cooling equivalent to the removal of 44 Gt of CO₂ emissions, more than the projected global emissions for 2025¹⁴. At the local scale, increasing the reflectivity of dark roofs can lower energy use by up to 20%¹³ and temperatures by up to 0.6 °C per 0.10 albedo increase¹⁷. However, at both of these scales, a major source of uncertainty in modeling the effects of increasing albedo is estimating the modifiable land area¹⁸.

Accurate estimation of the potential for cool roof implementation and the temperature effects of wide-spread adoption necessitates a robust baseline estimation of rooftop albedo. This requires albedo imagery with a spatial resolution sufficiently high to resolve urban features relevant to cooling interventions, i.e. roofs. Further, the albedo inference should have an accuracy and revisit cadence that is useful for urban planning and progress monitoring.

Accurate and globally available characterization of albedo can be derived from Sentinel-2 observations by applying the narrow-to-broadband coefficients from Bonafoni and Sekertekin¹⁹, however, the highest spatial resolution available is 10 meters. To effectively characterize the energy flux of a surface, the feature must be substantially larger than the sensor resolution²⁰. In practice, roofs generally need to exceed roughly 1000 m²-slightly larger than a 3  × 3 block of Sentinel-2’s 10 meter bands (red, green, blue, and near infrared (RBG and NIR))-to have a reasonable chance that the spatial median within the building footprint represents a mostly roof albedo. This threshold is only a rough rule of thumb, and several factors make accurate rooftop albedo retrieval challenging even for larger buildings: (i) Sentinel-2 geolocation is accurate to only about 10 m, (ii) building footprints seldom align cleanly with pixel boundaries, and (iii) the key shortwave infrared (SWIR) bands have coarser a 20 m resolution. As a result, many roofs – particularly small or irregularly shaped ones—are dominated by mixed pixels, which leads to biased or unstable Sentinel-2 albedo estimates. High resolution rooftop albedo inference from the National Agriculture Imagery Program (NAIP) multispectral (RGBN) aerial images was introduced by¹³ using atmospheric corrections and methods to radiometrically calibrate the NAIP cameras and more recently revisited by²¹ who used machine learning methods and Sentinel-2 observations for calibration. Unfortunately, aerial imagery, such as NAIP, is not available globally, and the long acquisition times lead to the variation of illumination conditions.

This paper presents simple, scalable, and observationally validated methods for inferring high-resolution albedo by combining the spatial information from globally available high-resolution satellite imagery with the multi-spectral information of Sentinel-2, using Sentinel-2 as the source of radiometric calibration and atmospheric correction. Our approach demonstrates a scalable approach to urban albedo inference, realizing the potential for global building-level albedo coverage with high repeat cadence. To minimize illumination disparities (and to ensure global coverage of all habitable cities), we use sun synchronous high resolution satellite imagery which shares the roughly 10:30 AM local ascending node equatorial crossing time as Sentinel-2 (known as the morning train or M-train). A central contribution of this work is the large-scale validation of rooftop albedo using contemporaneous airborne hyperspectral measurements. This validation establishes confidence in both Sentinel-2 albedo for broad-scale urban analysis and in our high-resolution inferences for fine-scale, building-level planning. We quantify the accuracy of two methods for inferring high-resolution albedo-convolutional calibration and calibration transfer-and evaluate them along with Sentinel-2 against direct aerial observations. Finally, we apply these methods to a cross-section of 12 global cities, quantifying their solar radiation management potential and estimating the albedo and temperature impacts of three city-scale cool roof implementation scenarios.

NEON data & validation

Direct ground-based albedo measurements (e.g., from flux towers or pyranometers) are sparse, making it difficult to validate large-scale satellite-derived albedo estimates, especially for urban areas, and validation is exceptionally rare within the published scientific literature. To address this challenge, we collaborated with the National Ecological Observatory Network (NEON) to acquire high-resolution hyperspectral aerial imaging. While NEON flights typically focus on rural ecological observation, we use data from hyperspectral aerial imaging flown over Boulder, CO on May 8, 2024 (Fig. 1). We also tasked a Pléiades Neo satellite acquisition to temporally align with the NEON hyperspectral flight, providing a powerful 1-m urban reference datasets. While this validation site represents a single city, it offers an exceptionally rare and valuable benchmark for direct validation of albedo estimates.

The NEON albedo data has a resolution of 1 meter and is derived by integrating surface reflectance across wavelengths in the 0.4 to 2.5 micron band, weighted by the global flux received at the ground²². Data are provided as 18 overlapping images from flightlines taken throughout the day, with solar elevations ranging from 42.65° to 61.45° (see Supplementary Table 1 & Supplementary Fig. 1).

To create the high-resolution albedo inference, we used two methods to radiometrically calibrate a high-resolution image with a low-resolution image—convolution calibration and calibration transfer. For both methods, we started with a temporal median of cloud and snow free low-resolution albedo images proximate in time to a high resolution image. Then, we co-registered the two images and applied the calibration methods, described in detail below, to preserve the spatial structure of the high-resolution image while forcing it to agree radiometrically with the low-resolution image.

For the high-resolution imagery, we made arrangements with Airbus in advance of the NEON flyovers to task the Pléiades Neo satellite constellation for dates on and near to the NEON flyover (May 8, 2024 and May 16 2024). We also analyzed Pléiades Neo imagery available for May 23, 2023. The Pléiades Neo constellation acquires imagery across six multispectral bands at a native 1.2 meter resolution and also captures a co-registered 30 cm panchromatic band. This configuration enables pansharpening, which fuses the two data types to create a final multispectral product with a high spatial resolution of 30 cm.

For the source of the radiometric calibration, we use Sentinel-2. Sentinel-2 image processing begins with conversion of observations from top-of-atmosphere to bottom-of-atmosphere reflectance through atmospheric correction using the Sen2Cor algorithm²³. To convert narrowband Sentinel-2 reflectances into broadband albedo, we applied the model from from Bonafoni and Sekertekin¹⁹.

We created 100-day Sentinel-2 cloud and snow masked temporal median albedo images centered on the Pléiades Neo image acquisition date. Temporal medians are a stable and globally scalable albedo estimate under changing cloud cover and ground conditions (see Supplementary Fig. 2). Additionally, serendipity allowed the acquisition of a cloud-free Sentinel-2 image from April 30, 2024. To validate the use of the temporal median images for calibration, we evaluated our methods using both the single image and temporal medians (Table 1 and Supplementary Table 2).

The validation study area is the intersection of the bounds of the Pléiades Neo and NEON datasets and covers much of the urbanized area of Boulder, CO, U.S.A. To compare the four datasets, we aggregated the albedo estimates to roof segments produced from digital surface models (DSMs) that were generated from each of the Pléiades Neo high-resolution images using the method of²⁴ (Fig. 1). The roof segmentation represents the state of the art in Google’s AI-generated, self-aligned digital surface mapping whose objective is to extend roof segmentation globally by using relatively inexpensive satellite imagery and advanced machine learning methods²⁴. Because the roof segment polygons, which enclose areas of constant surface normal, are derived directly from the Pléiades Neo imagery, each image date creates a slightly different set of roof segments that are self-aligned to the Pléiades Neo imagery. We calculated the median albedo within each roof segment for each albedo inference. Having the surface normal for each roof segment polygon allows us to correct for differences in illumination due to directional reflectance by calculating the surface albedo from the image albedo as described in the Methods section.

Image albedo refers to the observed reflectance captured by the imaging sensor and is interpreted as if the scene were flat. However, reflectance varies with solar angle, view geometry, surface geometry, and the surface bidirectional reflectance distribution function (BRDF). Surface albedo represents the intrinsic reflectivity of the material under standardized illumination conditions. We corrected for these angular effects using the known orientation and tilt of roof planes with the assumption the surfaces are Lambertian, enabling a more accurate estimation of true surface albedo (see Methods for further details). While information about the roof geometry provides valuable context for evaluating our methods and may be accessed through the Google Maps Platform Solar API²⁵, they are not publicly available. To validate our methods for a broader use case, we additionally calculated the median albedo of each dataset for building footprints acquired from OvertureMaps²⁶.

To minimize the measurement inaccuracies resulting from differences in solar illumination, surface, and sensor viewing geometry, as well as atmospheric disturbances, we used only the highest precision NEON estimates, which we defined as roof segments covered by more than one flightline having a standard deviation of less than 0.01 between the multiple albedo measurements. The Kolmogorov-Smirnov (KS) test reports a KS statistic of D = 0.1554 on a scale of 0 to 1, indicating that the distribution of the highest precision NEON observations is similar to the distribution of all NEON observations with a maximum distance of 15.54% between the two cumulative distribution functions. We conclude that these highest precision estimates are representative of the entire NEON albedo dataset and define these observations as ground truth.

To evaluate our methods, we compared four datasets (Fig. 2): albedo measured at 1 meter resolution by the NEON hyperspectral instrument, albedo estimated at 10 meter resolution using Sentinel-2 imagery, albedo inferred by calibrating a 30 cm Pléiades Neo image using Sentinel-2 imagery using the convolutional calibration method, albedo inferred by calibrating a 30 cm Pléiades Neo image using Sentinel-2 imagery using the calibration transfer method.

We began by evaluating the two methods of albedo inference: convolution calibration and calibration transfer. Convolution calibration uses a circular averaging kernel of radius 10 meters (the size of a Sentinel-2 pixel) to rescale each band of the high-resolution image by the ratio of the kernel convolutions of the calibrated and uncalibrated image bands. Albedo is then computed from the calibrated high-resolution bands. The alternative method, calibration transfer, uses a machine learning model trained (on an image-wise basis) on upscaled high-resolution imagery and targeting Sentinel-2 albedo to infer the high-resolution albedo.

To validate our methods we compare the values for 54,076 roof segments for which we have both NEON observations and inferred albedo estimates. We began with the most precise data: surface albedo estimates for roof segments generated from a single Sentinel-2 image (April 30, 2024) and a single Pléiades Neo image (May 8, 2024), the high-resolution image being acquired on the same date as the NEON flyover and the Sentinel-2 image being the cloud-free image most proximate. This limited, high-precision ground truth sample contains 440 roof segments where the NEON observations overlap with both the April 30 Sentinel-2 image and the May 8 Pléiades Neo image. Compared to NEON ground truth data, the calibration transfer method produces a slightly higher R-squared value than the convolution-based calibration, although its RMSE and bias are also marginally greater (Supplementary Table 2). This pattern holds whether the analysis is conducted on uncorrected (image albedo) or slope-corrected (surface albedo) reflectance values derived from individual roof segments. Despite these corrections, the accuracy of the surface albedo and the image albedo are nearly equivalent when validated against NEON reference data, with the goodness-of-fit metrics of the image albedo exceeding those of the surface albedo for May 8, 2024.

We expand our sample by also comparing the results of the two methods using the high-resolution images from May 23, 2023 (n = 11,202 roof segments) and May 16, 2024 (n = 15,396 roof segments). As the temporal gap between the high-resolution imagery and the NEON flyover increase, the goodness-of-fit of both methods decreases, although the relative performance of the methods remains the same, with the calibration transfer performing slightly better than the convolution calibration, and both inference methods outperforming the Sentinel-2 albedo estimates. Both the convolution calibration and calibration transfer methods are effective at filling in the details of the inferred albedo image by making use of the high resolution content of the uncalibrated image. These are only two possible models, notably there are in theory an infinite number of possible high-resolution albedo images that will radiometrically match Sentinel-2 when they are upscaled to 10 meter resolution, including established methods like the HySure model²⁷ which better preserves edges and spectral continuity at the expense of added computation. It is not a goal of this study to exhaustively explore the space of all calibration methods, but rather to pick two rationally constructed yet algorithmically very different models and compare these to ground truth comprising the aerial NEON albedo. The results show that an image processing approach and a machine learning approach give very similar results with divergent methodological approaches. However, the calibration transfer method had two drawbacks in that it required training a new model for each high resolution image and for high/low albedos, introducing a bias and regression towards the mean. We determined that the marginal improvement in accuracy is not a meaningful gain over convolution calibration, which is simple, fast, avoids bias at low and high albedos, and does not require training a model. Based on this validation and the relative ease of computation we choose convolution calibration as the preferred method of albedo inference and we use this method for the illustrative application that follows.

An objective of this work was to see whether one can calibrate Pléiades Neo imagery wherever and whenever it is acquired, making high resolution calibrated imagery available globally. However, the availability of cloud-free Sentinel-2 imagery precisely contemporaneous with high-resolution imagery is not guaranteed; additionally, glint, glare, shadows and ephemeral scene-features impact Sentinel-2 surface reflectance measurements. To minimize these effects, we calculate the temporal median albedos from the 50 days on either side (100-day window) of the high-resolution images to use as the calibration source for the high-resolution albedo inferences (see Supplementary Temporal Medians for details on the choice of a 100-day window). We then compare the performance of the albedo inferences against the Sentinel-2 albedo estimates when validated against the NEON ground truth (Table 1). The temporal medians are slightly less accurate compared to NEON, but still perform strongly. Like with the single-day inferences, for both the high-resolution albedo and Sentinel-2 imagery as the temporal gap between the NEON flyover and the centers of the temporal medians increases, the accuracy decreases. This is true regardless of whether or not geometric corrections have been made. Across all metrics and temporal resolutions, the high-resolution albedo outperforms Sentinel-2. Intercomparison results are similar regardless of whether one is comparing image albedo or surface albedo.

To assess how roof characteristics impact the accuracy of the data we examine the goodness of fit of the uncorrected image albedos by roof segment pitch, roof segment azimuth, and roof segment area (Fig. 3). We bin the pitch, azimuth, and area of the roof segments into deciles and calculate the R², RMSE, and bias values for the observations in each bin. For each roof segment variable, the convolutional high-resolution albedo outperforms Sentinel-2, with a bias closer to zero, lower RMSE values, and higher R² values. In particular, the high-resolution albedo inferences provide better inferences for small roof segments and are more robust to the effects of increased pitch. Expectedly, as the roof segment area increases, the Sentinel-2 imagery approaches the accuracy of the high-resolution inferences.

Finally, we examine the overall accuracy of the two datasets, including all roof segments from all images (n = 54,076) (Fig. 4). For the roof segments, both the high-resolution albedo inference and the Sentinel-2 albedo are biased high relative to the NEON albedo, in general overestimating albedos. However, the density plots show that the convolution calibration methods (top Pléiades Neo distribution in Fig. 4) effectively shifts the distribution of the Sentinel-2 albedo towards that of the NEON ground truth. The albedo inference explains almost 25% more of the variation in the NEON albedo than Sentinel-2 (R² value of 0.841 compared to 0.607), has a lower RMSE, and is less biased.

While the roof segment outlines provide valuable information on pitch and azimuth, they are not yet publicly available. In contrast, the building footprints provided by Overture Maps are globally available and aggregating albedo to the roof level is a likely use case for the data (Fig. 1). We evaluate the accuracy of the convolutional albedo inference and the Sentinel-2 albedo against the NEON albedo measurements when aggregated to the building footprint (Fig. 4). We calculate the median albedo for each building footprint in the study area (n = 10,767) for each dataset. Where a building footprint is covered by multiple NEON flightlines, we use the albedo value from the flightline with the higher solar elevation to more closely match the solar conditions of the satellite overpasses (63.6° solar elevation for the Pléiades Neo acquisition and 61.46° solar elevation for the Sentinel-2 overpass on May 8, 2024). The density plot (Fig. 4) shows that in addition to overestimating building footprint albedos, Sentinel-2 also regresses towards the mean, with a narrower distribution than observed in NEON. The convolution calibration methods both shift the distribution to the left and widen the distribution, making it more similar to that of the NEON albedo measurements. The goodness of fit is higher than for the roof segments, unsurprising because the values are averaged over larger areas than for the roof segments. While the accuracy of Sentinel-2 is higher for building footprints than for the roof segments (R² of 0.88 compared to 0.71), the high-resolution albedo inference still provides a considerable increase in accuracy, explaining almost 17% more of the building level variation in NEON albedo.

Boulder cool roof scenario evaluation

Having validated our methods for estimating the albedo of building footprints, we apply these methods to quantify the solar radiation management potential for Boulder. Specifically, we calculate the potential albedo change (Δ albedo) for the building footprints if they were converted to cool roofs using a target albedo value of 0.55. This target value represents a conservative aged albedo achievable using existing and widely available cool roofing materials and was chosen to align with the minimum 3-year aged solar reflectance requirements for low-slope roofs as specified in numerous U.S. building codes tracked by the Cool Roof Rating Council²⁸. For buildings with existing albedos exceeding the target value, their albedo potential is defined as zero because the building is already considered to have a cool roof. To quantify the potential contribution of a cool roof to the overall area albedo, we multiply the albedo potential by the area of the roof, which we term the effective area. This provides a measure of solar radiation management potential in units of area corresponding to the 100% reflective material equivalent. By summing the effective areas and then normalizing by the area of the city boundary, we can estimate the increase in city-wide average albedo that would be achieved by adoption of cool roofs. Using this approach, we calculated the effective area of each building footprint then estimated the overall increase in city-wide albedo. We additionally estimated the reduction in air temperature associated with increased albedo by using the average albedo cooling effectiveness (ACE) of 0.4 °C per 0.10 albedo increase, the median value from the¹⁷ synthesis.

We evaluated three scenarios, using the intersection of the Boulder city limits and the building footprints covered by the NEON flightlines as the city boundary: Scenario I, all roofs converted to cool roofs; Scenario II, large buildings (building footprints above the 90th percentile of area) converted to cool roofs; and Scenario III, buildings with the lowest albedos (buildings above the 90th percentile of albedo delta) converted to cool roofs.

Scenario I estimates the maximum albedo change possible for Boulder through cool roof implementation, and while likely not realistic, gives a sense of what is technically possible. Scenario II prioritizes buildings with large footprint areas. These buildings are very likely to have flat slope roofs making cool roofs a cost effective option and the data shows that implementation will have a large impact (Supplementary Fig. 1). Scenario III prioritizes the buildings with the greatest albedo potentials, meaning they have the lowest current albedos. Implementing cool roofs on low albedo buildings would reduce energy costs and interior temperatures, and a policy that prioritizes the darkest buildings could improve heat equity.

While most urban areas have roof fractions around 20%, the NEON data was collected over a largely suburban area. To account for this, the boundary used to estimate the impact of the cool roof scenarios is the intersection of the city boundary and the overture buildings for which we have NEON albedos. The area of this boundary is 48.85 km² and contains 10,635 buildings with a total footprint area of 3.24 km², or approximately 6.6% of the land area. Estimates for city-wide albedo change are low because Boulder is not densely urbanized and the data only covers part of the urban area.

The greatest change in albedo and therefore the largest temperature reductions come from Scenario I (Fig. 5 and Table 2). The effective areas calculated from the three datasets vary because they estimate different building albedos, however, when normalized over the city area the results are not meaningfully different. The large buildings scenario implements cool roofs on buildings with areas greater than the 90th percentile (398 m²). Like Scenario I, the differences in effective area are insignificant when normalized by the city area. However, for Scenario III, which implements cool roofs based on low existing albedos (albedo potential greater than the 90th percentile value of 0.435, meaning that cool roof implementation would increase the albedo of a building by at least this amount), the differences are much greater. While NEON finds 6855 high potential buildings, the high-resolution inference identifies 20% of this number, and Sentinel-2 only identifies less than 3% of this number. Although the convolution calibration method reduces the overestimation of low albedos, the results do not fully correct for the positive bias of Sentinel-2. We hypothesize that this is primarily due to undercorrection of the Sentinel-2 surface reflectance for atmospheric scattering, as highlighted in the Discussion section.

Global cities cool roof scenarios

We extend these methods for estimating the solar radiation management potential to a selection of 12 global cities (Fig. 6). These cities were chosen to represent a range of climatic, geographic, morphological, and economic conditions. We acquired the Overture Maps building footprints that intersect the city geometries as well as the available Pléiades Neo imagery for the cities. We generate both the Sentinel-2 albedo estimates and the high-resolution albedo inference using the 100-day medians centered on the available Pléiades Neo dates and then calculate the median albedo of each building footprint for each image. Where there are multiple images (either Sentinel-2 or Pléiades Neo) for a building footprint we take the median value so that we have one estimate each for the two albedo datasets per building footprint. Additionally, we calculate the area of each building footprint. We follow the same methods for calculating the cool roof scenarios as for Boulder.

The albedo change associated with Scenario I is the greatest for every city and is strongly dependent on the building fraction (Fig. 6). However, in many cases Scenario II accounts for over half of the increase from Scenario I, showing that prioritizing large buildings could be a highly effective policy for a city to pursue. The results from Scenario III show that prioritizing buildings with high albedo deltas, while potentially having positive effects for equity, has a much lower impact on the city-wide albedo. However, it is important to note that our experiment in Boulder shows that both the Sentinel-2 and the high-resolution inference fail to identify many buildings in this scenario, with the ground truth albedo identifying five times as many buildings with high albedo deltas and estimating a higher albedo change from Scenario III than Scenario II. It is possible that the results for other cities follow a similar pattern. The high-resolution inferences and the Sentinel-2 estimates predict quite similar albedo increases for Scenarios I and II, however, they predict very different increases for Scenario III. We hypothesize that this is due to the inability of Sentinel-2 to accurately resolve the albedo of smaller buildings due to its coarse spatial resolution. Importantly, the greater the increase in city-wide albedo, the greater the potential temperature reduction-our results show that cities could achieve up to a 0.5 °C cooling effect with full-scale implementation of cool roofs.

Our convolution calibration method combines the spatial information of high resolution imagery with the spectral and radiometric information from Sentinel-2 to create highly accurate estimates of rooftop albedo at a 30 cm resolution. Validating albedo data is a considerable challenge-our use of the NEON data allows us to evaluate the ability of our albedo inferences to accurately estimate the albedo of urban roofs. Sentinel-2 albedo is the highest resolution albedo data currently available to the public, and we show that while it adequately measures the spatial variation in urban albedo at city scales and for large buildings, it overestimates the albedo of dark roofs and underestimates the albedo of light roofs due to the scale mismatch between urban features and the 10 meter spatial resolution. The high-resolution inference methods we present here reduce this tendency by broadening the distribution of albedo values (Fig. 4). By all measures, the high-resolution inferences do a better job of estimating the ground truth albedo of roofs as measured by the NEON hyperspectral instrument than Sentinel-2.

Both Figs. 1 and 4 illustrate that, in comparison to NEON, the inferred albedo underestimates high albedos and overestimates the low albedos. One possible explanation for this stems from the original Sentinel-2 surface reflectance data which is used for calibration. The Sentinel-2 surface reflectance product relies on atmospheric correction which attempts to remove the effects of the atmosphere (gases, aerosols, etc.) to provide a more accurate representation of the surface reflectance. Two main effects from the atmosphere are: (1) scattering of extraneous light from the bulk of the atmosphere into the imager; and (2) scattering within the direct line of sight through the atmosphere column that prevents light leaving the surface from reaching the imager. These effects are corrected to first order by subtracting an offset from the albedo and applying a positive gain to the albedo, respectively. New versions of SEN2COR are regularly released and deployed as new processing baselines; as methods improve, calibration accuracy is expected to improve commensurately.

This experiment suggests that for this particular set of conditions during the NEON flyover, the agreement with NEON would be improved by a gain and offset correction to the Sentinel-2 image, suggesting that SEN2COR may have applied too small of a correction of both extraneous atmospheric light scattering and atmospheric column scattering. While we could have fit such a correction to this experiment, we elected to show the results as is because they represent how the inference is applied in general where we do not have ground truth data. With this single experiment, it is not possible to know whether undercorrection for the scattering effects described occurs systematically or stochastically, but hyperspectral measurements like those from NEON offer a powerful tool for evaluation.

Using the high-resolution albedo inferences to estimate the median albedo of roofs in 12 global cities, we estimate that cities can see up to a 0.5 °C cooling effect from full scale implementation of cool roofs. We also illustrate the utility of robust albedo estimates associated with building footprints to estimate the effects of different cool roof policy scenarios. We show that prioritizing the largest buildings is a highly effective policy pathway for reducing heat by implementing cool roofs. Critically, the city-wide change in albedo varies across the cities based on both their current albedos and the distribution of building sizes. For instance, in some cities the largest buildings are also the brightest, yielding a lower delta from implementing cool roofs on large buildings.

While the high-resolution albedo and the Sentinel-2 albedo produce similar results when estimating the albedo increase from cool roof implementation on all buildings (Scenario I) and large buildings (Scenario II), they produce quite different results at finer spatial scales, like when prioritizing buildings based on their existing albedo (Scenario III). The 10-m resolution of Sentinel-2 is inadequate to measure the tails of the existing building albedo distribution—consequently it fails to identify many buildings that would see the most improvement with a cool roof (high albedo potential). While the high-resolution albedo may still miss the extremes of the tails, it provides a stronger capability for estimating the albedo of very dark or very light buildings.

Both the data validation and the global cities analysis highlight that there are multiple use cases for urban albedo data, one of which can be adequately filled by the widely available Sentinel-2 data and another which requires higher resolution data like we present here. When estimating the albedo increase of implementation from scenarios that do not involve prioritizing buildings based on their existing albedo-cool roofs on all buildings or cool roofs on large buildings–the high-resolution albedo and Sentinel-2 give very similar estimates (Fig. 6a). Indeed, when aggregating across many buildings the high-resolution albedo and the Sentinel-2 albedo yield similar estimates because errors cancel and the area of buildings as a fraction of the urban area is the same for both (Supplementary Fig. 4). We conclude that for a general research use case, the Sentinel-2 data does a good job of estimating the albedo change from cool roof implementation at the neighborhood scale and larger. However, for a use case that involves the prioritization of individual buildings, the high-resolution albedo and the Sentinel-2 albedo tell very different stories (Fig. 7). The relatively coarse spatial resolution of Sentinel-2 causes a regression to the mean, underestimating the albedo of existing high albedo surfaces, and overestimating the albedo of dark surfaces. For urban planning use cases that require understanding variation at the individual building scale, then, the high-resolution albedo is necessary to prioritize buildings based on their current albedo. While a full cost-benefit assessment is beyond the scope of this work, the high-resolution albedo data produced here provide the foundational input needed for such analyses. By resolving building-level variation in reflective potential, these data enable future studies to link achievable cooling benefits to the economic and health returns of targeted cool-roof investments.

Thanks to the growing constellations of multispectral imagers-and the methods developed and described here-it is now possible to measure calibrated, building-level albedo for the world’s urban areas and pursue climate-adaptation measures informed by these observations. Critically, this study provides a large-scale validation of rooftop albedo from overhead satellite radiometry using contemporaneous airborne hyperspectral imagery. By demonstrating urban albedo inference using two sun-synchronous satellites-one for radiometric calibration and one for resolution enhancement-we show a scalable pathway to global, high-resolution albedo mapping with high repeat cadence.

We find that high-resolution RGBN imagery, even when uncalibrated, can be leveraged to infer rooftop albedo with an error of approximately 4% when calibrated to contemporaneous Sentinel-2 imagery. The frequent revisit time of Sentinel-2 makes this calibration feasible in most locations, although persistent cloud or snow cover may pose challenges. When accurate digital surface models are available, further correction for roof pitch and solar geometry produces albedo estimates that more accurately represent intrinsic rooftop optical properties rather than snapshot imaging conditions. Among sources of high-resolution imagery, Pléiades Neo offers a key advantage by sharing a similar sun-synchronous orbit with Sentinel-2, minimizing illumination differences between images captured under near-nadir conditions. For other high-resolution imagery sources, additional steps to harmonize the differing shadows in the two images may be necessary to achieve similar precision.

Finally, while our validation against NEON airborne data demonstrated strong agreement with satellite-based estimates, inherent differences in acquisition timing and shadow geometry underscore the challenges of exact cross-sensor comparisons. Additionally, because low-flying aircraft cannot match the global coverage and instantaneous capture capabilities of satellites, contemporaneous broad-area comparison will always be limited. Our May 8, 2024 experiment further suggests that residual scattering effects may contribute to undercorrection in Sentinel-2 reflectance retrievals, an observation that may be of interest to the developers of the SEN2COR algorithm.

Together, these results showcase a practical, globally-scalable framework for high-precision urban albedo estimation and highlight the value of combining calibrated satellite radiometry, high-resolution imagery, and airborne reference data. This work establishes the foundations for building-level climate-adaptation planning at citywide scale and points to new opportunities for data-driven prioritization of cool-roof interventions worldwide.

There are several limitations to our work that must be acknowledged. This study was validated exclusively in Boulder, Colorado, which may limit the generalizability of the model performance. Boulder’s built environment includes unique characteristics such as specific roof geometries, materials, and building height restrictions that may not be representative of many other urban areas. In particular, the city’s stringent height ordinances result in a lack of the kinds of high-rise structures common in denser metropolitan contexts and the building density of our validation area is much lower (approximately 6%) than typical urban areas (approximately 20%), potentially reducing the impacts of shadowing. Additionally, the surface properties of Boulder’s natural environment, including coloration and bidirectional reflectance (BRDF), are likely to differ from those in other geographic regions. Ground truth validation in multiple cities representing a diversity of surface characteristics and urban morphologies would contribute greatly to the literature.

Our analysis assumes that rooftop surfaces are Lambertian reflectors, a simplification necessitated by data limitations. Even at 0.3 m resolution, the Pléiades Neo imagery does not provide sufficient angular or spectral information to reliably characterize non-Lambertian behavior or to segment rooftops by material type. We convert image albedo-albedo assuming a flat reflective scene-into surface albedo, with the tilt-corrected value of the albedo consistent with the Lambertian assumption. We acknowledge that real urban materials may exhibit directional reflectance properties that could introduce minor systematic biases. Future research could explore methods to relax the Lambertian assumption by estimating material-specific BRDFs and incorporating fine-scale surface geometry. Advances in ray-tracing techniques like inverse rendering may eventually allow retrieval of BRDF and DSM parameters at sub-meter scales. Such developments could enable more accurate representation of complex urban surface reflectance, extending beyond the Lambertian approximation used in this study.

In dense urban areas, shadows cast by neighboring buildings can substantially influence observed radiance, especially under low solar angles. Our current workflow does not explicitly correct for pixel-level shading or incorporate building height information, which may result in locally biased albedo estimates-lower in shadowed regions and higher on adjacent illuminated surfaces. However, because Pléiades Neo and Sentinel-2 acquire imagery at nearly the same local time, the illumination geometry-and therefore the pattern of shadows-is effectively consistent between the calibration and target images. Therefore, the resulting high-resolution albedo estimates are most representative of the urban reflectance at 10:30 AM. Future work, including improved building orthorectification, could enhance accuracy by integrating solar geometry, building height, and footprint data to better account for these shadowing effects. Despite these limitations, our methods produce consistent, decision-relevant albedo estimates at actionable spatial scales.

Calibration imagery

We use the atmospherically corrected Sentinel-2 Level 2A (Sentinel-2) surface reflectance dataset as the source of calibration²⁹. Sentinel-2 was chosen because of its relatively high spatial resolution of 10 meters, compared to other radiometrically calibrated satellite data products like LANDSAT (30 m) and MODIS (1 km). Sentinel-2 also has a regular revisit time (3–5 days) and consistent operation close to nadir, with redundancy achieved with its 2-satellite constellation. The Sentinel-2 dataset is publicly and globally available and is maintained as an asset in Earth Engine³⁰. Because we are inferring solar spectrum albedo, we model the Sentinel-2 bands which cover the solar spectrum: B2, B3, and B4, which are the blue (490 nm), green (560 nm) and red (665  nm) visible bands;, B8 (842 nm), which is the near infrared band; and B11 (1610 nm) and B12 (2190 nm), which are the shortwave infrared bands. As was originally done by Bonafoni and Sekertekin¹⁹ we checked the broad spectrum solar albedo computation against the network of SURFRAD stations located in the United States.

Using every Sentinel-2 image tile that overlaps the high-resolution input image over a 100-day window centered on the date of the high-resolution image acquisition, we create a temporal median mosaic from Sentinel-2 imagery masked using the Cloud Score Plus dataset³¹ and the Sentinel-2 snow probability mask. Since clouds are unlikely to repeatedly appear in the same place every time a satellite flies overhead, a sufficiently large number of images will ensure full coverage of the high-resolution input image, and the median statistic is effective at eliminating any unmasked cloud and cloud-shadow artifacts.

High resolution input imagery

The geospatial input data used to infer solar albedo is Pléiades Neo (Pléiades Neo 3 and Pléiades Neo 4: © Airbus DS 2024) satellite imagery. The resolution of the input data (30 cm) is substantially higher than the resolution of the Sentinel-2 calibration data (10 m). The input imagery was taken by tasked satellite flyovers. This input imagery is a top of atmosphere measurement, and it does not typically have as much solar spectrum coverage as the calibration data. Before use as input data, the input imagery is subjected to processing including orthorectification and pansharpening.

The Pléiades Neo satellite is tasked to target specific regions of interest during each flyover, and thus its view zenith angle has variable obliquity depending on the orbit of the satellite and the timing of the imaging request. This results in variable tilt and displacements of the tops of tall buildings ranging from zero (nadir views) to many meters (tall buildings with oblique views). We tested the method in Batchu et al.²⁴ wherein rooftops viewed with view zenith angle v on buildings of height h are displaced by -h tan() at the viewing azimuth angle. This correction did not improve the accuracy of the high-resolution albedo estimates in Boulder and so we chose not to apply it for the global cities analysis.

Coregistration

The European Space Agency reports Sentinel-2 pointing accuracy to be about 20 meters³² and our own visual inspection confirmed that each Sentinel-2 flyover image will have pixels that jitter with respect to other flyover images by about  ± 10 meters. To avoid loss of resolution when we make composite images from multiple flyovers, we use our high resolution input imagery as a reference to which we co-register³³ the Sentinel-2 image mosaic before conducting additional processing.

Convolution calibration

A mathematical method to increase the spatial resolution (downscale) of the calibrated Sentinel-2 image is to apply a circular kernel convolutional calibration. The high resolution Pléiades Neo B, G, R, N, N, and N bands are calibrated using the B2, B3, B4, B8, B11 and B12 bands of Sentinel-2 respectively. Because there is only one infrared band in the high resolution image, it is replicated threefold. In this method, a circular averaging kernel, k_c(x, y) of radius 10 meters (the size of a Sentinel-2 pixel) is convolved, band-for-band with both the high resolution image h(x, y) and the low resolution image l(x, y). The calibrated image c(x, y) is then computed from:

This quotient of convolutions guarantees that k_c(x, y) ⊗ l(x, y) = k_c(x, y) ⊗ c(x, y); what this expression means is that if one averages over a circle at any point in either the low resolution or high resolution calibrated images, the results will agree. It is worth noting that meeting this criterion does not guarantee that the details within any averaging circle will precisely agree with ground truth measurements; empirically however, this simple mathematical approach produces images that appear to be very plausible renditions of the detailed narrow band reflectances of objects in the image. This method does not rely on sampling and model training which offers several advantages over trained models. First, there is no bias between c(x, y) and l(x, y) (by definition), whereas in a trained model, because albedo examples tend to be heavily concentrated around values of 0.2, without careful training data stratification, learned models tend to regress to the mean and hence over estimate dark albedos and underestimate light albedos. Second, because each high resolution image h(x, y) contains the artifacts of changing atmospheric conditions unique to that image, there is no simple model that can be trained to calibrate all high resolution images without incorporating an atmospheric model. Models can, however, be trained on a per image basis, as described below for the calibration transfer method.

Calibration transfer

In this machine learning method, a gradient boosted tree model is made by targeting a stratified sample of 10 meter resolution Sentinel-2 derived albedo pixels using as model inputs 10 meter resolution upscaled samples of the B, G, R, N and panchromatic bands from the high resolution image together with near neighbor pixels and semantic surface category probabilities. The model is trained, tested and validated on the upscaled high resolution image samples, and then applied pixel-wise to the high resolution image. This model relies on the assumption of scale invariance, namely that covariates that predict albedo at 10 meter resolution also represent the albedo at 30 cm resolution.

Surface albedo of roof segments

In order to find the outline, pitch and azimuth of each roof we rely on the methods described by Batchu et al.²⁴. This produces a set of roof segment polygons for which we can compute the median albedo. Because we have information about the surface geometry of each roof segment we are able to improve the accuracy of the albedo estimates by calculating the surface albedo from the image albedo. Sentinel-2 (and by extension our high-resolution albedo inferences) assumes surfaces are both Lambertian and flat. Because of these assumptions, roof segments belonging to the same building with identical roofing materials but with different geometries may have different albedo estimates because of differences in illumination. We calculate the surface albedo, α_s, from image albedo, α_i, for each of the four albedo estimates using the following equation:

where D is the magnitude of the direct solar irradiance, (widehat{n}) is the surface normal, (widehat{s}) is the unit vector pointing toward the solar disc, I is the magnitude of the indirect solar irradiance, γ is the zenith angle of the surface, and (widehat{k}) is the vertical unit vector. We assume a ratio of direct to diffuse radiation of 0.85 to 0.15 and use the solar geometry corresponding to the high resolution imagery (rather than the Sentinel-2 imagery) for correcting the high-resolution inferences (Table 3). Both satellites are sun synchronous for 10:30 am local solar time flyovers. The only solar geometry differences will stem from the obliquity of the Pléiades Neo satellite.

Data analysis was conducted using a combination of Google Earth Engine, Python, and R.

Reporting summary

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