High Spatial Resolution Building Characteristics for the Global South: Insights from the Google Open Buildings Temporal Dataset (2016-2023)

Overview on google open buildings temporal

Google Open Buildings Temporal Dataset⁴² is a collection of data showing how building presence, counts, and heights have changed over time in many parts of the world. It provides annually snapshots from 2016 to 2023 at a spatial resolution of 50 cm. Focusing on the Global South, the dataset covers areas in Africa, South and Southeast Asia, Latin America, and the Caribbean (see Figure 1).

Figure 1. The coverage of google OBT on the global south (pink tiles). The building statistics for three countries (annotated) were validated by comparing them against population data at the subnational level.

Locations selected for technical validation are marked as red dots.

To identify buildings, the Google Research team employed a deep learning technique⁴² to analyse a large corpus of Sentinel-2 imagery. A key task of this approach was building segmentation, generating both confidence masks and building centroid locations. To overcome the limited availability of human-derived building labels, a teacher-student learning framework⁵² was employed.²⁷ The teacher model, which was a large and computationally exhaustive model, trained using high-resolution (50 cm) imagery and existing human labels, generated additional training data for the student model. This effectively expanded the training dataset. The student model then performed building super-resolution segmentation using stacks of 32 Sentinel-2 images (10-m resolution) acquired around June 30 each year (typically 16 images before and 16 after). Multi-temporal stacks were proven to enhance segmentation accuracy, e.g., increase of mean intersection over union, from 72% to 77%, when full-stacks were used in the training instead of single-timeframe images.⁴² However, cloud cover frequently resulted in incomplete image stacks, potentially degrading the model’s performance.

Building height prediction followed a similar teacher-student approach. Height labels were derived by calculating the difference between the Digital Surface Model and the Digital Terrain Model, effectively isolating above-ground object height.⁵³ This approach captured the heights of both natural and man-made features. Therefore, to isolate building heights, a corresponding building confidence layer, derived from the segmentation process, was used to mask out non-building objects. Validation against ground truth data from North America, Europe, and Japan yielded a mean absolute error of 1.5 m for building height predictions.⁴² This indicates a reasonable level of accuracy across diverse geographic contexts, although further regional validation may be warranted.

The overall process yielded three output layers, accessible via Google Earth Engine⁵⁴: building presence, fractional count, and building height, all at a native resolution of 50 cm. The building presence layer, with values ranging from 0 to 1, represents the model’s confidence that a given grid cell belongs to a building. The fractional count layer (values ranging from 0 to approximately 0.2) encodes building centroid information; the grid cell with the highest fractional count within a building’s footprint corresponds to its centroid. Integrating the fractional count over a given area provides an estimate of the number of buildings within that area. Finally, the building height layer represents the height of buildings above ground level, capped at 100 m.

Computed layers

We computed six 100-m resolution layers characterising buildings for each year covered by the OBT dataset: building count, total perimeter, total footprint area, total volume, variance of building heights, and mean distance to buildings. For each layer, we computed annual data from 2016 to 2023. From these characteristics, other indicators can be derived, e.g., mean building heights, mean number of storeys, and the average complexity index.³¹

As illustrated in Figure 2, we utilised three layers from the OBT dataset and perform several processes, including thresholding, arithmetic calculation, convolution, and aggregation. Table 1 summarises input used and output layers produced in this study. To focus on building characteristics, we applied a masking procedure, excluding grid cells with building presence values below a threshold T. We explored three thresholds: T = 0.3, 0.4, and 0.5, resulting in three sets of output layers. While the Google Research team suggests a threshold of T = 0.34, acknowledging that the optimal value may vary by region, our approach allows for a broader analysis. This initial masking step was essential to focus specifically on building characteristics.

Figure 2. Workflow from OBT layers to building characteristics at 100-m spatial resolution.

Building count was simply the integral of building fractional counts (frac) in 100-m grid cells. A multiplication factor of 40,000 was used to adjust the total count obtained using reduceResolution() function in GEE. This factor was associated with the pyramiding policy implemented, i.e., the multiplication factor was the number of original 50-cm grid cells inside the final 100-m cell. Almost similar to this process, the sum of grid cells with building presence (pres) above T multiplied by 10,000 yielded the total area covered by the building in m². Next, replacing the confidence mask with the masked building height layer (h) prior to the aggregation yielded total building volume in m³. For practical reasons, the total building volume layer was multiplied by 0.1. All above layers were stored as 16-bit unsigned integers rasters, which provide the optimal balance between data size and dynamic range for the extracted values.

Calculating the total building perimeter required a different approach. Typical building footprint extraction involves segmenting the image, then vectorizing the segmented images and applying morphological corrections to fit simplified polygons, resulting in vector-based building footprint data.⁵⁵ However, because we were aggregating building characteristics to a lower resolution, we skipped the vectorization step and directly calculated the perimeter from the edges from the building confidence mask. We used a Laplacian-8 kernel to perform edge detection on this mask. Among several kernels available for edge detection in Earth Engine (e.g., Prewitt, Roberts, Sobel), Laplacian-8 was selected due to its sensitivity to the changes in image intensity and its ability to detect edges in all directions equally.⁵⁶ Lastly, the sum of the detected edge grid cells then provided our total building perimeter value for each 100-m cell.

Variance of building heights in each grid cell was computed using the formula summarised in Table 1. While a direct calculation of variance from individual building heights was possible, it was computationally burdensome. To reduce this computational effort, we adopted an alternative calculation: variance equals to the average of the squared height minus the square of the average height. Critically, the averaging part utilised the building fractional count as a weighting factor. This technique correctly isolated the contribution of each building’s height, preventing the building’s spatial area from unduly influencing the variance calculation.

Building proximity was calculated in Google Earth Engine by applying a Euclidean distance transform with a 400-m kernel radius to a defined building confidence mask. This process generated a raster representing the distance from each grid cell to the nearest building edge. Applying a radius limit of 400-m enabled efficient computation without losing too much information on sparsely built areas. Subsequently, the resulting distance raster was aggregated from 50-cm original resolution to a 100-m resolution grid using a mean reducer, providing the mean distance to buildings for each grid cell.

Additional processing

Six layers were computed tile-by-tile using Earth Engine Python API. This approach facilitated processing large datasets. The resulting EPSG:4326 raster were downloaded for local processing, including creating global mosaic and clipping to country boundaries that conformed to the WorldPop Global Demographic Data Project’s master grid⁵⁷ for consistency.

General characteristics

Figure 3 depicts the spatial distribution of building characteristics derived from OBT for a 25 km² region encompassing Flor Amarillo, Carabobo State, Venezuela. The visualization highlights: (1) a densely settled area in the southern sector, characterised by a high concentration of buildings; (2) commercial zones in the northern and eastern sectors, distinguished by above-average total building volume; and (3) a central band displaying substantial height variance, indicative of apartment and commercial complexes surrounded by low buildings. This analysis provides insight into the diverse urban morphology of the region.

To understand more, we checked the distribution of the building characteristics from the 183 sampled locations and assess the relationships between different characteristics. The characteristics were derived using threshold of 0.4. For every pair of building characteristics, we computed Pearson’s correlation coefficient (r) as a statistical measure of correlation. Accordingly, the associated univariate and bivariate distributions are depicted in Figure 5.

Figure 5. Density plots (darker means denser) highlighting correlations between building characteristics in 2020 (0.5 threshold) from 183 sample locations.

Pearson’s r is indicated at the top-right corner of each panel. The bottom panels display a univariate distribution of the characteristics with Q₁, Q₂, and Q₃ pinned on top of it.

In general, total area is the metric with the highest correlation with other characteristics (|r| > 0.6, except with height variance). This result is consistent with a simple expectation that the increase in built-up total area usually aligns with the increase of building count, perimeter, and volume. In more dense regions where built-up total area is larger, the average inter-building distance is lower.

Analysis reveals moderate to strong correlations among building count, total perimeter, total volume, and inter-building distance. The relationship between total area and perimeter exhibits a quadratic trend, with increase of data scatter at higher total areas. Meanwhile, lack of data points in the upper-right quadrant is observed, indicating potential limitations. Closely spaced buildings lead to perimeter underestimation due to blending effects during extraction. When adjacent structures are identified as overlapping, their shared boundaries are lost, reducing the calculated perimeter. While geometric edge detection, such as the Hough transform (San & Turker 2010), could alleviate these blending effects, we intentionally omitted this step to focus on broader 100-m grid aggregations and avoid computational burden. Furthermore, inter-building distance displays discrete values, a direct result of the 400-m radius distance kernel used in calculations. Spatial aggregation further reduces the granularity of these distance values.

Building height variance exhibits a distinct behaviour, showing minimal correlation with most other building characteristics. A moderate positive correlation exists between height variance and total volume (r = 0.45), suggesting that areas with larger buildings tend to display greater height variation. Conversely, the correlation between height variance and mean building height (i.e., total volume divided by total area) is considerably weaker (r = 0.28), indicating that average building height has less influence on height variability.

Univariate distributions reveal typical characteristics across sampled locations: a median building density of 14 buildings per hectare and a median building size of 150 m². The median building height, estimated at 5.1 m, approximates a two-storey structure. Assuming random sampling, these values offer a generalised representation of average building conditions within the Global South.

Comparison with other datasets

This study assessed a dataset of 183 randomly selected locations across the Global South ( Figure 1). For each 5x5 km² area around the designated locations, we extracted building characteristics from our dataset and compared them with the following established datasets:

The building characteristics for 2020 were compared against Google Polygons, Microsoft and GHSL, whereas the 2019 building metrics were compared against WSF3D to maximise temporal alignment. Figure 6 illustrates building characteristics derived from the different datasets for sampled locations. We calculated Pearson’s correlation coefficient (r) and the normalised root mean square difference (nRMSD) to compare these datasets quantitatively. High dataset comparability was indicated by an r value approaching 1 and a minimised nRMSD. Tables 3 and 4 summarise the r and nRMSD computed in the current comparative analysis. The following discussion highlights on the results associated with T = 0.5 as the dataset with the best comparability with other datasets, but the scores relevant to the other thresholds were also computed.

Figure 6. Comparison between building characteristics from 183 sample locations derived from OBT (vertical axis) and other datasets (horizontal axis).

The density plots are scaled logarithmically to increase clarity.

Our dataset, especially the one produced using T = 0.5, exhibited the strongest agreement with the Google Polygons dataset (panel b, e, and i of Figure 6). Building count showed a high correlation (r = 0.88, nRMSD = 0.44), while building area demonstrated the strongest correlation overall (r = 0.90, nRMSD = 0.42). Although building perimeter derived from OBT correlated well with Google Polygons (r = 0.85), a systematic deviation was observed (nRMSD = 0.71). Specifically, our method tended to underestimate building perimeter compared to the Google Polygons dataset, particularly at higher values. In densely populated areas, imperfect edge detection sometimes resulted in the blending of some buildings, leading to an underestimation of total building perimeter. At the right-end of the distribution (i.e., total building perimeter of around 800 m), our method underestimated the total building perimeter of about 20% compared to Google Polygons dataset. While lowering the confidence threshold was explored as a potential solution, it did not improve accuracy. Although this adjustment reduced the distinction between neighbouring buildings, it also had the unintended consequence of further decreasing the estimated perimeter.

The high degree of agreement between the Google Polygons and Temporal datasets was anticipated, given their shared source imagery and similar (though not identical) deep learning methodologies. This concordance also suggests that the processing steps in the current study were performed effectively.

A comparison of our data products with those derived from the Microsoft dataset is summarised in Tables 3 and 4. The Microsoft dataset contains fewer buildings than OBT, which is reflected in our data showing a significantly larger total building area. This suggests that OBT may have a higher building detection rate. Notably, OBT has demonstrated high accuracy in building counts, achieving an R² coefficient of variation of 0.91 and a mean absolute error of 5.67 when evaluated at 300 × 300 m² tiles.

Our data products showed a weak correlation with both the GHSL and WSF3D datasets. Figure 5, panel (f ), reveals a saturation effect in the GHSL data, with building areas at 100-m grid cells rarely exceeding 4,000 m², while OBT estimates can reach 8,000 m² or more. The correlation between OBT and WSF3D building areas was particularly weak (r = 0.15). These discrepancies likely arise from differences in input imagery, computational methods, and resampling processes.

A complex relationship exists between GHSL and building volumes from the OBT ( Figure 5, panel (j)). In rural areas, where building density is low, the OBT dataset shows a larger total volume of buildings than the GHSL dataset. Conversely, the GHSL dataset records a higher total volume of buildings when considering a broader scale. Despite this, the building volumes from these two datasets maintain a moderate positive correlation (r > 0.5). On the other side, the correlation between OBT and WSF3D building volumes was considerably weaker, with a normalised root mean square difference exceeding 1.

Similar to the OBT case, Sentinel-2 imagery serves as a primary input for estimating total built-up surface in both GHSL and WSF. However, the production procedures for these datasets differ. GHSL¹⁹ generates built-up surfaces via symbolic machine learning, trained on an older Landsat-based GHSL dataset and combined with other sources such as the European Settlement Map,⁶⁰ Facebook High-Res Settlement Data, and Microsoft Building Footprints. Different sources are selected depending on the availability of the datasets in the training regions. These variations in training sets directly influence the machine learning outputs. Furthermore, textural filtering of small gaps (1–2 m) faces constraints due to the limited resolution of Sentinel-2 images, which degrades the model’s ability to infer built-up surfaces. To produce GHSL height and volume layers, a 30-m DEM is combined with the built-up surface layer to estimate average net building height. This process occurs at a 250-m resolution before resampling to a final 100-m resolution. Despite using Sentinel-2 inputs, these procedures result in smoother spatial variations in terms of building height and volume. Finally, GHSL is stored in a Mollweide projection such that reprojecting the data to the EPSG:4326 used in our study potentially introduces discrepancies.

In the case of WSF3D, a Sentinel-2-derived vegetation index identifies impervious areas²⁰ as the starting point for computing built-up fraction and area. Meanwhile, TDX-DEM data serves as the main input for building height estimation. Within identified impervious areas, height variations in the 12-m TDX-DEM indicate building edges. These values are then spatially aggregated to a 90-m grid to provide the average building height. Consequently, any inaccurate identification of the built-up area leads to errors in height estimation. Much like the GHSL case, the transformation of the WSF3D raster to our specific grid introduces additional divergence.

It is crucial to recognise that the validation process exclusively employs satellite-derived datasets, with no inclusion of authoritative ground-surveyed references.⁶¹ Consequently, the presented results offer a relative validation against other datasets with similar features, and their interpretation should consider the potential for systematic biases and uncertainties inherent in remote sensing data. This mirrors the situation in other studies, such as a similar cross-comparison of building footprints covering Africa where ground-truth data was also unavailable.²⁴ Therefore, it is difficult to say which dataset is superior. Furthermore, different datasets should not be used interchangeably.

Variance of building heights

Publicly available data concerning variance of building heights, especially for Global South, could not be found. The Global Building Morphology Indicators published by Bijecky and Chow³¹ captured several cities, including Chennai (India), Nairobi (Kenya), and Kampala (Uganda) in the Global South. In this dataset, building indicators were computed from a building footprint dataset like the one from OSM and aggregated at administrative unit and grid-cell levels. Unfortunately, this dataset does not contain building height information for like-for-like comparison with our results.

Alternatively, to validate our method of computing building height variance, we combined the building height layer from OBT dataset and building footprints from Open Buildings Polygons to estimate the height for each building listed in the latter dataset. By doing this, we extended the vector dataset to include building height information so that vector-based aggregation could be performed as an alternative (see Figure 7 for illustration). For each 100-m grid cell, we selected buildings inside (or intersecting) the grid and computed the building height variance. This process was done for some sample locations in Honduras, Philippines, and Rwanda.

Figure 7. Illustration of how we computed variance of building heights using only OBT dataset (a, b, c) and in combination with Open Buildings Polygons vector dataset (d, e, f ).

By comparing building height variances computed using Equation 1 and the ones from the combined dataset, we demonstrated the validity of our method. Figure 8 displays how well-correlated the two were. The obtained Pearson’s correlation coefficient was 0.79 while the root mean square deviation was 45.76.

Figure 8. Building height variances at 183 sample locations computed using two different approaches.

We used 2020 data and 0.5 threshold for this analysis. Lighter colour represents more data points.

Correlation with population count estimates

In summary, a strong positive correlation (r > 0.8) was observed between building count and population estimate ( Figure 9). The relationships between three building characteristics and the population estimate at administrative unit level could be represented using exponential function (P = aX^b, with X represents building characteristic) pretty well. For most of the cases, the exponents were around 0.90, while the scaling factors differ by country.

Figure 9. Correlation between building count, total area, and volume with total population count at certain administrative units.

Exponential relation is displayed on the top-right corner of each panel, together with associated Pearson’s correlation coefficient (r). Dashed line represents one-to-one relationship between variables.

Population count at administrative unit level was affected by many factors and building characteristics like building count was only one of the factors. But, from these simplified empirical relationships we could learn some interesting points. The exponents were less than 1, implying sub-linear relationship between the variables. This implies that while population increases as building characteristics hike, the rate of population increase slows down. This could be due to factors like population density limits, differing land use policies, or increasing non-residential buildings in denser regions.

Meanwhile, the country-specific scaling factor (a) acts as a baseline population density. A country with a high value for a will have a greater population for the same level of building characteristics, reflecting cultural, economic, or policy factors that lead to higher population density, such as a preference for large households, smaller per-capita living space, or a greater proportion of residential vs. commercial buildings.

Different thresholds

Thresholding was a critical step in our process and was based on the building identification confidence scores provided by the dataset. These scores were uncalibrated,⁴² meaning that a value of 0.8 did not represent a literal 80% probability of a building's existence. Instead, the scores served primarily as a tool for relative ranking and thresholding. The reliability of these values was influenced by various environmental and technical variables, including cloud cover and imagery misalignment, and showed a systematic sensitivity to roof color. While Sirko et al.⁴² identified 0.35 as the optimal threshold for mean Intersection over Union (mIoU) and 0.42 for total built-up area, we recommend a more conservative threshold of 0.5. This choice accounts for the lack of site-specific ground truth data and aligns with the comparative analysis presented previously.

The validation described above was done using characteristics computed using T = 0.45. How do different thresholds affect the computed characteristics? Figure 10 visually summarises the answers. Higher thresholds resulted in lower building counts, areas, and volumes. This is because stricter thresholds tend to truncate building edges and exclude some fractional building counts. Conversely, a slight increase in building perimeter was observed with higher thresholds. This seemingly counterintuitive result likely stems from a balancing effect: while higher thresholds contract individual building sizes, they also reduce the blending of closely spaced buildings, thus potentially increasing the overall measured perimeter.

Figure 10. Comparison between building characteristics in sampled locations year 2020, derived using different thresholds.

One-to-one relationship is represented by dashed line, while the best linear fit is marked with solid lines.

Spatial consistency

In this study, we aggregated fine-resolution OBT data to a coarser 100-m resolution via spatial aggregation and additional computations. This final scale was selected to ensure compatibility with the WorldPop Global Demographic Data Project, facilitating its use in various applications at regional and global scopes. Because data aggregation can significantly influence analytical outputs and subsequent decision-making, it is critical to quantify the scale effects during spatial integration. For this purpose, we utilised GEE and extracted six building characteristics in Flor Amarillo, Carabobo State, Colombia (see Figure 3), across different spatial scales, from 5 m to 100 m. We compared pixel values at 100 m to assess the sensitivity of the result to the extraction scale and computed the Root Mean Square Difference (RMSD).

The assessment demonstrates high compatibility between the 100 m pixel values and those derived from finer resolutions. As illustrated in Figure 11, the median RMSD for each building characteristic remains within acceptable thresholds, validating the consistency of the upscaling process. Notably, the median differences are relatively small, being roughly equivalent to the variance of a single building. A systematic decrease in RMSD is indeed observed for perimeter. Extraction at smaller scales tends to overestimate the total building perimeter due to excessive edge detection. However, this discrepancy is not a cause for concern. The 100-m total perimeter estimates produced in this study are validated against metrics derived from high-fidelity building footprint polygon data (see Figure 6).

Figure 11. Root Mean Square Difference between building characteristics derived at 100-m and smaller resolutions.

Temporal consistency

The temporal consistency of individual building detections presents a notable challenge. Over successive time periods, the same physical structure may be assigned varying confidence scores or height estimations. Consequently, aggregated metrics such as building counts within a defined area can also exhibit instability. Several factors contribute to this temporal variability, including: (i) the presence of cloud cover, which can obscure building features and impact detection accuracy; (ii) subtle misalignments between the input satellite imagery acquired at different times, potentially leading to inconsistencies in feature extraction; and (iii) a reduced availability of Sentinel-2 imagery, particularly during the 2016-2017 period, which limits the temporal density of observations and can affect the robustness of time-series analyses.

To evaluate temporal stability of the dataset produced, we extracted building characteristics (0.4 confidence threshold) from 183 sample locations mentioned before and examined the observed fluctuations.

We defined x_t as building characteristics (e.g., building count) at time t, and its relative change as Δ = (x_t/x _t-1) − 1. Changes where |Δ| was below a defined tolerance were considered insignificant. To identify temporal fluctuation over an 8-year span (2016-2023), we examined the signs of the minimum and maximum Δ values calculated for each 100-m grid cell. Fluctuation was determined if sign (min(Δ)) ! = sign (max(Δ)), signifying that both positive and negative relative changes occurred, excluding any changes deemed insignificant (|Δ| < 10%).

Following the principle outlined previously, we quantified temporal stability by calculating the percentage of grid cells exhibiting fluctuation for a given tolerance level. As illustrated in Figure 12, a lower tolerance threshold results in a higher proportion of grid cells identified as fluctuating. Specifically, for building area, volume, perimeter, and height variance (varh), approximately half of the sampled grid cells show a temporal fluctuation of 5% or greater. At this same level of tolerance, around 44% of grid cells fluctuate in building count, while only about 17% display fluctuation in average building distance. The steepness of the curves in Figure 10 also indicates the temporal stability of these aggregated building characteristics. Notably, even at a higher tolerance of 20%, roughly 30% of sampled grid cells still experience fluctuation in building count, total area, and total perimeter. However, building characteristics related to height (i.e., total volume and height variance) demonstrate lower temporal stability. For height variance, the percentage of fluctuating cells remains above 40% even at a relatively high tolerance of 50%.

Figure 12. Percentage of grid cells in 183 sampled locations experiencing fluctuation at different levels of tolerance.Statistical summary of temporal fluctuation in the dataset (sampled from 183 selected locations) assuming 20% tolerance.

Given the inherent temporal fluctuations observed even in the aggregated dataset, further processing is necessary before utilizing this temporal data. To mitigate these inconsistencies and achieve a smoother representation of building characteristic evolution over time, we explored fitting polynomial functions to the temporal data. The outcomes of this smoothing approach are illustrated in Figure 13. In this figure, the percentage of fluctuating cells is based on 20% tolerance level while the nRMSD is based on the deviation between original data and the smoothed one. Fitting the first order polynomial (linear function) results in a significant reduction in fluctuating grid cells, but it introduces a large nRMSD. As normally expected, fitting higher order polynomials reduces the deviation between model and data, but the reduction of fluctuating cells needs to be sacrificed. Fitting second order polynomial to the data seems to be an optimal strategy to improve temporal stability while appreciating the original data.

Figure 13. Statistical summary of temporal fluctuation in the dataset (sampled from 183 selected locations) assuming 20% tolerance.

Percentages of grid cells with fluctuations in original data and the smoothed ones (using first, second, and third order polynomials) are depicted in orange bars. Emerging normalised root mean square differences ( nRMSD) are portrayed in red bars.

Uncertainty analysis

Temporal fluctuations in the dataset can also be used to quantify stochastic errors embedded in the input data set as the basis for uncertainty estimates at 100-m aggregates. For this purpose, we selected buildings from 183 sampled locations and extracted time series data representing the average building fractional count, height, and presence within a 100-meter circular buffer zone around each location, from 2016 to 2023. Values representing the building edge and building mask (where presence exceeds T = 0.5) were also extracted. The selection criteria required buildings to be present in 2016, as verified by both the OBT data and the World Settlement Footprint.²⁰

Assuming these locations did not undergo significant destruction or redevelopment during the study period, we expect their fractional count, height, and presence values to remain constant. Multi-temporal data can be regarded as the product of repeated measurements. Consequently, any observed temporal variations in the data are attributed to uncertainties within the OBT dataset.^62,63 Standard deviation of the values (fractional count, etc.) at pixel level were computed and then spatially aggregated over the 100-m circular buffer zones to get more statistically meaningful uncertainty estimates.

Figure 14 summarises the distributions of standard deviations representing temporal variations for the four metrics, including building mask generated through thresholding of building presence layer using T = 0.5. We utilise the median of these distributions as the quoted uncertainties at pixel level. The uncertainty in building height is 1.53 m, which agrees with the mean absolute error of 1.50 m reported by the OBT producer.⁴² Estimated uncertainties for other parameters are summarised in Table 5.

Figure 14. Temporal variations, parameterised as standard deviation, of building fractional count, height, mask, and perimeter extracted from 183 sample locations. Cumulative distribution functions (CDF) are indicated by black lines. Median values (Q₂) are indicated as well.

During the calculation process to obtain building characteristics ​​at 100-m resolution, uncertainty at the pixel level surely propagates. Assuming that the uncertainty of each variable are independent, uncertainty propagation can be calculated using the variance formula. Except for the mean distance to buildings, where the values are highly dependent to the composition and configurations of buildings, the uncertainty estimates on a 100-m grid are summarised in Table 5. A factor of $\sqrt{N}$ , with the number of original 0.5-m pixels $N=\mathrm{40,000}$ was accounted when propagating uncertainties through spatial summation and then multiplied by other factor representing the contribution of each pixel, i.e., 0.25 m² for area, 0.25 m³ for volume, and 0.50 m for perimeter. Considering the complexity of the formula used to compute height variance, its uncertainty was estimated using bootstrap technique based on the uncertainties in height and fractional count. The figures summarised in Table 5 are the first order estimates of the uncertainties. The actual values may vary over diverse geographical settings.

Usage notes

Our post-processing yielded a 100-m resolution raster of building characteristics, defined by six fundamental parameters. As indicated in literature,³¹ aggregating building metrics or morphology at this scale offers critical insights into urban form. These metrics, including building count and total volume, facilitate a deeper understanding of both urban and rural development patterns. While the 2016–2023 OBT dataset enables the analysis of short-term residential trends, researchers must account for fluctuations inherent in multi-temporal satellite imagery. Subtle geometric misalignments between multi-temporal satellite acquisitions can introduce inconsistencies during feature extraction. Persistent cloud cover frequently results in incomplete image stacks, obscuring morphological features and degrading the accuracy of the underlying deep learning detection models. These fluctuations propagate from individual building footprints into the aggregated 100-m grid cells. Consequently, aggregated metrics can exhibit significant inter-annual fluctuations. In its raw form, the dataset requires post-processing, such as polynomial fitting or smoothing, to ensure reliability for time-series applications.

While our calculated building characteristics demonstrate high precision in stable, well-developed areas (Table 5), significant fluctuations emerge in regions undergoing rapid growth or decline. This is particularly critical in the Global South, where small cities and rural settlements serve as hotspots for settlement expansion.⁶⁴ To accurately quantify regional development rates using this dataset, data fluctuations must be mitigated through robust smoothing techniques. Consequently, further investigation into optimal smoothing methodologies and their subsequent impact on results is warranted. Alternatively, computing a multi-year average prior to estimating growth would also be useful. Creation of spatio-temporal matrix⁴⁴ becomes another option to make use of the temporal data we produced, while maintaining monotonic growth. Though this matrix is not directly applicable to some metrics like total area and volume, spatio-temporal matrix provides information about when a particular grid cell started to be inhabited.

Tracking building heights, volumes, and 3D morphology provides deeper insights into building space inequality and urban microclimate. It is important in our endeavour to monitor progress toward the UN's Sustainable Development Goal 11 in relation to sustainable cities and communities establishment. Building volume per capita serves as a robust indicator of socio-economic well-being in certain contexts,²⁶ while spatial inequality across populations can highlight regions requiring intervention to alleviate overcrowding and improve housing. Our analysis of the relationship between population density and different building characteristics reveals that, for sampled countries in the Global South, building count serves as a more reliable proxy for population distribution than either total area or volume.

Urban microclimate modeling frequently identifies building height distribution as a critical factor. For instance, building volume density accounts for approximately 60% of the variance in local air temperatures in Beijing Metropolitan Region.⁴⁵ This metric can be directly derived from our data product by dividing total building volume by 100 m² grid area. Polycentricity, which defines the presence of multiple urban activity centers, is associated with a reduction in the Urban Heat Island (UHI) effect.^65,66 Identification of such centers can be performed using any of the building characteristic available in our dataset. Beyond thermal impacts, variations in building height also influence the accumulation of air pollutants, specifically the concentration of particulate matter.⁴⁵

Lastly, remote-sensing-based building identifications often suffer from spatial uncertainties and building adhesion, which results in multiple closely-spaced buildings being represented as a single block. This pixel-to-pixel classification disregards the true morphology and regularized boundaries of individual structures, resulting in generalized, blob-like shapes. As a result, this blending inherently causes a systematic underestimation of the true building perimeter.⁶⁷ The building perimeter is a required mathematical input for calculating the façade-to-site ratio (FSR), another key of urban morphology parameters. Consequently, underestimating the perimeter of dense buildings leads to artificially low FSR values, miscalculation of aerodynamic roughness, and slight underestimate the severity of UHI effects.⁴⁸

Despite its promising potential across urban and environmental disciplines, the practical versatility of this dataset requires empirical testing in varied geographical contexts. By making this data product openly available, we encourage its application in independent studies, ensuring that it undergoes the rigorous scrutiny necessary to establish its reliability.