Mapping the world population one building at a time
High resolution datasets of population density which accurately map sparsely-distributed human populations do not exist at a global scale. Typically, population data is obtained using censuses and statistical modeling. More recently, methods using remotely-sensed data have emerged, capable of effectively identifying urbanized areas. Obtaining high accuracy in estimation of population distribution in rural areas remains a very challenging task due to the simultaneous requirements of sufficient sensitivity and resolution to detect very sparse populations through remote sensing as well as reliable performance at a global scale. Here, we present a computer vision method based on machine learning to create population maps from satellite imagery at a global scale, with a spatial sensitivity corresponding to individual buildings and suitable for global deployment. By combining this settlement data with census data, we create population maps with meter resolution for countries. We validate our method, and find that the building identification has an average precision and recall of and , respectively and that the population estimates have a standard error of a factor or less. Based on our data, we analyze percent of the world population, and show that percent lives within km of the nearest urban cluster. The resulting high-resolution population datasets have applications in infrastructure planning, vaccination campaign planning, disaster response efforts and risk analysis such as high accuracy flood risk analysis.
Accurate information on global population distribution is crucial to many disciplines. A population and housing census is the traditional tool for deriving small-area detailed statistics on population and its spatial distribution. However, censuses are time-consuming, and the spatial resolution is naturally set by the census enumeration areas (EA), which lack fine-grained information about the aggregation of population. The sizes of the EAs vary by many orders of magnitude from country to country, ranging from hundreds of square meters in urban areas to tens of thousands of square kilometers in low population areas, resulting in an average spatial resolution of a census unit of 33 km at a global scale. Recently, multiple higher resolution maps of human-made built up areas have emerged [15, 16], most notably the Global Human Settlement Layer (GHSL)[3, 4], the Global Urban Footprint (GUF)[5, 6], the WorldPop project, Landscan and Missing Maps project. However, none provide a scalable solution with high accuracy in rural areas.
Over the past decade high-resolution (sub-meter) satellite imagery has become widely available, enabling the global collection of recent and cloud-free earth imagery. Additionally, in the past years, the surge in research on computer vision and in particular convolutional neural networks (CNN) have enabled bulk processing of imagery in a rapid manner [10, 11, 12]. The combination of these methods enables the global analysis of high-resolution imagery as a promising method for detecting individual buildings; combining building estimates with available census data to produce updated and higher resolution population maps; and offering alternative, state-of-the-art population estimates in the absence of census data. Various approaches using machine learning have been demonstrated on small areas [13, 14], yet a method which allows global mapping has remained elusive.
Here, we present a method to generate high-resolution population density maps at a global scale (see figure 1) (referred to as High Resolution Settlement Layer (HRSL)). By using various CNN architectures, we identify human-made buildings in high-resolution satellite imagery, for a wide range of terrains, seasons and climates, even under poor image-quality conditions. Under the assumption that buildings act as a proxy for where people live, we obtain population estimates on a country-wide level, with 1x1 arcsecond resolution ( at the equator) and sensitivity to individual buildings, enabling accurate studies of population aggregation in rural areas.
To enable global analysis we develop a building detection model that is country-agnostic, fast, and works with easily obtainable binary labeled data. Our pipeline consists of several steps; we extract 64x64 pixel images (’patches’) around detected straight lines using a conventional edge detector, which reduces the amount of data for classification by times. A portion of those candidates are sampled across all countries and labeled as training and evaluation data for the CNNs (see Supplemental Information (SI)). The computer vision models are trained on a single machine with four GPUs, whereas the classification runs on Facebook’s infrastructure on a CPU cluster. This approach allows us to process in hours.
We use three different types of CNN (see figure 2): a classification model based on the SegNet  architecture in order to perform binary classification (building/no-building); a feedback neural network (FeedbackNet) performing weakly-supervised segmentation of the satellite images enabling us to obtain building footprints , and a denoising network which is capable of improving the quality of the source data by removing high-frequency noise from the satellite imagery (see SI). For all models the tradeoff between good performance at a global scale (larger networks) and the run time of the networks to perform global analysis (smaller networks) determined the limit in performance of the models.
The encoder-decoder style SegNet, is customized to perform the classification at the level of a patch. The encoder (a convolutional sub-network) is used to extract abstract information about the input, and the decoder (a deconvolutional sub-network) is trained to upsample the output of the encoder into a spatially meaningful probability map representing the possibility of house existence in the input. The probabilities generated by the decoder are averaged over all spatial locations within the patch to derive the final classification. This yields high accuracy and a reduced false positive rate on a global scale compared to other methods we explored, however, at the expense of loss of spatial details such as the boundaries and shapes of individual buildings.
To recover the building shapes in addition to classification we use a separate CNN to perform image semantic segmentation [20, 21]. Traditional methods, such as region proposal  based CNNs , are not able to handle small-sized foreground objects. Moreover, these methods require a large amount of pixel-wise labeled training samples following a supervised learning setting. Obtaining such a large volume of labeled data is currently intractable to achieve at a global scale because of it being time-consuming to acquire, as well as it having a large possibility to accumulate errors in supervision. To facilitate a generalized and scalable model, we employ the weakly-supervised learning that takes the abundant and easy-to-get image level categorical supervision (binary labeling) into training, and perform pixel-level prediction during deployment . The methodology is motivated by the feedback mechanism in human cognition  and recent advances of computational models in Feedback Neural Networks , which deactivates the non-relevant neurons within hidden layers of neural networks and achieve pixel-wise semantic segmentation. Both models are trained on binary labeled (building/no building) patches, randomly sampled from all countries and seasons, covering both rural areas and urban areas.
We have ran our model on 18 countries across 3 continents, spanning 29% of the world’s population. For nearly all countries we analyze over 95% of the landmass, of which contains a building at a arcsecond resolution. The coverage is limited by the availability of cloud-free, high-resolution satellite imagery (see SI) and nosignificant differences in performance between countries is observed, unless the source imagery is of poor quality.
A well characterized accuracy of the dataset is crucial for its further use. We consider the accuracy of the building identification and the accuracy of the population redistribution over buildings separately, with a focus on the former, which is the main topic of our work; the latter will be addressed briefly. Since a global ground truth dataset is not available four independent analyses are performed, each testing different potential errors in our dataset. Statistical errors and systematic errors are treated separately, the former having random origin and the latter originate e.g. from repetitive errors, such as large rocks, boats or mountain ridges (see Figure 3) being falsely classified as buildings. Systematic errors are more challenging to quantify but are of particular concern since they can result in clusters of false positives potentially misinterpreted as a settlement. To quantify statistical errors, we first study the accuracy of the computer vision model by comparing our classification results with a reference (human labeled) dataset. Averaged over 18 countries the SegNet achieves a precision and recall of , ( are true positive, false positive and false negative counts respectively) on a global dataset where only of the randomly sampled patches contains a building. This analysis tests the performance of our model, however, it doesn’t capture errors due to clouds, missed straight edge detection or missing source imagery.
To address this, a second analysis is performed comparing the identified buildings to the GPS coordinates of the households interviewed for the Malawi Third Integrated Household Survey (IHS3). This dataset is obtained by interviewing a nationally-representative sample of households, and thus, is independent of artifacts of remote sensing. The incidence of finding a household is measured within a distance of at most 100m from the nearest populated pixel in our building dataset, thereby accounting for the limited accuracy of the survey location data (see SI). At the national-level, percent of the IHS3 households coincide with a populated pixel in our building dataset.
In order to quantify the systematic errors we compare our data against two datasets (GUF and GHSL) generated using independent data sources and methodologies. The three datasets are compared at the 1 arc-second resolution of the binary classification to create all possible combinations of absence or presence of a settled area (see figure 3c). The areas where our classification disagrees with both the GUF and GHSL classifications represented as contiguous vectors are visually inspected using the satellite imagery to identify false positive and negative units. Country totals of agreement and disagreement as a proportion of country area are shown in Table 1 (see SI for additional countries). Four test cases of Malawi, Ghana, Vietnam and Haiti are presented since they cover a wide range of geographies and image quality.
We find that the visually confirmed systematic errors contribute to less than 1% of the total landmass and only a few percent of the country built up area. Additionally, in general areas of systematic false negatives occur more often than areas of systematic false positives. In order to assess the impact of larger clusters of disagreement which could yield more severe systematic areas we consider the largest areas of disagreement. These errors contribute to less than of the built up area for all 4 countries. The GUF and GHSL often miss buildings which are discovered in ours.
Finally, we perform an analysis comparing the three datasets (HRSL, GUF, GHSL) against a single region near Blantyre, Malawi which has been completely manually labeled, at a resolution of 2x2 arcsecond to allow for misalignment between the datasets (see figure 3c). The northern urban area (above the upper dotted line) is captured fairly well in all three datasets expressed in the recall values of for the GHSL/GUF and HRSL respectively. However, for the southern rural area (below the lower dotted line) the recall values are for the GHSL/GUF and HRSL respectively, demonstrating the superior accuracy of our dataset in rural areas. In conclusion, all errors are estimated to be a few percent or less, indicating a high degree of confidence in the building dataset. Additionally, our results are superior over existing datasets in rural areas.
After assessing the accuracy of the buildings, we turn to the population redistribution over the buildings. Population estimates are obtained using a minimally modeled approach of proportional allocation [26, 27]. We take two approaches, first, the population is distributed equally to settled areas in the binary classification that fall inside census units and second, the population is distributed proportionally to the fraction of built up area within a 1 arcsecond gridcell (see SI for details). The population allocations of the second method are expected to be more accurate, however, since currently no method exists to validate the population data globally at sub-arcsecond lengthscales we limit our analysis to the first method. The largest uncertainty of the population estimates originates from census data obtained at a too coarse administrative unit. The impact is estimated by comparing the population estimates performed on a coarse administrative unit to known population counts of finer administrative units (see SI). The error (standard deviation) in population estimate is a factor of 2 or less depending on the available country specific census data. Finally, the errors are slightly larger in urban areas (2.5 or less) compared to rural areas. Our data is well suited for a more sophisticated population allocation, which is an active area of research, however, this is beyond the scope of this work.
Our results represent a large improvement over the state-of-the art in globally robust building identification via satellite imagery, and thus population estimation, particularly for rural areas. As mentioned in the introduction, population estimates are a crucial interdisciplinary concern, and a cursory evaluation of our data hints at the nature of insights to come. For example, the improved rural accuracy enables the study the population distribution with respect to the nearest urbanized center. Definitions of urban versus rural vary widely, and are an active research area. Using the definition of Ref.  for urban clusters; clusters with a population density of at least and a minimum population of we calculate the distribution of the rural population as a function of the distance to the nearest urban gridcell. Figure 4 shows the distribution of the countries we have analyzed, and shows that of this population lives within of the nearest urban cluster. This proximity of rural population to urban clusters provides guidelines for telecommunication infrastructure development.
In conclusion, our model achieves high accuracy and is applicable at a global scale. The errors in building identification are estimated to be at the few percent level on a nation wide scale, while maintaining an unprecedented sensitivity to individual buildings. The uncertainty in population estimates originates from the underlying census data and the method of proportional allocation which is a topic of future research.
We thank D. Liu and T. Huang for insightful discussions on the denoising and feedback networks, and C. Deuskar and B. Stewart for insightful discussions on urbanization. All satellite imagery ©2016 DigitalGlobe. Funding for World Bank affiliates has been provided by The World Bank Living Standards Measurement Study ? Integrated Surveys on Agriculture (LSMS-ISA) project, and The World Bank Strategic Research Program Projects: (i) Use of Satellite Data and CDRs for Census-Independent Spatial Distribution of Populations and (ii) Survey Imputation for Improved Poverty and Shared Prosperity Diagnostics. Funding for CIESIN affiliates has been provided by Facebook. The boundary and census data collection were developed with funding from the National Aeronautics and Space Administration under Contract NNG13HQ04C for the Socioeconomic Data and Applications Distributed Active Archive Center (DAAC). Funding for Facebook affiliates was provided by Facebook, Inc.
References and Notes
-  United Nations Department of Economic and Social Affairs (2015)
-  Balk, D.L., U. Deichmann, G. Yetman, F. Pozzi, S.I. Hay, and A. Nelson , Advances in Parasitology (2006) 62 p. 119-156.
-  M. Pesaresi, et al. (2015) European Commission, Joint Research Centre (JRC) [Dataset] PID: http://data.europa.eu/89h/jrc-ghsl-ghs_built_ldsmt_globe_r2015b
-  M. Pesaresi, et al., Publications Office of the European Union UR 27741 EN (2016).
-  Esch, T., et al., ISPRS Journal of Photogrammetry and Remote Sensing (2017) 134 pp. 30-42. https://doi.org/10.1016/j.isprsjprs.2017.10.012
-  Esch, T., et al., IEEE Transactions on Geoscience and Remote Sensing, (2011) 49, pp. 1911-1925. https://doi.org/10.1109/TGRS.2010.2091644.
-  A. J. Tatem, Sci. Data 4, 170004 (2017).
-  B. Bhaduri, E. Bright, P. Coleman, M. L. Urban, GeoJournal 69, pp. 103 (2007).
-  Red Cross, Missing Maps project, http://download.geofabrik.de/africa/malawi-latest.osm.pbf. [Online; accessed 1-Aug-2017].
-  A. Krizhevsky, I. Sutskever, G. E. Hinton, Advances in neural information processing systems (2012), pp. 1097–1105.
-  C. Szegedy, et al., Proceedings of the IEEE conference on computer vision and pattern recognition (2015), pp. 1–9.
-  K. He, X. Zhang, S. Ren, J. Sun, Proceedings of the IEEE conference on computer vision and pattern recognition (2016), pp. 770–778.
-  F. Pacifici, M. Chini, W. J. Emery, Remote Sensing of Environment (2009) 113, pp. 1276
-  J. Yuan, (2016) arXiv:1602.06564
-  D. Potere, A. Schneider, S. Angel and D. L. Civco, International Journal of Remote Sensing (2009) 30, 24 pp. 6531-6558
-  P. Gamba, M. Herold, CRC Press (2009)
-  Y. Ban, P. Gong, C. Giri, ISPRS J. Photogramm. Rem. Sens. (2015) 103, pp. 1-6.
-  V. Badrinarayanan, A. Kendall, R. Cipolla, (2015) arXiv:1511.00561 (2015).
-  X. Liu, A. Zhang, T. Tiecke, A. Gros, T. S. Huang, (2016) arXiv:1612.02766 (2016).
-  J. Long, E. Shelhamer, T. Darrell, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2015), pp. 3431–3440.
-  L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, (2014) arXiv:1412.7062
-  J. R. Uijlings, K. E. Van De Sande, T. Gevers, A. W. Smeulders, International journal of computer vision (2013) 104, pp. 154
-  S. Ren, K. He, R. Girshick, J. Sun, Advances in neural information processing systems (2015), pp. 91–99.
-  R. Desimone, J. Duncan, Annual review of neuroscience 18, 193 (1995).
-  C. Cao, et al., Proceedings of the IEEE International Conference on Computer Vision (2015), pp. 2956–2964.
-  A. De Sherbinin, S. Adamo, United Nations Expert Group Meeting on Strengthening the Demographic Evidence Base for the Post-2015 Development Agenda. New York, UNDESA. (2015).
-  E. Doxsey-Whitfield, et al., Papers in Applied Geography 1, 226 (2015).
-  By combining satellite imagery with census data, there is no association with any information that would lead to re-identification of individuals, and the granularity of the data is not enough for individual identification. Additionally, no Facebook user data has been used in this project.
-  Dijkstra, Lewis, H. Poelmann (2014), vol. European Commission Urban and Regional Policy. Working paper 1 (2014).
See pages - of supplementary_information_v6.pdf