Source-linked AI summary
Mapping the world population one building at a time
Tobias G. Tiecke, Xianming Liu, Amy Zhang, Andreas Gros, Nan Li, Gregory Yetman, Talip Kilic, Siobhan Murray, Brian Blankespoor, Espen B. Prydz, Hai-Anh H. Dang
TL;DR
Accurate global population distribution data are important, but existing census units can have an average spatial resolution of 33 km and sparse populations remain difficult to map. The paper presents a high-resolution global population-density mapping method using building-focused analysis, denoising neural networks, and satellite imagery; the approach processes 0.5 Mkm2 in approximately 24 hours, while accuracy remains limited by population input data and redistribution.
Problem
Accurate information on global population distribution is crucial, while census units can have an average spatial resolution of 33 km at a global scale.
Method
The paper presents a method to generate high-resolution global population-density maps, focusing on building identification and using an Image Denoising Neural Network to address noise in satellite imagery.
Results
0.5 Mkm2 can be processed in approximately 24 hours on a CPU cluster.
Takeaways & Limitations
The analysis indicates that population-estimate accuracy is limited by available population input data and its redistribution rather than building identification.
Takeaways & Limitations
Accuracy is limited by the available population input data and its redistribution to building structures.
Abstract
from arXiv · showhide
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 ~30 meter resolution for 18 countries. We validate our method, and find that the building identification has an average precision and recall of 0.95 and 0.91, respectively and that the population estimates have a standard error of a factor ~2 or less. Based on our data, we analyze 29 percent of the world population, and show that 99 percent lives within 36 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.
1. Source satellite imagery
The study uses 50 cm satellite imagery to detect individual buildings globally, while denoising networks address image noise and support large-scale processing. Imagery timing and cloud-free availability constrain coverage.
- 50 cm imagery provides the resolution needed to detect individual houses.
- Cloud-free coverage combines imagery collected over multiple years, with 90% of the data taken between 2011 and 2015.Seasonal performance differences are mitigated when training data includes sufficient seasonal variation.
- The DeepDenoiser addresses spatially dependent noise in satellite-image mosaics before building detection.Noise can arise during image normalization or calibration and produce systematic errors.
- The denoising approach achieves similar performance to BM3D with at least 10 times greater speed efficiency.This efficiency enables denoising at large scale.
- Applying the denoising network recovers buildings omitted from noisy imagery in western Lagos.
2. Evaluation of neural networks
The neural-network evaluation uses labeled satellite-image patches sampled globally and reports country-level precision, recall, and coverage. Label quality is especially challenging in rural imagery, while incomplete coverage reflects missing cloud-free imagery.
- 150K 64x64 patches were sampled for building-detection training and evaluation.Training used balanced positives and negatives, while testing used spatially random samples matching the real-world distribution.
- Disputed labels were removed, but rural imagery still produced substantial labeling errors because buildings are sparse and image quality is poor.
- Precision, recall, and coverage are reported separately for each country in Table S1.
- Coverage below 100% results from the unavailability of cloud-free satellite imagery.
3. Allocating population to buildings
Population is allocated to settled areas identified from satellite imagery by combining 1 arc-second building classifications with census-unit boundaries and counts. The allocation is deliberately proportional and minimally modeled.
- Census-unit boundaries and population estimates from GPWv4 distribute population onto binary settlement classifications at approximately 30x30m resolution.
- The source census data characteristics differ across Ghana, Haiti, Malawi, and Vietnam.
- Population is allocated equally among settled areas within each census unit.The proportional model creates a population-count surface for each country.
- The minimally modeled approach is chosen because it is understandable, usable as an independent variable, and compatible with additional modeling.
4. Validation of population redistribution
Validation compares high-resolution population estimates with census-based enumeration-area counts using regressions across countries and administrative resolutions. Results indicate that population-input coarseness, rather than building identification, limits redistribution accuracy.
- Validation design: Enumeration-area validation sums pixel-level population estimates and compares them with updated census-based population counts.
- Regression analysis: The regression uses census enumeration-area population projections as y and summed high-resolution map estimates as P.α is the constant term and ε is the error term.
- Effect of census resolution: In Malawi, traditional-authority and district-level input data substantially reduce the accuracy of enumeration-area population estimates.The analysis tests simulated coarser input-data scenarios against the highest-resolution available census input.
- Scope: Vietnam validation excludes areas whose satellite imagery was unusable because of perpetual cloud cover.
- Validation results: 1.6 and 2.1 are the estimated population-data error factors for Malawi using traditional-authority-level and district-level inputs, respectively.Urban performance is slightly worse than rural performance, attributed to greater diversity in building heights and uses.
- Conclusion: Population-estimate accuracy is limited by the available population input data and its redistribution rather than by building identification.Future improvements include building-type classification and additional geospatial data.
5. Validation of building localization with household survey data
Household survey comparisons show close agreement with the high-resolution population map, including at national scale and under a 100-meter location tolerance.
- The validation used confidential GPS dwelling coordinates from Malawi’s Third Integrated Household Survey and measured proximity to mapped populated pixels.The survey included 12,271 households distributed across 768 enumeration areas from 31 districts.
- 1.7 percent of Malawian households were more than 100m from the nearest populated pixel at national level.The tolerance accounts for GPS limitations, measurements outside structures, and gridding imprecision.
- 0.4 percent of households appeared in pixels with no mapped population when positive identification allowed a 500-meter distance.
6. Comparison with GUF and GHSL
The building dataset was compared with GUF and GHSL at the building-map resolution, using agreement analysis and optical-image review of remaining disagreements.
- GUF and GHSL provide independent settlement classifications based on different data sources and methods.GUF uses synthetic-aperture radar imagery and unsupervised classification; its source imagery has approximately 12-meter resolution.
- The three datasets were compared at 1 arc-second resolution by enumerating settled-area presence or absence combinations.Single-pixel disagreements and cloudy source-image areas were removed before review.
- Remaining contiguous disagreements with both comparison classifications were checked against DigitalGlobe optical imagery for false-positive and false-negative units.Country-level agreement and disagreement proportions are reported in Table S4.
- Approximately 7% of global data gaps were filled with more recent imagery from 2013 and 2014.