Estimating global economic well-being with unlit settlements

The study addresses how to spatially identify and monitor economic well-being using globally available satellite data, with a focus on areas lacking detectable nighttime radiance. Despite progress in reducing global poverty, nearly a billion people lack reliable and affordable electricity, and many countries—especially in sub-Saharan Africa—have limited, low-quality economic statistics. Aid allocation often misses the poorest regions within countries, underscoring the need for spatially precise indicators of poverty and well-being. Nighttime lights (NTL) from VIIRS DNB correlate with economic activity, and the World Settlement Footprint (WSF) provides detailed global settlement mapping. However, conventional approaches struggle in regions with little or no detectable light, and some recent methods depend on proprietary datasets or have interpretability challenges. The authors propose a complementary and scalable approach that leverages the percentage of unlit settlement footprint—derived by combining VIIRS DNB nighttime lights and WSF—to infer relative wealth classes across low- and middle-income countries.

Prior research has consistently shown strong relationships between satellite-derived nighttime radiance and economic indicators such as GDP and economic growth. The VIIRS DNB sensor, with improved spatial and radiometric accuracy over DMSP-OLS, enables analysis of lighting at neighborhood scales and primarily captures light from human settlements. The World Settlement Footprint (WSF) offers a high-resolution, global inventory of built-up areas. While NTL-based proxies perform well where lighting exists, their utility is limited in impoverished regions with little detectable light. Alternative approaches include linking NTL with daytime imagery, leveraging mobile phone data for poverty estimation, and socioecological frameworks, though many require upscaling and proprietary data. Recent deep learning methods using publicly available satellite imagery have improved poverty prediction in Africa but pose interpretability challenges for policy uptake. The current work builds on this body of research by focusing on the share of unlit settlements as a simple, interpretable indicator tied to economic well-being.

Data sources: The study used the 2015 annual average VIIRS DNB nighttime lights composite (vcm-ntl, version 1) from the Earth Observation Group, which removes background noise, solar/lunar contamination, cloud-degraded data, and non-anthropogenic lighting (e.g., fires, flares) to retain anthropogenic radiance. The composite is provided at ~500 m grid spacing (downscaled from 750 m). Settlement data were obtained from the World Settlement Footprint 2015 (WSF2015), resampled to 500 m to report percentage of settlement area per pixel. The WSF integrates Landsat-8 and Sentinel-1 imagery and has an average accuracy of 86%, though it can miss very small, dispersed, or low-profile structures and extremely poor dwellings. Definitions and stratification: Rural and urban areas were delineated using the GHSL GHS-SMOD (1 km, 2015), which refines the Degree of Urbanization method by combining population size, density, and built-up area density to classify cities, towns/semi-dense areas, and rural areas. Unlit settlement determination: For each 500 m pixel, WSF settlement area was labeled as lit or unlit according to whether the corresponding VIIRS DNB pixel indicated anthropogenic radiance. Using country boundaries (Natural Earth), the total WSF area (km²) per country was summarized as lit or unlit, and aggregated to continents. Development indicators analysis: World Bank indicators (2015 GDP per capita PPP; electric power consumption averaged 2014–2016; secondary school enrollment averaged 2014–2016; and 2015 urban population share) were regressed bivariately against unlit settlement shares by continent, using logit transforms for percentages and log transforms for GDP and electricity consumption. Due to multicollinearity, indicators were analyzed separately. Household survey data: DHS microdata were used to represent household wealth via an asset-based index. Household recode files (wealth factors) were linked to household cluster geocoordinates. To enable cross-survey comparability, the authors harmonized wealth indices by running a PCA across all surveys and also constructed electricity-free variants (excluding assets requiring grid electricity, or excluding any electricity-using assets) using SVDimpute to handle missing records. Data selection and spatial linkage: The analysis included surveys conducted after 2010 with all eight anchor points and complete representation of five wealth classes. Household-level wealth classes were reduced to three categories (poorer, average, richer) by taking the modal wealth class per cluster. To accommodate DHS coordinate displacement, buffers of 2 km (urban) and 5 km (rural; with 1% up to 10 km) were applied around cluster centroids, and the percentage of unlit settlement area within buffers was computed. Classification and validation: A Naïve Bayes classifier was used to predict the wealth class (poorer, average, richer) from discretized percentages of unlit settlements. Class probabilities were estimated from contingency tables with small-count corrections to avoid division by zero. Accuracy was evaluated using 10-fold cross-validation, reporting overall and class-specific accuracies, as well as separate results for rural and urban subsets. Mapping: Country-wide maps of relative wealth classes were generated by applying the Naïve Bayes classifier to gridded unlit settlement percentages, assigning each pixel to the wealth class with the highest posterior probability and using the top-class probability as a confidence measure. Benchmarking: In Nigeria, the mapped wealth classes were compared against a deep learning-based wealth map (within 775 level-2 districts) and the Subnational Human Development Index (SHDI; 37 subnational units). The study also assessed how the share of unlit settlements predicts the SHDI Income Index while controlling for nighttime lights.

• Globally, 19% of the 2015 settlement footprint had no detectable associated nighttime radiance. The largest shares of unlit settlement footprint were in Africa (39%) and Asia (23%). When restricted to rural areas, unlit shares rose to 65% in Africa and 40% in Asia. Notable unlit settlement areas also exist in some developed countries; Europe had 16% unlit settlement footprint.
• Across 49 countries in Africa, Asia, and the Americas, the percentage of unlit settlements is strongly associated with lower DHS-based economic well-being. Using Naïve Bayes classification of wealth classes (poorer, average, richer) based on unlit share, accuracies were: Africa 93%, Asia 86%, Americas 86%; overall 86.6%, with rural 88.2% and urban 85.1%.
• Country-scale maps for Bangladesh, Cambodia, Nigeria, and Uganda highlight extensive poorer (typically high unlit percentages and low settlement coverage) regions, with richer areas concentrated around capitals and dense urban centers.
• In Nigeria, the mapped wealth classes align closely with a recent deep learning wealth index and with the SHDI; the share of unlit settlements predicts the SHDI Income Index.
• Bivariate analyses with World Bank indicators by continent (not detailed in main text) further contextualize relationships between unlit settlement shares and development metrics.

The findings demonstrate that substantial portions of settlement infrastructure worldwide, particularly in rural Africa and Asia, lack detectable nighttime lighting, and that the share of unlit settlement footprint is a simple, interpretable proxy for relative economic well-being. Technical and observational factors influence detectability: VIIRS DNB’s native 750 m resolution can miss small settlements; BRDF effects and the post-midnight (~1:30 am) overpass time reduce observed anthropogenic lighting; the rise of LED lighting and energy-saving policies further dim signals; natural light sources (airglow, moonlight) can mask faint emissions; and urban skyglow can artificially elevate apparent radiance in adjacent areas. These factors tend to overestimate unlit areas in both developed and developing contexts, though in developing countries many small, off-grid rural settlements would remain undetected even with earlier overpasses. On the settlement side, WSF2015, while accurate overall, likely underestimates scattered, low-profile, or extremely poor dwellings (e.g., thatched houses), implying the true extent of unlit settlements in developing countries is potentially higher than observed. Urban heterogeneity (e.g., partially lit slums) and single bright sources within a pixel can complicate classification. Despite these caveats, the approach agrees with independent benchmarks (deep learning wealth maps, SHDI) and provides a transparent, scalable indicator for prioritizing aid and infrastructure. The clear rural-urban disparity in unlit shares underscores the need to target rural electrification (grid and off-grid) to improve incomes, health, and education, and to track progress toward SDG 7 and broader poverty-reduction goals.

By combining VIIRS nighttime lights with the World Settlement Footprint, the study quantifies the global extent of unlit settlements (19% of settlement footprint in 2015) and shows that the share of unlit settlement area is a robust, interpretable predictor of relative economic well-being across 49 countries, achieving high classification accuracy. The method enables country-scale mapping of wealth classes to identify poorer regions for targeted interventions and to monitor changes over time as developing countries electrify and developed countries potentially reduce light emissions. Future work should improve detection of small and extremely poor dwellings (e.g., through enhanced settlement products, crowdsourcing, and in-situ data), refine corrections for BRDF and overpass-time biases, integrate multiple nighttime light products, and routinely update maps to monitor progress toward SDGs and inform policy.

Key limitations include: (1) VIIRS DNB overpass at ~1:30 am and sensor resolution (750 m native) reduce detection of smaller or dimmer settlements; (2) BRDF, natural light (airglow, moonlight), and urban skyglow can bias radiance measurements, leading to over- or underestimation of lit/unlit areas; (3) single bright objects can classify entire pixels as lit, while partial lighting within urban areas (e.g., slums) may be missed; (4) WSF2015 can miss dispersed, low-profile, or extremely poor dwellings and transient structures, likely underestimating unlit settlements in developing regions; (5) DHS cluster displacement necessitates buffering, introducing spatial uncertainty; (6) classification relies on discretized unlit percentages and Naïve Bayes assumptions; and (7) homeless populations and some informal settlements may not be represented, affecting generalizability.