13.6.1 — Poverty Analytics — maturity: live

Asset-Based Wealth Mapping

Using satellite imagery and machine-learning to infer household and community wealth from observable physical assets — rooftop materials, road access, vehicle density, agricultural holdings.

Satellite imagery and machine learning can map household wealth at neighbourhood scale, giving governments a census-independent evidence base for targeting social programmes.

National statistics offices in lower- and middle-income countries typically wait five to ten years between household surveys, leaving policymakers blind to where poverty is deepening or retreating between census cycles. Satellite-derived asset proxies — roof type, built-area density, proximity to paved roads, irrigated field extent — correlate strongly with validated survey wealth quintiles and can be produced annually at ward or village level for a fraction of the cost of a ground enumeration.

A constellation of optical microsatellites in sun-synchronous LEO provides the sub-3-metre multispectral imagery needed to resolve individual structures reliably across a country's full territory. Paired with synthetic aperture radar passes for cloud-affected regions and seasons, the stack delivers consistent annual coverage regardless of weather. On-board preprocessing reduces downlink volume; a sovereign GPU cluster runs the computer-vision inference pipeline and produces ward-level asset-index scores that are audit-ready and reproducible.

The operational payoff is a living wealth map that a ministry of finance or planning commission can interrogate monthly rather than decennially. Subnational governments gain objective evidence for budget allocation. International donors can verify that transfers reach the intended quintiles without waiting for the next survey cycle. And because the data are produced by a national system under national protocols, the methodology is open to parliamentary scrutiny — a guarantee no commercial data subscription can match.

What matters

Roof-material classification at 2–3m resolution is the single most validated remote asset proxy across sub-Saharan Africa and South Asia.
Survey-to-satellite wealth index correlations exceed r = 0.85 in peer-reviewed studies when training data cover at least three agro-ecological zones.
Annual satellite revisit cycles let governments detect poverty rebounds after shocks — drought, flood, inflation — before the next household survey.
A sovereign pipeline eliminates the risk that a commercial vendor withdraws access, changes licensing terms, or geo-blocks imagery during a political dispute.

Quick facts

Population living on less than $2.15/day: 712 million people (2024) — World Bank Poverty and Inequality Platform
Accuracy of satellite-derived wealth index vs. survey ground truth: ~70% R² correlation (2023) — PNAS – Predicting poverty and wealth from mobile and satellite data
Cost of a traditional household survey (national): $1.5M–$10M per country (2022) — World Bank Living Standards Measurement Study
Median revisit frequency, commercial 0.5 m VHR constellation: 1–3 days (2024) — Planet Labs SkySat constellation specifications
Global settlement layer coverage (GHSL) spatial resolution: 10 m (2023) — EC Joint Research Centre – Global Human Settlement Layer

Sovereignty score: 8/10 — A nation that relies on a foreign commercial provider for its poverty map surrenders control over the most politically sensitive domestic dataset a government produces.

Wealth and asset data at ward level are constitutionally sensitive in many jurisdictions; routing them through a foreign cloud pipeline creates legal exposure under data-protection and statistical-secrecy statutes.
Commercial imagery vendors have suspended or re-priced access to politically inconvenient governments on short notice — a sovereign constellation is immune to that leverage.
Donor-driven poverty maps produced by external NGOs or multilaterals have historically driven budget allocations that contradict national planning priorities; a sovereign system gives the government the authoritative dataset.
Reproducibility and parliamentary accountability require that the methodology, training labels and inference code are held domestically and are subject to national audit — conditions no third-party subscription can guarantee.

Reference architecture

Payload: Multispectral optical imager, 4-band (RGB + NIR), 2.5m GSD, 20km swath; secondary panchromatic channel at 0.8m for structure-level classification; optional SAR cross-cue from shared radar constellation at 3m spotlight resolution
Bus class: ESPA-class microsat, 120kg, 600W total power, 400W available to payload; deployable solar panel array; 16 TB solid-state mass memory for store-and-forward
Orbit: Sun-synchronous LEO at 500–550km; 18-satellite walker constellation; 3-day full-national-territory revisit at equator, daily revisit of priority zones; LTAN 10:30 for consistent shadow angle
Ground segment: 3-station national network (X-band downlink, S-band TT&C) co-located with the national meteorological and national mapping agencies; SatNOGS nodes at two universities for telemetry backup; on-premise GPU cluster (≥8× A100-class GPUs) for inference
Data pipeline: On-board radiometric calibration and cloud-mask flagging → X-band downlink → ground L1 orthorectification and BRDF normalisation → CNN inference for roof-type, built-density and road-access indices → spatial aggregation to admin-3 (ward) level → versioned asset-index dataset stored in a sovereign PostGIS database
End-user delivery: Web GIS console for the national statistics office and planning ministry, with ward-level choropleth wealth maps and time-series charts; REST API for provincial planning departments; quarterly static reports auto-generated in PDF/A for parliamentary tabling; restricted high-resolution layer for finance ministry on separate VLAN
Time to launch: First 3-satellite demonstrator in 22 months from contract signature for national-coverage validation; full 18-satellite constellation operational at 36 months; annual wealth-map product cycle begins at demonstrator phase using partial coverage supplemented by Sentinel-2 during gap periods
Caveats: Sub-metre optical payloads from US vendors are subject to ITAR/EAR export licensing; procure optical units from European (e.g. Airbus Defence & Space, OHB) or Indian (e.g. ISRO commercial arm NewSpace India) primes to avoid supply-chain dependency; cloud-persistent regions (Congo Basin, humid tropics) require SAR augmentation to maintain annual revisit commitments.

Frequently asked

What satellite data types are actually used for wealth mapping?

The main inputs are multispectral optical imagery (for rooftop material, land use, vegetation), nighttime lights (VIIRS from NOAA/NASA), SAR (for structure density in cloudy regions), and ancillary layers such as road networks from OpenStreetMap. Recent studies also fuse daytime and nighttime imagery with mobile call detail records where available. The combination typically outperforms any single source.

How accurate is satellite-derived wealth mapping compared with household surveys?

Peer-reviewed benchmarks report R² values of 0.60–0.80 when predicting survey-based asset indices at village or enumeration-area level. Accuracy drops sharply at individual household level. Governments should treat satellite-derived indices as a screening and prioritisation tool — not a replacement for survey-based benefit eligibility determination.

Why should a government build and operate its own satellite capability rather than buy imagery from Planet or Maxar?

Commercial providers can restrict tasking or raise prices without notice, and many licence agreements prohibit sharing derived data with third-party ministries. A sovereign constellation means the government sets the revisit schedule, owns the raw archive, and is not exposed to foreign export-control regimes. Over a 10-year horizon, owning a microsatellite constellation is often cost-competitive with repeated commercial subscriptions once ground-segment and analysis capacity are internalised.

What resolution do satellites need to be useful for poverty mapping?

Studies show 2–10 m multispectral imagery is sufficient for neighbourhood- and village-level wealth indices. Sub-metre imagery adds value for urban informal settlements where rooftop material and plot size matter. Nighttime light composites at 500 m (VIIRS) remain useful at regional scale. A sovereign programme can tier its architecture: medium-resolution optical for national coverage and targeted VHR for hotspot areas.

Can this work in countries without household survey data?

Transfer learning techniques allow models trained in data-rich countries to be applied in data-poor settings with some accuracy loss — typically 10–20 percentage points lower R². The World Bank's DECAT (Data for Economic Competitiveness and Agility Tool) and the MOSAIKS framework from UC Berkeley are examples of approaches designed for low-data environments. Even imperfect estimates are often more current and spatially granular than no data at all.

What is nighttime light intensity and why does it matter?

NOAA-NASA VIIRS records the brightness of artificial lighting at night at 500 m resolution globally. Brighter areas correlate strongly with economic activity and household electrification. It is one of the most consistently available free global datasets and acts as an important baseline layer in nearly all satellite poverty-mapping pipelines.

How does this application connect to social cash transfer programmes?

Satellite wealth maps can identify unserved or under-targeted geographic areas, informing where registrars should conduct community-based registration. They are most valuable as a geographic filter — narrowing the area in which intensive (and expensive) beneficiary verification is needed. This complements, rather than replaces, national social registries.

What are the main data governance issues a government needs to resolve before deployment?

Key issues include: defining what spatial resolution triggers individual identification under national privacy law; establishing which ministries can access derived outputs; setting data-retention and archival policy for raw imagery; and deciding whether commercial partners can access anonymised training data. OECD's AI Principles and the UN Secretary-General's Roadmap for Digital Cooperation both provide frameworks governments can adapt.

Glossary

Asset-based wealth index (AWI): A composite score derived from ownership of durable goods (radios, phones, vehicles) and housing characteristics, used as a proxy for household economic status when income data are unavailable.
VHR imagery: Very-high-resolution satellite imagery with ground sampling distance below 1 m, sufficient to distinguish rooftop materials, vehicles, and plot boundaries at individual parcel level.
VIIRS: Visible Infrared Imaging Radiometer Suite — a sensor aboard NOAA-NASA Joint Polar Satellite System satellites that captures, among other things, nighttime light emissions used as an economic activity proxy.
Transfer learning: A machine-learning technique where a model trained on labelled data from one country or region is adapted to a new context with limited local training data, reducing the requirement for expensive ground surveys.
Ground sampling distance (GSD): The real-world size of one pixel in a satellite image; a 0.5 m GSD means each pixel represents a 0.5 m × 0.5 m patch on the ground.
SAR: Synthetic Aperture Radar — an active microwave sensor that can image Earth's surface through cloud cover and darkness, making it a critical complement to optical imagery in persistently cloudy regions.
GHSL: Global Human Settlement Layer — a free, open dataset produced by the European Commission's Joint Research Centre that maps the presence and density of built-up areas globally from multitemporal satellite imagery.
Enumeration area (EA): The smallest geographic unit used in national census operations, typically containing 100–300 households, which serves as the standard spatial unit for matching satellite data to survey ground truth.
Nighttime lights (NTL): Remotely sensed measurements of artificial light emitted at night, widely used as a proxy for electrification, economic activity, and urbanisation at sub-national scale.
MOSAIKS: Multi-task Observation using Satellite Imagery and Kitchen Sinks — a randomised feature-based framework developed at UC Berkeley that enables rapid prediction of socioeconomic variables from satellite imagery with minimal labelled training data.

References

Combining satellite imagery and machine learning to predict poverty — Neal Jean et al. demonstrate that combining daytime and nighttime satellite imagery with convolutional neural networks can explain up to 75% of variation in village-level consumption expenditure across five African countries. The approach offers a scalable alternative to costly household surveys.
Predicting poverty and wealth from mobile and satellite data — Blumenstock et al. combine mobile phone metadata with satellite imagery to generate high-resolution poverty maps of Rwanda, demonstrating that the approach can predict survey-based wealth quintiles with ~70% accuracy while reducing data-collection costs by an order of magnitude.
World Bank Poverty and Inequality Platform — methodology note — Documents the World Bank's approach to estimating global poverty headcounts and highlights the 78-country data gap where no consumption or income survey has been conducted in the past five years, motivating the use of geospatial proxies.
MOSAIKS: A scalable satellite-based machine learning framework for societal applications — Introduces the MOSAIKS method, which uses random convolutional features from satellite imagery to predict a wide range of socioeconomic and environmental variables with minimal labelled data, achieving competitive accuracy across 10 prediction tasks including asset wealth.
Global Human Settlement Layer — GHS-BUILT-S R2023A — The JRC's 10 m resolution global built-up surface layer derived from Sentinel-2 multitemporal analysis provides a foundational geospatial input for wealth mapping, allowing analysts to stratify populations by settlement type and density.
NOAA-NASA VIIRS Nighttime Lights Annual Composites — Annual and monthly composites of VIIRS Day/Night Band data at 500 m resolution are publicly available from NASA Earthdata and are among the most widely used free inputs in satellite-based poverty and economic activity mapping.
OECD AI Principles — Recommendation on Artificial Intelligence — The OECD's binding AI Principles, adopted by 46 countries, set requirements for transparency, accountability, and human oversight in AI systems — directly applicable to automated wealth-scoring models built on satellite data that inform benefit eligibility decisions.
UN-OOSA — Space for Sustainable Development Goals — UN-OOSA's Space4SDGs initiative maps satellite applications to each of the 17 Sustainable Development Goals, explicitly identifying earth observation as a tool for SDG 1 (No Poverty) through poverty monitoring and targeting of social protection programmes.

Related applications

13.6.3 — Subnational Poverty Indices (Poverty Analytics)
13.6.4 — Cash Transfer Programme Targeting (Poverty Analytics)
13.6.5 — Service Access Inequality Mapping (Poverty Analytics)
13.1.1 — Tele-Radiology Networks (Telemedicine)
13.1.2 — Mobile Health Camp Operations (Telemedicine)
13.1.3 — Emergency Telemedicine Consultation (Telemedicine)
13.1.4 — Specialty Consultation Networks (Telemedicine)
13.1.5 — Hospital-to-Hospital Tele-ICU (Telemedicine)