Nighttime light as a proxy for health infrastructure reliability and energy poverty
VIIRS Day/Night Band radiance time series can expose chronic power instability at health facilities long before a ministry's grid reports register a fault. This page explains the physics, the products, and the honest limits of the method.
Sensors
- VIIRS Day/Night Band (Suomi NPP / NOAA-20): 750 m nadir resolution, daily global revisit, panchromatic low-light band sensitive from 0.5 to 0.9 µm. The primary workhorse for nightly radiance time series. Detects radiance as low as roughly 2 × 10⁻¹¹ W cm⁻² sr⁻¹ under optimal conditions, sufficient to register a single petrol generator powering a rural clinic.
- NASA Black Marble VNP46 product suite: Monthly and daily composites derived from VIIRS DNB with atmospheric, lunar and stray-light corrections applied. VNP46A2 daily gap-filled product and VNP46A3 monthly composites are the standard inputs for facility-level radiance trend analysis. Distributed via NASA Earthdata.
- DMSP-OLS archive (NOAA, 1992–2013): Coarser resolution (roughly 2.7 km) and no onboard calibration, but provides a two-decade baseline for electrification trends prior to the VIIRS era. Useful for establishing whether a facility location was electrified at all before 2012, despite significant inter-satellite calibration challenges.
- Landsat 8 / 9 OLI panchromatic band: 15 m panchromatic resolution at night, 16-day revisit per satellite (8-day combined). Not designed for low-light work, but under very dark-sky conditions can distinguish individual lit structures within a facility compound that VIIRS conflates with neighbours. Sparse temporal coverage limits its use to validation snapshots rather than continuous monitoring.
Why a radiance reading is a health outcome proxy
Reliable electricity is not a comfort amenity for a health facility. It is a clinical prerequisite. Vaccine cold chains require continuous refrigeration between 2°C and 8°C; a single overnight outage can render an entire immunisation stock unusable. Surgical theatres depend on stable lighting. Oxygen concentrators, autoclaves, foetal monitors and blood-bank refrigerators all fail silently when voltage drops below operating thresholds. In settings where these failures go unrecorded, satellite radiance is often the only independent witness.
The VIIRS Day/Night Band detects emitted visible and near-infrared light at night with sufficient sensitivity to register a cluster of fluorescent tubes or a small diesel generator. When a facility's nightly radiance drops to background levels, or oscillates erratically across consecutive overpasses, the signal is consistent with a power outage or chronic load-shedding. That inference is not proof of a clinical failure, but it is a statistically tractable flag that warrants ground follow-up.
What a time series reveals that a snapshot cannot
A single night's radiance image tells you almost nothing useful about infrastructure reliability. The analytical value comes from the temporal dimension. Plotting nightly DNB radiance at a fixed facility coordinate across twelve months reveals patterns that are invisible to any static assessment: seasonal dimming correlated with dry-season generator fuel shortages, sudden step-changes coinciding with grid extension or collapse, and persistent near-zero readings that indicate a facility is effectively unelectrified despite appearing on government infrastructure maps.
The NASA Black Marble VNP46A2 daily product, which applies lunar-phase, cloud-mask and atmospheric corrections before delivery, is the standard input for this kind of analysis. Monthly VNP46A3 composites smooth out cloud gaps but can obscure short outage events lasting only a few days. For health-facility monitoring, daily granules are preferable even though cloud cover in tropical regions can leave gaps of two to three weeks during peak wet seasons. Those gaps are themselves informative: a facility that disappears from the radiance record every wet season may be relying on solar-battery systems that are undersized for cloudy periods.
The DMSP-OLS archive extends the analysis back to 1992, though inter-satellite calibration offsets and the absence of onboard gain control mean DMSP data require careful cross-calibration before being joined to a VIIRS time series. The transition year around 2012 to 2014 is particularly noisy.
The contamination problem in dense areas
At 750 m nadir resolution, a single VIIRS DNB pixel covers an area larger than many urban health facility compounds. In any town with commercial streets, market lighting or petrol stations near a clinic, the pixel radiance integrates all of those sources together. A facility can appear brightly lit on every night of the year simply because a busy road runs alongside it, regardless of whether the clinic itself has power.
This is the method's most significant honest limit. In dense urban settings, VIIRS DNB cannot reliably attribute radiance to a specific building. Landsat panchromatic night imagery at 15 m can, in principle, resolve individual structures within a compound, but its 16-day revisit makes it unsuitable for detecting short outages. The practical workaround is spatial stratification: restrict facility-level radiance analysis to rural and peri-urban locations where the nearest significant light source is at least one to two VIIRS pixels away. In dense cities, the method should be replaced or supplemented by ground-reported grid-stability data or high-resolution commercial satellite tasking.
A secondary contamination source is moonlight. The Black Marble products apply lunar-phase corrections, but residual errors around full moon can introduce false radiance spikes of 10 to 30 percent in some environments. Any anomaly detection algorithm applied to the time series should account for the lunar cycle.
Linking radiance stability to published health outcome evidence
The correlation between facility electrification and health service delivery is well-documented in the public health literature. Studies drawing on Demographic and Health Survey data and WHO facility assessments have found that facilities with reliable power are significantly more likely to offer delivery services, maintain cold-chain vaccines and perform basic surgery. The satellite radiance approach does not generate new health outcome data; it provides a scalable, near-real-time proxy for the electrification variable that those studies identify as critical.
Energy poverty mapping at the national scale, using VIIRS DNB composites, has been published by groups including the World Bank's ESMAP programme and the Energising Finance initiative. These typically use annual or monthly composites to classify settlements by electrification rate. The facility-level application described here is a more targeted variant: instead of characterising whole settlements, it anchors the analysis to the GPS coordinates of known health posts, clinics and hospitals from databases such as the WHO Service Availability and Readiness Assessment or national HMIS registries.
Building the analytic pipeline
The core workflow extracts nightly VNP46A2 radiance values at each facility coordinate, applies a cloud-validity mask, and constructs a gap-filled time series using harmonic regression or linear interpolation across cloudy periods. Anomaly detection then flags nights where radiance falls more than a defined threshold below the facility's rolling baseline. A threshold of two standard deviations below the twelve-month mean, applied over at least three consecutive valid overpasses, reduces false positives from single-night cloud-mask errors.
The output is a facility-level outage probability score updated nightly, delivered as a GIS layer or structured feed that a ministry of health or development-finance institution can overlay on its facility registry. Facilities scoring above a defined outage-frequency threshold in any rolling ninety-day window enter an alert queue for field verification. Satellize runs this class of analytics on open VIIRS and Black Marble products for clients requiring sovereign, non-commercially-shared analysis pipelines; the Tonga crop-estimation programme uses a structurally similar temporal anomaly approach on a different sensor stack.
Calibration against ground truth is essential before any operational deployment. Even a small sample of facilities with independently verified power logs, gathered over one to two months, allows the radiance threshold parameters to be tuned to local conditions. Without that calibration step, the false-positive rate in areas with variable street lighting can be high enough to overwhelm a field verification team.
What the method cannot tell you
Radiance absence does not distinguish between a grid outage, a deliberate shutdown, a facility that has closed permanently, or a facility that never had grid power and is simply not lit at night. All four conditions produce the same near-zero DNB reading. Interpretation requires a facility status layer that records whether a site is operational, and ideally some record of its nominal power source.
The method also cannot detect voltage instability that leaves lights on but equipment inoperable. A generator running at 180 volts may illuminate a ward while simultaneously damaging refrigeration compressors. Radiance is a presence-of-light signal, not a power-quality signal. For voltage-sensitive equipment like vaccine refrigerators, radiance stability is a necessary but not sufficient indicator of safe operating conditions. Pairing the satellite signal with even sparse ground-reported voltage logs substantially improves diagnostic accuracy.
Typical figures
| Primary sensor spatial resolution | 750 m nadir (VIIRS DNB); degrades to ~1600 m at swath edges |
| Temporal revisit | Daily global coverage (Suomi NPP + NOAA-20 combined); cloud gaps of 1–21 days common in tropical wet seasons |
| Radiance detection floor | Approximately 2 × 10⁻¹¹ W cm⁻² sr⁻¹ under low-lunar conditions (VIIRS DNB) |
| Spectral range | VIIRS DNB: 0.5–0.9 µm panchromatic; Landsat 8/9 pan: 0.503–0.676 µm |
| Archive depth | VIIRS: 2012–present; DMSP-OLS: 1992–2013 (requires inter-satellite calibration) |
| Standard data product | NASA Black Marble VNP46A1 (daily at-sensor), VNP46A2 (daily corrected), VNP46A3 (monthly composite) |
| Latency (standard product) | VNP46A2 typically available 1–3 days after acquisition via NASA Earthdata |
| Minimum detectable facility signal | Isolated rural facility with generator lighting detectable; urban facilities contaminated by adjacent sources at 750 m resolution |
| Delivery formats | HDF5 (NASA native), GeoTIFF (reprocessed), facility-level CSV time series, GIS vector alert layer |
| Landsat night validation resolution | 15 m panchromatic; 16-day revisit per satellite (8-day combined Landsat 8+9) |
Analytics Satellize can run
| Facility-level nightly radiance time series | Point extraction from VNP46A2 daily corrected product at registered facility GPS coordinates, with cloud-validity masking | Structured CSV or GeoPackage time series per facility, updated nightly |
| Outage event detection | Rolling-baseline anomaly detection (2-sigma threshold over 90-day window) applied to gap-filled radiance series; minimum three consecutive valid low-radiance overpasses to flag | Alert feed (JSON or GIS layer) listing facility ID, outage start date, duration in valid overpasses, and confidence score |
| Chronic low-power classification | Annual radiance percentile scoring per facility relative to peer facilities of similar type and region; facilities in bottom decile flagged as chronically underlit | Annual facility-level energy-poverty risk tier map (GeoTIFF + tabular report) |
| Electrification trend analysis (multi-year) | Harmonic regression and Mann-Kendall trend test on monthly VNP46A3 composites at facility locations, with optional DMSP-OLS pre-2012 baseline splice | Trend report per facility or district: direction, magnitude and statistical significance of radiance change over the study period |
| Urban contamination risk screening | Spatial buffer analysis comparing facility-pixel radiance to surrounding 3×3 pixel neighbourhood; facilities where >40% of neighbourhood radiance originates outside the facility footprint are flagged for manual review | Facility registry annotation layer indicating analysis confidence class (rural-isolated, peri-urban, urban-contaminated) |
| Ground-truth calibration report | Comparison of satellite-derived outage flags against independently collected facility power logs or grid-operator records for a calibration subset; threshold tuning and false-positive rate estimation | Calibration report with recommended detection thresholds and expected precision/recall for the specific country context |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.