Nighttime light anomaly detection at dark-sky heritage buffer zones
VIIRS Day/Night Band and the DMSP-OLS archive let analysts track artificial-light intrusion into legally protected dark-sky buffers around heritage sites and ancient observatories, separating permanent infrastructure from fires and festivals.
Sensors
- VIIRS Day/Night Band (Suomi NPP / NOAA-20): Panchromatic low-light band, 750 m nominal resolution at nadir, daily global revisit. Detects radiance in the 0.5–0.9 µm range down to roughly 2×10⁻⁹ W/cm²/sr, making it sensitive to single construction-site floodlights in otherwise dark areas. The primary instrument for radiance time-series analysis.
- DMSP-OLS archive: Operational Linescan System data from 1992 to 2013, roughly 2.7 km spatial resolution, providing the only decadal baseline for light-growth trend analysis before VIIRS. Saturation at high-radiance urban cores limits its dynamic range, but it is indispensable for establishing pre-development baselines at buffer zones.
- Sentinel-2 MSI: 10 m resolution optical imagery in visible and near-infrared bands, 5-day revisit at mid-latitudes with two satellites. Used for contextual validation: once VIIRS flags a new light source, Sentinel-2 daytime imagery identifies the physical structure responsible, whether a building, road, or temporary camp.
- VIIRS Nightfire (VIIRS NTL ancillary product): Sub-pixel fire and high-temperature combustion detection derived from VIIRS shortwave infrared bands. Distinguishes open fires and festival bonfires from the steady broadband emission of electric lighting, reducing false positives in ephemeral-event classification.
What a dark-sky buffer is, and why satellites are the only practical auditor
Many countries and international bodies now impose legally binding dark-sky buffer zones around ancient observatories, megalithic sites and UNESCO-listed landscapes where astronomical alignment is part of the heritage significance. Jodrell Bank's buffer in the UK, the Roque de los Muchachos zone in the Canary Islands, and the International Dark-Sky Places programme all define radiance thresholds that new development must not breach. Ground-based photometry can verify compliance at a fixed point, but it cannot monitor a buffer perimeter of tens or hundreds of square kilometres continuously. A satellite with a daily revisit can.
The practical audit problem is that buffer zones are often in remote terrain where enforcement agencies have neither the staff nor the vehicles to patrol regularly. A single new warehouse, road depot or agricultural polytunnel with overnight lighting can raise the sky background by a measurable fraction before anyone on the ground notices. VIIRS detects that change within days of it appearing.
The physics of what VIIRS actually measures
The VIIRS Day/Night Band is a panchromatic low-light sensor covering roughly 0.5 to 0.9 micrometres. Its design heritage comes from the DMSP-OLS but with far better radiometric calibration and a dynamic range that avoids the saturation problems that made OLS data difficult to use near bright sources. At 750 m nadir resolution, a single pixel integrates radiance from an area slightly under 0.6 km². That is coarse enough to miss individual buildings but sensitive enough to detect the aggregate output of a construction compound or a new petrol station.
The detection floor matters. Published characterisation of the VIIRS DNB places its minimum detectable radiance at approximately 2×10⁻⁹ W/cm²/sr under good atmospheric conditions. A single 1,000-watt floodlight at 750 m range produces a radiance well above that threshold in an otherwise dark pixel. This is why the instrument is useful for buffer-zone monitoring rather than just urban mapping: it was designed for exactly this regime of isolated, low-level sources against a dark background.
Separating a festival from a factory: ephemeral versus persistent light
The central analytical challenge is not detection but classification. A midsummer bonfire festival, a temporary film shoot, or a military exercise can all produce a VIIRS anomaly that looks, in a single night's data, identical to a new permanent installation. Misclassifying an ephemeral event as infrastructure growth generates false alerts and erodes trust with the heritage authority that commissioned the monitoring.
The standard approach uses temporal persistence filtering across a rolling window, typically 30 to 90 nights. A genuine new installation shows radiance in the same pixel on most clear nights; a festival shows one or two nights of elevated signal followed by a return to baseline. VIIRS Nightfire adds a spectral dimension: combustion sources have a characteristic shortwave-infrared signature absent from electric lighting. Combining persistence with spectral class reduces false positives substantially, though it does not eliminate them entirely. Clouds are the remaining enemy: a persistent cloud deck over a buffer zone can produce data gaps of two to three weeks in humid climates, during which a new installation could become operational without triggering an alert until the sky clears.
Building a decadal baseline from DMSP-OLS
VIIRS data begin in 2012 for Suomi NPP and 2018 for NOAA-20. For sites where the heritage designation or the buffer zone predates those years, the DMSP-OLS archive, which runs from 1992 to 2013 with annual composites available from NOAA and the Colorado School of Mines VIIRS Nightfire team, provides the only satellite record of pre-VIIRS light conditions. The two datasets are not directly comparable in radiance units: OLS saturates over bright sources and uses a different spectral response. Intercalibration methods published in the remote-sensing literature use regression against overlapping years (2012 to 2013) to stitch the records together into a single radiance index.
The result is a time series stretching back three decades for most sites. That is long enough to show whether a buffer zone was genuinely dark at the time of designation or already compromised, which has legal relevance when a heritage authority is pursuing enforcement action. It also reveals slow trends that would be invisible in a short monitoring window: a gradual brightening of 5 to 10 percent per year, compounding over a decade, represents a significant cumulative impact even if no single year triggers an alert threshold.
Contextual validation with Sentinel-2: naming the source
A VIIRS anomaly tells you that light exists and roughly where. It does not tell you what is producing it. Sentinel-2 daytime imagery at 10 m resolution closes that gap. Once a persistent anomaly is confirmed in the VIIRS time series, a Sentinel-2 scene acquired on a cloud-free day over the same location will typically reveal the physical structure: a rooftop, a car park, a polytunnel, a new access road. That identification is what a heritage enforcement officer needs to initiate a site visit or a planning inquiry.
The workflow is sequential rather than simultaneous. VIIRS runs continuously as a screening layer. Sentinel-2 is tasked reactively when the screening layer raises an alert. This is more cost-effective than attempting to detect new construction from optical imagery alone, where the signal is a building rather than its light output and change detection is complicated by seasonal vegetation and shadow variation. Satellize applies this two-stage approach in its analytics work, including the kind of radiance-threshold monitoring that underpins its Tonga crop-estimation programme's contextual land-use layers.
One honest limit: Sentinel-2 carries no night-imaging capability. It can identify the daytime structure but cannot directly image the light source in operation. For that, the only freely available night-imaging option remains VIIRS at 750 m. Commercial very-high-resolution night imagery exists but is expensive and not routinely available for heritage monitoring budgets.
What an operational monitoring system looks like in practice
A functional dark-sky buffer monitoring service has three layers. The first is a nightly VIIRS radiance ingest, clipped to the buffer-zone polygon, with cloud-mask applied and anomaly scores computed against the site's historical baseline. The second is a persistence filter that elevates an anomaly to an alert only after it appears on a defined number of nights within a rolling window. The third is an automated Sentinel-2 tasking trigger that requests the next available cloud-free acquisition over the alert pixel.
Latency from event to alert is typically three to ten days, depending on cloud cover and the persistence threshold chosen. That is fast enough to catch new construction before it is complete and potentially before planning permission has been formally assessed. Alert outputs are most useful as GIS layers overlaid on the legal buffer-zone boundary, delivered to the heritage authority's existing mapping environment rather than as standalone reports that require specialist interpretation.
Typical figures
| Primary sensor spatial resolution | VIIRS DNB: 750 m at nadir; DMSP-OLS: ~2.7 km; Sentinel-2 MSI (contextual): 10 m |
| Revisit frequency | VIIRS: daily global coverage (Suomi NPP + NOAA-20 combined gives two passes per night over most latitudes); Sentinel-2: 5-day at equator with both satellites |
| Minimum detectable radiance (VIIRS DNB) | ~2×10⁻⁹ W/cm²/sr under clear-sky, low-lunar-illumination conditions |
| Spectral bands used | VIIRS DNB: 0.5–0.9 µm panchromatic; VIIRS SWIR (Nightfire): 1.6 and 2.25 µm; Sentinel-2: 10 visible/NIR/SWIR bands at 10–20 m |
| Archive depth | DMSP-OLS annual composites: 1992–2013; VIIRS DNB: 2012–present (Suomi NPP), 2018–present (NOAA-20) |
| Alert latency (event to notification) | 3–10 days depending on cloud cover and persistence-filter window length |
| Cloud impact | Persistent cloud cover can produce data gaps of 2–3 weeks in humid climates; monthly compositing mitigates but does not eliminate this |
| Typical delivery format | GeoTIFF radiance anomaly rasters, GeoJSON alert polygons, time-series CSV per pixel, optional WMS feed into heritage-authority GIS |
| Coverage | Global; no geographic restriction on buffer-zone location |
Analytics Satellize can run
| Baseline radiance profile per buffer zone | Median compositing of VIIRS DNB over a user-defined reference period, with lunar-illumination and atmospheric correction applied following published NOAA calibration procedures | GeoTIFF baseline raster and summary statistics table, delivered once at programme start |
| Nightly anomaly score layer | Z-score deviation of each night's VIIRS DNB pixel from the site baseline, masked for cloud and moonlit conditions | Nightly GeoTIFF anomaly grid clipped to buffer-zone polygon, ingested into client GIS or dashboard |
| Persistent new-light alert | Temporal persistence filter: anomaly must exceed threshold on a configurable proportion of nights within a 30- or 90-night rolling window before alert is raised | GeoJSON point or polygon alert with centroid coordinates, first-detection date, mean anomaly radiance, and persistence score |
| Ephemeral-event classification | VIIRS Nightfire shortwave-infrared combustion index cross-referenced against persistence score to flag fires and festival events and suppress false alerts | Alert metadata field indicating probable source class (electric light, combustion, ambiguous) alongside each persistent-light alert |
| Decadal light-growth trend report | Intercalibrated DMSP-OLS to VIIRS radiance index time series using overlap-year regression; Mann-Kendall trend test applied per pixel within buffer zone | PDF trend report with per-pixel trend maps, annual radiance index chart, and statistical significance table; suitable for regulatory submission |
| Source-structure identification | Sentinel-2 MSI change detection triggered by persistent VIIRS alert; comparison of pre- and post-anomaly daytime scenes at 10 m to identify new built structures | Annotated Sentinel-2 image chip with suspected source structure outlined, delivered within 5–15 days of alert depending on cloud-free acquisition availability |
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.