Household solid-fuel combustion mapping as indoor air quality proxy
Where household surveys are thin, combining VIIRS nighttime fire detections, MODIS active fire radiative power, and census fuel-use priors lets analysts estimate district-level biomass and coal combustion intensity as a proxy for indoor PM2.5 burden.
Sensors
- VIIRS Day/Night Band (DNB) and I-band fire product (Suomi-NPP / NOAA-20): 375 m pixel resolution active fire product (VNP14IMG); detects fires at night using thermal anomaly in I4 (3.74 µm) and I5 (11.45 µm) bands. Twice-daily overpass at most latitudes. Fire radiative power (FRP) reported per pixel. Minimum detectable FRP is roughly 5 MW under good conditions, which means a single large open cooking fire cluster may be detectable but a shielded stove is not.
- MODIS Terra and Aqua (active fire, MOD14 / MYD14): 1 km resolution active fire product with FRP retrieval; four daily overpasses when both platforms are used together. The coarser pixel means many small household fires are subpixel and only aggregations of fires in a compound or village produce a detectable thermal anomaly. Archive runs from 2000, giving over two decades of seasonal combustion patterns.
- Sentinel-5P TROPOMI (CO column): Carbon monoxide total column at approximately 5.5 km x 7 km pixel (post-2019 reprocessing improves to 5.5 km x 3.5 km). Daily global coverage. CO is a direct tracer of incomplete combustion from biomass and coal. Useful for identifying persistent combustion plumes over districts where household solid-fuel use is suspected but fire detections are suppressed by roofing.
- Landsat 8 / 9 OLI-TIRS: 30 m multispectral imagery for settlement delineation and built-up density mapping. Roof material classification (metal, thatch, tile) from shortwave infrared bands informs the prior probability that cooking fires are shielded from satellite view. 16-day revisit per platform; combined 8-day revisit when both are tasked.
- WorldPop / GPWv4 population grids (ancillary): 100 m gridded population estimates derived from census data and building footprint models. Not a satellite sensor, but essential for converting fire radiative power density into per-capita combustion burden estimates at district level.
What fire radiative power actually tells you about cooking
Fire radiative power is a measure of the rate at which a fire releases energy as thermal radiation. VIIRS retrieves it at 375 m; MODIS at 1 km. For open biomass fires, published work has established approximate conversion factors between FRP and fuel consumption rate, typically in the range of 0.3 to 0.8 kg of dry matter per megawatt-second depending on fuel type. That relationship is the foundation of the spatial interpolation approach: accumulate FRP detections over a season, weight by population density and census-derived solid-fuel use fractions, and you get a district-level combustion intensity index that correlates with measured indoor PM2.5 in settings where ground truth exists.
The key word is 'open'. VIIRS and MODIS fire products were designed for wildfire and agricultural burning. Household cooking fires are far smaller and, in most urban and peri-urban settings, occur under a roof. The thermal signal is attenuated or entirely blocked. This is not a minor caveat: in sub-Saharan African cities, where charcoal stoves are used inside concrete or metal-roofed structures, satellite fire products will systematically undercount combustion. The method is most reliable in rural settings where cooking happens in open courtyards or under light thatch.
Building the prior: census fuel fractions and settlement morphology
No satellite product alone can tell you what fuel a household burns. The method depends on combining fire detections with a spatial prior derived from national census data or Demographic and Health Survey (DHS) microdata on solid-fuel use rates. Where census data are available at district level, solid-fuel use fractions (wood, charcoal, crop residue, dung, coal) can be disaggregated and assigned to population grid cells. Where census data are absent or stale, fuel-use models trained on nighttime light intensity, building density and income proxies provide a substitute, though with wider uncertainty.
Landsat-derived settlement maps contribute here in a specific way. Roof material classification from shortwave infrared reflectance distinguishes metal from thatch at 30 m resolution with reasonable accuracy in dry conditions. Metal roofs suppress fire detection probability; thatch roofs do not. Incorporating a roof-type layer into the detection model adjusts the FRP-to-combustion conversion upward in areas where shielding is likely, partially correcting the systematic undercount. It is an imperfect correction, but it is better than ignoring the bias entirely.
Carbon monoxide as a cross-check when fires are invisible
TROPOMI's CO column product offers a partial workaround for the shielding problem. Incomplete combustion from biomass and coal produces CO regardless of whether the fire is visible to a thermal sensor. In areas with dense household solid-fuel use, CO columns are persistently elevated above background, particularly in the early morning hours when cooking activity peaks and boundary layer mixing is shallow. TROPOMI's 5.5 km pixel cannot resolve individual streets, but it can distinguish districts with high household combustion from those dominated by traffic or industry when combined with NO2 and formaldehyde columns as additional tracers.
The honest limit here is that CO columns integrate the entire atmospheric column, not just surface sources. Industrial point sources, vehicle fleets and biomass burning episodes all contribute. Disentangling household combustion from these other sources requires either a source attribution model or careful selection of periods and locations where household burning dominates. In rural settings with little industry, the attribution is cleaner. In mixed urban environments, it is genuinely ambiguous.
The hard limits: what orbit cannot see
It is worth being direct about the ceiling on this method. VIIRS at 375 m has a minimum detectable FRP of roughly 5 MW. A single charcoal brazier produces perhaps 0.5 to 2 kW of radiative power. You need thousands of simultaneously burning stoves within a single pixel to produce a detectable signal, and even then the signal competes with background thermal emission from warm rooftops. Urban charcoal stove use is, for practical purposes, invisible from orbit.
The method therefore functions as a district-level burden index, not a household-level measurement. It is most useful for prioritising intervention areas, tracking seasonal combustion cycles and estimating relative exposure differences between districts. It cannot replace ground-based PM2.5 monitoring or household surveys for absolute exposure quantification. Governments planning clean-cooking programmes need this satellite-derived layer as a spatial allocation tool, not as a substitute for the monitoring networks that should follow.
From detection to district health burden: the interpolation workflow
The operational workflow has four steps. First, aggregate VIIRS and MODIS fire detections over a defined season (typically the dry season, when household biomass burning peaks) and compute cumulative FRP per grid cell. Second, apply the roof-type correction factor derived from Landsat classification to adjust for detection probability. Third, weight the corrected FRP density by population count and solid-fuel use fraction from census priors to produce a combustion intensity index in units of megawatt-hours per capita per season. Fourth, convert to an estimated PM2.5 exposure increment using published emission factors for the relevant fuel mix, with uncertainty bounds that honestly reflect the propagation of errors from each step.
Output uncertainty is substantial. Published studies applying similar methods in South Asia and sub-Saharan Africa report order-of-magnitude agreement with ground measurements at district level, but individual district estimates can be off by a factor of two or three. That is sufficient for prioritisation and trend monitoring. It is not sufficient for epidemiological dose-response modelling without ground validation. Satellize incorporates this workflow into district-level combustion burden reports for public health agencies; the Tonga crop-estimation programme is a separate example of how the same open-constellation analytics infrastructure applies to different estimation problems in data-sparse environments.
Seasonal patterns and the value of the long archive
MODIS Terra has operated since 2000. That archive allows analysts to reconstruct two decades of seasonal combustion cycles at 1 km resolution, identifying whether household burning intensity is increasing, decreasing or shifting geographically as fuel markets and urbanisation patterns change. In regions where liquefied petroleum gas subsidies have been introduced, a measurable decline in dry-season FRP density in peri-urban areas is detectable in the record, providing independent evidence of programme uptake without relying solely on self-reported survey data.
Seasonal timing matters for health burden too. Peak household combustion during cold dry seasons coincides with the lowest boundary layer heights, concentrating pollutants near the surface. Integrating VIIRS FRP time series with ERA5 reanalysis boundary layer height data produces a seasonally adjusted exposure index that better reflects actual inhalation burden than annual averages. This is a published methodological refinement, not a proprietary technique, and it meaningfully improves the correlation between satellite-derived indices and measured PM2.5 at monitoring stations.
Typical figures
| VIIRS active fire spatial resolution | 375 m (I-band product, VNP14IMG) |
| MODIS active fire spatial resolution | 1 km (MOD14 / MYD14) |
| VIIRS / MODIS combined revisit | Up to 4 overpasses per day at mid-latitudes (Terra + Aqua + Suomi-NPP + NOAA-20) |
| TROPOMI CO pixel size | Approximately 5.5 km x 3.5 km (post-2019 reprocessing); daily global coverage |
| Landsat settlement mapping resolution | 30 m; 8-day combined revisit (Landsat 8 + 9) |
| Minimum detectable FRP (VIIRS) | Approximately 5 MW under clear-sky, night conditions; shielded stoves below detection threshold |
| MODIS archive depth | February 2000 to present (Terra); July 2002 to present (Aqua) |
| District-level output uncertainty | Factor of 2 to 3 in absolute PM2.5 increment estimates; relative ranking between districts more reliable |
| Latency (near-real-time fire products) | VIIRS and MODIS NRT products available via FIRMS within approximately 3 hours of overpass |
| Coverage | Global; cloud cover does not affect thermal fire detection as severely as optical bands, but thick smoke can attenuate signal |
Analytics Satellize can run
| District combustion intensity index | Seasonal FRP accumulation (VIIRS + MODIS) weighted by population density and census solid-fuel use fractions, with Landsat roof-type detection probability correction | GIS raster layer (GeoTIFF) at 1 km resolution, delivered per season with uncertainty bounds |
| Estimated household PM2.5 exposure increment | FRP-to-fuel-consumption conversion using published emission factors for wood, charcoal and coal; propagated to PM2.5 using IPCC-cited emission factor ranges | District-level tabular report with low/central/high estimates and documented uncertainty sources |
| CO column anomaly map for combustion hotspot identification | TROPOMI CO column monthly composites with background subtraction and source-type screening against co-located NO2 and HCHO columns | Monthly GIS layer flagging districts with persistent CO elevation attributable to household combustion |
| Roof-type and settlement density classification | Landsat 8/9 SWIR-based supervised classification of metal vs. thatch/other roofing; built-up fraction from OLI band ratios | 30 m classification raster used as input to FRP correction model; also delivered as standalone GIS layer |
| Seasonally adjusted exposure index | VIIRS FRP time series combined with ERA5 boundary layer height reanalysis to weight combustion intensity by atmospheric mixing depth | Monthly time-series CSV and raster stack per district; suitable for input to health burden models |
| Multi-year combustion trend analysis | MODIS 20-year FRP archive aggregated to annual dry-season totals per district; Mann-Kendall trend test for monotonic change detection | Trend report with significance levels; designed to evaluate clean-cooking programme impact independently of survey data |
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.