Disproportionate air pollution burden mapping across income and race strata
Satellite-derived NO2 and PM2.5 surfaces, joined to census demographic data, can quantify whether low-income and minority communities bear pollution burdens that exceed their share of emission sources. The method is powerful and the limits are real.
Sensors
- Sentinel-5P TROPOMI: Tropospheric NO2 columns at 3.5 × 5.5 km (reprocessed; original 7 × 3.5 km before August 2019), daily global revisit. The primary satellite source for column-integrated NO2 used in equity studies.
- Aura OMI: NO2 and SO2 columns at roughly 13 × 24 km, daily revisit since 2004. Lower spatial resolution than TROPOMI but a 20-year archive that anchors long-term trend analysis and pre-dates TROPOMI by 14 years.
- MODIS Terra and Aqua: Aerosol optical depth (AOD) at 3 km (MAIAC algorithm) and 10 km (standard product), twice-daily combined revisit. AOD is the satellite input most commonly converted to surface PM2.5 via geophysical models or machine-learning calibration.
- Landsat 8 and 9 OLI: 30 m multispectral imagery, 16-day revisit per satellite (8-day combined). Used as ancillary land-use and land-cover input: road density, green space fraction and impervious surface area are standard covariates in PM2.5 downscaling models.
What the satellite actually measures, and what it does not
TROPOMI measures NO2 in the tropospheric column above the surface. What a person breathes is surface concentration, not column burden. The two correlate, but the relationship varies with boundary-layer height, temperature and local meteorology. Published studies, including the widely cited 2021 analysis by Kerr and colleagues in Science Advances, use TROPOMI column NO2 as a proxy for surface exposure precisely because no other instrument offers comparable spatial coverage at daily frequency. That is a pragmatic choice, not a perfect one.
PM2.5 is not measured directly from orbit at all. The standard approach converts MODIS MAIAC AOD to surface PM2.5 using either a chemical transport model ratio (the GEOS-Chem or MERRA-2 approach) or a machine-learning model trained on ground monitor data. The resulting grids typically carry uncertainties of 2 to 5 micrograms per cubic metre in urban areas, larger in regions with sparse ground truth. Equity conclusions drawn from these grids carry that uncertainty forward into every disparity estimate.
The spatial join: where the method gains power and where it bends
The core workflow is a spatial join: pollution rasters are aggregated to census-tract or census-block-group boundaries, then merged with American Community Survey (US) or Eurostat NUTS-level (EU) demographic tables. The resulting dataset supports regression of mean pollution exposure on income quintile, racial composition, or deprivation index, with or without controlling for proximity to emission sources.
Three tensions deserve explicit acknowledgement. First, census tracts average roughly 4,000 residents across areas that can span several square kilometres in suburban settings. A TROPOMI pixel at 3.5 km fits reasonably inside a large tract but straddles boundaries in dense cities, introducing spatial mismatch error. Second, the ecological fallacy applies: a census tract with 60 percent minority population and high mean NO2 does not tell you which individuals within that tract are most exposed. Third, unmeasured confounders, particularly building stock, occupational exposure and time-activity patterns, are invisible to any satellite sensor.
Despite these limits, the approach has produced findings consistent enough across independent datasets to be taken seriously. A 2019 study by Tessum and colleagues in PNAS found that in the United States, non-Hispanic Black Americans were exposed to 56 percent more PM2.5 from consumption than they generated through consumption. That figure used satellite-derived PM2.5 as the exposure surface. The directional finding has been replicated with different sensors and different demographic stratifications.
European evidence and the regulatory context that demands it
The EU's Zero Pollution Action Plan and the revised Ambient Air Quality Directive explicitly require member states to assess whether pollution burdens fall disproportionately on vulnerable populations. Satellite data fills a critical gap here: ground monitor networks in most European cities are too sparse to resolve intra-urban variation at the neighbourhood scale. A 2022 analysis using TROPOMI NO2 over 13 European cities found consistent patterns of higher NO2 in lower-income districts, with the gradient steepest in cities where industrial zones abut residential areas without buffer land.
OMI's archive back to 2004 is particularly useful in Europe for demonstrating that disparities pre-date current regulatory frameworks, strengthening the case that they are structural rather than incidental. The archive also shows where disparities have narrowed, which is equally important for evaluating whether specific interventions have worked.
Downscaling: getting from 3.5 km to something a planner can use
A TROPOMI pixel at 3.5 × 5.5 km is coarse relative to a city block. Several published downscaling approaches exist. Land-use regression (LUR) uses satellite-derived covariates, including road density from Landsat-derived impervious surface maps, green space fraction and distance to industrial facilities, to predict sub-kilometre NO2 at census-block level. The model is trained on whatever ground monitor data exist and then applied spatially. In cities with dense monitor networks the approach produces credible sub-500 m surfaces. In data-sparse settings the uncertainty expands substantially and should be reported explicitly.
A second approach uses high-resolution commercial imagery to classify micro-scale land use, then applies that classification as a spatial disaggregation weight on the TROPOMI column. This is faster than LUR but assumes that land use is the dominant driver of within-pixel variation, which is not always true near point sources.
Building a defensible disparity metric
The output of an equity analysis is typically a disparity ratio: the population-weighted mean exposure of a defined demographic group divided by the population-weighted mean exposure of a reference group, or of the full population. A ratio above 1.0 indicates excess burden. Reporting the ratio alone is insufficient. Analysts should also report the absolute concentration difference, because a ratio of 1.3 applied to a low-background city is very different from the same ratio applied to a heavily polluted one.
Sensitivity analysis matters more here than in most remote-sensing applications. The disparity estimate should be re-run using different PM2.5 conversion models, different demographic aggregation levels and different years. If the directional finding is consistent across those permutations, it is credible. If it flips with a different AOD-to-PM2.5 model, the data do not yet support a strong conclusion.
Satellize runs this workflow, including the spatial join, downscaling and sensitivity reporting, on open-constellation data as a structured analytics product. The Tonga crop-estimation programme uses a comparable spatial-join approach at a different scale, pairing satellite-derived vegetation indices with ground-reference plot data.
What the record cannot settle
Satellite data can establish that a disparity exists and quantify its magnitude at the population level. It cannot establish causation, identify the responsible emitters with legal precision, or measure indoor exposure, which for many pollutants exceeds outdoor exposure. It also cannot capture temporal patterns: a community adjacent to a facility that operates primarily at night may show lower mean column NO2 than a daytime reading implies.
The honest position is that satellite-derived equity analysis is a screening tool. It is rigorous enough to direct regulatory attention and to prioritise where ground-level monitoring networks should be expanded. It is not rigorous enough to substitute for those networks in enforcement proceedings. Any report that presents satellite disparity ratios without those caveats is overselling the method.
Typical figures
| NO2 spatial resolution (TROPOMI) | 3.5 × 5.5 km (post-August 2019 reprocessing) |
| NO2 spatial resolution (OMI) | 13 × 24 km standard; 13 × 12 km zoom mode |
| AOD spatial resolution (MODIS MAIAC) | 1 km (MAIAC); 10 km standard Dark Target / Deep Blue |
| Revisit frequency | Daily (TROPOMI, OMI, MODIS); 8-day combined (Landsat 8+9) |
| Archive depth | OMI from 2004; MODIS from 2000; TROPOMI from 2018; Landsat from 1972 |
| Cloud limitation | TROPOMI and OMI retrievals flagged at cloud radiance fraction above 0.5; MODIS AOD invalid over bright surfaces and under cloud |
| Typical PM2.5 retrieval uncertainty | 2 to 5 µg/m³ in urban areas with moderate ground-monitor density |
| Downscaling floor (LUR approach) | Approximately 100 to 500 m, dependent on covariate resolution and monitor density |
| Demographic join granularity | Census tract (US, ~4,000 residents) or NUTS-3 (EU); block-group feasible with LUR downscaling |
| Delivery formats | GeoTIFF pollution grids, GeoPackage or Shapefile with joined demographic attributes, CSV disparity tables, PDF analytical report |
Analytics Satellize can run
| Census-tract NO2 exposure surface | TROPOMI tropospheric column aggregated to tract centroids; seasonal and annual averaging; cloud-flag filtering | GeoTIFF raster and tract-level GeoPackage with mean, 90th-percentile and trend columns |
| PM2.5 surface concentration grid | MODIS MAIAC AOD converted to PM2.5 via published geophysical model ratio (GEOS-Chem or MERRA-2); optional ML calibration against available ground monitors | Annual-mean PM2.5 GeoTIFF at 1 km, with uncertainty band layer |
| Disparity ratio report by income quintile | Population-weighted mean exposure calculated per quintile from joined ACS or Eurostat income data; ratio and absolute difference computed against full-population mean | PDF analytical report with sensitivity tables across two PM2.5 models and two aggregation levels |
| Disparity ratio report by racial or ethnic group | Same spatial-join workflow; demographic weights from decennial census or ACS 5-year estimates; follows Tessum et al. (2019) PNAS methodology | PDF report and CSV table suitable for regulatory submission |
| Long-term trend layer (2004 to present) | OMI NO2 column time series at tract level; Mann-Kendall trend test per tract; TROPOMI spliced from 2018 with inter-calibration offset | GeoPackage with per-tract trend slope, p-value and confidence interval; time-series chart pack |
| Downscaled intra-urban NO2 surface | Land-use regression using Landsat-derived impervious surface, road density and NDVI as covariates; trained on available monitor data; applied to TROPOMI column | Sub-500 m GeoTIFF with explicit uncertainty raster; block-group join table |
| Emission-source proximity layer | Buffer analysis around permitted facility locations (EPA NEI, E-PRTR or national register); intersected with demographic grid to separate proximity-driven from background-driven burden | GIS layer and summary table distinguishing facility-adjacent from background-exposed tracts by demographic group |
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.