Long-term PM and NO2 exposure mapping for mental health burden studies

Chronic exposure to PM2.5 and NO2 is associated with depression, anxiety and cognitive decline. Satellite-derived multi-year exposure grids from TROPOMI, MODIS AOD and OMI are now standard inputs for cohort studies, but exposure misclassification remains the central methodological problem.

Sensors

Sentinel-5P TROPOMI: Provides daily global NO2 tropospheric column density at 3.5 × 5.5 km (reprocessed to 3.5 × 3.5 km from August 2019). Also retrieves aerosol layer height and UV aerosol index. Operational since October 2017, giving a multi-year archive now exceeding seven years.
MODIS Terra and Aqua (AOD): Dark Target and Deep Blue aerosol optical depth retrievals at 3 km and 10 km resolution, twice-daily overpass. The backbone of the van Donkelaar geophysical regression model that converts columnar AOD to near-surface PM2.5 estimates. Archive runs from 2000, enabling 20-year exposure surfaces.
OMI (Aura): Ozone Monitoring Instrument, operational since 2004. Provides NO2 column retrievals at roughly 13 × 24 km, historically important for trend studies predating TROPOMI. Row-anomaly issues after 2009 reduce spatial coverage but the long record is irreplaceable for decade-scale epidemiological cohorts.
MISR (Terra): Multi-angle Imaging SpectroRadiometer retrieves aerosol optical depth and particle shape information at 4.4 km resolution. Multi-angle geometry allows some discrimination between coarse and fine-mode aerosol, useful for separating dust from combustion PM2.5 in mixed-source environments.

Why mental health researchers turned to orbit

The epidemiological case for a link between chronic air pollution and psychiatric outcomes has strengthened considerably over the past decade. Large cohort studies in the UK Biobank, Danish national registers and several US Medicare datasets have reported associations between long-term PM2.5 or NO2 exposure and elevated rates of depression, anxiety disorder and accelerated cognitive decline. The proposed biological pathways include neuroinflammation, oxidative stress and disruption of the blood-brain barrier, none of which requires exotic mechanisms.

The practical problem for researchers is exposure measurement. Ground monitoring networks are sparse, unevenly distributed and biased toward urban cores. Satellite-derived exposure surfaces fill the spatial gaps. They are not perfect substitutes for personal exposure monitors, but they cover entire national populations at consistent spatial resolution, and their archives now span two decades. That combination is what makes them attractive as cohort-level exposure inputs.

How a multi-year mean exposure grid is actually built

The most widely used PM2.5 surfaces come from the van Donkelaar group's geophysical regression approach, which combines MODIS, MISR and SeaWiFS aerosol optical depth retrievals with a chemical transport model (GEOS-Chem) to estimate near-surface PM2.5 at 0.01-degree resolution (roughly 1 km at mid-latitudes). Annual mean grids are publicly archived. The method does not directly measure particles in the boundary layer; it infers them from the total atmospheric aerosol column, then uses the model to apportion that column vertically. That inference step introduces uncertainty that varies by season, region and aerosol type.

For NO2, TROPOMI tropospheric column densities are averaged over 12-month windows. Because TROPOMI retrieves a column from the surface to the tropopause, a land-use regression step is usually applied to downscale from the 3.5 km pixel to street level. That regression requires local ground truth: without it, the satellite column and the concentration a person actually breathes can differ by a factor of two or more in cities with heavy traffic corridors.

Building a multi-year mean is straightforward in principle: stack annual grids, compute the pixel-wise mean over the study period, then assign each cohort participant the value at their residential postcode centroid. The statistical complexity comes later, when analysts must account for the fact that this assignment is a proxy, not a measurement.

Exposure misclassification: the problem that does not go away

Residential address at enrolment is a poor proxy for lifetime exposure. People move. They commute. They spend roughly a third of their lives in workplaces, vehicles and indoor environments where the outdoor satellite column has limited relevance. Classical non-differential misclassification biases effect estimates toward the null, meaning published associations between pollution and mental health outcomes are probably underestimates. That is reassuring in one sense and methodologically uncomfortable in another.

Street-canyon gradients are a separate issue. A TROPOMI pixel at 3.5 km resolution averages over a large area that includes parks, motorways and back streets. Within that pixel, roadside NO2 can be two to five times higher than the neighbourhood mean. Studies that assign the pixel mean to all residents within it will underestimate exposure for people living on arterial roads and overestimate it for those in green spaces. The error is spatially correlated, which violates the independence assumptions of standard regression.

Satellite columns also fail on cloudy days. TROPOMI has no valid retrieval through thick cloud, so monthly or annual means are composites of clear-sky days only. If pollution episodes are systematically cloudier or clearer than average days, the annual mean is biased. High-aerosol-load days are sometimes cloud-contaminated, which can cause the surface to underestimate peak exposures.

What the archive depth actually buys you

MODIS AOD extends back to 2000, OMI NO2 to 2004, TROPOMI NO2 to late 2017. For a cohort enrolled in 2010 with a ten-year follow-up, the researcher can construct a pre-enrolment exposure history using OMI and MODIS, then continue with TROPOMI after 2017. The sensors are not directly interchangeable: OMI's coarser footprint and the row anomaly require careful harmonisation, typically via bias-correction against ground monitors in overlapping periods.

This multi-sensor continuity matters because mental health outcomes, particularly cognitive decline, are thought to respond to cumulative lifetime exposure rather than recent ambient concentrations. A ten-year mean exposure surface is a better biological hypothesis than last year's annual mean. The archive makes that hypothesis testable at national scale without recruiting a single additional participant.

Statistical assumptions researchers must state explicitly

Any analysis linking satellite exposure grids to health records rests on at least four assumptions that should appear in the methods section, not be buried or omitted. First, that the chemical transport model vertical apportionment is approximately correct for the study region and season. Second, that clear-sky bias in the satellite composite does not systematically correlate with the health outcome. Third, that residential address is a reasonable centroid for individual exposure, given the population's mobility patterns. Fourth, that confounders correlated with both pollution and mental health, including socioeconomic deprivation, green space access and noise, are adequately measured and adjusted.

The last assumption is the hardest. Deprived neighbourhoods tend to have higher pollution and higher rates of depression. Satellite data can map the pollution gradient precisely; it cannot measure the deprivation gradient with the same precision unless additional layers are incorporated. Studies that adjust only for area-level deprivation indices may leave substantial residual confounding.

What a research team can commission

A typical engagement produces annual-mean PM2.5 and NO2 grids for a defined study region and period, delivered as georeferenced rasters at 0.01-degree resolution with associated uncertainty estimates. Those grids are then joined to participant postcodes to generate individual-level exposure assignments ready for import into statistical software. Sensitivity analyses using different averaging windows, different sensor inputs or different land-use regression models can be run systematically rather than ad hoc.

Satellize runs this class of analysis on open constellations including TROPOMI, MODIS and OMI, applying published retrieval and regression methods to client-defined geographies and time windows. The Tonga crop-estimation programme is a different domain, but the underlying workflow, stacking multi-year satellite observations into a spatially consistent analytical surface, is the same. Research teams that want to interrogate the uncertainty in their exposure assignments, rather than treat the grid as ground truth, will find that explicit sensitivity testing is where most of the methodological value lies. A good first step is to share your cohort geography and enrolment dates so the archive coverage and cloud-fraction statistics can be assessed before any modelling begins.

Typical figures

Spatial resolution (PM2.5 surface)	~1 km (0.01 degree) for van Donkelaar AOD-regression product; native MODIS AOD at 3 or 10 km
Spatial resolution (NO2 column)	3.5 × 3.5 km (TROPOMI, post-August 2019); ~13 × 24 km (OMI)
Revisit / temporal sampling	Daily global coverage (TROPOMI, MODIS); clear-sky retrievals only; effective monthly sampling varies by cloud climatology
Archive depth	PM2.5 surfaces from 2000 (MODIS); NO2 columns from 2004 (OMI) and 2017 (TROPOMI)
Spectral bands used	UV-VIS (NO2 retrieval, 405-465 nm); VIS-NIR-SWIR (AOD retrieval); multi-angle geometry (MISR)
Minimum detectable concentration (NO2)	TROPOMI detection limit ~0.03 × 10¹⁵ molecules/cm² per pixel; surface-level sensitivity depends on boundary-layer height
Cloud limitation	No valid retrieval at cloud radiance fraction above ~0.5 (TROPOMI); high-aerosol episodes may be cloud-contaminated
Uncertainty (PM2.5 surface)	Typically ±20-40% at annual mean level in validation against ground monitors; higher in dust-dominated or biomass-burning regions
Delivery format	GeoTIFF raster grids; CSV exposure assignment tables keyed to postcode or participant ID; NetCDF for time-series stacks

Analytics Satellize can run

Annual-mean PM2.5 exposure surface	Van Donkelaar AOD-to-PM2.5 geophysical regression applied to MODIS Dark Target and Deep Blue retrievals, calibrated against available ground monitors	GeoTIFF raster at 0.01-degree resolution per study year, with per-pixel uncertainty band
Annual-mean tropospheric NO2 column grid	TROPOMI L2 offline product aggregated to annual means, gap-filled for cloud days using climatological weighting	GeoTIFF raster at 0.01-degree resolution; accompanying cloud-fraction and valid-day-count layers
Participant-level exposure assignment table	Point-in-raster extraction at residential postcode centroid; optional buffer averaging for address uncertainty	CSV with participant ID, annual PM2.5 and NO2 values, valid-retrieval count, and flag for high-uncertainty pixels
Multi-sensor NO2 time series (OMI + TROPOMI harmonised)	Bias-correction of OMI retrievals against TROPOMI in 2017-2019 overlap period; blended annual series from 2004	Per-pixel annual NO2 time series in NetCDF; harmonisation coefficients documented in accompanying metadata
Sensitivity analysis ensemble	Systematic variation of averaging window (1, 3, 5, 10 years), sensor input and land-use regression specification to bound exposure misclassification	Tabular summary of effect-estimate stability across ensemble members; GIS layers for each scenario
Deprivation-pollution co-exposure layer	Overlay of annual PM2.5 and NO2 grids with publicly available area-level socioeconomic indices to identify confounding structure	GIS layer with bivariate pollution-deprivation classification; summary statistics by study area quintile

Sources

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.