Prenatal NO2 exposure mapping for adverse birth outcome risk
TROPOMI measures tropospheric NO2 at 3.5 km resolution with near-daily global coverage. Combined with land-use regression and trimester-resolved residential records, those columns can estimate gestational exposure windows linked to preterm birth and low birthweight at neighbourhood scale.
Sensors
- Sentinel-5P TROPOMI: Measures tropospheric NO2 column density in the 405–465 nm band at 3.5 × 5.5 km nadir pixel (3.5 × 3.5 km since August 2019 processor update). Near-daily global revisit; data averaged to monthly composites to suppress retrieval noise from cloud cover and single-overpass variability. Level-2 offline products typically available within 3 hours of sensing.
- Landsat 8 / 9 OLI: 30 m multispectral imagery used to derive land-use ancillary variables (road density proxies via impervious surface mapping, vegetation fraction, urban morphology) that feed the land-use regression downscaling step. 16-day repeat per satellite; combined Landsat 8 and 9 gives roughly 8-day revisit at mid-latitudes.
- ERA5 reanalysis (ECMWF): Hourly planetary boundary-layer height (PBLH) at approximately 31 km grid spacing. PBLH controls the vertical dilution of surface NO2; including it as a predictor in downscaling models materially reduces bias in winter inversion episodes when surface concentrations decouple from column totals.
- OpenStreetMap / national road network data: Traffic-related NO2 is strongly distance-dependent. Road-class buffers (motorway, trunk, primary) at 100 m, 300 m and 500 m intervals are standard LUR predictor variables, grounding the spatial disaggregation in geometry rather than extrapolation.
What the column actually measures, and why that matters for a pregnancy
TROPOMI retrieves the total tropospheric NO2 column: the vertically integrated number of molecules between the surface and the tropopause, expressed in mol/m². It does not directly measure what a pregnant woman breathes at pavement level. That distinction is not a footnote; it is the central methodological problem the downscaling chain exists to solve.
The epidemiological literature linking NO2 to adverse birth outcomes, including preterm birth (gestational age below 37 weeks) and low birthweight (below 2,500 g), is substantial. A widely cited meta-analysis published in Environment International found odds ratios for preterm birth in the range of 1.05 to 1.10 per 10 µg/m³ increment in NO2 exposure across trimester windows. Those estimates were built on ground-monitor studies, not satellite columns. Translating satellite data into health-relevant exposure estimates therefore requires an explicit and honest downscaling chain, not a direct correlation.
The downscaling chain: from 3.5 km pixel to census tract
Land-use regression (LUR) is the published workhorse for this step. The method regresses ground-monitor NO2 concentrations against spatial predictor variables (road-class buffers, population density, land-cover type, elevation, PBLH) to produce a fine-scale prediction surface. TROPOMI monthly composites enter as an additional predictor, providing the regional background gradient that ground monitors alone cannot capture at national scale. The combined model can push effective spatial resolution toward 100–200 m in well-monitored cities, though the honest figure depends entirely on monitor density.
The trimester-resolved component assigns each residential address a mean estimated concentration for each of the three trimesters, derived by averaging the relevant monthly LUR surfaces. First-trimester windows are typically weeks 1–12 of gestational age; the exact mapping requires a known or estimated last menstrual period date from birth records. Where those dates are uncertain or absent, the exposure estimate inherits that uncertainty directly. No satellite processing step can recover a missing clinical date.
ERA5 boundary-layer height enters as a monthly mean field, interpolated to the LUR grid. Its role is to adjust for the fraction of the column that reflects surface conditions rather than residual aloft. In summer, deep convective mixing means the column correlates reasonably well with surface NO2. In winter anticyclones, a shallow boundary layer traps surface pollution while the column may look moderate; ignoring PBLH in those conditions can underestimate surface exposure by a factor of two or more in valley cities.
Honest limits: what this method cannot see
Indoor versus outdoor exposure is invisible to any satellite. Pregnant women in high-income countries spend roughly 85–90% of their time indoors, according to time-use surveys. A woman living next to a motorway but working in a filtered-air office accumulates a very different actual dose than the outdoor concentration at her address implies. The method estimates ambient outdoor exposure at the residential location. It is a proxy, not a personal dosimeter.
Cloud cover is a persistent retrieval problem. TROPOMI requires cloud radiance fraction below roughly 0.5 for a valid NO2 retrieval. In persistently cloudy regions (tropical coasts, monsoon belts, northern Europe in winter), monthly composites may rest on as few as four to eight valid overpasses rather than the theoretical 28–30. The resulting composite is noisier, and the effective revisit for a clean observation can stretch well beyond the nominal near-daily figure.
Causal inference from cross-sectional or ecological analyses of this kind faces confounding from socioeconomic status, maternal health, smoking, indoor heating fuel and access to antenatal care. Satellite NO2 exposure maps are inputs to epidemiological studies; they are not, by themselves, evidence of causation. Any report that presents the spatial correlation between NO2 and birth outcomes as a causal estimate without explicit confounding adjustment is overstating what the data support.
Building the exposure-outcome dataset: what the analyst actually needs
The minimum viable dataset on the health side is a geocoded birth registry with gestational age, birthweight, maternal residential postcode or census tract, and estimated or recorded last menstrual period date. On the satellite side, the analyst needs monthly TROPOMI Level-3 offline NO2 products (available from the Copernicus Data Space Ecosystem), a Landsat-derived land-use layer for the study region, ERA5 monthly PBLH fields, and a road network file. Ground NO2 monitors are needed to calibrate the LUR; the model is unreliable in areas with fewer than roughly 20 monitors per 10,000 km².
Temporal alignment is non-trivial. A birth recorded in March 2023 with a gestational age of 38 weeks implies conception around June 2022. The analyst must pull TROPOMI composites from June, July and August 2022 for the first trimester, then the subsequent months for the second and third. Automating this linkage for tens of thousands of births across multiple years is where the analytic pipeline earns its keep.
What the output looks like in practice
The primary deliverable is a trimester-resolved NO2 exposure surface at census-tract or postcode level, typically expressed in µg/m³ estimated surface concentration, with uncertainty bounds derived from leave-one-out cross-validation of the LUR model. A secondary deliverable is a linked exposure-outcome table ready for regression analysis, with each birth record assigned three trimester exposure estimates and a flag for data quality (number of valid TROPOMI overpasses in the composite, monitor coverage score).
Satellize builds this pipeline on open constellations, the same Sentinel-5P, Landsat and ERA5 data described here, and can deliver trimester-resolved exposure grids as GeoTIFF layers or PostGIS-compatible tables for integration with a client's health information system. The Tonga crop-estimation programme demonstrated the same underlying principle of linking temporally resolved satellite composites to ground-truth outcome data; the exposure-outcome linkage here is epidemiological rather than agronomic, but the data engineering is structurally similar. Speak to the analytics team about study-region scoping before commissioning a birth-records linkage, because monitor density in the target area determines whether the LUR is publishable or merely indicative.
Typical figures
| TROPOMI NO2 pixel size (post-Aug 2019) | 3.5 × 5.5 km (across-track × along-track at nadir) |
| TROPOMI revisit | Near-daily global; monthly composites typically use 4–28 valid overpasses depending on cloud cover |
| TROPOMI Level-2 offline latency | Approximately 3 hours after sensing; monthly composites available within days of month end |
| Downscaled LUR output resolution | 100–200 m in well-monitored cities; census-tract aggregation recommended where monitor density is low |
| Landsat land-use ancillary resolution | 30 m; 8-day effective revisit (combined Landsat 8 and 9) |
| ERA5 PBLH grid spacing | ~31 km; hourly temporal resolution, aggregated to monthly means for LUR |
| TROPOMI archive depth | May 2018 to present (Sentinel-5P launched October 2017) |
| Minimum detectable NO2 column (TROPOMI) | ~0.5–1.0 × 10¹⁵ mol/cm² (single retrieval); monthly averaging reduces noise substantially |
| Spectral band (NO2 retrieval) | 405–465 nm (visible, DOAS differential optical absorption spectroscopy) |
| Delivery formats | GeoTIFF exposure grids, PostGIS tables, CSV birth-record linkage files with uncertainty flags |
Analytics Satellize can run
| Monthly trimester-resolved NO2 exposure surface | TROPOMI Level-3 monthly composite combined with Landsat-LUR downscaling and ERA5 PBLH correction | GeoTIFF raster stack per trimester month, with per-pixel uncertainty estimate from LUR cross-validation |
| Birth-record exposure linkage table | Automated temporal alignment of geocoded birth registry to trimester-specific monthly composites; three exposure values assigned per birth record | CSV or PostGIS table ready for epidemiological regression, with data-quality flags (valid overpass count, monitor coverage score) |
| LUR model calibration and validation report | Stepwise or LASSO regression of ground-monitor NO2 against spatial predictors; leave-one-out cross-validation R² and RMSE reported | PDF methodology report with model coefficients, predictor importance ranking and spatial residual maps |
| High-exposure neighbourhood risk map | Percentile ranking of census-tract mean gestational NO2 across the study region; overlay with deprivation index where available | GIS layer (GeoPackage or Shapefile) for public health authority GIS systems, with accompanying data dictionary |
| Seasonal and inter-annual NO2 trend analysis | Mann-Kendall trend test on monthly TROPOMI composites across the study period (2018 to present); breakpoint detection for policy evaluation | Time-series chart package and trend-significance table; useful for before-and-after evaluation of traffic or emissions interventions |
| Cloud-gap quality assessment | Per-pixel count of valid TROPOMI retrievals in each monthly composite; flagging of census tracts where fewer than eight valid overpasses underpin the exposure estimate | Quality-flag raster and summary table; allows epidemiologists to exclude or down-weight unreliable exposure estimates in regression |
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.