Urban heat island intensity mapping for preterm birth heat exposure
Urban heat islands impose a thermal burden that regional weather stations systematically underestimate. Landsat and MODIS land surface temperature data, disaggregated to neighbourhood scale, can quantify that extra exposure during the critical third trimester.
Sensors
- Landsat 8/9 TIRS: Thermal Infrared Sensor bands 10 and 11 (10.6–12.5 µm) deliver land surface temperature at 100 m native resolution, resampled and sharpened to 30 m in standard products. Sixteen-day repeat cycle; daytime and nighttime acquisitions available depending on orbit crossing time.
- MODIS Terra/Aqua (MOD11/MYD11 LST): Daily LST at 1 km resolution using split-window algorithm. Terra crosses the equator at approximately 10:30 local time, Aqua at approximately 13:30, giving two daytime and two nighttime overpasses per day. Useful for seasonal compositing and gap-filling where Landsat cloud contamination is severe.
- Sentinel-3 SLSTR: Sea and Land Surface Temperature Radiometer provides LST at approximately 1 km in the thermal channels (10.85 µm and 12.0 µm) with a roughly one-day revisit at mid-latitudes from the dual-satellite constellation. Useful for capturing rapid heat events but cannot resolve sub-neighbourhood variation.
- ERA5 reanalysis (ECMWF): Hourly 2-metre air temperature at approximately 31 km grid spacing. Not a satellite sensor, but the standard rural reference baseline against which urban LST anomalies are computed. Essential for constructing the UHI intensity metric and for linking satellite observations to epidemiological exposure windows.
Why a thermometer on a rooftop is not a proxy for a pregnant woman's exposure
Conventional meteorological stations are sited to represent regional air temperature: grass surfaces, ventilated screens, locations chosen to avoid local heat sources. A woman in her third trimester walking to a clinic through a dense urban neighbourhood experiences something quite different. Asphalt, concrete and dark roofing materials absorb shortwave radiation and re-emit it as longwave heat, raising surface temperatures well above the surrounding rural baseline. The difference between that urban surface temperature and the rural equivalent is the urban heat island intensity, and it is the quantity that matters for exposure assessment.
Published epidemiological work, including studies using data from US cities and South Asian megacities, has associated ambient heat exposure during the third trimester with elevated risk of preterm birth and, at higher temperature thresholds, stillbirth. The mechanism is not fully resolved, but physiological stress responses, dehydration and reduced placental blood flow are plausible pathways. The satellite contribution is not to settle the physiology; it is to produce spatially disaggregated exposure estimates that a single weather station simply cannot provide.
What Landsat thermal bands actually measure, and what they do not
Landsat 8 and 9 TIRS records upwelling radiance in two thermal infrared channels. After atmospheric correction and emissivity estimation, the standard USGS Collection 2 product delivers land surface temperature (LST) at 30 m spatial resolution. That is fine enough to distinguish a park from an adjacent housing block, or a tree-lined street from an unshaded market square. The archive runs continuously from 2013 for Landsat 8 and 2021 for Landsat 9, with earlier Landsat 5 and 7 thermal data extending retrospective analysis back to the 1980s.
The critical caveat is that LST is not air temperature. Over impervious surfaces, LST can exceed the 2-metre air temperature by 10–20°C on clear summer afternoons. That positive bias means LST-derived UHI intensity overstates the thermal burden experienced at pedestrian level. The relationship between LST and screen-height air temperature varies with wind speed, humidity and surface cover, and it is not constant across cities or seasons. Epidemiological studies that use LST directly as the exposure variable should account for this; studies that use LST only to spatially disaggregate station-measured air temperature (a downscaling approach) are on firmer methodological ground.
Constructing the third-trimester exposure window
A preterm birth study requires not a single snapshot but a characterisation of heat exposure across a specific gestational window, typically weeks 28 to 36. That means matching satellite acquisitions to the calendar dates of each pregnancy in a cohort, which in practice means building seasonal composites rather than relying on individual overpasses.
The standard approach is to stack all cloud-free Landsat LST scenes within the relevant months and compute the median or 90th-percentile LST per pixel. MODIS daily LST fills temporal gaps where Landsat cloud cover is persistent, using spatial downscaling methods (such as the DisTrad or TsHARP algorithms, both published in the peer-reviewed literature) to sharpen MODIS 1 km values toward Landsat 30 m spatial detail. ERA5 hourly 2-metre air temperature provides the rural reference baseline: the UHI intensity for each pixel is then the difference between the downscaled urban LST and the ERA5 value for the nearest rural grid cell. Seasonal compositing over a three-month window typically requires 4 to 8 usable Landsat scenes at mid-latitudes; in tropical humid cities, cloud cover can reduce that to 1 or 2, which increases uncertainty substantially.
The rural reference problem
Defining what counts as rural is less obvious than it sounds. The standard approach selects pixels beyond a defined urban boundary with low impervious surface fraction, but peri-urban agriculture, irrigation and land-use change can all alter the reference temperature. Some studies use a fixed annular buffer around the city; others use land-cover classification to select vegetated reference pixels. The choice affects the computed UHI intensity by a non-trivial margin, and different studies are not always directly comparable as a result.
ERA5 reanalysis air temperature is an alternative reference that sidesteps the pixel-selection problem, but it introduces its own uncertainty because reanalysis products blend model output with observations and carry their own biases in data-sparse regions. For health-outcome studies, the honest position is to report sensitivity to the reference definition rather than presenting a single UHI intensity figure as definitive.
From raster to birth record: linking spatial data to health outcomes
The analytic pipeline connects a geocoded birth record (or a residential address from a prenatal care register) to the median LST or UHI intensity value within a defined buffer, typically 500 m to 1 km, around the home address during the relevant gestational weeks. That linked dataset feeds a regression model controlling for maternal age, parity, socioeconomic status and access to cooling. The satellite layer is the exposure variable; the epidemiological model is the client's domain.
Satellize can deliver the geospatial exposure layer: a 30 m gridded UHI intensity raster, seasonally composited, with uncertainty estimates derived from the number of cloud-free scenes used and the sensitivity of the rural reference definition. The Tonga crop-estimation programme demonstrated the same seasonal compositing logic in a different context, building confidence in the compositing pipeline for small-island and data-sparse environments. The exposure layer is a GIS deliverable; the health analysis sits with the research team or public health authority that holds the birth records.
Honest limits of the method
LST-to-air-temperature bias is the largest known systematic error. Studies that treat LST as a direct proxy for experienced heat without a correction function will overestimate exposure, particularly over dense urban cores. Night-time LST (available from MODIS and, with lower frequency, from Landsat night-time acquisitions) is a better correlate of overnight air temperature and may be more relevant to physiological heat stress than peak daytime values.
Cloud cover is the operational constraint. Tropical and monsoon-affected cities, which include many of the highest-burden settings for preterm birth, can have cloud fractions exceeding 80 per cent in the wet season. Compositing over longer windows reduces noise but may mix climatically distinct periods. Sentinel-3 SLSTR's higher revisit partially compensates, but at coarser spatial resolution. Where fewer than three cloud-free Landsat scenes exist within a trimester window, the composited LST carries wide uncertainty and should be flagged as such in any downstream health analysis.
Typical figures
| Spatial resolution (Landsat TIRS) | 100 m native thermal; 30 m in USGS Collection 2 standard product |
| Spatial resolution (MODIS LST) | 1 km (MOD11A1 / MYD11A1 daily products) |
| Revisit (Landsat 8 + 9 combined) | 8 days at equator; fewer usable scenes in high-cloud-cover regions |
| Revisit (MODIS Terra + Aqua) | Up to 4 overpasses per day; daily composite products standard |
| Thermal spectral bands | Landsat TIRS: 10.6–11.19 µm and 11.50–12.51 µm; MODIS: 11 µm and 12 µm split-window channels |
| LST accuracy (published) | ±1–2 K for Landsat TIRS under clear-sky conditions after atmospheric correction; similar for MODIS split-window |
| ERA5 reference grid | ~31 km horizontal resolution; hourly temporal resolution; available from 1940 to present |
| Archive depth | Landsat thermal: 1984 (Landsat 5 TM) to present; MODIS LST: 2000 to present |
| Cloud contamination limit | LST retrievals invalid under cloud; compositing over 3-month windows typically requires ≥3 clear scenes for reliable median |
| Deliverable format | GeoTIFF raster (EPSG:4326 or client CRS), seasonal composite with pixel-level scene-count metadata |
Analytics Satellize can run
| Third-trimester UHI intensity raster | Median LST composite from all cloud-free Landsat scenes in the gestational window, differenced against ERA5 rural reference; TsHARP downscaling applied where MODIS gap-fill is used | 30 m GeoTIFF per trimester window, with pixel-level scene-count and uncertainty band layers |
| Residential exposure score | Zonal statistics within 500 m buffer around geocoded address points; outputs mean, 90th-percentile and peak UHI intensity for each location | CSV or GeoPackage linked to anonymised address identifiers, ready for regression ingestion |
| Night-time LST composite | MODIS MYD11 Aqua night-time overpasses composited monthly; night-time values used as closer correlate of overnight air temperature than daytime peak | 1 km monthly raster stack covering the study city and reference buffer zone |
| Impervious surface fraction layer | Landsat-derived NDBI and NDVI used to classify impervious fraction per 30 m pixel; used to define rural reference pixels and stratify UHI intensity by surface type | 30 m classified raster with impervious fraction 0–100%, updated annually |
| Seasonal heat anomaly flag | Comparison of current-season LST composite against 10-year MODIS climatological baseline; pixels exceeding 1.5 standard deviations flagged as anomalously hot | Binary anomaly mask overlaid on UHI intensity raster, delivered as GIS layer with anomaly magnitude attribute |
| Uncertainty report | Bootstrap resampling of available clear-sky scenes to estimate composite stability; sensitivity analysis across two rural reference definitions (annular buffer vs. land-cover-classified pixels) | PDF technical annex quantifying confidence intervals on UHI intensity by neighbourhood, suitable for peer-review supplementary material |
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.