Urban heat health vulnerability mapping for excess mortality risk
Satellite land surface temperature at 30–70 m resolution, combined with census social vulnerability data, pinpoints city blocks where heatwave excess mortality risk is highest. The method is powerful and honest about its limits: LST is not air temperature.
Sensors
- ECOSTRESS (ISS-mounted): Thermal infrared radiometer delivering land surface temperature at approximately 70 m resolution. Irregular revisit due to ISS orbital precession, typically 1–5 days at mid-latitudes, with overpass times that vary across the diurnal cycle, making it uniquely capable of capturing peak afternoon heat as well as pre-dawn minimum temperatures within the same city.
- Landsat 8/9 TIRS: Thermal Infrared Sensor bands 10 and 11 at 100 m native resolution, resampled to 30 m in standard products. Sixteen-day revisit per satellite, eight days when both are operational. Archive extends to 1984 for trend analysis. Cloud cover is the principal operational constraint; a single overpass lost to cloud in summer can miss a heatwave peak entirely.
- Sentinel-2 MSI: No thermal band, so it does not measure LST directly. Its 10 m resolution in visible and near-infrared bands is used to derive impervious surface fraction and normalised difference vegetation index as proxies for canopy cover, feeding the vulnerability index rather than the temperature layer. Five-day revisit at the equator.
- MODIS Terra/Aqua (MOD21/MYD21): LST at 1 km resolution with twice-daily overpasses. Useful for city-wide UHI magnitude and multi-year climatology but cannot resolve intra-urban block-level differences. Best used as a calibration anchor or temporal gap-filler rather than the primary spatial layer.
Why block-level resolution changes the public-health question
A 1 km MODIS pixel averages across a park, a car park and a social-housing estate simultaneously. The average may look tolerable. The estate is not. During the 2003 European heatwave, excess mortality in Paris was concentrated in specific arrondissements where elderly residents lived in top-floor flats without air conditioning, surrounded by stone and asphalt. A city-wide temperature map would not have found them. A 30 m Landsat TIRS pass would have.
ECOSTRESS at 70 m and Landsat TIRS at 30 m are the two sensors that currently provide intra-urban thermal differentiation at a scale meaningful for public-health intervention. The difference between a shaded street with mature trees and an exposed concrete courtyard 80 metres away can exceed 10°C in surface temperature on a calm sunny afternoon. That gap corresponds to measurable differences in physiological heat stress for anyone spending time outdoors, and to substantially different indoor temperatures in poorly insulated buildings.
LST is not air temperature. The gap matters, and it varies.
Land surface temperature is what the sensor sees: the radiative emission from the surface itself, whether that is roofing felt, grass or tarmac. Air temperature at 2 m, which is what the body actually experiences and what weather stations report, is related but not identical. The conversion from LST to a meaningful human-exposure estimate depends on surface emissivity, which varies between roughly 0.92 for dense vegetation and 0.97 for water, and on sky-view factor, which describes how much of the sky hemisphere a given point on the ground can see. A narrow alley with tall buildings on both sides has a low sky-view factor; it traps longwave radiation at night and may stay hotter than an open plaza even if its daytime LST appears similar.
Emissivity must be estimated, usually from land-cover classification or from the NDVI-based methods published by Sobrino and colleagues. Errors in emissivity of 0.01 translate to LST errors of roughly 0.5°C, which is small at the extremes of a heatwave but non-trivial when comparing similar surface types. Users should treat block-level LST comparisons as directionally reliable and quantitatively approximate, not as thermometer readings.
Building the vulnerability index: what goes in and why
Surface temperature is one input. The vulnerability index combines it with at least three other data layers. Impervious surface fraction, derived from Sentinel-2 spectral unmixing or supervised classification, captures how much of a block is sealed against evaporative cooling. Tree canopy cover, also from Sentinel-2 or from GEDI lidar canopy height products where available, captures shading and latent heat flux. These two together describe the physical heat exposure environment at block level.
The social side of vulnerability comes from census data: the proportion of residents aged over 65 (who have impaired thermoregulation), the proportion living alone (who may not be checked on during an event), housing type and construction era (pre-1970s masonry without cavity insulation retains heat overnight), and income proxies that correlate with air-conditioning ownership. Combining physical and social layers into a composite index requires explicit weighting decisions. There is no universally correct weighting; the published literature, including the work underpinning the US CDC Social Vulnerability Index and European adaptations, generally uses equal weighting across normalised sub-indices as a transparent default, with sensitivity testing across alternatives.
The output is a ranked map of city blocks by composite vulnerability. It does not predict mortality counts. It identifies where the risk is disproportionately concentrated relative to the city average, which is the operationally useful question for emergency planners deciding where to open cooling centres or deploy welfare checks.
Honest limits of the satellite input
Cloud is the most immediate problem. Heatwaves in continental climates often arrive under clear skies, which is fortunate. Coastal and tropical cities may experience humid heat events under partial cloud, and a cloudy Landsat overpass during the critical peak-temperature day is simply lost. Multi-sensor compositing, using ECOSTRESS passes alongside Landsat, reduces but does not eliminate this gap.
Revisit is the second constraint. Landsat's 16-day repeat means the thermal snapshot may be days old when a heatwave strikes. ECOSTRESS's variable overpass time is an advantage here: its ISS orbit produces passes at different times of day across successive visits, which improves the chance of capturing a peak-heat overpass during a multi-day event, but it cannot be tasked to a specific overpass time. Neither sensor provides the real-time thermal monitoring that a dense ground network would offer. The satellite layer is best understood as a structural vulnerability map, updated seasonally or annually, rather than a live heat alert.
Finally, the vulnerability index is a spatial prioritisation tool, not an epidemiological model. Excess mortality during a specific event depends on factors the index cannot see: whether cooling centres were open, whether a social media warning reached residents, whether the event duration exceeded acclimatisation thresholds. Interpret the map accordingly.
From map to municipal action
The practical use case is pre-event planning, not real-time response. A city public-health authority commissions the vulnerability map in spring, before the heatwave season. The map identifies the ten or twenty highest-risk census blocks. Welfare check rosters, cooling centre locations and outreach budgets are allocated against that ranked list. When a heatwave warning is issued by the national meteorological service, the response is already spatially targeted rather than city-wide and diffuse.
A secondary use is infrastructure investment prioritisation. Urban greening programmes, cool-roof subsidies and shade-structure grants have limited budgets. A vulnerability map makes the case, in spatial terms, for directing those resources to the blocks where the physical and social exposure coincide. Satellize produces these composite indices for government clients as GIS-ready layers with accompanying methodology reports; the analytics pipeline runs on open Sentinel and Landsat data, with ECOSTRESS accessed through NASA's Earthdata portal. The Tonga crop-estimation programme demonstrated the same principle at national scale: open satellite data, rigorously processed, answering a specific government planning question.
Typical figures
| Primary LST spatial resolution | 30 m (Landsat 8/9 TIRS, resampled from 100 m native); 70 m (ECOSTRESS) |
| Impervious/canopy layer resolution | 10 m (Sentinel-2 MSI visible and NIR bands) |
| Revisit (Landsat 8 + 9 combined) | 8 days at equator under clear sky; 16 days per satellite |
| ECOSTRESS revisit | Irregular, approximately 1–5 days at mid-latitudes; overpass time varies across diurnal cycle |
| LST accuracy (published) | ±1–2°C for Landsat TIRS under clear sky; emissivity uncertainty adds ±0.5°C per 0.01 emissivity error |
| Thermal spectral bands | Landsat TIRS: 10.6–11.19 µm (band 10), 11.50–12.51 µm (band 11); ECOSTRESS: five bands 8.28–12.13 µm |
| Archive depth | Landsat thermal: 1984–present (TM, ETM+, TIRS); ECOSTRESS: 2018–present; Sentinel-2: 2015–present |
| Cloud impact | Any cloud or thick haze invalidates thermal retrieval for affected pixels; no gap-fill from radar (SAR carries no thermal signal) |
| Minimum spatial unit for reliable block comparison | Approximately 90 m × 90 m (3 × 3 Landsat pixels) to reduce per-pixel noise in composite statistics |
| Delivery formats | GeoTIFF (LST and index layers), GeoPackage (block-level ranked polygons), PDF methodology report |
Analytics Satellize can run
| Block-level land surface temperature composite | Multi-date clear-sky median or peak-percentile compositing from Landsat TIRS and ECOSTRESS, with split-window or single-channel LST retrieval and NDVI-based emissivity correction | GeoTIFF raster at 30 m, clipped to city boundary, with seasonal peak and mean layers |
| Impervious surface fraction map | Spectral unmixing or random-forest classification of Sentinel-2 10 m bands, trained on publicly available urban land-cover datasets | GeoTIFF fraction layer (0–1 scale) at 10 m, aggregated to census block polygons |
| Tree canopy cover layer | NDVI thresholding and texture analysis from Sentinel-2, optionally supplemented by GEDI canopy height where ISS coverage permits | GeoTIFF canopy fraction at 10 m, aggregated to block level |
| Composite heat-health vulnerability index | Normalised combination of LST percentile rank, impervious fraction, canopy deficit, and census social indicators (age, isolation, housing era, income proxy) using equal-weight default with documented sensitivity variants | GeoPackage of ranked census blocks with per-indicator scores and composite score, ready for GIS import |
| Priority intervention zone report | Threshold-based selection of highest-composite-score blocks, with descriptive statistics on dominant vulnerability drivers per zone | PDF report with maps, block-level tables and plain-language findings for non-specialist municipal audiences |
| Annual vulnerability trend update | Year-on-year comparison of LST composite and canopy cover layers to detect greening or densification changes; social indicators updated when new census data are released | Updated GeoPackage and one-page change summary flagging blocks with materially worsened or improved scores |
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.