Urban heat island mapping from thermal infrared
Thermal infrared satellites measure land surface temperature across urban fabrics, revealing heat-island cores, cool corridors, and the stark contrast between asphalt and tree canopy. Resolution limits are real and matter for city planning.
Sensors
- Landsat 8/9 TIRS: Two thermal bands (Band 10 at 10.6–11.19 µm, Band 11 at 11.50–12.51 µm) at 100 m native resolution, resampled to 30 m in delivered products. 16-day revisit per satellite; combined Landsat 8 and 9 constellation halves that to roughly 8 days. The primary workhorse for city-scale LST time series.
- ASTER TIR: Five thermal bands (8.125–11.65 µm) at 90 m spatial resolution with on-board blackbody calibration. Better emissivity separation than single-band sensors, but ASTER acquires on request rather than continuously; archive density varies strongly by city. The instrument is ageing and acquisition scheduling has become less predictable since 2019.
- ECOSTRESS (ISS): Four TIR bands centred near 8.29, 8.78, 9.2, and 10.49 µm at approximately 70 m spatial resolution, finer than Landsat TIRS. Critically, the ISS non-sun-synchronous orbit means ECOSTRESS captures multiple times of day, including afternoon peak-heat acquisitions that fixed sun-synchronous sensors miss. Revisit is irregular, ranging from same-day to gaps of weeks for a given city, which limits its use for systematic monitoring.
- Sentinel-3 SLSTR: Two TIR channels (10.85 µm and 12.0 µm) at 1 km spatial resolution, with a 1–2 day revisit. Useful for regional heat-stress context and time-series analysis of large metropolitan areas, but the 1 km pixel renders individual city blocks invisible. Best treated as a synoptic backdrop rather than a planning tool.
Why asphalt radiates and grass does not
Every surface above absolute zero emits longwave radiation. The Stefan-Boltzmann law governs the total flux; for the temperature ranges found in cities (roughly 280–340 K), peak emission falls between 8 and 14 µm, squarely in the thermal infrared window that satellite sensors are designed to detect. Land surface temperature (LST) derived from these measurements is not the same as air temperature recorded by a weather station. It is the radiometric temperature of the surface itself: a dark asphalt car park in July can exceed 60 °C while the air two metres above it sits at 35 °C.
The urban heat island effect is the aggregate result of several physical processes. Impervious surfaces absorb more shortwave solar radiation than vegetated ground and release it slowly as longwave heat after sunset. Reduced evapotranspiration, because there is little moisture to evaporate, removes a major cooling mechanism. Waste heat from vehicles, air conditioning and industrial processes adds a further anthropogenic load. Satellite thermal imagery does not separate these contributions, but it maps their combined surface expression with precision, which is what city planners and public-health agencies actually need.
The emissivity problem no one mentions in the brochure
A satellite sensor measures radiance, not temperature. Converting radiance to LST requires knowing the emissivity of each surface, the fraction of theoretical blackbody radiation it actually emits. Bare concrete has an emissivity of roughly 0.92–0.96; dry soil around 0.93; dense vegetation close to 0.98; standing water near 0.99. A mixed pixel containing rooftop, road and a strip of trees will have an effective emissivity that is the area-weighted average of its components, and that average is not directly observable.
Two standard correction approaches exist in the public literature. The NDVI-based emissivity method uses a coincident shortwave vegetation index to estimate fractional vegetation cover and then assigns emissivities by land-cover class. The temperature-emissivity separation (TES) algorithm, developed for ASTER, solves a system of equations across multiple TIR bands simultaneously to retrieve both temperature and emissivity without assuming land cover. TES is more physically rigorous but requires at least five TIR bands; Landsat TIRS, with two usable bands, cannot apply it directly. This is not a minor caveat. Emissivity errors of 0.01 translate to LST errors of approximately 0.5 K, which is meaningful when the policy question is whether a neighbourhood is 2 K or 3 K hotter than a reference park.
What 100 metres actually resolves, and what it does not
Landsat TIRS pixels are 100 m before resampling. A single pixel therefore covers 10,000 square metres, roughly the footprint of a medium city block. Individual buildings, narrow street canyons, small parks and green roofs are subpixel features whose thermal signal is averaged into the surrounding fabric. This means Landsat LST maps are reliable for neighbourhood-scale and district-scale heat-island analysis, and for tracking change in those patterns over years or decades. They are not reliable for auditing a single building's roof temperature or identifying a 20-metre-wide cool corridor between two streets.
ECOSTRESS at approximately 70 m offers a modest improvement in spatial detail and, more usefully, samples at different times of day. Peak urban surface temperatures typically occur between 14:00 and 16:00 local time. Landsat's fixed equatorial crossing time of approximately 10:00–10:30 local time means it routinely misses the daily maximum. An afternoon ECOSTRESS acquisition over the same city can show LST values 8–12 K higher than the morning Landsat pass, which changes the apparent severity of the heat island considerably. The trade-off is that you cannot guarantee an ECOSTRESS acquisition on a given date.
For the finest spatial detail, airborne TIR cameras or drone-mounted sensors can reach sub-metre resolution, but they are expensive, cover small areas, and are outside the scope of satellite-based monitoring. The honest position is that satellite thermal data answers city-wide and district-wide questions well, and block-level questions poorly.
Reading the thermal map: cores, corridors and the park effect
A well-processed LST map of a city typically shows a clear hierarchy. Dense commercial and industrial districts form the hottest cores, often 4–8 K above the rural reference temperature in mid-latitude cities during summer, though published values vary widely by climate zone, city morphology and season. Major roads radiate outward as warm corridors. Residential suburbs with tree cover sit several degrees cooler. Parks, particularly those larger than a few hectares, show a measurable cool island effect that extends a short distance beyond their boundaries into adjacent streets.
The analytical value is in the spatial relationships, not just the absolute temperatures. Overlaying LST with population density data identifies which communities bear the greatest heat exposure. Overlaying with green-space maps (a separate analysis, covered in the Urban green-space audit page in this library) quantifies the cooling deficit in underserved districts. Overlaying with impervious surface fraction shows how strongly surface sealing correlates with temperature, which gives planners a defensible basis for setting targets in local development plans.
Archive depth and change detection
Landsat's thermal archive extends to Landsat 5 TM (operational from 1984), giving a potential 40-year LST record for cities covered by cloud-free imagery. In practice, Landsat 5 and 7 TM/ETM+ thermal bands were single-channel at 120 m and 60 m respectively, with different calibration characteristics than TIRS, so multi-decadal comparisons require careful inter-sensor normalisation. The USGS provides surface temperature products for Landsat Collection 2 that apply consistent atmospheric correction across the archive, which significantly reduces the burden on the analyst.
Meaningful change detection requires careful scene selection: same sensor, same season, similar atmospheric conditions and similar solar geometry. Comparing a July 2010 Landsat 7 scene with an August 2023 Landsat 9 scene without correcting for these variables will produce apparent temperature changes that are partly artefact. Satellize's analysts apply scene-pairing protocols and cross-validate against meteorological station records before drawing conclusions about long-term urban warming trends. The Tonga crop-estimation programme uses similar multi-temporal calibration discipline, applied to a very different sensor suite, but the underlying rigour is the same.
Honest limits and where the method breaks down
Cloud cover blocks thermal infrared completely. In humid tropical and monsoon climates, finding cloud-free imagery during the hottest months, precisely when heat-island analysis is most urgent, can be genuinely difficult. Compositing multiple acquisitions over a season reduces the problem but introduces temporal averaging that blurs the peak-heat signal.
Atmospheric water vapour absorbs and re-emits TIR radiation, introducing errors that must be corrected using either concurrent radiosonde profiles or modelled atmospheric data. The USGS Collection 2 LST product uses MERRA-2 reanalysis for this correction; errors in the reanalysis propagate into the LST. In arid cities with very low humidity, this matters less. In coastal or tropical cities, it can introduce uncertainties of 1–2 K that are difficult to eliminate entirely.
Finally, LST is a surface property. It tells you the temperature of the roof, not the indoor temperature of the building below it, and not the air temperature experienced by a pedestrian. Translating LST into human heat-stress metrics requires combining it with air temperature, humidity and wind data from other sources. Satellite thermal imagery is a powerful input to that analysis, not a complete answer on its own.
Typical figures
| Spatial resolution (Landsat 8/9 TIRS) | 100 m native, delivered at 30 m resampled |
| Spatial resolution (ECOSTRESS) | ~70 m |
| Spatial resolution (ASTER TIR) | 90 m |
| Spatial resolution (Sentinel-3 SLSTR) | 1 km |
| Revisit (Landsat 8 + 9 combined) | ~8 days at equator, weather permitting |
| Revisit (ECOSTRESS) | Irregular; hours to weeks depending on ISS orbit and tasking |
| Spectral bands used | 8–14 µm thermal infrared window; split-window or multi-band TES methods where available |
| LST retrieval accuracy (Landsat Collection 2) | ±1–2 K under clear-sky, low-humidity conditions; larger errors in humid or hazy conditions |
| Archive depth | Landsat thermal back to 1984 (Landsat 5 TM); ASTER from 2000; ECOSTRESS from 2018 |
| Minimum detectable heat-island signal | ~1 K above reference, subject to atmospheric correction quality |
Analytics Satellize can run
| City-wide LST map | Single-channel or split-window LST retrieval from Landsat TIRS, with NDVI-based emissivity correction and USGS Collection 2 atmospheric correction | GeoTIFF LST layer at 30 m, delivered per acquisition with cloud-mask applied |
| Heat-island intensity report | Zonal statistics comparing LST of urban land-cover classes against a rural reference polygon; percentile distributions by district | PDF report with ranked district table and map, updated seasonally or on request |
| Multi-year LST trend layer | Scene-paired, same-season compositing across Landsat archive with inter-sensor normalisation; linear trend fitting per pixel | GeoTIFF trend raster (K per decade) with uncertainty band layer |
| Cool-corridor and park-effect analysis | Spatial buffer analysis around green-space polygons; LST transect extraction and regression against distance from vegetation edge | GIS layer showing cooling radius by park, plus summary statistics for planning input |
| Afternoon peak-heat composite (ECOSTRESS) | Aggregation of available ECOSTRESS acquisitions in afternoon window (13:00–17:00 local); quality filtering by cloud fraction | Best-available afternoon LST composite GeoTIFF with acquisition date metadata |
| Heat-exposure equity overlay | LST layer intersected with census population grid; identification of high-LST, high-density zones lacking green-space within 500 m | Priority-zone shapefile for planning department use, with supporting methodology note |
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.