Thermal infrared land surface temperature retrieval
Thermal infrared sensors measure emitted radiance from the Earth's surface, which is converted to land surface temperature through emissivity correction and atmospheric adjustment. Resolution, revisit and cloud cover set hard limits that no processing chain can overcome.
Sensors
- Landsat 8/9 TIRS: Two thermal bands centred near 10.9 µm and 12.0 µm, native detector resolution 100 m resampled to 30 m in distributed products. 16-day single-satellite revisit, extended to roughly 8 days with both Landsat 8 and 9 operating together. Free archive back to 2013 (Landsat 8) via USGS.
- ECOSTRESS (ISS-mounted): Five thermal bands spanning 8.3 to 12.5 µm at approximately 70 m ground sampling distance. ISS orbital precession gives variable overpass times, which is scientifically useful for diurnal temperature studies but means revisit at a given site is irregular, typically every few days to a week. Captures land surface temperature and evapotranspiration products.
- ASTER TIR: Five thermal bands from 8.125 to 11.65 µm at 90 m resolution aboard Terra. On-demand acquisition rather than systematic coverage; the instrument has operated since 1999, giving a deep archive for change analysis. ASTER's temperature-emissivity separation algorithm is a published reference method.
- Satellogic thermal band: Commercial smallsat constellation offering a thermal channel alongside multispectral imagery, enabling same-pass optical and thermal comparison. Tasking on demand. Published ground sampling distance is approximately 1 m for optical bands; thermal resolution is coarser, consistent with detector physics at smallsat aperture sizes.
What the sensor is actually measuring
Every surface above absolute zero emits thermal radiation. In the 8 to 14 µm atmospheric window, the atmosphere is relatively transparent, and spaceborne sensors can detect that emitted radiance directly. This is fundamentally different from reflected sunlight. Thermal sensors work at night just as well as by day, which is one of their practical advantages over multispectral optical instruments.
The raw quantity recorded is at-sensor radiance: a mix of surface emission, atmospheric emission, and atmospheric absorption along the path. Converting that to land surface temperature requires two corrections. First, emissivity: real surfaces are not perfect blackbodies. Dry bare soil emits at roughly 0.95 of the theoretical maximum; dense vegetation is close to 0.98; metal roofing can drop below 0.85. If emissivity is assumed incorrectly, the temperature estimate shifts by 1 to 3 K, which is enough to invalidate urban heat island comparisons or irrigation stress assessments. Second, the atmospheric profile: water vapour and trace gases absorb and re-emit in the thermal window. Correcting for this requires either a radiosonde profile, a reanalysis product such as MERRA-2 or ERA5, or a split-window algorithm that exploits the differential absorption between two adjacent thermal bands. Landsat TIRS and ECOSTRESS both support split-window approaches, though the method introduces its own uncertainty of roughly 1 to 2 K under high-humidity conditions.
Resolution and revisit: the trade-off no one escapes
Detector physics creates a direct tension between spatial resolution and signal-to-noise ratio in the thermal infrared. Smaller pixels collect less radiance, so thermal sensors are almost always coarser than their optical siblings on the same platform. Landsat TIRS illustrates this: the optical bands resolve 30 m natively, while TIRS detectors are 100 m, resampled in the standard product to match the 30 m grid. The resampling sharpens the geometry but does not recover spatial detail that was never measured.
ECOSTRESS at 70 m is the current best freely available resolution for systematic thermal coverage. Commercial thermal smallsats can improve on this, though published specifications should be read carefully: a 1 m optical sensor paired with a thermal channel does not deliver 1 m thermal data. At the aperture sizes typical of smallsats, thermal ground sampling distances of 3.5 to 5 m are the current practical frontier for commercial systems.
Revisit is the other constraint. Landsat's 8-day effective revisit (with both satellites) sounds adequate until a heat-stress event lasts 36 hours, or an industrial discharge happens between passes. ECOSTRESS offers more frequent but irregular coverage. For time-critical applications, the honest answer is that no single free-access thermal constellation provides daily sub-100 m coverage globally. Combining sensors or accepting coarser resolution from MODIS or VIIRS (1 km, near-daily) is often the practical compromise.
Cloud cover is not a processing problem
Thermal emission from the surface cannot pass through cloud. A thick cumulus deck radiates at its own temperature, which is what the sensor records. No algorithm, no matter how sophisticated, reconstructs surface temperature beneath opaque cloud from a single thermal observation. This is a physical limit, not a data-quality issue.
The practical consequences are significant for tropical and maritime applications. In persistently cloudy regions, Landsat may yield only a handful of cloud-free thermal acquisitions per year. Temporal compositing helps: taking the minimum or maximum temperature across a season to map cold irrigation signatures or peak urban heat. But composites lose the time-stamp that makes many applications useful. Radar-based proxies for soil moisture can partially substitute in some use cases, though they measure a different physical quantity. Buyers should ask, before commissioning any thermal analytics programme, what the cloud-free acquisition frequency actually is for their region of interest.
Emissivity separation: the hidden variable
Land surface temperature and emissivity cannot both be retrieved from a single thermal measurement. Every retrieval algorithm has to make an assumption or use additional information to break this degeneracy. The three main approaches in operational use are: the split-window algorithm (exploits differential absorption across two bands, requires known emissivity); the temperature-emissivity separation method developed for ASTER (uses multiple bands and the empirical relationship between emissivity minimum and spectral contrast); and the single-channel method (uses an external atmospheric profile and a land-cover-derived emissivity map).
Each method has a published uncertainty budget. ASTER TES is generally cited at 1.5 K or better under good atmospheric conditions. Landsat TIRS single-channel retrievals using ERA5 profiles achieve similar accuracy over homogeneous surfaces; accuracy degrades over mixed pixels containing water bodies, urban materials and vegetation within a single 100 m cell. Users applying thermal data to sub-pixel targets should be cautious: a 100 m pixel containing 20% rooftop and 80% grass reports an area-weighted radiance, not the rooftop temperature.
What the numbers are actually used for
Urban heat island mapping is the most common application. Comparing land surface temperature across land-cover classes within a city, or tracking how a district warms over a decade as vegetation is replaced by impervious surface, is well within the capability of Landsat TIRS. The 30 m resampled product resolves individual city blocks in most urban grids.
Agricultural stress detection uses the relationship between canopy temperature and evapotranspiration. A stressed crop, unable to cool itself by transpiring water, runs warmer than a well-irrigated neighbour. ECOSTRESS was designed partly for this purpose and produces evapotranspiration products directly. For irrigation management at field scale, 70 m resolution is borderline adequate for large fields and insufficient for small-holder plots typical of South and Southeast Asia.
Industrial heat source monitoring, volcano thermal anomaly tracking, and wildfire perimeter mapping all use thermal data, though at different intensity scales. Active fire detection typically relies on MODIS or VIIRS at 375 m to 1 km, not Landsat, because the near-daily revisit matters more than spatial resolution for a moving perimeter. Geothermal and industrial anomalies are detectable at Landsat resolution if the heat source occupies a meaningful fraction of a pixel and the temperature contrast above background is at least a few Kelvin.
Satellize runs thermal retrievals on Landsat and ECOSTRESS data as part of its analytics stack, applying split-window and single-channel methods with ERA5 atmospheric correction. The Tonga crop-estimation programme draws on multispectral vegetation indices rather than thermal bands, but the same pipeline architecture supports thermal stress layers where acquisition conditions allow.
Choosing the right sensor for the question
The decision tree is shorter than it looks. If the target is large (city, agricultural district, industrial zone), cloud-free acquisitions are plausible in your region, and daily revisit is not required, Landsat TIRS at 30 m resampled offers a free, well-calibrated, 10-year archive and should be the starting point. If diurnal variation matters, ECOSTRESS adds value through its variable overpass time, though the irregular revisit complicates time-series analysis.
If the target is smaller than roughly 50 m, or if you need same-pass optical and thermal comparison for material identification, commercial tasking is the only current option, and the thermal resolution of the specific commercial system needs to be confirmed against your minimum detectable target size before procurement. If your region has persistent cloud cover for more than half the year, build that into the expected data yield from the outset. A thermal analytics programme that assumes cloud-free monthly acquisitions in equatorial West Africa will underdeliver.
Typical figures
| Spatial resolution (Landsat TIRS) | 100 m native detector, resampled to 30 m in standard products |
| Spatial resolution (ECOSTRESS) | ~70 m ground sampling distance |
| Spatial resolution (ASTER TIR) | 90 m, five bands from 8.125 to 11.65 µm |
| Revisit (Landsat 8+9 combined) | ~8 days at equator, cloud-free acquisitions vary by region |
| Revisit (ECOSTRESS) | Irregular, typically every few days to ~1 week per site due to ISS precession |
| Spectral bands | Landsat TIRS: 10.6–11.2 µm and 11.5–12.5 µm; ECOSTRESS: 5 bands 8.3–12.5 µm; ASTER: 5 bands 8.125–11.65 µm |
| Temperature retrieval accuracy (typical) | 1 to 2 K under clear-sky, low-humidity conditions; degrades to 2 to 3 K under high atmospheric water vapour |
| Cloud penetration | None. Thermal infrared is fully blocked by opaque cloud |
| Archive depth | Landsat TIRS from 2013 (L8) and 2021 (L9); ASTER from 1999; ECOSTRESS from 2018 |
| Delivery formats | GeoTIFF (standard), NetCDF for time-series stacks, GIS-ready raster layers |
Analytics Satellize can run
| Urban heat island temperature map | Split-window or single-channel LST retrieval from Landsat TIRS with ERA5 atmospheric correction; land-cover-stratified emissivity assignment | Seasonal GeoTIFF raster with per-pixel LST and anomaly layer relative to urban-rural baseline |
| Agricultural thermal stress index | Crop Water Stress Index derived from canopy temperature relative to air temperature, using ECOSTRESS or Landsat thermal retrievals | Field-level stress classification layer in GIS format, delivered within 48 hours of cloud-free acquisition |
| Industrial heat anomaly alert | Temporal baseline subtraction on Landsat TIRS time series; anomaly flagging where pixel temperature exceeds seasonal mean by a configurable threshold | Alert report with coordinates, anomaly magnitude and acquisition timestamp |
| Multi-year urban warming trend | Landsat TIRS archive compositing (annual cloud-free maximum or mean) with linear trend fitting per pixel | Decadal trend map (K per year) as GeoTIFF with accompanying summary statistics report |
| Evapotranspiration estimation | ECOSTRESS Level-3 ET product or Penman-Monteith energy balance approach combining LST, NDVI and meteorological reanalysis | Monthly or seasonal ET raster (mm per day) for specified area of interest |
| Cloud-gap frequency assessment | QA band analysis across Landsat and Sentinel-2 archive for target region to quantify usable thermal acquisition frequency before programme commitment | Site suitability report with monthly cloud-free acquisition probability and recommended sensor strategy |
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.