Satellite land surface temperature for heatwave mortality lag-exposure modelling

Weather-station networks are too sparse to reconstruct the spatially continuous heat-exposure histories that distributed lag non-linear models require. MODIS and Landsat land surface temperature time series fill that gap, letting health authorities calibrate city-specific alert thresholds against locally observed mortality.

Sensors

MODIS MOD11A1 / MYD11A1 (Terra and Aqua): Daily land surface temperature at 1 km spatial resolution, retrieved in thermal infrared bands 31 and 32 (approximately 10.78–12.27 µm) via a split-window algorithm. Terra overpass is around 10:30 local solar time and Aqua around 13:30, giving two daytime and two night-time observations per day. Cloud cover voids pixels; in persistently cloudy tropical cities gap-filling is required and can introduce uncertainty of 1–3 K.
Landsat Collection 2 Surface Temperature (Landsat 8 and 9 TIRS): Land surface temperature at 100 m (resampled to 30 m) in a single thermal band (band 10, 10.6–11.19 µm), with a 16-day revisit per satellite and an 8-day combined revisit for Landsat 8 plus 9. Absolute accuracy is cited by USGS as within approximately ±2 K under clear-sky conditions. Infrequent revisit limits its standalone use for daily lag modelling; it is most valuable for calibrating MODIS pixels and mapping intra-urban thermal heterogeneity.
ERA5 reanalysis (ECMWF): Hourly near-surface air temperature and dewpoint at roughly 31 km grid spacing. ERA5 does not measure land surface temperature directly but provides the meteorological covariates (wind, humidity, boundary-layer height) needed to bias-correct satellite LST toward apparent air temperature and to construct wet-bulb or apparent-temperature exposure metrics. Free via the Copernicus Climate Data Store.
VIIRS Day/Night Band (Suomi-NPP / NOAA-20): Visible near-infrared radiance at approximately 750 m resolution, used as a proxy for population density and economic activity in cities where census data are coarse or outdated. Nighttime light intensity correlates with built-up density, helping weight LST exposure estimates by the population actually present in each pixel.

Why the two-to-three-day lag changes everything about the data requirement

Epidemiological studies of heatwave mortality, including the landmark analyses of the 2003 European event and subsequent multi-city studies in Asia and Latin America, consistently find that excess deaths peak one to three days after the hottest day. The distributed lag non-linear model (DLNM) framework, developed by Gasparrini and colleagues and now standard in environmental epidemiology, estimates this lag structure statistically. To do that honestly, it needs a temperature exposure time series at the location of each death or postcode, not at the nearest weather station.

In many low-income cities, weather stations are spaced tens of kilometres apart and are biased toward airports and administrative centres. A station at an airport records the temperature over concrete and open tarmac; a resident two kilometres away in a dense informal settlement with metal roofing and no tree cover may experience a land surface temperature several degrees higher. That gap is not a modelling nuisance. It is the signal the health authority is trying to measure.

What MODIS gives you and where it fails

MODIS MOD11A1 is the workhorse for daily LST time series. Its 1 km pixel is coarse enough to average across mixed surfaces but fine enough to distinguish a park from an adjacent residential block in most cities. The archive runs from 2000 for Terra and 2002 for Aqua, giving more than two decades of data against which to fit lag models and validate alert thresholds. That archive depth matters: a threshold calibrated on a single anomalous summer will overfit.

The honest limits are significant. Cloud cover voids pixels entirely; in monsoon-affected cities, the hottest weeks of pre-monsoon season may be well observed, but a city in the humid tropics can lose 60–80 percent of daily observations to cloud in summer months. Gap-filling methods such as temporal interpolation, spatial regression against ERA5, or machine-learning reconstruction from partial observations each introduce their own uncertainty. Published studies typically report gap-filled LST errors of 1–3 K under moderate cloud conditions, rising steeply when cloud persists for more than three consecutive days. Any mortality model built on gap-filled data should propagate that uncertainty explicitly rather than treating the filled values as observed.

Landsat as a calibration instrument, not a daily thermometer

Landsat's 100 m thermal band resolves the intra-urban thermal heterogeneity that MODIS cannot. A single MODIS pixel over a dense city may contain a market square, a school with a metal roof, a small park, and a canal. Landsat separates them. This matters for exposure assignment: a study that assigns a single MODIS LST to all residents within a square kilometre will misclassify the park-adjacent population as more exposed than they are and the canal-adjacent population as less exposed.

The practical workflow is to use Landsat scenes from clear-sky days to build a disaggregation model, relating the MODIS 1 km value to the distribution of Landsat 100 m values within it as a function of land cover fractions. That model is then applied to MODIS daily data to produce synthetic 100 m daily LST fields. The 16-day revisit means that in any given summer, a city may have only four to six usable Landsat scenes for calibration. That is usually enough to build a stable disaggregation relationship, provided the land cover does not change substantially between scenes.

Building the exposure history: from pixel to person

A DLNM requires a time series of exposure at the unit of analysis, typically a postcode, district, or hospital catchment. The construction steps are: extract daily MODIS LST for each spatial unit, gap-fill using ERA5-anchored interpolation, disaggregate to finer resolution where Landsat calibration data are available, then compute the exposure metrics the model needs. Common metrics include the daily maximum LST, the overnight minimum (which governs physiological recovery), and the cumulative heat load over a rolling window.

Population weighting is non-trivial. Gridded population products such as WorldPop or LandScan distribute census totals across space, but they are updated infrequently and can be substantially wrong in fast-growing informal areas. VIIRS nighttime light provides a higher-frequency proxy for relative density that can adjust population weights between census years. The resulting exposure estimate is still an approximation. Researchers should test sensitivity to alternative weighting schemes before reporting lag-structure results to health authorities.

Calibrating alert thresholds: what the satellite record can and cannot settle

Once a DLNM is fitted with satellite-derived exposure histories, the output is a city-specific exposure-response function relating LST to relative mortality risk at each lag day. Health authorities can use this to set alert thresholds: for example, the LST value at which the two-day-lagged mortality risk exceeds a chosen relative-risk level. Because the satellite archive extends back to 2000, the model can be fitted on historical heatwave years and validated on held-out years, giving some honest estimate of out-of-sample performance.

What the satellite record cannot settle is causation between LST and mortality as opposed to air temperature. Land surface temperature and air temperature are physically related but not identical: LST is a radiometric skin temperature of the surface, while the body responds to air temperature and humidity. The conversion between the two depends on surface emissivity, wind speed, and boundary-layer mixing, all of which vary by location and time of day. Studies that use LST as a direct mortality predictor without this conversion are making an implicit assumption that the LST-to-air-temperature relationship is spatially uniform within the city. That assumption should be tested, not accepted by default.

Satellize applies this workflow for public-health clients using open Sentinel, Landsat, and MODIS data, with ERA5 bias correction and VIIRS population proxies. The Tonga crop-estimation programme demonstrated the same underlying approach of combining coarse-daily and fine-occasional satellite records to produce spatially continuous time series where ground observations are sparse.

Honest limits of the whole approach

Three limits deserve plain statement. First, LST is a clear-sky measurement. During the heatwave itself, if skies are partly cloudy, the hottest pixels may be the ones most likely to be missing. Gap-filling during the event window, not just around it, is where uncertainty is highest and where it matters most for the mortality model.

Second, the MODIS 1 km pixel is too coarse to resolve the thermal environment of a single street or courtyard. For studies focused on individual-level exposure rather than district-level, satellite LST is a useful covariate but not a substitute for dense ground sensors or high-resolution commercial thermal imagery.

Third, the DLNM framework estimates association, not mechanism. A well-fitted model with satellite exposure data can tell a health authority that mortality in their city rises approximately X percent per degree of two-day-lagged LST above a threshold. It cannot tell them whether the driver is dehydration, cardiovascular strain, or disrupted sleep from overnight heat. That distinction matters for intervention design, and it requires data the satellite cannot provide.

Typical figures

MODIS LST spatial resolution	1 km (MOD11A1 / MYD11A1)
MODIS LST revisit	Daily (Terra ~10:30, Aqua ~13:30 local overpass; two daytime, two night-time observations per day under clear sky)
Landsat TIRS spatial resolution	100 m native, delivered resampled to 30 m (Collection 2)
Landsat revisit	16 days per satellite; approximately 8 days combined for Landsat 8 and 9
MODIS LST absolute accuracy (clear sky)	Approximately ±1 K under clear-sky conditions per MODIS validation studies; gap-filled values typically ±1–3 K
Landsat ST absolute accuracy	Within approximately ±2 K under clear-sky conditions (USGS Collection 2 specification)
ERA5 grid spacing	Approximately 31 km; hourly temporal resolution
MODIS archive depth	Terra from February 2000; Aqua from July 2002
Spectral bands used	MODIS bands 31–32 (10.78–12.27 µm, split-window); Landsat 8/9 band 10 (10.6–11.19 µm)
Cloud impact	LST pixels voided under cloud; tropical cities may lose 60–80% of summer observations without gap-filling

Analytics Satellize can run

Daily 1 km LST time series per city district	MODIS MOD11A1 extraction with ERA5-anchored temporal gap-filling using linear interpolation or random-forest reconstruction	NetCDF or GeoTIFF stack covering the requested historical window, with per-pixel data-quality flags and gap-fill fraction metadata
Disaggregated 100 m daily LST field	Landsat-MODIS spatial disaggregation using land-cover-fraction regression fitted on clear-sky Landsat scenes	Daily 100 m GeoTIFF rasters for the study period, with uncertainty band derived from the disaggregation model residuals
Population-weighted heat-exposure index by administrative unit	VIIRS nighttime-light-adjusted population weighting applied to disaggregated LST; rolling maximum and overnight minimum computed per unit	CSV time series per administrative unit, formatted for direct ingestion into R dlnm or Python statsmodels DLNM pipelines
Exposure-response function and lag-structure report	Distributed lag non-linear model (DLNM) fitted on satellite LST exposure histories and provided mortality counts, following the Gasparrini cross-basis framework	PDF report with exposure-response curves, lag-specific relative-risk estimates, and recommended alert thresholds at client-specified risk levels
Near-real-time LST anomaly alert	Daily MODIS acquisition compared against the 2000-to-present percentile distribution for the same calendar week; anomaly flagged when pixel exceeds the 90th or 95th percentile threshold	Daily GeoJSON alert layer and optional email digest, with two-day forward mortality-risk annotation derived from the fitted exposure-response function
Bias-correction layer aligning LST to local air temperature	Regression of MODIS LST against co-located station air temperature and ERA5 boundary-layer variables to produce a city-specific LST-to-Tair conversion surface	Gridded correction coefficients (GeoTIFF) and validation statistics against held-out station observations

Sources

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.