Forest evapotranspiration and rainfall interception estimation
Forest canopies return 30–70% of precipitation to the atmosphere through evapotranspiration and interception. Thermal sensors and energy-balance models now quantify that flux from orbit, giving water managers a number rather than an assumption.
Sensors
- ECOSTRESS (ISS-mounted): Thermal infrared radiometer at 70 m spatial resolution, measuring land surface temperature (LST) in five TIR bands (8–12.5 µm). Revisit is irregular, roughly 1–5 days at mid-latitudes depending on ISS orbital precession. The 70 m pixel is the finest routine thermal product available from orbit for forest AET work.
- MODIS MOD16A2: Global 8-day composited AET product at 500 m, derived from the Penman-Monteith equation driven by MODIS LST, LAI and albedo inputs. Covers 2000 to present; suited to catchment-scale water-balance accounting where sub-pixel forest detail is less critical than temporal depth.
- Landsat 8/9 TIRS: Thermal Infrared Sensor provides 100 m LST (resampled to 30 m in distribution) in two bands (10.6 and 12.5 µm). 16-day revisit per satellite; combined Landsat 8 and 9 gives an 8-day cycle. Long archive from 2013 (Landsat 8) supports multi-year AET trend analysis.
- Sentinel-2 MSI: No thermal capability, but 10 m red-edge and NIR bands supply NDVI, EVI and LAI inputs that sharpen the vegetation-resistance terms in SEBS and METRIC models. 5-day revisit at the equator. Cloud contamination in humid tropical forests can reduce usable acquisitions to fewer than 12 per year in some regions.
What a forest does with rain before it reaches the ground
A closed-canopy forest is not a passive surface. When rain falls, the canopy intercepts a fraction, holds it on leaf and bark surfaces, and returns it directly to the atmosphere by evaporation before any of it infiltrates the soil. Published interception studies across forest types report a range of 10–40% of gross rainfall lost this way, with tropical broadleaf forests typically at the high end and open boreal stands at the low end. Add transpiration through stomata and the total water flux leaving the canopy, actual evapotranspiration (AET), routinely exceeds 60–70% of annual precipitation in humid tropical zones.
That flux matters enormously for downstream water supply, reservoir yield forecasting and flood attenuation modelling. It also matters for carbon accounting, because stomatal conductance links CO₂ uptake to water loss. The problem historically was measurement: eddy-covariance flux towers give precise local readings but cover perhaps a few hundred hectares. Satellite-based AET estimation scales that to millions of hectares, at the cost of some precision.
How energy-balance models turn temperature into a water flux
The core physics is the surface energy balance: net radiation equals the sum of sensible heat flux, latent heat flux and soil heat flux. Latent heat flux is evapotranspiration in energy units. If you can measure or model the other three terms, AET falls out by residual. That is what SEBS (Surface Energy Balance System) and METRIC (Mapping EvapoTranspiration at high Resolution with Internalised Calibration) do.
Land surface temperature from a thermal sensor is the critical input. A canopy that is transpiring freely stays cool relative to air temperature; water stress or sparse cover raises LST. ECOSTRESS at 70 m resolves individual forest patches and canopy gaps that a 500 m MODIS pixel would average away. Sentinel-2 NDVI and LAI retrievals sharpen the aerodynamic and surface-resistance terms. The models are not black boxes: SEBS is documented in the peer-reviewed literature and METRIC was developed and published by the University of Idaho. Both carry uncertainty on the order of 10–20% at the daily scale under good conditions, rising sharply under cloud or when atmospheric correction of LST is imprecise.
Interception specifically is harder to retrieve directly from orbit. The standard approach is to run a Gash analytical interception model, parameterised with canopy cover fraction (from Sentinel-2) and free throughfall coefficient, then subtract modelled interception from gross AET to separate the two components. The Gash model requires rainfall input from gauge networks or gridded products such as CHIRPS or ERA5, adding another uncertainty source.
Honest limits: cloud, revisit and the 70 m floor
Thermal remote sensing and cloud cover are fundamentally incompatible. A single overcast day means no LST retrieval, and in wet tropical forests, cloud-free ECOSTRESS or Landsat TIRS acquisitions can be rare enough to force reliance on 8-day or monthly composites rather than daily values. Monthly AET totals are more reliable than daily estimates, but they obscure the within-month variability that matters for drought stress detection.
The 70 m ECOSTRESS pixel is the finest available routine thermal product, but it still averages across a heterogeneous canopy. A 70 m pixel over a forest edge will blend canopy and clearing signals. Landsat TIRS at 100 m (native) is coarser still. For forests fragmented into patches smaller than a few hundred metres, both sensors will underestimate AET because cooler, wetter forest interior pixels are diluted by warmer surroundings. MODIS at 500 m is effectively useless for fragmented landscapes and is best reserved for large, contiguous forest blocks or whole-catchment budgets.
Atmospheric correction of thermal data over humid forests is non-trivial. Water vapour in the atmospheric column attenuates the surface signal, and humid tropical regions have high and variable column water vapour. Errors of 1–2 K in LST propagate to AET errors of roughly 0.5–1 mm per day, which is significant at the daily scale.
Catchment water balance: what the numbers are actually used for
The practical output of forest AET estimation is a spatially distributed water-balance component that planners cannot get any other way at scale. Watershed managers use it to close the balance equation: precipitation minus AET minus change in storage equals runoff. If AET is wrong, runoff predictions are wrong, and reservoir operating rules based on those predictions carry hidden error.
Governments managing forested catchments as water-supply infrastructure, particularly in tropical regions where forests cover the headwaters of major rivers, need AET estimates to evaluate the hydrological consequences of deforestation or reforestation. A forest cleared for agriculture typically reduces AET, which sounds like more water available downstream, but the reduction in interception also increases storm-peak runoff and erosion. The satellite-derived AET map makes that trade-off spatially explicit rather than a matter of expert opinion.
Satellize has applied similar energy-balance and vegetation-index pipelines in its analytics work, including the Tonga crop-estimation programme, where separating crop water use from background evaporation was a necessary pre-processing step. Forest AET is a more demanding version of the same physics.
What a useful delivery looks like
A credible forest AET product for a water-management client needs at minimum: a spatial resolution appropriate to the catchment's forest fragmentation (ECOSTRESS or Landsat TIRS for fragmented landscapes, MODIS MOD16 for large contiguous blocks), a temporal cadence matched to the client's planning cycle (monthly for annual water budgets, 8-day for seasonal drought monitoring), and an uncertainty estimate that is not buried in a methods appendix.
Interception maps derived from Gash-model parameterisation add value when paired with rainfall data, because they separate the two components of AET that have different management implications. Interception is lost to the atmosphere regardless of soil moisture; transpiration is linked to forest health and productivity. Conflating them in a single AET number obscures decisions about forest management.
Delivery formats that actually get used are GeoTIFF time-series for GIS integration, tabular catchment-aggregated monthly totals for hydrological model ingestion, and a short interpretive note explaining which months had insufficient cloud-free thermal acquisitions and how gap-filling was handled. An honest product flags its own gaps.
Typical figures
| Finest thermal spatial resolution (routine) | 70 m (ECOSTRESS); 100 m native / 30 m resampled (Landsat 8/9 TIRS) |
| AET product spatial resolution (global) | 500 m (MODIS MOD16A2) |
| Vegetation index input resolution | 10 m (Sentinel-2 MSI, NDVI / EVI / LAI) |
| Revisit (thermal) | Irregular 1–5 days (ECOSTRESS, ISS-dependent); 8-day combined Landsat 8+9; 1–2 day MODIS |
| Temporal compositing period (MOD16) | 8-day and monthly composites; annual product also available |
| AET estimation uncertainty (daily, good conditions) | Approximately 10–20% with SEBS or METRIC; higher under cloud or high atmospheric water vapour |
| Archive depth | MODIS MOD16: 2000 to present; Landsat TIRS: 2013 to present; ECOSTRESS: 2018 to present |
| Spectral bands used | TIR 8–12.5 µm (LST); NIR / red-edge (vegetation indices); shortwave for albedo |
| Minimum detectable AET difference | Approximately 0.5 mm/day at monthly aggregation under cloud-free conditions; daily estimates less reliable |
| Delivery formats | GeoTIFF (per-scene and composited), catchment-aggregated CSV, interpretive PDF noting data-gap periods |
Analytics Satellize can run
| Monthly forest AET map | SEBS or METRIC energy-balance model driven by ECOSTRESS or Landsat TIRS LST, Sentinel-2 LAI/NDVI, and ERA5 meteorological reanalysis | GeoTIFF time-series at 70 m or 30 m, one layer per month, with per-pixel uncertainty flag |
| Catchment water-balance component table | Spatial aggregation of AET rasters within user-defined watershed polygons; precipitation input from CHIRPS or ERA5 | Monthly CSV: precipitation, AET, estimated runoff residual, data-availability score per catchment |
| Rainfall interception fraction map | Gash analytical interception model parameterised from Sentinel-2 canopy cover fraction and free throughfall coefficient; gross rainfall from gridded product | Seasonal GeoTIFF showing interception as percentage of gross rainfall, with model parameter report |
| Multi-year AET trend analysis | Mann-Kendall trend test on MODIS MOD16A2 annual AET stack (2000 to present) to detect long-term change linked to land-cover or climate shifts | Trend-magnitude raster and summary report flagging statistically significant pixels at p < 0.05 |
| Forest versus cleared-land AET contrast | Stratified comparison of AET values within forest and recently cleared polygons derived from land-cover classification; paired t-test for significance | Statistical summary table and annotated map for use in hydrological impact assessments |
| Seasonal drought-stress indicator | Evaporative Fraction (EF = LE / (LE + H)) from SEBS; anomaly relative to 10-year MODIS climatology flags periods of suppressed transpiration | 8-day raster anomaly product and catchment-level alert when EF drops more than one standard deviation below climatological mean |
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.