Irrigation withdrawal detection and agricultural water accounting
Satellite-derived evapotranspiration lets water managers infer irrigation timing and volume from orbit, without a meter in the ground. GRACE-FO gravity data cross-checks cumulative groundwater abstraction at basin scale.
Sensors
- Landsat 8/9 TIRS: Thermal Infrared Sensor captures land-surface temperature at 100 m native resolution (resampled to 30 m in distribution products). Revisit is 16 days per satellite, 8 days combined for Landsat 8 and 9. The thermal band is the primary input for field-scale surface-energy-balance models such as METRIC and SEBAL.
- Sentinel-3 SLSTR: Sea and Land Surface Temperature Radiometer provides thermal imagery at 1 km resolution with a roughly 1-2 day revisit. Insufficient for individual field attribution but useful for regional ET mapping and filling temporal gaps between Landsat acquisitions.
- MODIS MOD16 ET: NASA's operational global ET product derived from MODIS Terra and Aqua, at 500 m resolution and 8-day composites. Useful for basin-scale water-balance analysis and long time-series back to 2000, though too coarse to resolve fields smaller than roughly 25 hectares reliably.
- GRACE-FO: Gravity Recovery and Climate Experiment Follow-On detects changes in terrestrial water storage by measuring micro-scale variations in the distance between twin satellites. Spatial resolution is approximately 300-400 km, so it cannot attribute abstraction to individual aquifers. Monthly solutions provide the basin-scale groundwater depletion signal that corroborates ET-based withdrawal estimates.
What a warm field at midday is actually saying
When a crop is actively transpiring, it draws water from the soil and releases it through stomata, cooling the canopy in the process. A thermal sensor passing overhead at midday will record a cooler surface temperature over a well-irrigated field than over a dry or rainfed one nearby. Surface-energy-balance models, specifically SEBAL (Surface Energy Balance Algorithm for Land) and METRIC (Mapping EvapoTranspiration at high Resolution with Internalized Calibration), exploit this contrast. They partition net radiation into sensible heat, latent heat and soil heat flux, solving for actual evapotranspiration as the residual.
The critical step for irrigation accounting is separating actual ET from what rainfall alone would have produced. If a rainfed reference field exists in the same image, or if a soil-water-balance model provides a rainfed ET baseline, the excess ET can be attributed to applied irrigation. Over a season, integrating that excess gives an estimate of total water applied per field or per irrigation district. The physics is well-established. The uncertainty chain is not short.
Where the uncertainty accumulates
Landsat TIRS acquires a single thermal snapshot per 16-day orbit cycle per satellite. Cloud cover over an irrigated region during that window means the observation is lost entirely. Combining Landsat 8 and 9 halves the gap to roughly 8 days, but a sustained cloudy period can still leave a multi-week blind spot at a critical irrigation event. Sentinel-3 SLSTR can fill the temporal gap but at 1 km resolution, which averages across field boundaries and introduces mixing errors in fragmented landscapes.
The energy-balance models themselves carry published errors of roughly 10 to 30 percent in daily ET estimates at the field scale, depending on wind conditions, surface heterogeneity, and the quality of ancillary meteorological inputs. Seasonal cumulative errors tend to be smaller because random errors partially cancel, but systematic biases from poorly characterised aerodynamic roughness or atmospheric correction do not cancel. Attribution of a specific ET signal to a specific crop requires cadastral parcel boundaries and crop-type maps. Without those, the satellite can tell you water was applied somewhere in a pixel; it cannot tell you whether it was wheat, cotton or a market garden.
GRACE-FO as the basin-scale auditor
ET-based methods estimate water consumed at the surface. They say nothing directly about how much groundwater was pumped versus how much came from surface canals or stored soil moisture. GRACE-FO addresses a different question: is the total water stored in a basin going up or down, and by how much?
By removing the contributions of soil moisture and surface water (modelled or observed separately), GRACE-FO residuals can be interpreted as groundwater storage change. Studies in the Indus, Central Valley and Arabian Peninsula have used this approach to quantify multi-year depletion trends. The limitation is resolution: a GRACE-FO signal integrates over hundreds of thousands of square kilometres, so it cannot distinguish one over-abstracted aquifer from a neighbouring one. It is most useful as a cross-check, confirming that season-by-season ET-derived withdrawal estimates are accumulating into a depletion trend that matches gravity observations, or flagging a discrepancy that warrants investigation.
Turning ET maps into water accounts
A practical water-accounting workflow runs in three layers. At the field scale, Landsat-derived ET maps (produced via METRIC or SEBAL) are intersected with cadastral parcel boundaries to assign seasonal water consumption to individual licence holders or irrigation units. At the district scale, those field-level estimates are aggregated and compared against canal delivery records or metered pump logs, if they exist, to identify unexplained gaps. At the basin scale, GRACE-FO trends and MODIS MOD16 long-run ET climatologies provide the envelope within which the field-scale numbers must plausibly sit.
Water authorities in several arid-region jurisdictions have used variants of this approach to move from self-reported abstraction to satellite-corroborated accounts. The satellite does not replace metering; it provides an independent line of evidence that makes under-reporting visible. Where cadastral data is incomplete or parcel boundaries are disputed, the method degrades to district-level attribution, which is still useful for policy but insufficient for individual licence enforcement.
What this cannot do, stated plainly
Thermal-band ET methods do not detect irrigation that occurs at night or during overcast periods covered by cloud. They do not distinguish drip irrigation (which suppresses ET relative to flood irrigation, because less water evaporates from bare soil) from surface flood irrigation without supplementary information on irrigation method. They cannot resolve fields smaller than roughly two to four Landsat pixels reliably, which is approximately 0.6 to 3.6 hectares depending on field geometry and sensor noise.
GRACE-FO cannot attribute groundwater depletion to irrigation specifically; urban pumping, industrial extraction and natural aquifer discharge all contribute to the gravity signal. Separating them requires independent data. Satellize's analytics work for the Kingdom of Tonga crop-estimation programme demonstrated how open-constellation imagery can be combined with ancillary ground data to improve agricultural estimates in data-sparse environments, the same principle that applies here when cadastral records are thin.
Getting from pixels to a defensible number
The most common failure mode in satellite water accounting is presenting ET maps as if they were direct measurements of abstraction, without acknowledging the modelling steps between the thermal radiance and the volumetric figure. A defensible workflow documents the meteorological inputs, the reference ET method (FAO-56 Penman-Monteith is standard), the calibration approach for the energy-balance model, and the uncertainty bounds on the seasonal total.
Validation against eddy-covariance flux towers, where they exist in or near the study region, is the accepted method for quantifying model error. Published validation studies for METRIC in irrigated agriculture in the western United States report seasonal ET errors of around 3 to 5 percent against flux-tower measurements, under clear-sky conditions with good meteorological data. In regions without flux towers and with frequent cloud, honest uncertainty ranges are wider, often 15 to 25 percent on a seasonal basis. Buyers of this analysis should ask for those ranges explicitly, and be sceptical of any water-accounting product that does not provide them.
Typical figures
| Thermal spatial resolution (Landsat 8/9 TIRS) | 100 m native, resampled to 30 m in standard products |
| Thermal spatial resolution (Sentinel-3 SLSTR) | 1 km |
| Revisit cadence (Landsat 8 + 9 combined) | ~8 days at equator under clear sky |
| Revisit cadence (Sentinel-3 SLSTR) | ~1-2 days globally |
| GRACE-FO spatial resolution | ~300-400 km (equivalent spherical harmonic degree 60-96) |
| GRACE-FO temporal resolution | Monthly mass-change solutions |
| ET model seasonal accuracy (METRIC, validated sites) | ~3-5% vs flux towers under clear-sky, good met data; 15-25% in data-sparse or cloudy regions |
| Minimum field size for reliable ET attribution | Approximately 0.6-3.6 ha (2-4 Landsat pixels, geometry-dependent) |
| Archive depth (Landsat thermal) | Landsat 8 from 2013; Landsat 9 from 2021; earlier thermal from Landsat 5/7 TM/ETM+ back to 1982 at coarser quality |
| MODIS MOD16 ET archive depth | 2000 to present, 8-day composites at 500 m |
Analytics Satellize can run
| Seasonal ET maps at field scale | METRIC or SEBAL surface-energy-balance model applied to Landsat 8/9 TIRS and OLI imagery with FAO-56 Penman-Monteith reference ET | GeoTIFF raster layers per growing season, clipped to area of interest, with per-pixel uncertainty estimates |
| Irrigation withdrawal volume per parcel | Residual ET (actual minus rainfed baseline) integrated over season and intersected with cadastral parcel boundaries | Tabular report (CSV or GIS polygon layer) with seasonal water consumption in mm and cubic metres per parcel |
| Irrigation event detection | Change detection in NDVI and land-surface temperature time series between successive Landsat acquisitions, flagging anomalous cooling or greening consistent with a recent irrigation application | Event log with date, field identifier and confidence score; delivered as GIS alert layer |
| Basin-scale water-balance summary | MODIS MOD16 long-run ET climatology combined with precipitation data (CHIRPS or ERA5) to compute basin P minus ET residual as proxy for net water storage change | Annual water-balance report with time-series charts and comparison against GRACE-FO terrestrial water storage anomalies |
| Groundwater depletion trend corroboration | GRACE-FO mascon solutions (JPL RL06 or CSR RL06) with GLDAS soil-moisture and surface-water corrections removed to isolate groundwater storage change signal | Basin-level groundwater storage anomaly time series (monthly), with confidence intervals, formatted as chart and data table |
| Crop-type-stratified water use | ET maps intersected with satellite-derived crop classification (Sentinel-2 multitemporal NDVI phenology) and ancillary cadastral or survey data | Water-use breakdown by crop type per district, as GIS layer and summary table; accuracy dependent on quality of ancillary crop data provided |
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.