Blue-water and green-water partitioning at basin scale
Splitting the terrestrial water budget into blue water (runoff and groundwater) and green water (evapotranspiration) tells water managers where consumption is going and whether abstraction is sustainable. The method fuses GRACE-FO gravity, GPM rainfall and MODIS/VIIRS ET, and works best for basins above roughly 10,000 km².
Sensors
- GRACE-FO (Gravity Recovery and Climate Experiment Follow-On): Measures monthly changes in terrestrial water storage (TWS) as gravity anomalies. Spatial resolution is effectively 300–400 km (roughly equivalent to a ~3° spherical harmonic truncation at degree 60–96 depending on mascon solution). Minimum detectable TWS change is approximately 1–2 cm water equivalent over large areas. Provides the closing term in the water-balance equation.
- GPM IMERG (Global Precipitation Measurement Integrated Multi-satellitE Retrievals): Merged passive-microwave and infrared precipitation product. Final Run latency is approximately 3.5 months; Early Run available within hours. Spatial resolution 0.1° (~11 km at equator), 30-minute temporal resolution. Global coverage 60°N–60°S, with partial coverage to ±90°. Uncertainty is highest over complex terrain and in arid regimes where rain rates are low.
- MODIS MOD16 ET: Global evapotranspiration product derived from MODIS land-cover, albedo and LST inputs using the Penman-Monteith algorithm. Delivered at 500 m spatial resolution with 8-day and monthly composites. Separates total ET into transpiration, soil evaporation and interception evaporation, which is essential for partitioning productive from non-productive green-water fluxes. Cloud gaps affect composites in persistently overcast regions.
- VIIRS VNP21 (Land Surface Temperature and Emissivity): Day/night LST at 375 m resolution from Suomi-NPP and NOAA-20, used to constrain energy-balance ET models as an alternative or complement to MODIS. Improves spatial detail in heterogeneous landscapes but shares the cloud-contamination limitation of all thermal-infrared retrievals.
- Sentinel-2 MSI (supporting land-cover and NDVI input): 10 m multispectral imagery used to produce high-resolution land-cover and vegetation-fraction maps that disaggregate coarser ET products within the basin. 5-day revisit at mid-latitudes with two satellites. Not a direct water-balance sensor, but improves the spatial attribution of green-water consumption to specific land-use classes.
Two fates for every raindrop
Every millimetre of precipitation that falls on a basin does one of two things. It either leaves via a channel or percolates to groundwater, becoming blue water available for extraction and downstream flow. Or it evaporates from soil or is transpired through plant stomata, becoming green water that is consumed in place and is, for practical purposes, gone. The ratio between these two fates determines whether a basin can support more irrigation, more cities or more ecological flow. In water-scarce regions, managers often know neither quantity with confidence.
The satellite approach to partitioning rests on the water-balance identity: change in storage equals precipitation minus evapotranspiration minus runoff. If three of those four terms are measured independently, the fourth can be inferred. In practice, GRACE-FO supplies the storage change, GPM IMERG supplies precipitation, and MODIS MOD16 or an energy-balance model supplies ET. Runoff, and by extension blue-water availability, becomes the residual. The elegance is real. So is the accumulated uncertainty.
What each sensor actually measures, and where it struggles
GRACE-FO detects the integrated mass of all water in a column: snow, soil moisture, groundwater, surface water, even biomass water. It cannot distinguish between these stores without ancillary data. The mascon solutions published by JPL and CSR resolve features at roughly 300–400 km, which sets the hard lower bound on basin size. A basin of 8,000 km² will alias signal from its neighbours. Monthly temporal resolution also means that sub-monthly storage swings, important during intense monsoon events, are smoothed out.
GPM IMERG is accurate enough over large areas and long periods, but its 0.1° grid conceals orographic gradients that matter enormously in mountain-fed basins. Validation studies over the Tibetan Plateau and the Andes consistently show biases of 20–40% in high-altitude cells. MODIS MOD16 ET carries its own structural uncertainty: the Penman-Monteith parameterisation depends on stomatal conductance look-up tables by land-cover class, and those tables are global averages. Actual ET in a dryland shrubland or a flooded rice paddy can depart substantially from the model's assumption. Cloud cover is the operational enemy of both MODIS and VIIRS thermal products; monthly composites in humid tropical basins can carry gaps exceeding 30% of pixels.
The combined uncertainty on the blue-water residual is not simply the sum of individual errors. It is their quadrature sum, and in practice the 95% confidence interval on annual blue-water flux in a mid-sized basin is often ±15–30% of the mean. That is not a reason to abandon the method. It is a reason to report it honestly and to use multi-year averages rather than single-year snapshots for policy decisions.
Productive transpiration versus wasted evaporation: the distinction that matters for agriculture
Not all green water is equal. Transpiration through crop stomata produces biomass. Evaporation from bare soil between rows, from flooded paddy surfaces, or from irrigation canals produces nothing. Separating these two fluxes is where the analysis earns its keep for agricultural water managers.
MOD16 partitions total ET into its components using fractional vegetation cover derived from NDVI. The separation is approximate, but it is directionally useful. A basin where bare-soil evaporation accounts for 40% of total ET has a very different improvement opportunity than one where transpiration dominates. Fusing MOD16 partitioned ET with Sentinel-2 crop-type maps at 10 m allows the analysis to attribute green-water consumption to specific crops and irrigation schemes, identifying where water productivity (yield per unit of green water consumed) is lowest. This is the analytic layer most directly relevant to irrigation policy.
Closing the balance when groundwater abstraction is unmetered
In many water-scarce basins, groundwater pumping is the dominant consumptive use and is almost entirely unmetered. Farmers drill, pump and report nothing. National statistics, where they exist, are often years out of date. GRACE-FO is the only tool that can observe the net result: a declining gravity signal that, once corrected for soil moisture and surface-water changes using land-surface model outputs (GLDAS or ERA5-Land are commonly used), isolates the groundwater storage trend.
Combining a GRACE-derived groundwater depletion rate with the water-balance residual allows an approximate attribution of blue-water consumption between surface withdrawals (visible in reservoir-level and river-flow data) and groundwater abstraction (inferred from the gravity residual). The method has been applied in published literature to the North China Plain, the Central Valley of California and the Indus basin, among others. In each case the gravity signal confirmed that official abstraction estimates were materially understated. The satellite cannot tell you which farmer is pumping, but it can tell a regulator that the aggregate rate is unsustainable, and it can do so without requiring a single gauge or meter in the field.
Honest limits and what they mean for basin size and timescale
The 10,000 km² floor on basin size is a practical rule of thumb, not a physical law. Some studies have pushed GRACE closure to smaller basins using statistical downscaling, but the uncertainty bands widen rapidly below that threshold and the results should be treated as indicative rather than operational. River basins in the range of 50,000–500,000 km² are the sweet spot: large enough for GRACE to resolve cleanly, small enough that GPM precipitation gradients are meaningful.
Annual and multi-annual averaging is almost always more defensible than monthly closure. Monthly water-balance residuals in semi-arid basins can be dominated by the timing mismatch between a rainfall event and the GRACE observation window. Five-year trend analysis, by contrast, smooths these artefacts and reveals the structural shift in blue-green partitioning that matters for long-term planning. Satellize runs this analysis on open constellations for clients where basin-scale water accounting is a regulatory or investment requirement; the Tonga crop-estimation programme demonstrated a related approach to agricultural water-use attribution from the same sensor stack.
One further caution: the water-balance approach gives fluxes, not stocks. It tells you how fast groundwater is declining, not how much remains. Combining GRACE trends with hydrogeological estimates of aquifer volume is necessary before any statement about years-to-depletion can be made responsibly.
From water balance to a number a minister can act on
The analytic output of a basin-scale partitioning study is, at its most useful, a time series of annual blue-water availability and green-water consumption, broken down by land-use class, with uncertainty bounds. That time series can be compared against licensed abstraction volumes to identify the gap between what is permitted and what physics allows. It can be fed into an economic model to estimate the cost of the current trajectory. It can be updated annually with new GRACE, GPM and MODIS data without additional field campaigns.
The translation from satellite signal to policy-relevant number requires choices about which land-surface model to use for soil-moisture correction, which ET algorithm to apply, and how to handle data gaps. Those choices should be documented and, where possible, validated against the gauges that do exist. A study that presents a single deterministic estimate without uncertainty quantification is doing the client a disservice. The satellite record is good enough to be genuinely useful. It is not good enough to be treated as ground truth.
Typical figures
| GRACE-FO spatial resolution (effective) | 300–400 km (mascon solutions, JPL RL06 / CSR RL06) |
| GRACE-FO temporal resolution | Monthly; gaps exist due to battery management (typically 1–3 months per year) |
| GRACE-FO minimum detectable TWS change | ~1–2 cm water equivalent over areas ≥ ~200,000 km² |
| GPM IMERG spatial resolution | 0.1° (~11 km at equator) |
| GPM IMERG temporal resolution / latency | 30-minute accumulations; Early Run ~4 h, Final Run ~3.5 months |
| MODIS MOD16 ET spatial resolution | 500 m; 8-day and monthly composites |
| VIIRS VNP21 LST resolution | 375 m, daily (cloud-permitting) |
| Minimum practical basin size for GRACE closure | ~10,000 km² (indicative floor; uncertainty rises sharply below this) |
| Archive depth | GRACE/GRACE-FO: April 2002–present (with gap 2017–2018); GPM IMERG: June 2000–present (merged with TRMM); MODIS MOD16: 2000–present |
| Typical combined uncertainty on annual blue-water residual | ±15–30% of mean flux (basin-dependent; improves with multi-year averaging) |
Analytics Satellize can run
| Annual blue-green water partition | Water-balance closure: ΔTWS (GRACE-FO mascon) + ET (MOD16) + P (GPM IMERG Final Run); soil-moisture correction via GLDAS or ERA5-Land | Annual GIS layer and tabular report showing blue-water availability and green-water consumption with ±1σ uncertainty bounds, per basin sub-unit |
| Groundwater storage trend | GRACE-FO TWS anomaly minus GLDAS soil-moisture and snow-water-equivalent terms; linear trend fitted over user-specified period | Time-series chart and GIS polygon with depletion rate (cm/year water equivalent) and 95% confidence interval |
| ET partitioning by land-use class | MOD16 component ET (transpiration, soil evaporation, interception) intersected with Sentinel-2-derived crop-type or land-cover map | Tabular breakdown of productive transpiration vs. non-productive evaporation by land-use class; annual and seasonal aggregates |
| Water productivity index | Crop yield estimate (from Sentinel-2 NDVI phenology) divided by MOD16 seasonal transpiration, per field or irrigation district | Raster layer and ranked district table showing yield per unit of green water consumed |
| Multi-year trend attribution | Mann-Kendall trend test on annual blue-water residual and GRACE groundwater anomaly time series; breakpoint detection to identify step changes coinciding with land-use or climate shifts | Trend report with statistical significance, breakpoint dates and candidate attribution narrative for policy briefing |
| Abstraction gap assessment | GRACE-inferred net groundwater depletion compared against reported licensed abstraction volumes; residual attributed to unmetered pumping | Regulatory briefing document quantifying the gap between permitted and satellite-inferred abstraction, with uncertainty ranges |
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.