Methodology.
The same principles apply to every module, from crop yield and water balance to groundwater, rivers, reservoirs and ground subsidence. The platform is in closed beta with design partners, so this page describes the method, not results.
From satellite archive to delivered figure.
Four stages, the same for every module.
Acquisition
Radar and optical data, an orbital water-mass signal, climate reanalysis, official statistics and ground measurements are gathered for each region, wherever they exist.
Fusion
The signals are aligned for each location and compared with one another.
Modelling
Process models and machine learning turn the aligned signals into estimates.
Delivery
Estimates are delivered as ranges, with their uncertainty and the version of the method, in reports and through APIs.
What the method rests on.
Implementation details (algorithms, parameters and data sources) are shared with design partners under NDA.
Physics first, learning second
Where a process model exists, such as crop growth, soil water or catchment hydrology, it runs first. Machine learning is trained on the residual: what the physics leaves unexplained. It corrects the models; it does not replace them.
Several sensors for each location
Radar sees through cloud, optical sensors read vegetation, and an orbital water-mass signal tracks the water stored in the ground. Where the signals disagree, for example when the surface looks fine while the water mass beneath it is falling, the case is reviewed rather than averaged away.
Downscaling under physical constraints
A coarse signal is brought down to district level and, where the inputs allow, to the parcel. Terrain, soils and recharge constrain the result, so it is more than interpolation. The uncertainty of this step is stated with the figure. Where no downscaling is applied, values keep the native resolution of the sensor; only geometry, such as parcel boundaries, is sharpened.
Long records
Models are built and checked on 20 or more years of official statistics, satellite archives and climate reanalysis, wherever such records exist. Where a record is shorter, the uncertainty is wider and the figure says so.
Ranges, not single numbers
A result is a range: a lower, a central and an upper estimate. The uncertainty is calibrated, and the aim is that a stated range contains the observed value about as often as it claims to. Figures are updated as new satellite passes arrive.
“Unavailable” is an answer
When a source is down, or a gap cannot be filled from real data, the figure is shown as unavailable. There are no silent fallbacks and no synthetic values standing in for missing data.
How these methods are checked against independent measurements is described on the Validation page.