This is the machinery behind every number in the dashboard: each source that goes in, how a gridded satellite field becomes one figure for one river basin, the eight stages in the order they run, the four equations being solved, what the answer was tested against, and the limits that stay whether or not they are convenient. Nothing here is a summary of a method written elsewhere — the counts and spans on this page are read out of the same build the dashboard uses.
Fourteen sources, every one public and cited in the dashboard’s reference list. Three of them are the same quantity produced by three different groups; four are there only to test the answer, and never enter the pipeline that produces it.
| Source | What it provides | Coverage |
|---|---|---|
| Satellite storage | ||
| CSR RL06.3 mascons | Total water storage, monthly | Apr 2002 – Jun 2026 |
| NASA GSFC RL06 v2.0 mascons | The same quantity, processed separately | to Mar 2026 |
| NASA JPL RL06.3 v04 CRI mascons | The same quantity, a third time | to Jul 2026 |
| Climate | ||
| GPCP v2.3 | Precipitation | to Jun 2026 |
| GHCN-CAMS | Air temperature | monthly |
| Land surface | ||
| CPC soil moisture | Modelled soil water | monthly |
| NCEP/DOE Reanalysis 2 | Snow water equivalent | monthly |
| Geometry | ||
| HydroBASINS level 3 | 202 basin outlines | fixed |
| Surface water | ||
| GloLakes | Satellite lake storage | 2,185 lakes in 49 basins |
| US Bureau of Reclamation | Gauged reservoir storage | Colorado |
| Check | ||
| Jasechko et al. 2024 | In-situ well levels worldwide | 41,866 wells, 2003-2022 |
| CGWB and State Groundwater Boards | Well levels, India | 6,843 wells, 2013-2019 |
| Payne et al. 2025 | Land subsidence measured by radar | 79 regions |
| O'Neill, Rodell & Loomis 2026 | A separately produced trend field | 155 basins compared |
| Screening layer, not a check | ||
| Hasan et al. 2023 | Land subsidence predicted by a model, worldwide | 57 of 202 basins |
GRACE and GRACE-FO measure the pull of the water beneath them, and the three processing centres turn those same observations into monthly maps of stored water. Each map is averaged over each basin outline separately, weighted by area, in an equal-area projection — so a basin’s figure is its own water rather than a pixel near it. The measurement resolves about 300 km, which is why the unit of analysis is a river basin and not anything smaller.
Where the three centres disagree, that difference measures their processing choices. It is not three independent looks at the planet, and it is not treated as one.
Nothing later in the chain can reach back and change something earlier. The classifier never sees the wells; the wells are used only to test what it already decided.
Four, and no more than four. Each one is standard, and each is written out here so that the method can be checked rather than taken on trust.
Nothing here was checked against itself. Wells, sinking ground, published studies and a separately produced satellite field all sit outside the pipeline that produced the classes. Where a check contradicts a result, that basin’s confidence is lowered and the reason is shown — a check can never raise a confidence grade.
A tool is worth as much as what it refuses to claim. These limits are on the page because they are real, not because they are comfortable.
GRACE resolves total storage at roughly 300 km. Every result describes a whole river basin — never an aquifer, a district or a field. The satellite grid drawn on this page is at that resolution, not at a prettier one.
The first mission ended its science operations in October 2017 and GRACE-FO launched in May 2018. Trends are fitted across the gap and nothing is modelled into it, so the gap stays visible in every chart.
Under the formal definition (Madani 2026), water bankruptcy needs insolvency and irreversibility. A storage record alone cannot establish irreversible damage, so the strongest label this tool gives is Severe insolvency. A predicted land-subsidence layer is now shown on each basin sheet as a first step on the irreversibility side, but it is a model prediction rather than a measurement and it enters no class, threshold or confidence grade.
The model is fitted to observed rainfall and temperature, and observations carry natural and human-forced variability together. Both sides of the fit contain the forced signal, so the split reported is between what observed weather explains and what it does not — not between natural and forced.
CSR, GSFC and JPL process the same GRACE and GRACE-FO observations. Their spread measures processing and regularisation choices, and understates the true uncertainty. All three also apply the same glacial isostatic adjustment model, so that spread carries no GIA uncertainty at all.
Every threshold is published and every class is tested against alternatives, but the final values for any basin must be set locally, with the people who manage it.
The same six steps, and the method worked on one real basin, as a ninety-second film recorded from the dashboard while it runs — not re-drawn for video, so it cannot show anything the tool does not do.