We weighed 27 of Earth’s great water reserves from orbit. 15 are being drained faster than rain can explain.

The debtors: 15 regions where rainfall cannot explain the loss

Sorted by box total. Northwest India (Punjab–Haryana–Rajasthan) leads at 13.6 km³ a year — Rodell et al. 2009, Nature put the groundwater there on the map in 2009; a quarter-century later the drawdown has not stopped. “Beyond rainfall” is the structural component: the trend left after storage is allowed to follow the region’s own rainfall history.

The lakes are a different story — and the method knows it

The single biggest number on this page is not an aquifer. The Caspian Sea and margin box is losing 30.4 km³ a year — the world’s largest lake, falling as evaporation outruns the Volga. Our rainfall test refuses to attribute it (the fit coefficient goes negative, because a lake fed by a distant river does not obey local rain), so it is classified “declining — attribution out of reach” rather than dressed up as structural depletion. The Aral basin gets the same honest label. A method that cannot say “this one is out of my reach” should not be trusted on the ones it does call.

Where the method says the opposite

A league table of decline is only credible if it can also print the other column. 6 regions are stable — including the Guaraní and Great Artesian, two of the planet’s largest fossil reserves, and the Nubian sandstone, where the box mean barely moves. 4 are gaining: Lake Victoria basin (+11.4 km³/yr — record lake levels, and our sign control), the Southern Murray–Darling basin (+2.9 km³/yr — the Millennium drought ended and the ledger shows the recovery), the Kalahari and the Congo. Not everything is a crisis, and the instrument can tell the difference.

Method — and the controls that gate it

1. Mass: CSR RL06.3 GRACE/GRACE-FO mascons (Save et al. 2016; University of Texas CSR), 257 monthly grids, April 2002 to May 2026. Glacial isostatic adjustment already removed (ICE6G-D). Each region is a published lat/lon box; cosine-weighted means; the 2017–18 gap between the two missions is carried, not interpolated.
2. Trend and interval: ordinary least squares with a moving-block bootstrap (24-month blocks, 1,000 draws) — monthly storage is strongly autocorrelated, and naive error bars would flatter us.
3. The rainfall question: NASA POWER monthly precipitation (satellite-informed model reanalysis — named as such, it is not a rain gauge), averaged over a 2° grid per region, 2002–2025. Storage is regressed on cumulative rainfall anomaly plus time; the time term is the structural residual.
4. Controls, printed: 5 literature-documented depletion regions (nw-india, north-china, arabia, central-valley, iran) had to classify structural — all did. 2 wet basins (amazon, congo) had to come out clean — both did. The gain control (victoria) had to show a rise — it did, at +2.6 cm/yr.
5. Magnitude anchor: our Northwest India (Punjab–Haryana–Rajasthan) box loses 13.6 km³/yr of total water over 2002–2026; Rodell et al. 2009 estimated 17.7 ± 4.5 km³/yr of groundwater for Rajasthan–Punjab–Haryana over 2002–2008. Different window, different quantity, same order — consistent, not identical, and that is the honest comparison.

What we did not find, and what stays open

• Zero regions classified “rainfall-explained decline”. That surprised us, and we report it rather than tune it away: over 24 years, droughts mean-revert — a region whose loss is truly rainfall-driven tends to recover (the Southern Murray–Darling basin did exactly that, and shows as gaining). A quarter-century decline that persists is, by then, structural almost by definition.
• Central Chile is the edge case: megadrought country, structural component -0.33 cm/yr with an interval of -0.60 to -0.05 — barely clear of zero. If any row on the debtor list migrates when the record lengthens, it is this one.
• The Tarim box borders the Tien Shan and Kunlun: glacier loss bleeds into its mascons, so its structural label mixes ice and groundwater. Flagged, kept, discounted.
• Boxes are not aquifers. The High Plains box averages a depleting south against a stable north; the Arabian box holds several systems. Box totals are conservative for concentrated hotspots and say nothing about any single wellfield.
• Lake Mead, for scale: 32.2 km³ at full pool. The 15 structural regions lose ~65 km³/yr — 2.0 Lake Meads a year, every year, for as long as this record runs.

Full table with coordinates, intervals, fit coefficients and classifications: regions.csv. Data: CSR RL06.3 mascons (NASA/DLR/GFZ missions, University of Texas processing) · NASA POWER. This page is dated 2026-08-08; the mascon record it rests on runs April 2002 to May 2026. We did not produce the measurements; we produced the join.

How can a satellite weigh water?

GRACE and its successor GRACE-FO are twin satellites flying about 200 km apart, measuring the distance between themselves to better than the width of a red blood cell. When they fly over a region that has lost water mass, gravity tugs them slightly differently and the gap changes. Two decades of those wobbles, processed into monthly mass grids by the Center for Space Research in Texas, are the raw material here. It is the only instrument humanity has that weighs whole regions.

Is this groundwater?

Not exactly — the satellites weigh everything: soil moisture, snow, rivers, lakes and groundwater together. That is why this page says "water mass" and reserves "groundwater" for regions where published studies did the attribution (northwest India, the North China Plain, California, Arabia). It is also why the Caspian and Aral rows are labelled as lakes: their losses are surface water, and our rainfall test correctly refuses to attribute them.

What does "structural" mean, exactly?

We reconstruct each region's rainfall history (NASA POWER monthly reanalysis, 2002–2025) and let storage follow it: wet spells refill, droughts draw down. "Structural" is the decline left over after that bookkeeping — water leaving faster than rainfall variability explains. Mostly that is pumping; in warming regions rising evaporative demand contributes too. It is a residual, not an accusation — which is why every number ships with its interval and the fit coefficient.

What would prove this wrong?

The controls are printed on the page: four regions with literature-documented depletion had to classify as structural, two wet basins had to come out clean, and Lake Victoria — at record-high levels — had to show a gain. All passed. The full per-region table with coordinates, intervals and classifications is downloadable, so anyone with the same public data can re-run the sums.

More from the same discipline: Sinking Britain · the ground under the sea-level record · the subsidence-claims forecast · Sinking, unexplained (US). Site-specific ground-motion measurement on demand: GroundTruth.