An interactive geospatial analysis and data visualization tool that screens river basins for water insolvency, the measurable first pillar of water bankruptcy.
Three mascon solutions, independently processedTested on basins built to contain no water useMethod published in full, with a citable data record
Surface water is like a checking account meant for regular use, while groundwater functions as a savings account … Yet today, many countries are living off this savings account without a plan to repay the debt.
Each figure is that basin's storage trend over 2002–2026, in cubic kilometres a year — a long-run average, not a live reading. Tap a name to open it.
Each answer is a picture rather than a paragraph. Four of the nine are no, or not yet, and they are drawn as plainly as the rest. Where the method or dataset behind an answer is published work, the star opens that paper.
Both, in that order. Only the first box is a measurement. The third is a subtraction, which is why the page calls it beyond climate and never calls it water use.
It does not know. The remainder holds all six of these together and this tool does not separate them. That is why every label reads explained or not explained by observed weather, and never human or natural.
For 66 of the 155 assessed basins. For the other 89 there is no independent instrument reaching them at all. The page calls this the weakest thing about it rather than leaving it to be found.
Not the severe ones. In 1,530 worlds built to contain no water use at all, the classifier called severe insolvency zero times. It does misfire into the milder classes, and those rates are above rather than in a footnote.
Its mistakes are misses, not inventions. A loss slow enough can pass unseen, and in five basins it would at any rate tested.
No. The ticks are the places inside one reading, and the reading cannot be taken apart into them. The smallest thing this page can speak about is a basin.
Mostly not, and where it does the page says so. Every basin was reclassified 288 ways. The sixth severe insolvency basin is the one to be careful with, and it is marked on its own sheet.
No. Storage falling and a country running short are two different statements, and this page only makes the first. Two basins with the same trend can face completely different pressure.
Not yet. The screening itself is being written up. What it is built from is published work, and every star on this page opens the paper behind the thing it sits on.
In 2026, a report from the UN University's Institute for Water, Environment and Health described many river basins and aquifers as being in “water bankruptcy”: they are spending more water than nature can refill. That raised a simple question: which basins, how fast, and is it the weather or something else?
What is measured here. Formally, water bankruptcy has two parts: insolvency (use persistently exceeds what nature renews) and irreversibility (damage that cannot be undone on human timescales). A system can be insolvent without being bankrupt. AquaSolvency screens the insolvency pillar, basin by basin; it does not rule on bankruptcy. Definition: Madani (2026), Water Resources Management 40:78.
Rain is a basin's income; groundwater, lakes and soil water are its savings. Which basins are spending their savings?
Satellites have weighed Earth's water monthly since 2002. Add rain and temperature, and each basin gets a balance sheet.
Anyone can read a draining tank in seconds. Every number and uncertainty stays one click away for experts.
Aqua is water; solvency is paying your bills. Can each basin keep paying its water bills?
— Dr. Pradeepika (Praddy) Kaushik · Geospatial & Data Visualization Expert | GeoAI | EO Expert
Satellites weigh the water in every river basin each month. Watch storage fall (red) and rise (blue) against each basin's own 2002–05 level, and see how much the declining basins have drained.
Timeline: red = declining basins with high or medium confidence (—), blue = all other assessed basins (—), km³ against 2002–05. Ice-dominated and contaminated-signal basins are left out. Storage includes groundwater, soil water, snow, lakes and reservoirs. News notes are context, not results of this analysis. Satellite gravity records start in 2002.
Tap a basin, or type a river or country, to open its balance sheet.
The 12 largest yearly losses (high and medium confidence). Solid: not explained by observed rainfall and temperature. Light: explained by them.
Blue line: storage if only rain and temperature mattered. Shaded gap: the part the rainfall-and-temperature model cannot account for. It may include water use, glacier melt in mountains, land-use change and model error.
The satellite weighs snow, soil moisture, surface water and groundwater together. Snow and soil moisture are taken out everywhere, and lake and reservoir storage wherever a record exists; what is left is the groundwater estimate. This is what that separation finds in the largest storage losses on this page — and where it could not be done at all.
The hero says rainfall and temperature do not explain most of the measured loss. That sentence leaves three questions open: where the unexplained part actually sits, whether it is getting worse, and whether any of it is permanent. Nothing here is a new measurement — each view is a different cut of the fields already on every basin page.
The two bars share one axis and are deliberately not stacked. The beyond-climate term can exceed the total loss, which happens where the climate record points to storage rising and it fell anyway. Stacking would have had to clamp that case out of sight. “Beyond climate” is what the rainfall-and-temperature model cannot account for: it may include water use, glacier melt in mountain basins, land-use change, and error in the model itself. This view does not separate them.
Click a basin to open its balance sheet. Both trends are Theil–Sen slopes on the same record, over a split fixed at 2015 for every basin. A basin that is falling faster is not necessarily losing more water than its neighbours — it is losing it faster than it used to.
| Basin | Permanent | Share of subsiding area | Fastest |
|---|
Irreversibility is the second pillar of the formal definition of water bankruptcy, and this page measures it only here, only in part, and only where radar has been flown. A basin missing from this table has not been shown to be undamaged — in almost every case it has not been looked at. The page does not rule on bankruptcy for any basin.
A basin authority does not arrive asking whether storage is falling. They arrive asking whether they can act on it, and if not, what is missing. This splits the whole measured loss by what actually stands behind each result.
This says what the evidence can support. It does not say what anyone should decide. Who reduces use, by how much, and when is not a question a storage record can answer — those belong to the people who manage the basin and the people who depend on it. What the record can do is say which of those conversations currently has independent measurement behind it, and which does not.
Around the world, teams are measuring, refilling and sharing water better. Here are a few of them. Tap a place.
A small sample of the many teams working on water, not a ranking. Figures come from each organisation's own pages or cited studies (years as stated), checked September 2026.
One dot per basin, placed left to right by how fast its stored water is changing. Regroup them to see where the differences actually sit. Hover for the name, click to open that basin's balance sheet.
Small screen? Tap Expand to open this full size — you can pan across it.
Dot size is basin area. Ice-dominated and contaminated basins are shown hollow: their signal is not a water-use signal. Position is the Theil–Sen trend in millimetres of water a year.
Every basin on this page carries a number. Not every number has something standing behind it. This places the two side by side: how much water a basin is losing, against how many independent measurements — a monitoring well, radar, altimetry, a published ground study — reach it. Sources built from this same satellite record are not counted, however well they agree.
One dot per declining basin. Hover for the name; click to open its balance sheet. The figures below are computed from the table on this page, not entered by hand.
A register records that a monitoring point exists and who operates it, never what it reads. Where that is the case, what is missing is not the measurement but an openly licensed, per-station record that can be aggregated to a basin and compared with these trends.
This is a statement about the evidence available, not a ranking of effort or intent. Many of these basins lie in countries that monitor their groundwater carefully. It closes where records are opened, and until then the basins above rest on the satellite record alone and should be read that way.
Two basins where the balance sheet changes the conversation. Sub-basins are re-analysed with the same method.
Two further mascon solutions, a separately produced satellite trend map at five to ten times the spatial detail, land subsidence measured by radar, a rainfall test that needs no model, published studies recomputed over their own periods, an out-of-sample test of the climate model, 288 reruns of every basin under other defensible choices, 48 ways of preparing the satellite record, and a perfect-model test that runs the whole classifier on synthetic basins containing no water use at all.
Well levels are the thinnest of these, and the headline number is not the measure of them. The 41,866 wells in the global compilation reach 30 of the 202 basins and leave only 22 with enough coverage to score, where the two rankings agree moderately (Spearman ρ = 0.44, p = 0.04); the 6,726 Indian wells score eight basins, and there the rankings do not agree significantly (ρ = 0.33, p = 0.42). Both are reported basin by basin below, including where they disagree.
Four checks that reach the ground, and how far each one gets. The second bar on every card is the one that decides the first: strong agreement across a tenth of the basins is not strong validation.
moderate agreement, p = 0.04, against 41,866 wells
only 22 of 202 basins have enough wells to score
pre-monsoon round agrees; the January round agrees in 5 of 8
8 of 202 basins, and a basin counts only when both rounds agree
sinking faster than 10 mm/yr, every mapped region inside a basin this tool flags
Iran only — this is where the radar has been flown
of the 155 assessed basins have no independent instrument reaching them at all
66 of 155 have at least one independent check
The check above asks whether a basin is called beyond-climate because the model failed. This asks the reverse: where the model claims the weather explains the decline, can the model be trusted in that basin at all? Skill is the out-of-sample correlation — trained on 2002–2014, scored on 2015–2026.
A screening tool is only useful if a dry spell, a draw-down and a basin in lasting trouble do not all come out looking the same. The classes themselves are set by thresholds on the climate-explained share, the depth of the loss and recovery, so showing that the classes differ on those measures would only restate the definition. Everything below is evidence the classifier never consults.
Each basin keeps its real rainfall and temperature, trends and all, and has no water use by construction. Anything the method reports there is a false alarm. This is why the weaker labels are written as prompts to look, not as findings.
A drought is not read as a draw-down. The obvious objection is that a basin might be called beyond-climate simply because the climate model fails there. If that were happening, model skill would be worse in those basins, and it is not: the median out-of-sample correlation is 0.750 for the beyond-climate classes against 0.745 for Drought deficit, which is no separation at all (AUC 0.54, p = 0.74). The direct test is the perfect-model world, where each basin keeps its real rainfall and temperature, including their real trends, and has no water use by construction: across 1,530 runs, Severe insolvency never occurred once. Drought deficit occurred in 6.3% of those runs and some deficit class in 21.7%, which is why the weaker labels are written as prompts to look rather than findings.
A lasting decline looks different from a passing one. The severity rule uses the late period, so the early half of the record is free evidence. All six Severe insolvency basins decline in both halves, and 23 of 27 Structural deficit basins, against three of eight Drought deficit basins (Fisher exact p = 0.007). The shape differs too: the drought basins have the steeper early decline and the shallower late one, which is a deficit easing, while the beyond-climate basins get worse in the second half.
The number underneath is not a mascon artefact. Against the separately produced Level-1B trend field, every deficit class agrees in sign for every basin, with a median difference of 0.27 mm/yr in the Severe insolvency basins. And the labels hold up under other choices: four of the six Severe insolvency basins hold that class in all 288 reruns, one in 94% and one in 60%, and the class survives all 243 threshold settings and all 48 ways of preparing the record, while Drought deficit survives about a third of the reruns. The strongest label is the steadiest one, and the weakest label is the one that moves.
Every basin now carries its own vote. Rather than assert one word, each basin sheet shows how the 288 reruns actually fell. Across the 202 basins, 106 carry the same class in every single rerun, 17 fall below half, and in 11 the published class is not the most common outcome — those are the basins where the confidence and contamination rules act on top of the classifier, and each one says so on its own sheet. Of the six Severe insolvency basins, four hold that class in all 288 reruns, one in 94% and one in 60%, with the remainder falling to the milder Structural deficit rather than out of deficit altogether.
Where a lake sits in the signal, it is separated rather than assumed. Four basins were held out because a named lake or reservoir is part of the storage signal and could have been the whole story. Three now have a 10-day altimetric level record, courtesy of the NASA/GSFC Global Water Measurements programme, and where the conversion is sound the lake is removed from every solution before the record is analysed, so the basin is assessed on what is left. Lake Balkhash turns out to be about 15% of its basin's loss: without it the basin still falls 5.75 mm/yr, and it is now classed a structural deficit, the same class in all 288 reruns. The Sobradinho reservoir has been filling slightly while its basin fell, so it explains none of that loss, and São Francisco is now classed a drought deficit — though only just, with the reruns split 41% drought deficit against 40% structural deficit, which its sheet shows. The Aral Sea cannot be corrected this way at all: its surface area collapsed over the same period, so a level series alone gives no volume, and it stays out of the assessment with that reason stated. Two basins therefore move from excluded to assessed, 153 to 155, and nothing else in the run changes: of the 207 records analysed — the 202 basins, four sub-basins and one no-use benchmark region — 205 return trends identical to the previous version. The weak point is the constant area the conversion assumes: a reviewer has pointed to a database that publishes surface area and volume variation alongside water level, which would drop that assumption and could give even the Aral Sea a volume change. That is the next version's work, and until it is done the assumption stays stated wherever the correction is used.
What it misses. An injected loss reaches 80% detection at 2 mm/yr in the median basin, and of the 148 basins that reach 80% detection at all, 90% do so by 5 mm/yr. Five basins never reach it at any rate tested. The errors this tool makes are therefore misses rather than inventions.
Where the ground itself is sinking. Irreversibility needs evidence that storage capacity has been destroyed rather than emptied, and when aquifer sediments compact that capacity does not come back. No open global measurement of it exists. A published global prediction does: a model trained on radar from 47 regions and mapped worldwide at about 2 km. Read onto these basins it puts sinking ground worth printing in 57 of the 202, led by the Hai basin of the North China Plain, where 18.8% of the basin — about 64,000 km² — is predicted to sink at 1 cm a year or more. It does not track the storage classes: across those 57 basins the predicted share has no relation to the storage trend (Spearman rho −0.13, p = 0.33) or to the beyond-climate rate (−0.20, p = 0.14), and the deficit basins are no different from the rest (median 0.51% against 0.28%, p = 0.51). That independence is the point of a second axis and also the warning that comes with it: the layer neither confirms nor contradicts a class, and a high predicted share over a gaining basin, as in Java and on the south-east China coast, is not a contradiction — a basin can be refilling overall while a city inside it compacts. Eight basins carry both a deficit class and 1% or more predicted sinking ground, and those eight are where the two halves of the definition would first be tested together. Where measured radar also exists the prediction and the measurement disagree in both directions and by factors of three to five, partly because that radar was processed over Iran only, so a basin crossing the border is measured in part. Both numbers are printed side by side on the basin sheet, the prediction is labelled a prediction, and none of it enters a class, a threshold or a confidence grade.
Measured against the formal definition, indicator by indicator. Water bankruptcy has a published definition (Madani, 2026), and it names five kinds of indicator. Rather than claim the definition loosely, this is where this dataset sits against each one. Three of the five it does not touch at all, and saying so is the point of the table.
| Indicator in the definition | What this dataset provides |
|---|---|
| Chronic water deficit — withdrawals above renewable availability | Nothing. No withdrawal, irrigation or allocation data enters this tool anywhere. It sees the consequence in storage, never the abstraction. |
| Groundwater mining — persistent decline in groundwater storage, particularly where pumping exceeds natural recharge and subsidence is observed | Two of the three parts. A groundwater trend per basin with its interval, the part observed weather cannot account for, a confidence grade, and a subsidence layer beside it on 57 basins. Pumping against recharge is the third part, and nothing here measures either quantity: the decline is the consequence, not the abstraction that caused it. |
| Manifested irreversible impacts — drying lakes and wetlands, subsidence, salinisation, desertification | Two of several. Lake and reservoir storage is now reported as a trend in its own right for 49 basins, and subsiding ground for 57. Wetland extent, perennial-river loss, salinisation and biodiversity are absent. |
| Structural over-commitment — entitlements beyond what is hydrologically available | Nothing. This needs legal and allocation records, which are not satellite-visible. |
| Valid narratives of permanent scarcity — official recognition that past deliveries cannot be met | Nothing. This is a documentary judgement, not a measurement. |
So this dataset covers part of two indicators and none of the other three. That is why every class here is written as insolvency and never as bankruptcy, and why each basin sheet shows the two conditions side by side with the gaps left visible. The definition is deliberately agnostic about numeric thresholds, which vary by basin and ecosystem; the thresholds here are screening conventions in exactly that spirit, to be set locally.
One part of the definition this dataset can speak to: water bankruptcy is a property of human–water systems, not of natural ones. The climate model here gets part of the way to that distinction — it reports how much of a storage trend the basin's own observed rainfall and temperature account for, and how much they do not, with an interval and an out-of-sample test rather than an assumption. What is left over is not the human share: glacier melt, land-use change and error in the model itself sit in it too, which is why every label on this page reads explained or not explained by observed weather and never human or natural.
Two things it does not measure at all. Water stress is demand against supply, and no withdrawal, irrigation, population or allocation data enters this tool anywhere, so two basins with the same storage trend can face completely different stress and nothing here can tell them apart. And irreversibility, the second half of water bankruptcy, is still not assessed: persistence is shown above, permanence is not. The layer just described is a screening prompt on a separate axis, not a measurement of lost capacity, and the only measurement in hand remains radar-observed subsidence in three of the six Severe insolvency basins. This screening measures insolvency, and stops there.
Ground evidence is still thin where it matters most. Open piezometer networks cover six of the 33 beyond-climate basins and none of the six Severe insolvency basins, because those six sit in countries with no public well data. What they carry instead is stated basin by basin: radar-measured subsidence in three, a rainfall test that needs no model in one, published in-situ observation wells for part of the basin in one, and a published recharge-against-abstraction assessment in one. Those published studies are shown on the basin sheet and left out of the confidence grade, because a citation should not quietly promote a basin. Six Severe insolvency basins and eight Drought deficit basins are small groups, and the contrast between them is a result on this record rather than a general law.
A separately produced satellite trend field was published in 2026 by O’Neill, Rodell and Loomis Published. It measures the same planet a different way and has never seen this tool. The same 202 basins are measured twice below — it plays on its own, and you can stop it at any step.
Every basin AquaSolvency screens, coloured by the storage trend it measures from the monthly CSR, GSFC and JPL mascons. Red is loss, blue is gain. This is the picture the rest of the dashboard is built on.
O’Neill, Rodell and Loomis regress the satellites’ Level-1B ranging data directly, at about 111 km instead of the 235–350 km a monthly mascon field carries. Their field was averaged over these same outlines. Watch the colours: almost nothing moves.
Each basin leaves the map and takes its place by its two numbers — this tool along the bottom, the separately produced estimate up the side. A basin on the diagonal is one the two methods agree about.
Correlation 0.96 across the basins this tool assesses. The shaded band is where the two land within 1 mm a year of each other, which is most of them.
Severe insolvency is reserved for a basin whose storage has fallen past a full year’s natural swing, has not recovered since 2021, is still falling, and is mostly not explained by its own rainfall and temperature. The separately produced field reads every one of them as a loss.
Sixteen basins disagree on the direction of change. Every one is classed Stable or Gaining, and none is steeper than 3.4 mm a year — the two methods are arguing about noise, not about loss.
Confidence is set from basin size, how far the three satellite solutions disagree, how stable the class is under 288 alternative settings, and the climate model’s out-of-sample skill. None of that knows anything about this product — and yet High-confidence basins agree most closely. Confidence starts from basin size and larger basins are measured better by both methods, so this is not the label acting alone; the ordering is what matters.
The basins this tool declines to assess — ice-dominated, or with a gravity signal distorted by ice, post-glacial rebound, the Caspian or earthquakes — disagree more than twice as much as the rest, and they hold every one of the large gaps. Those rules were written long before this comparison existed.
Trends read from monthly mascons are supposed to run small, because heavy smoothing spreads a compact signal over ground that is not changing. The paper puts that penalty worst below about 100,000 km². River basins are mostly far larger, and it shows: across the losing basins the median ratio between the two estimates is 1.01. Where they do differ in the strongest class, this tool is the conservative one.
O’Neill, Rodell & Loomis, PNAS 2026 Published
O'Neill, Rodell and Loomis (PNAS, 2026) estimate water storage trends a different way: instead of fitting a line to monthly mascon maps, they regress the satellites' Level-1B inter-satellite ranging data directly, which resolves about 111 km instead of the 235–350 km a monthly mascon field carries. It is one number per place for 2002–2025, so it cannot replace the monthly series this tool needs — but it can be averaged over the same basin outlines and compared. Neither method has seen the other — but both read the same GRACE and GRACE-FO observations, so this is a check on processing and aggregation, not a second, independent measurement of the Earth. What it can establish is that these basin trends are not an artefact of monthly-mascon smoothing or of how this tool averages over a basin. What it cannot test is anything the two share: the GRACE systematic errors, and the glacial-isostatic-adjustment model, which is ICE‑6G_D in every product used here. Four things are tested below, and each one tests a decision this tool makes: the confidence label, the exclusion rules, the class boundaries, and whether reading trends from monthly mascons biases them low.
Confidence is set from basin size, how far the three satellite solutions disagree, how stable the class is under 288 alternative settings, and the climate model's out-of-sample skill — none of which knows anything about this product. Where the tool says High, the separately produced estimate agrees most closely. Confidence starts from basin size and larger basins are measured better by both methods, so this is not the label acting alone; the ordering is what matters to a user.
Every basin is one dot, placed by how far the two estimates differ. The vertical line is each group's median. Hover a dot for the basin; select it to open it.
Every loss class agrees on the direction in every basin. All of the disagreements sit in Stable and Gaining, where the trend is near zero and the two methods are arguing about noise rather than about loss.
For every assessed basin a synthetic world was built that contains, by construction, no water use at all: the basin's own seasonal cycle, a climate response driven by its real rainfall and temperature including their real trends, and noise resampled from its real residual. The classifier was then run on it unchanged. Across 1,530 such runs it produced no severe insolvency, 2.0% structural deficit, 6.3% drought deficit and 12.0% declining, against 64.4% correctly stable. Spurious gains (12.7%) and spurious declines (12.0%) came out at the same rate, so the method does not lean towards finding deficits. Injecting a known loss into the same world gives the other half of the picture: the median basin reaches 80% detection at 2 mm/yr, and of the 148 basins that reach 80% detection at all, 90% do so by 5 mm/yr — the remaining five never reach it at any rate tested, and a loss in those basins can go unseen. One class fails: chance alone would produce about 0.8 hidden deficits across the 153 basins this test covered — the assessed set at the time it was run, two short of today's 155 — and exactly one is observed — so that single label is a prompt to look, not a result.
Recomputed with each of the three GRACE solutions separately over each study's own period. Boundaries and processing differ, so agreement within about a third is a pass. One benchmark differs in sign; it is shown, not hidden.
| Study | Published | AquaSolvency | Verdict |
|---|
Every limit below is already stated somewhere on this page or in the deposited record. Scattered, a reader meets them one at a time, if at all. Gathered, they can be read before the numbers are used rather than discovered afterwards. All the figures here are computed from the same tables the rest of the page is drawn from.
Nothing in this dataset is measured at a national boundary. Every figure here is a basin's figure attributed to a country by the share of the basin inside it, and the measurement resolves about 300 km. Country outlines and names are Natural Earth's; like any such layer they follow one cartographic convention, and this page does not adjudicate any boundary.
Every number on this page answers one particular set of choices: one solution ensemble, one set of rules, one standard of evidence. If you would have chosen differently, this is where you find out whether it matters. Usually it does not. Choosing a single mascon solution instead of the ensemble moves the class on between 33 and 41 of the 155 assessed basins — a fifth to a quarter of them — and that is not something you should have to take on trust.
Disagree with a chart on this page? Draw your own. Every number here is the one in the results table — nothing is smoothed, filled in or generated for the picture, and a basin with no value is named as missing rather than drawn as zero. Take the figure as a PNG and the rows behind it as a CSV.
One tank per basin. Filling up or running dry? The table below has every number.
Each tank's level is the basin's storage change since 2002–05, measured in its own normal yearly swing; very large changes are compressed so no tank looks completely empty or full. The number under each tank is the change in km³. The tank is a picture of change, not of how full the basin really is. Low-confidence basins are marked. Hover a tank for numbers; select it to open the basin. The search and filters below also apply.
| Basin | Status | Confidence | Trend mm/yr | km³/yr | Change km³ | Loss ÷ swing | Beyond climate | Groundwater | Robustness | Trajectory | Area km² |
|---|
Every number on this page is also a file. Static JSON, served from the same address, so there is no server to go down and no rate limit to hit. Pull it into your own product, your own map, or your own report.
/api/v1/meta.json
what this covers, what it does not claim, and when it next updates/api/v1/schema.json
every field, its unit and its meaning/api/v1/basins.json
all basins, summary fields, one request/api/v1/basins/<id>.json
one basin in full — Basin as JSON on any basin opens itThe monthly series is not distributed in bulk. It is plotted on every basin page, and the full series for a basin is sent on request, for checking, replication or review — the same way the deposited files are handled while the manuscript is in preparation. Everything needed to quote, check or cite a result is in the files above.
<id> is the HydroBASINS identifier and does not change
between versions. Paths and field meanings under /api/v1/ do not change either:
fields may be added, but a breaking change goes to /api/v2/ and v1 stays where it
is. Null means not known — it never means zero.
checks.independently_measured
Every basin has a permanent link of the form
#basin=<id> that reopens the page on it. Cite this basin in the basin
panel writes the citation, with that link, the basin's own numbers and the retrieval date.
A report that cites the whole dataset makes its reader hunt; one that cites the basin lands
them on it.
Results, classifications, time series and text are CC BY 4.0 — reuse them, including commercially, with attribution to Kaushik, P. (2026). AquaSolvency. The analysis code is not distributed under that licence. Third-party inputs keep their own licences; the References section lists every one.
Open satellite and climate data, and a method written out in full. Every step is below, including what the measurement can and cannot see. The data record, the thresholds and every uncertainty interval are here, so the work can be checked on its substance.
Before any result, the machinery: every source that goes in, how a satellite grid becomes a number 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.
Read it as one page instead · download the PDF · watch the 90-second film
Open data · open method · every source cited
Nothing below is a diagram of the method. Every line is the actual record of one basin, and every number beside it is read straight out of that basin’s row in the results table, so what the film shows and what the table holds cannot drift apart.
Worked on
GRACE and GRACE-FO measure the pull of the water beneath them. Three groups — the Center for Space Research, NASA Goddard and NASA’s Jet Propulsion Laboratory — turn those same observations into monthly maps of stored water, each making its own choices about how to regularise and constrain the solution. All three are averaged over the basin’s outline. Where their trends differ, that difference measures the choices, not the measurement: they are not three independent looks at the planet.
The three become one ensemble record. Eleven months in the middle have no satellite at all: the first mission ended its science operations in October 2017 and GRACE-FO launched in May 2018. The gap is left open. Trends are fitted across it, but no modelled value is ever inserted into it, so a reader can see exactly where the instrument was silent.
Every basin fills in its wet season and empties in its dry one, and in a monsoon basin that swing is far larger than any trend. The seasonal cycle is fitted with harmonics and removed, and its peak-to-trough range is kept as the basin’s annual swing — later used as the yardstick for how deep a loss has to be before it counts as more than one hard year.
The trend is estimated with Theil–Sen on annual means, which takes the median of all pairwise slopes and so cannot be dragged by one extreme year. Significance uses the Hamed–Rao modified Mann–Kendall test, which allows for the fact that one year in a water record resembles the next — ignoring that would make almost everything look significant.
Total storage holds soil moisture, snow, surface water and groundwater together. Modelled soil moisture and snow water equivalent are subtracted, and so is lake and reservoir storage wherever a record exists — gauged storage in the Colorado, satellite lake storage elsewhere. What is left is the groundwater estimate. Basins holding very large managed lakes with no usable record keep a flag instead of an estimate.
Before anything is called a water-use problem, the basin’s own weather is given every chance to explain it. An antecedent-climate model is fitted to the basin’s rainfall and temperature anomalies with a memory of up to sixty months, so a long dry spell can still be felt years later. The model is trained on the early record and scored on the later one, so its skill is measured on years it has never seen.
The part the climate model accounts for is set aside, and what remains is the beyond-climate rate, with a 90% interval from a block bootstrap that resamples whole two-year blocks rather than single months. That remainder is not read as pumping alone: it holds diversions, reservoir operations, land-use change, glacier melt and the model’s own error together, and this tool does not separate them.
A label is only worth as much as its thresholds. Each of the five thresholds is moved to three settings — 243 combinations — and the class is counted again; the whole analysis is then rerun 288 further ways, and the satellite record prepared 48 different ways. A basin that changes class easily is marked down in confidence rather than quietly reclassified. And the strongest label this tool will give is insolvency: bankruptcy in the formal sense also requires irreversible damage, which a storage record alone cannot establish.
An antecedent-climate model (after Humphrey et al., 2017) is fitted to detrended storage, so it learns only how storage answers wet, dry, hot and cool spells:
S′(t) = a·Σ P′(t−j)e−j/τ + b·Σ T′(t−j)e−j/τ, a ≥ 0, b ≤ 0
P′ = GPCP v2.3 precipitation anomaly, T′ = GHCN-CAMS temperature anomaly (vs 1981–2010), τ = memory, 1–60 months. The remainder is beyond climate: pumping, diversions, reservoir operations, land-use change, glacier melt and model error. Its 90% interval comes from a 150-member block bootstrap (24-month blocks). Model skill is tested out of sample: trained on 2002–2014, scored on 2015–2026.
All input data are public and cited, and every step of the method, every threshold and every uncertainty interval is written out on this page so the work can be judged on its substance now.
The derived dataset — the basin table, the monthly series, the data dictionary, the schema and every robustness and validation table behind the numbers on this page — is deposited as a citable record on Zenodo. Its files are held closed while the manuscript is in preparation, so the record carries the authorship and the date while the contents stay with the author. Researchers and basin agencies who want to check a specific result can ask for a file before then, and those requests are answered.
Every threshold is published together with the test that shows how stable it is. Basin managers who check one against their own records are what carries a threshold from screening to decision grade — and every accepted correction is credited.
Published with the reviewer's permission. A review appears here only after the author releases it, and a name appears only where the reviewer has agreed in writing — otherwise the review stands on its own. Corrections are published whether or not they flatter the tool.
Plain answers to what people ask most. If your question isn't here, or you have a correction, an idea or ground data to share, send it below.
No question matches. Ask it in the form and it may be added here.
Who develops it, why it exists, and the other tools that came out of the same work.
Geospatial & Data Visualization Expert | GeoAI | EO Expert
I work with satellite and geospatial data, and with the question of how to make a measurement legible to the person who has to act on it. AquaSolvency is developed, written and maintained by me; the analysis, the method and the dashboard are all open, so anyone can check them or disagree with them.
ORCID 0000-0002-0767-3983
The measurements that would tell you whether a river basin is living beyond its means already exist, and most of them are open. They are just not in a form a basin manager, a journalist or a student can pick up and use.
In 2026 a report from the UN University's Institute for Water, Environment and Health described many basins and aquifers as being in “water bankruptcy”. That is a strong phrase, and it deserved a check anyone could inspect. Formally, bankruptcy means insolvency plus irreversible damage; the first of those can be measured from space, basin by basin: which basins, how fast, and is it the weather or something else?
AquaSolvency is that check. Every basin is measured with three mascon solutions processed independently at CSR, NASA GSFC and NASA JPL, the classifier is run on synthetic basins built to contain no water use at all to see what it invents, and every threshold is published — so the people who manage these basins can test them against their own records and change them.
Want to collaborate, or point out something wrong? The feedback form reaches me directly, and every accepted contribution is credited.
Ownership, citation and terms of reuse.
Author: Dr. Pradeepika Kaushik · ORCID 0000-0002-0767-3983
Results, classifications, time series, visualisations and text © 2026 Dr. Pradeepika Kaushik, licensed under CC BY 4.0: use, share and build on them, with credit.
Analysis and dashboard code © 2026 Dr. Pradeepika Kaushik. All rights reserved.
Input data remain under their providers' terms (see References), including the HydroSHEDS attribution below.
v2.7, built . What changed since the last build: an evidence layer that records what every outside source says about every basin and, just as deliberately, where nothing says anything; a recovery test comparing 2002–2017 with 2018–2026 once the shift between the two satellite missions is removed; measured subsidence split into elastic and permanent compaction where radar reaches; a European archive of 15,760 monitoring wells, which reaches fifteen basins and seven that had no independent measurement at all; a bound on abstraction from an independent model; an answer, rather than a shrug, for the three basins whose average and whose ground point different ways; and this country view.
How often it is refreshed. The satellite record gains a month roughly every month, and this page is rebuilt when enough has accumulated to move something — in practice two to four times a year, and whenever a correction is found. The build date in the footer is the one to check; a version number only moves when a new record is deposited, so two different builds can both read v2.7. Corrections are made as soon as they are found rather than held for a release.
From this page, now: the basin table as CSV or JSON from the Basin-level results section, a one-basin brief from any balance sheet, and the citation and full reference list from the panels beside this one.
In the deposited record: the basin table with all its fields, the monthly series, the data dictionary, a machine-readable schema (Frictionless datapackage.json), the robustness and validation tables — 243 classifier-threshold combinations, 288 analysis reruns, 48 record-preparation variants, the well and subsidence checks, the synthetic-basin test, the recomputed published benchmarks and the independent trend cross-check — and the written method as it stands here.
The record is deposited on Zenodo with the DOI above. Its files are held closed while the manuscript is in preparation, so the record carries the authorship and the date while the contents stay with the author. Researchers and basin agencies who need a file to check a specific result are welcome to ask, through the request button on the record or the form below.
Also in it: the basin polygons as GeoPackage and GeoJSON, with their attribute tables rebuilt at this version, and the monthly series as a single JSON file. The outlines are HydroBASINS, lightly simplified, because GRACE resolves about 300 km and finer detail carries no information.
AquaSolvency stands on open data and decades of shared science: the GRACE and GRACE-FO missions (NASA, DLR, GFZ), CSR at the University of Texas, NASA Goddard, NOAA PSL, HydroSHEDS, the USGS, the US Bureau of Reclamation and Natural Earth, and the researchers whose methods and studies are listed in the references. Thank you to everyone who shares water data openly.
Answers read from this page's own table and text · no language model · works offline