High Mountain Asia is often described as a water tower because its snowfields and glaciers collect water at altitude and release it into some of the world’s great rivers. But a second reservoir sits out of sight beneath the ground. It helps farmers through dry seasons, supports cities when surface supplies falter and smooths some of the wild year-to-year swings in mountain weather.

That hidden reserve is shrinking. A study published in Environmental Research Letters estimates that High Mountain Asia lost groundwater at an average rate of 24.2 billion tonnes a year between 2003 and 2020. About 69 percent of the region showed a declining trend.

The losses were concentrated where the consequences are hardest to ignore: the densely populated, heavily irrigated basins downstream of the mountains. The Ganges–Brahmaputra basin accounted for an estimated 12.8 billion tonnes a year, more than half the regional net loss. Declines also stood out in the Indus, Amu Darya and Tarim basins.

The future part of the study is less straightforward. Melting glaciers and thawing permafrost can add water to the ground, partly concealing the underlying decline for several more decades. In the researchers’ projection, that temporary help weakens around the 2060s. Groundwater loss then accelerates.

It sounds like a simple countdown. It is not. The 2060s are a modelled transition under a specified climate pathway and a deliberately simplified assumption about future water use. Understanding the result means separating what satellites observed, what researchers inferred and what the model projected.

The Asian Water Tower is a network, not a tank

High Mountain Asia stretches across the Tibetan Plateau, the Himalaya, the Hindu Kush, the Karakoram, the Pamir and the Tianshan. Snow, ice and rain from these highlands feed river systems including the Indus, Ganges, Brahmaputra, Amu Darya, Tarim, Yangtze and Yellow.

Calling this vast region a water tower is useful, but only up to a point. A tower suggests one tank with one outlet. High Mountain Asia is a web of basins with different monsoons, elevations, geology, glaciers, farming systems and national borders. Water gained on the interior Tibetan Plateau does not simply become water available to a farmer beside the lower Indus.

The climate-and-pumping overlap has appeared in earlier research. ScienceBlog reported in 2023 that warming could sharply worsen groundwater depletion in India if irrigation practices continued on a business-as-usual path. The new reconstruction places that pressure inside a much larger mountain-to-basin system.

Some high-elevation inland areas gained groundwater. Across the inner Tibetan Plateau, the study estimated a combined increase of about 3.51 billion tonnes a year, plausibly supported by precipitation, glacier melt, thawing permafrost and exchanges between expanding lakes and groundwater. Those gains reduced the net regional loss. They did not cancel the severe declines in the large downstream basins.

How satellites weighed water they could not see

The backbone of the analysis came from NASA and German gravity missions: GRACE, which operated from 2002 to 2017, and GRACE Follow-On, launched in 2018. Each mission uses a pair of spacecraft flying one behind the other. As the leading satellite approaches a slightly stronger gravitational pull, it speeds up by a minute amount and changes its distance from its partner. Instruments measure those changes with extraordinary precision.

Month by month, the measurements reveal changes in mass across broad areas of Earth. Water is heavy, so a region that gains or loses a large quantity of water changes the local gravity field. As NASA explains, the resulting terrestrial-water estimate includes groundwater, soil moisture, surface water, snow and ice.

This is also the instrument behind the broad changes discussed in ScienceBlog’s 2025 report on NASA’s nearly two-decade view of shifting global water patterns. The High Mountain Asia study asks a narrower and more difficult question: how much of the combined regional mass change can reasonably be assigned to groundwater?

This is the most important methodological boundary in the paper: GRACE does not directly photograph aquifers or measure the depth of a particular well. It weighs the combined water account over a large area.

To isolate groundwater, Kai Liu and colleagues subtracted estimates for canopy water, soil moisture, snow, glaciers, permafrost and lakes from the total terrestrial-water signal. Think of a household statement that gives only the combined balance of several accounts. If the balances of all but one account can be estimated, the remainder can be assigned to the last. The answer can be highly informative, but uncertainty in each subtraction flows into the remainder.

The team averaged two independently processed GRACE “mascon” products, from the University of Texas Center for Space Research and NASA’s Jet Propulsion Laboratory. Short missing periods were filled, including the longer gap between GRACE and GRACE Follow-On. The researchers also propagated uncertainties from the total-water signal and the component estimates.

That procedure is one reason to retain the word estimated beside 24.2 billion tonnes. It is a regional reconstruction, not a meter reading. It is also one of the few practical ways to obtain a consistent view across mountains and borders where well records are sparse, inaccessible or measured under different systems.

The wells supplied a crucial reality check

The researchers did not leave the satellite-derived result floating on its own. They compared it with records from 2,554 groundwater wells that had at least eight continuous years of observations during the study period. Most of those wells were in India and were concentrated in the Ganges–Brahmaputra and Indus basins.

Of the 2,554 wells, 2,047, or about 80 percent, showed statistically significant declining trends. That agreement supports the broad picture of depletion in the two most heavily monitored downstream basins.

It does not turn the network into uniform ground truth for all of High Mountain Asia. A cluster of wells in accessible agricultural plains cannot validate every aquifer beneath the plateau, and a well measures a local water level rather than regional water mass. The two methods observe different scales. Their convergence is reassuring precisely because their weaknesses are different.

The distinction between level and storage also matters. A metre of decline in one aquifer need not represent the same volume of water as a metre in another. Storage depends on the thickness, area and physical properties of the water-bearing formation. The gravity record helps with the regional mass change; wells reveal what people encounter locally.

Where 24.2 billion tonnes a year was lost

One billion tonnes of freshwater occupies roughly one cubic kilometre. The annual net loss estimated here is therefore about 24.2 cubic kilometres, close to ten million Olympic swimming pools. That comparison conveys scale, but it can also mislead if it makes the water seem as though it vanished from Earth.

The figure is a change in groundwater storage. Water pumped for irrigation may enter crops, evaporate, run into rivers or eventually reach the ocean. Other groundwater may drain naturally or fail to be replenished after weak precipitation. The study does not say that all 24.2 billion tonnes was pumped, nor that it all had the same destination.

Basin results show how a regional total can conceal very different stories. The Ganges–Brahmaputra loss was estimated at 12.8 billion tonnes a year. The Amu Darya lost about 3.3 billion and the Tarim about 2.1 billion tonnes annually. The Indus also showed a major decline, though the study’s central message is the shared exposure of the large, intensively irrigated downstream basins rather than a league table.

The consequences of a falling water table do not wait for an aquifer to empty. ScienceBlog’s earlier report on aquifer depletion and crop yields showed how the growing difficulty of extracting groundwater can reduce production during drought. High Mountain Asia’s farming systems and aquifers are different, but the basic sequence matters here too: less accessible water can become a practical limit long before the regional storage reaches zero.

Climate, ice and pumping leave overlapping fingerprints

Working out that storage declined is easier than assigning a percentage to each cause. Rainfall, temperature, glacier melt, permafrost thaw, irrigation and urban demand interact. A dry year can reduce recharge and increase pumping at the same time. Meltwater can raise a river, seep into an aquifer and encourage more water use downstream.

The team used a lightweight Transformer, a machine-learning architecture designed to detect relationships and time lags in sequences. They fed it six broad groups of influences: precipitation, temperature, shortwave solar radiation, glacier-mass change, permafrost dynamics and human activity. An interpretation method known as SHAP then estimated how much each input contributed to the model’s groundwater result.

Across the analysis, direct climate drivers accounted for about 47 percent of groundwater variability. Cryospheric processes, meaning glaciers and permafrost, added about 15 percent. Human activity contributed as much as 38 percent of declines, with the largest influence in irrigation-intensive downstream basins.

Those numbers are not a direct causal experiment. They describe how the trained model divided the explanatory weight among correlated inputs. They also do not impose the same split on every basin or year. “Up to 38 percent” should not quietly become “38 percent everywhere.”

Still, the spatial pattern makes physical sense. Groundwater withdrawals for rice and cotton cultivation, industry and domestic supply are substantial in the Ganges–Brahmaputra, Indus and Amu Darya basins. The model found the human signal strengthening after roughly 2008 to 2010, consistent with growing demand and irrigated production.

Why melting ice can make the decline look slower

A warming climate can worsen water security and temporarily increase water supply at the same time. That is the apparent paradox at the centre of the projection.

As glaciers lose mass, some meltwater reaches rivers quickly. Some moves through soils and fractured rock, recharging groundwater. Thawing permafrost can also open pathways that allow more water to infiltrate. For a time, those additions can partially offset groundwater removed by wells or lost through natural drainage.

This is not free, renewable recharge. It is closer to meeting everyday expenses by drawing down a long-held savings account. The cash flow may look healthier while the account still contains capital, even though the underlying reserve is being spent.

Earlier glacier studies have found the same uncomfortable timing above ground: extra melt can swell warm-season runoff before that contribution tapers as the ice reservoir loses mass. The new study extends that logic underground. A pulse of meltwater can soften a groundwater decline without making the combined ice-and-water system sustainable.

There is another complication. More recharge is not automatically available where and when people need it. Water may enter a deep or disconnected formation, arrive during a flood, or flow through a basin without replenishing the wells used by communities and farms. Regional storage and practical supply are related, not interchangeable.

What changes around the 2060s

For the future analysis, the researchers trained their model on the 2003–2020 reconstruction and supplied climate projections from eight CMIP6 Earth-system models. They focused on SSP2-4.5, a middle-of-the-road pathway in which global emissions do not follow either the most optimistic or the highest trajectory.

Under that setup, groundwater storage continued to decline across High Mountain Asia through the end of the century. In vulnerable downstream basins, glacier melt and permafrost thaw slowed the decline around mid-century. Once the rate of added meltwater weakened and warmer conditions increased evaporation and plant demand, depletion accelerated again after roughly the 2060s.

The temporary buffer was especially visible in the Ganges–Brahmaputra, Indus and Amu Darya basins. The model did not predict identical behaviour everywhere. Parts of the inner Tibetan Plateau and upper Amu Darya retained positive storage trends, another reminder that “High Mountain Asia” is not one hydrological unit.

Nor did every climate model agree equally well on the timing or size of changes in the Indus and Ganges–Brahmaputra. Those are exactly the places where human water use adds uncertainty. The 2060s should therefore be read as a broad modelled transition, not a year circled on the calendar and not a deadline after which every aquifer suddenly worsens.

The projection deliberately freezes one moving target

The phrase “if current water use continues” needs care. In the baseline projection, the researchers held human water use at historical 2003–2020 levels. That choice helps isolate the climate signal. It does not claim that population, crops, pumping technology, regulation or water demand will remain unchanged for eighty years.

If pumping grows, some basins could decline faster. If irrigation becomes more efficient, crop patterns change, managed recharge expands or withdrawals are constrained, the path could improve. Efficiency can also have rebound effects if saved water is used to expand irrigated land. None of those social choices can be forecast with the same inputs used to project temperature and precipitation.

The team ran sensitivity tests with different water-use assumptions and found that the broad direction of the long-term trends persisted, while the most human-dominated downstream basins changed more. That is a useful result, but it does not remove the uncertainty.

There is a second limit. The model learned relationships during an 18-year historical window and then extended them into climates outside that experience. Transformer models are good at representing lags and interactions, but they do not repeal the usual difficulty of extrapolation. The study offers a structured risk scenario, not a literal preview of every decade through 2099.

Storage is not the same as usable, safe water

Gravity satellites estimate changes in water mass. They cannot tell whether the remaining groundwater is fresh or saline, shallow or expensive to lift, moving rapidly or trapped in formations that recharge over centuries. They do not measure arsenic, microbial contamination or the condition of a particular village pump.

Aquifers can also suffer damage before their regional storage approaches zero. Pumping lowers pressure, raises energy costs and can cause land to subside. It may draw poorer-quality water into a well field. A broad trend of decline therefore matters even though the research does not identify a single threshold at which the system “runs out.”

The study’s estimate that more than 430 million people face heightened water risk should be understood in this regional sense. It does not mean 430 million people will lose water on a particular date. It means large populations live in basins where the storage buffer is weakening and where climate and withdrawals can reinforce each other.

The useful warning arrives before the buffer is gone

The immediate value of the research is not that it produces a dramatic 2060s deadline. It is that three very different kinds of evidence line up: gravity-based mass changes, thousands of declining wells and a future model in which the ice-derived reprieve does not last.

There are also reasons for humility. Groundwater is the remainder after several uncertain water components are subtracted. Monitoring wells are unevenly distributed. Attribution depends on a machine-learning model, and the projection assumes one emissions pathway and a simplified future for human water use.

Yet those limitations do not make the underlying pattern easy to dismiss. The most heavily populated downstream basins were already losing groundwater during the observed period. A temporary increase in recharge from melting ice is not evidence that the loss has been solved. It may make the balance sheet look less severe while another part of the system is being depleted.

That gives water managers a window, not a grace period in which nothing needs to change. Better well monitoring, shared basin data, crop and irrigation choices, limits on withdrawals and deliberate recharge all take time to design and enforce. Because the major rivers cross borders, cooperation matters as much as measurement.

The clearest way to read the study is to picture two reserves. High Mountain Asia stores water in ice above the ground and in aquifers below it. For several decades, loss from the first may partly disguise loss from the second. By the time that disguise fades, both buffers will be smaller. The sensible moment to respond is while the temporary water is still arriving, not after it stops.

The study was led by researchers at the Aerospace Information Research Institute of the Chinese Academy of Sciences. An institutional summary and the study’s basin graphic are available from the Chinese Academy of Sciences.