Satellite Gravity Data Sharpened by AI Reveals Alarming Groundwater Loss in the Ganga Basin

Beneath the vast alluvial plains of the Ganga Basin, one of the most densely populated regions on Earth, water is disappearing faster than anyone can measure it directly. Groundwater sustains hundreds of millions of people across northern India, feeding agriculture, industry and household wells, yet the network of monitoring wells that should track its decline is patchy, unevenly distributed and often incomplete. Now a team of researchers at the Indian Institute of Remote Sensing, part of the Indian Space Research Organisation, has found a way to squeeze far more detail out of a pair of aging satellites that measure gravity from orbit, producing the sharpest picture yet of where the basin’s hidden water reserves are vanishing and when drought struck hardest.

The satellites in question belong to the Gravity Recovery and Climate Experiment, known as GRACE, a joint NASA and German Aerospace Center mission that operated from 2002 to 2017 and was later continued by its successor, GRACE Follow-On. Rather than photographing the surface, GRACE tracks the distance between two orbiting spacecraft with extraordinary precision. When the mass of water stored in soils, aquifers and surface reservoirs shifts beneath them, the resulting gravitational tug subtly alters that distance. By measuring these minute changes, scientists can calculate how much water a region gains or loses each month, a quantity called the change in terrestrial water storage. The technique has transformed hydrology, but it comes with a fundamental catch: the gravity signal is smeared across enormous areas, so a single GRACE data point represents a block of land roughly 150,000 to 200,000 square kilometers in size. For a basin where a farmer’s tube well in one district may be draining an aquifer while a neighboring district recharges after monsoon rains, that blur hides exactly the local detail that water managers need most.

In a study published in Environmental Monitoring and Assessment, Chethan Varadaganahalli Anandagowda, Bhaskar R. Nikam and Suresh Kannaujiya tackled this resolution problem by teaching statistical models to translate the coarse satellite signal into fine-grained maps. Their approach, known as downscaling, exploits the fact that water storage changes are closely linked to variables that satellites and models can observe at much finer scales, including rainfall, soil moisture, evapotranspiration, temperature and terrain. The researchers built monthly climatology-based regression models that relate these high-resolution environmental datasets to the GRACE measurements, then used the learned relationships to estimate water storage change on a grid of 0.25 degrees by 0.25 degrees, roughly 25 to 28 kilometers per side. That is a dramatic sharpening of the effective resolution, bringing the satellite picture down to a scale where individual districts and aquifer systems begin to separate.

The heart of the method was a comparison of several statistical and machine learning approaches, including multiple linear regression, partial least squares regression and random forest modeling. The random forest technique, an ensemble method that builds hundreds of decision trees and averages their predictions, emerged as the clear winner. It achieved a Pearson correlation coefficient of 0.78 against validation data, meaning it captured more than three quarters of the variance in the observed storage signal, with a root mean square error of 12.3 centimeters of equivalent water height. In practical terms, the model could reproduce the month-to-month ebb and flow of the basin’s water storage with an accuracy that makes the downscaled maps genuinely useful rather than merely illustrative. The team validated their results using k-fold cross validation, a technique that repeatedly trains the model on most of the data and tests it on the withheld remainder, guarding against the model simply memorizing patterns rather than learning real physical relationships.

With the sharpened storage maps in hand, the researchers turned to the question that matters most for the Ganga Basin: what is happening to its groundwater specifically. GRACE measures all water stored on and below the land surface, so isolating the groundwater component requires subtracting the contributions of soil moisture, surface water and canopy storage. The team used products from the Global Land Data Assimilation System, a NASA modeling framework that simulates land surface water fluxes, together with satellite soil moisture records from the European Space Agency’s Climate Change Initiative, to strip away these shallower reservoirs. What remained was an estimate of groundwater storage change, which they then compared against observations from monitoring wells across the basin, converted to storage terms using specific yield values for the aquifer material.

The agreement between the space-based and ground-based estimates was striking, with both showing a consistent declining trend across the basin. But the fine-resolution data revealed something the coarse GRACE footprint had smoothed away: the decline is not uniform, and in some places it is far steeper than previous studies had suggested. Over the Delhi National Capital Region and Western Uttar Pradesh, regions where explosive population growth and intensive irrigation place enormous pressure on the aquifers, the downscaled estimates showed groundwater storage falling at rates of 4.69 to 7.16 centimeters of equivalent water height per year. Those figures are more severe than the trends reported in earlier basin-wide analyses, and they pinpoint the western Ganga plains as the epicenter of the basin’s groundwater crisis. The finding matters because it demonstrates that coarse satellite data, by averaging heavily depleted zones with less stressed ones, can systematically understate the intensity of depletion in the worst-affected areas.

The study went beyond storage trends to tackle drought, using multiple drought indices computed from the water storage record to identify major hydrometeorological drought events across the basin. Drought indices condense complex hydrological information into standardized measures of severity, allowing events to be compared across time and space. The analysis identified several significant drought episodes during the satellite era, but one stood out above the rest: the period from August 2015 to October 2017, which the researchers identified as the most intense drought in the record, persisting for a remarkable 27 consecutive months. That two-and-a-quarter-year event, which coincided with consecutive weak monsoon seasons, placed sustained stress on both surface reservoirs and the aquifers that farmers turned to when canals and rivers ran low. Because the downscaled data resolves the drought’s footprint at district scale, it shows which parts of the basin bore the brunt of that prolonged deficit rather than blurring the event into a basin-wide average.

The implications extend well beyond academic hydrology. India’s Central Ground Water Board assesses the country’s groundwater resources annually, but its estimates depend on the density and quality of well observations, which vary widely across states. Satellite-derived storage change offers an independent check on those assessments, and the downscaling approach demonstrated here brings that check down to a scale where it can inform district-level management decisions, from regulating well permits to planning managed aquifer recharge. The researchers note that their findings enhance understanding of local-scale groundwater abstractions and can contribute to sustainable water resource management in a basin where climate change is expected to make the monsoon more erratic even as demand continues to grow. Every centimeter of equivalent water height in these maps represents billions of cubic meters, so a decline of seven centimeters per year in a heavily farmed region translates into a withdrawal of water that took centuries or millennia to accumulate.

There are also lessons here for how the global scientific community monitors water in data-sparse regions. The Ganga Basin is hardly unique: aquifers from the North China Plain to California’s Central Valley to the Middle East are being drawn down faster than they refill, and in many of these places the observational infrastructure is thin. The success of the random forest approach, which outperformed classical regression methods by capturing nonlinear relationships between climate variables and storage change, suggests a transferable recipe. The datasets the team relied on, including GRACE mass grids processed at the University of Texas Center for Space Research, CHIRPS rainfall estimates, GLDAS model outputs, ESA soil moisture records and Indian well observations, are all freely available, meaning the workflow could be replicated for other stressed basins at little cost. As the GRACE Follow-On mission continues to deliver gravity data and machine learning tools grow more capable, the gap between what satellites can see and what water managers need to know keeps narrowing. For the Ganga Basin, that narrowing gap has just exposed an uncomfortable truth: the water is leaving faster, and in more places, than the blur of the old maps ever let on.

Subject of Research: Downscaling GRACE satellite gravity data with machine learning to assess groundwater depletion and drought in the Ganga Basin

Article Title: Downscaling GRACE storage change to analyze groundwater dynamics and drought evaluation in the Ganga Basin

Article References: Anandagowda, C. V., Nikam, B. R., & Kannaujiya, S. (2026). Downscaling GRACE storage change to analyze groundwater dynamics and drought evaluation in the Ganga Basin. Environmental Monitoring and Assessment, 198(11), Article 1146. https://doi.org/10.1007/s10661-026-16002-9

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16002-9

Keywords: GRACE, groundwater, Ganga Basin, downscaling, random forest, machine learning, drought, terrestrial water storage, remote sensing, Delhi-NCR, Uttar Pradesh, water resources