Separating cumulative and differential land subsidence with machine learning

Land subsidence driven by unsustainable groundwater extraction has become one of the most damaging and least visible geohazards of the twenty-first century, silently sinking cities and agricultural plains from Jakarta to Mexico City. In Iran, where decades of aquifer overexploitation have pushed many basins to the brink, a new study offers what its authors describe as a scalable, machine learning-based early-warning framework capable of forecasting where and how fast the ground will sink. The research, published in the journal Natural Hazards, goes beyond simply predicting total subsidence: it decouples cumulative displacement from differential subsidence rates, revealing a fundamental geomechanical dichotomy that changes how such hazards should be modeled and managed.

The team, led by Amirreza Shahmiri of Sultan Qaboos University together with Masoud Ebrahimi Derakhshan and Seyed Mostafa Siadatmousavi of Iran University of Science and Technology, assembled an unusually rich decade-long dataset spanning 2015 to 2025. At its core are displacement time series derived from Sentinel-1 interferometric synthetic aperture radar, or InSAR, a satellite technique that measures millimeter-scale ground deformation by comparing the phase of radar signals returned from the same points on Earth’s surface across repeated passes. The researchers supplemented these radar-derived displacement measurements with hydro-climatic variables from the ERA5-Land reanalysis dataset, which provides gridded estimates of precipitation, soil moisture, and evapotranspiration, as well as static soil properties from the ISRIC SoilGrids database. All of the data were processed openly, drawing on platforms such as Google Earth Engine and the COMET-LiCS Sentinel-1 InSAR processing facility.

The methodological design is notable for its breadth. Four machine learning algorithms were evaluated head-to-head: ElasticNet, a regularized linear regression; LightGBM and XGBoost, two gradient-boosted decision tree methods prized for their speed and ability to capture nonlinear interactions; and BiLSTM, a bidirectional long short-term memory neural network designed to learn temporal dependencies in sequential data. Rather than committing to a single feature set, the authors tested three input scenarios: a History-Only configuration relying purely on past displacement, an Environmental-Only configuration using climate and soil variables, and a Combined scenario integrating both. Validation was deliberately rigorous, employing a multi-city split, an in-city temporal split in which the models were trained on the first 80 percent of the record and tested on the final 20 percent, and a demanding Leave-One-City-Out cross-validation scheme in which an entire basin was held out from training to test whether the models generalize to unseen terrain.

The results exposed what the authors call a critical geomechanical dichotomy, and it is here that the study makes its most consequential contribution. For the prediction of cumulative displacement, the simple linear ElasticNet model outperformed its sophisticated competitors, achieving a coefficient of determination greater than 0.99. The explanation, illuminated by SHAP, or SHapley Additive exPlanations, analyses, is that cumulative subsidence is dominated by autoregressive momentum: the accumulated sinking to date largely dictates the trajectory that follows, a signature of the slow, largely irreversible compaction of aquifer sediments once pore pressures collapse. In other words, for total displacement, the past is overwhelmingly the best predictor of the future, and a linear model with access to historical displacement captures nearly all of the explainable variance.

Differential subsidence, by contrast, the month-to-month variability in the rate at which the ground sinks, tells a fundamentally different story. Predicting these dynamic, non-linear rates favored the gradient-boosting algorithms LightGBM and XGBoost, and the SHAP attribution shifted decisively toward environmental fluxes. Hydrological variables such as soil moisture dynamics and precipitation-related terms modulate the monthly rate of deformation, reflecting the elastic and inelastic responses of the aquifer system to seasonal recharge and pumping cycles. This is the physical fingerprint of an aquifer being squeezed and, occasionally, partially recovering, and it cannot be captured by a model that assumes linear momentum alone.

Perhaps the most sobering finding emerged from the Leave-One-City-Out cross-validation, which probed the limits of spatial transferability. Cumulative subsidence trajectories transferred remarkably well to basins the models had never seen, achieving a mean R² of 0.96, precisely because their autoregressive memory travels with them: a sinking basin, wherever it is, keeps sinking along a similar momentum-driven path. Differential-rate models, however, collapsed when applied to unseen cities, with mean R² values of 0.06 or lower. The reason is that these rates depend on localized, basin-specific signatures, combinations of soil texture, aquifer architecture, and extraction patterns that do not generalize across regions. The practical implication is stark: early-warning systems built around differential rates cannot simply be exported from one basin to another; they must be trained, at least in part, on local data.

Having established these foundations, the team extended the framework to genuine forecasting. Multi-horizon predictions were generated for lead times ranging from one to sixty months, accompanied by quantile-based uncertainty bounds that communicate not just a single predicted value but a probabilistic envelope of plausible outcomes. The skill of the models decayed gracefully with horizon length, remaining high out to a full year, with an R² of 0.938 at the twelve-month mark. From these forecasts the researchers produced rate-based risk maps that delineate persistent high-risk nuclei beneath the irrigated plains of Bardsir and Neyshabur, two Iranian agricultural heartlands where irrigation demand has historically driven intense groundwater withdrawal. These hotspots, the study suggests, are not transient anomalies but durable zones of deformation that will remain hazardous under current management regimes.

The choice of Iran as the case study is hardly incidental. Satellite analyses published in recent years, including work in Science Advances by Haghshenas Haghighi and Motagh, have documented widespread, accelerating subsidence across Iranian plains as aquifers are drawn down faster than they can recharge, a situation aggravated by climate change and drought. Subsidence of this kind is largely permanent: when clay-rich aquitard layers compact, their pore structure collapses and they cannot be re-inflated even if water levels recover. The resulting damage includes fissuring of the land surface, cracking of roads, buildings and pipelines, and heightened vulnerability to flooding as the terrain loses gradient. Because the deformation unfolds slowly and without visible surface expression, however, it often escapes political attention until the damage is done, which is precisely why predictive monitoring tools carry such practical weight.

The study also carries lessons for the growing field of geohazard machine learning more broadly. Much of the existing literature applies a single algorithm to a single target and reports impressive in-sample statistics. By systematically varying the predictive target, the feature scenario, and the validation scheme, this work demonstrates that such one-size-fits-all evaluations can be misleading. A model that looks nearly perfect for cumulative displacement may be structurally unsuited to differential rates, and apparent accuracy on one city may conceal a total inability to generalize to the next. The authors argue that their multi-scenario, multi-target framework provides a template that can be replicated in other subsidence-prone regions, using freely available Sentinel-1 radar data, ERA5-Land climate reanalysis, and SoilGrids soil properties, with no need for expensive in-situ instrumentation as a prerequisite.

In line with the movement toward open and reproducible science, the code and dataset underpinning the study have been released openly on GitHub, allowing other researchers and water agencies to adapt the pipeline to their own basins. The authors, who received no external funding for the work, position the framework as a decision-support tool for localized early-warning monitoring, one that can inform sustainable groundwater management and infrastructure risk planning in tandem. As aquifer depletion accelerates under climate stress in arid and semi-arid regions worldwide, the ability to distinguish the inexorable momentum of cumulative sinking from the environmentally modulated fluctuations of differential rates may prove essential for knowing not only where the ground will sink, but when, how fast, and what can still be done about it.