Global land subsidence mapping reveals widespread loss of aquifer storage capacity

Excessive groundwater pumping leads to various negative consequences, including land subsidence, aquifer depletion, arsenic contamination, saltwater intrusion, and infrastructure damage. Quantifying groundwater storage loss is challenging due to sparse monitoring networks and inconsistent data availability. While remote sensing offers potential for global-scale assessment, no existing dataset provides a direct estimate of available groundwater storage and loss. Land subsidence, a visible effect of groundwater depletion caused by aquifer material compaction, offers a way to quantify storage loss, particularly in unconsolidated confined aquifers. In-situ measurements exist but lack the spatial density for regional or global studies. Interferometric Synthetic Aperture Radar (InSAR) provides high-resolution subsidence data, but processing is computationally intensive and susceptible to noise. Process-based models offer another approach, but creating a global land subsidence model requires extensive data not readily available. This study uses a machine learning approach to bridge these data gaps and provide a global-scale estimate of land subsidence and associated groundwater storage loss.

Previous research has attempted to connect groundwater storage loss with subsidence at regional scales and to map global subsidence susceptibility. However, no prior study has quantified the global magnitude of subsidence and associated groundwater storage loss at a high resolution. Existing methods face challenges in characterizing groundwater storage loss globally with sufficient resolution for local studies. Regional studies using InSAR have been conducted, but they are limited in geographical coverage. Process-based models offer global estimates of groundwater resources, but a global land subsidence model has not been developed due to the lack of comprehensive geomechanical and hydrogeologic datasets needed to capture the complex, nonlinear processes involved.

This study employs a machine learning method using a random forest algorithm to map pumping-induced land subsidence at a high spatial resolution (~2 km) globally. The model was trained using a comprehensive dataset of InSAR and Global Navigation Satellite System (GNSS)-based land subsidence data from 47 regions worldwide, incorporating hydrologic, land use, climatic, and geologic datasets as predictor variables. These datasets included remotely sensed and model-based data representing proxies for the principal drivers of land subsidence. Data preprocessing involved downscaling/upscaling datasets to a uniform 2-km resolution using the nearest neighbor algorithm. Hydrologic datasets included precipitation, evapotranspiration, and soil moisture. Land use datasets included global irrigation area data and gridded population data, smoothed using a Gaussian filter for noise reduction and normalization. Geologic datasets included a 'normalized clay indicator,' created by combining high-resolution clay content data with unconsolidated material thickness, and a 'confining layer' dataset derived from a digital elevation model (DEM) to identify areas with likely confined aquifer conditions. The model's hyperparameters were optimized using a 10-fold cross-validation approach to enhance model accuracy and prevent overfitting. Class weights were adjusted to address the imbalance in the training dataset, where the majority of samples represented minimal subsidence. A land use filter was applied to the model's output to remove predictions over areas with both low irrigation density and low population density, which helped reduce noise. Model performance was evaluated using the F1-score and a leave-one-area-out (LOAO) accuracy test. The LOAO test involved training the model multiple times, each time excluding a different region, to assess the model's ability to generalize across diverse areas. Finally, the subsidence map was used to estimate permanent groundwater storage loss by assuming average subsidence values for different subsidence classes.

The random forest model generated a high-resolution global map of land subsidence, classifying subsidence into three classes: <1 cm/year, 1–5 cm/year, and >5 cm/year. Significant subsidence was mapped in East Asia (China, Taiwan, Vietnam, Philippines), South Asia, Central Asia, and parts of Europe and North America, aligning with existing InSAR-based studies. Subsidence was also predicted in previously unstudied regions in Afghanistan, Turkmenistan, Uzbekistan, Azerbaijan, Syria, Morocco, Algeria, Tunisia, and parts of Australia and South America. The model estimated a global permanent groundwater storage loss of -17 km³/year due to aquifer consolidation (with lower and upper bounds of -11.5 and -22 km³/year respectively), with China, the United States, and Iran accounting for a major portion of this loss. Analysis revealed that about 73% of the mapped subsidence occurs over croplands and urban areas. Partial dependence plots showed that the model captured the expected relationships between land subsidence and key drivers, including high clay content, high irrigation density, low precipitation, and low soil moisture. The model generally demonstrated good agreement with existing subsidence data, although overestimation occurred in some areas due to uncertainties associated with input variables.

The findings highlight the widespread extent of groundwater-induced land subsidence globally, impacting critical areas like croplands and urban centers. The high-resolution map reveals both the spatial extent of known subsiding areas and previously unknown areas under groundwater stress. The model's capacity to identify subsidence even in the absence of prior studies emphasizes its usefulness in guiding future research and informing sustainable groundwater management practices. The quantified global groundwater storage loss provides a crucial metric for understanding the impacts of excessive groundwater pumping. The use of readily available remote sensing datasets makes the approach scalable and applicable to other regions. While the model overestimates subsidence in some regions, it provides a valuable first-order estimate of the global scale of the problem and helps identify areas needing further investigation.

This study presents a novel machine learning approach for mapping global land subsidence and quantifying associated groundwater storage loss at a high spatial resolution. The results reveal the widespread extent of this issue, particularly over cropland and urban areas. The model provides a valuable tool for identifying regions under groundwater stress and informing sustainable water management strategies. Future research should focus on refining input datasets to reduce model uncertainty, exploring the impacts of subsidence on various sectors, and developing strategies for mitigating the effects of land subsidence.

The model's accuracy is limited by the uncertainties inherent in the input datasets used. Imprecise delineation of depositional settings in certain regions and limited depth of clay content data can introduce uncertainty. Additionally, the model only estimates subsidence related to aquifer system compaction from groundwater pumping; it may not fully capture subsidence from other sources. The classification of subsidence into three classes is a simplification, and more granular analysis may be needed for localized assessments. The model's performance varies across different regions, potentially reflecting the diversity in hydrogeological conditions and data availability. The LOAO test indicated that the model might not accurately detect subsidence in regions with unique characteristics not well-represented in the training dataset.