A novel physics-informed machine learning framework has been developed to enhance the prediction of high-resolution land subsidence in the Greater Houston Area, a region facing significant geotechnical challenges due to intensive groundwater extraction. This study, published in the International Journal of Geo-Engineering, employs a convolutional neural network (CNN) architecture that strategically integrates Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) observations with fundamental principles of soil mechanics, specifically Terzaghi’s consolidation theory. By embedding geomechanical constraints directly into the learning process, the researchers generated a high-fidelity susceptibility map that identifies localized deformation hotspots. The framework, trained on 155 Sentinel-1 SAR images spanning January 2020 to March 2023, achieved a predictive accuracy of 0.9862 as measured by the Area Under the Receiver Operating Characteristic (AUROC) curve. This hybrid approach offers critical insights for urban infrastructure planning and flood-risk mitigation, particularly in coastal megacities where groundwater-induced deformation threatens structural integrity and coastal resilience.
The study’s methodological foundation relies on the synthesis of remote sensing data with spatial conditioning factors to characterize the complex, nonlinear drivers of surface deformation. Researchers processed a comprehensive dataset of 10,652 persistent scatterer points, which were validated through integration with a network of 54 continuously operating GPS reference stations, resulting in a root mean square error (RMSE) of 2.1 mm/year and a correlation coefficient of 0.96. The hybrid CNN model incorporates twelve distinct mapped conditioning factors, covering topographic, geological, hydrogeological, and anthropogenic domains. These inputs, which include clay thickness, groundwater-level altitudes, and proximity to mapped growth faults, were resampled to a standardized 90-meter grid to harmonize the spatial support of disparate datasets. By formalizing these variables, the model moves beyond purely empirical association-based approaches, providing a robust mechanism-based framework for assessing regional subsidence susceptibility.
At the core of this investigation is the physics-informed architecture that constrains machine learning predictions with classical soil mechanics. The researchers implemented a preprocessing module that calculates pixelwise effective-stress change based on groundwater-level decline, as well as compaction potential and consolidation time factors derived from the consolidation theory. These derived physical parameters serve as additional input channels to the CNN, forcing the internal feature extraction process to respect the mechanical relationships between water pressure reduction and clay-matrix compression. Furthermore, the model uses a custom physics-constrained loss function, which includes penalty terms for any predictions that violate the governing principles of consolidation. This dual approach—incorporating geomechanical variables as predictors and enforcing them through the loss function—ensures that the final susceptibility mapping is physically consistent while retaining the high empirical performance typical of deep learning architectures.
The Greater Houston Area presents a complex hydrogeological setting characterized by thick sequences of poorly consolidated Quaternary and Tertiary sediments. The region’s aquifer system—comprising the shallow Chicot, intermediate Evangeline, and deep Jasper units—contains highly compressible clay layers that react significantly to changes in effective stress. As groundwater is extracted for municipal and industrial purposes, the loss of fluid pressure increases the effective stress within these clay-rich confining units, leading to both elastic and inelastic deformation. The study specifically addressed the spatial variability of these processes by including cumulative clay-thickness data from thousands of borehole records, which allowed the model to distinguish between zones of high and low subsidence susceptibility. This geological complexity necessitates sophisticated spatial prediction techniques that can account for the long-term, time-dependent response of the subsurface materials.
Comparative performance analysis underscores the substantial advantage of the physics-informed model over conventional machine learning methods. The researchers evaluated the framework against baseline models including Random Forest, Support Vector Machine, and Logistic Regression, using identical training and validation strategies. The proposed CNN architecture outperformed the strongest conventional baseline, the Random Forest model, by 7.42 percentage points in AUROC, with concurrent improvements in precision and recall metrics. This performance gain demonstrates that explicitly encoding consolidation mechanics into the model architecture significantly enhances generalization capability, especially when transfering to heterogeneous spatial conditions. The use of five-fold spatial block cross-validation, which includes geographically separated validation blocks and 5-km exclusion buffers, further ensures that the model’s reported accuracy reflects its ability to predict subsidence in regions distinct from its training data, rather than being an artifact of localized spatial autocorrelation.
Spatial analysis reveals that approximately 18.66% of the 9,900-km² study area exhibits high to very high land subsidence susceptibility, covering nearly 1,848 square kilometers. The maximum observed deformation rates, reaching 24.3 mm/year in western Harris County near the Katy Prairie region, highlight the persistence of high-rate subsidence hotspots. These results align with broader regional trends and confirm that despite significant regulatory progress by groups such as the Harris-Galveston Subsidence District and the Fort Bend Subsidence District, localized deformation continues in areas experiencing rapid suburban development or those underlain by particularly thick compressible clay sequences. The susceptibility map serves as a diagnostic tool for identifying these areas, where residual compaction and the delayed response of the aquitard system to historical groundwater extraction continue to manifest as surface deformation.
Integrated gradients attribution analysis provided key insights into the relative importance of the conditioning factors driving the model’s predictions. The high ranking of these physics-informed inputs confirms the success of the methodology in capturing the underlying mechanical drivers of subsidence. By ranking these variables through attribution analysis, the researchers were able to quantify the contribution of specific physical constraints, thereby increasing the interpretability of the deep-learning outcomes and confirming the framework’s scientific validity.
While the study demonstrates significant advancements in regional subsidence screening, the researchers noted essential limitations regarding the nature of the model’s outputs. The susceptibility map classifies the region into five probability-based classes, but it should not be interpreted as a deterministic forecast of absolute subsidence velocity at a specific site in future years. Because the framework does not explicitly isolate elastic rebound, primary inelastic compaction, and secondary compression—or “creep”—into separate model outputs, it does not prescribe a quantitative partition of these components. Achieving such a decomposition would require long-term coupled data assimilation of borehole-extensometer records and aquifer-specific hydraulic-head histories. Consequently, the current framework is optimally designed for regional susceptibility screening and risk-prioritization, identifying where monitoring and policy enhancement are most urgently needed.
The study also examined the effectiveness of current groundwater-management frameworks by analyzing the spatial susceptibility patterns across regulated jurisdictional areas. The findings indicate that while central Houston has stabilized due to long-term groundwater reduction policies, significant moderate-to-high susceptibility zones persist in suburban areas where infrastructure growth coincides with high-compressibility geological units. This persistent vulnerability illustrates the challenge of managing subsidence in a rapidly expanding megacity where geological heterogeneity creates diverse responses to regulatory pressure. By delineating susceptibility classes and mapping them against regulatory area boundaries, the research provides decision-makers with a granular view of where existing groundwater-reduction strategies have proven effective and where residual subsidence potential warrants additional attention or localized management interventions.
The methodology developed here offers a transferable approach for other coastal metropolitan regions facing similar geohazards, provided that applications are locally parameterized with representative data. To apply this physics-informed framework elsewhere, a region would typically require a multi-temporal SAR stack, locally representative indicators of groundwater stress, a hydrostratigraphic representation of compressible subsurface units, and independent geodetic data for validation. As high-resolution remote sensing technologies continue to evolve, the integration of these data-driven and physics-constrained workflows will become increasingly central to sustainable groundwater management and urban resilience. This research confirms that blending mechanical theory with advanced pattern recognition provides a robust path forward for geomechanical prediction in the face of climate variability and ongoing urban expansion.
Subject of Research: Physics-informed machine learning application in high-resolution regional land subsidence susceptibility mapping.
Article Title: A physics-informed machine learning framework for high-resolution land subsidence prediction in the Greater Houston Area
Article References: Kim, Y. J., Chowdhury, M. S., An, J., & Choi, S. (2026). A physics-informed machine learning framework for high-resolution land subsidence prediction in the Greater Houston Area. International Journal of Geo-Engineering, 17(1), Article 39. https://doi.org/10.1186/s40703-026-00298-0
Image Credits: AI Generated
DOI: 10.1186/s40703-026-00298-0
Keywords: land subsidence, physics-informed machine learning, Greater Houston Area, PS-InSAR, groundwater extraction, consolidation theory, convolutional neural networks, geotechnical engineering, spatial susceptibility, Sentinel-1, aquifer compaction, hydrogeology, geomechanical modeling, urban planning, soil mechanics, coastal subsidence, remote sensing
News Source: Katie Riggs. (November 16, 2026). Physics-Informed AI Framework Improves Land Subsidence Prediction in Greater Houston. Scienmag.



