Search for a command to run...
Assessing the relationship between groundwater dynamics and land subsidence remains challenging in urban basins where groundwater observations are incomplete, discontinuous, or irregularly sampled, and where deformation responses may vary through time rather than follow a single stationary pattern. This study develops and demonstrates an integrated geospatial framework for investigating groundwater-related land subsidence under such data-constrained conditions in the Bangkok Metropolitan Region, Thailand, one of Southeast Asia’s most important subsidence-prone urban alluvial basins. The framework combines Sentinel-1 Persistent Scatterer InSAR (PS-InSAR) for spatially dense surface-deformation monitoring, machine-learning-based reconstruction of incomplete groundwater records using eXtreme Gradient Boosting (XGBoost), and temporal coupling analysis using time-lagged cross-correlation and wavelet coherence. PS-InSAR processing was performed on 156 Sentinel-1 IW-mode SLC images acquired between January 2018 and February 2025 to derive deformation time series and vertical-rate estimates across Bangkok and the surrounding provinces. To evaluate the reliability of the deformation product, PSI-derived rates were compared with independent CORS-based GNSS observations and supplementary site-based estimates using spatial collocation within a 200 m radius and rate-based error metrics. The resulting validation indicates acceptable regional agreement, with RMSE values of 1.79 mm/year against CORS data and 1.13 mm/year against supplementary site-based estimates, supporting the use of the PSI dataset for regional interpretation and subsequent groundwater-coupling analysis. Groundwater data were obtained from the Department of Groundwater Resources monitoring network. However, substantial temporal gaps limited direct hydrogeological analysis: although 495 wells were initially available, only 69 met the 50% completeness threshold for further use. To address this limitation, missing groundwater observations were reconstructed using XGBoost within a time-aware validation framework designed to avoid leakage from future observations. The reconstruction model achieved an RMSE of approximately 0.044 m on held-out timestamps containing observed groundwater measurements, indicating that the imputed groundwater trajectories were sufficiently consistent for analytical comparison with InSAR-derived deformation. Detailed time-series coupling analysis was then conducted for three wells in Mueang Nakhon Pathom District, where reconstructed groundwater series were temporally matched to averaged PS-InSAR deformation records at shared acquisition dates. Conventional time-lagged cross-correlation indicated only weak to modest linear coupling, with wide lag confidence intervals, suggesting considerable uncertainty in any fixed lead–lag interpretation. In contrast, wavelet coherence revealed intermittent, scale-dependent, and spatially variable coupling, with stronger and more persistent coherence for some wells than for others. Because the dominant short-period coherence lies close to the effective sampling interval of the matched series, these short-period results were interpreted conservatively. Overall, the findings indicate that groundwater–subsidence coupling in the study area is non-stationary and spatially heterogeneous rather than governed by a single stable lag or stationary correlation. The main contribution of this study is therefore methodological rather than algorithmic: it establishes a defensible and transferable framework for integrating PS-InSAR, machine-learning-based groundwater reconstruction, and time–frequency analysis to investigate subsidence processes where deformation observations are dense but groundwater records are fragmented.