An InSAR–Machine Learning Framework for Ground Deformation Modeling and Scenario-Based Projections in the Ararat Valley (Armenia)

Ground deformation resulting from groundwater extraction is a critical yet frequently under-monitored process in semi-arid agricultural regions. Although Interferometric Synthetic Aperture Radar (InSAR) is a key tool for detecting surface deformation, its integration with predictive modelling frameworks remains limited, especially in basins with limited in situ hydrogeological data. This research introduces a data-driven framework that integrates InSAR time series with machine learning techniques to examine the temporal dynamics of ground deformation and its sensitivity to hydroclimatic variability. The proposed framework is applied to the Ararat Valley (Armenia), a transboundary agricultural basin characterized by semi-arid climatic conditions, strong seasonal variability in precipitation and evapotranspiration, and intensive groundwater-dependent irrigation. These conditions make the valley particularly sensitive to groundwater stress and associated land deformation, while the limited availability of long-term groundwater observations poses challenges for conventional hydrogeological analyses. Surface deformation is extracted from Sentinel-1 imagery using a time-series InSAR approach and combined with satellite-derived hydroclimatic and environmental variables, such as precipitation, temperature, evapotranspiration, vegetation dynamics, and soil moisture. Long Short-Term Memory (LSTM) neural networks are utilized to model non-linear temporal relationships between deformation and environmental drivers, enabling the capture of delayed and cumulative responses to hydroclimatic forcing. For exploratory future assessments, additional machine learning and empirical models estimate potential trajectories of vegetation and soil moisture based on regional climate projections, which are then incorporated into the deformation modelling framework. The methodology is designed to be scalable and transferable, facilitating deformation analysis in regions with sparse or unevenly distributed groundwater observations. Instead of prioritizing site-specific calibration, the framework emphasizes process representation and scenario exploration. A dedicated InSAR validation strategy, involving the comparison of deformation signals from ascending and descending Sentinel-1 acquisition geometries (ASC versus DESC), is used to assess the internal consistency and robustness of the InSAR-derived time series. This work advances methodological development and highlights the potential of integrating satellite-based deformation monitoring with machine learning to enhance groundwater-related risk assessment under evolving hydroclimatic conditions in poor monitored regions.