Groundwater Level Retrieval Using Temporal Integration of Sentinel-1 InSAR Time-Series and Recurrent Neural Networks

Changes in Groundwater Level (GWL) in confined aquifers can cause ground surface deformation, which can have significant implications. These movements can be captured in Sentinel-1 Interferometric Synthetic Aperture Radar (InSAR) time-series data. This means that InSAR deformation time-series data reflects GWL changes and can be used to estimate GWL values. Hence, this paper proposes a new method to estimate GWL from InSAR deformation time-series.  The method uses a temporal window of InSAR displacement values centered on a specific time, t, which includes observations from a defined period before and after t, and retrieves GWL for an earlier time, t–Δt, where Δt is the delay between GWL changes and surface deformation. By leveraging temporal patterns embedded in the InSAR data, a more accurate and timely estimation of GWL is retrieved. To model the temporal relationships inherent in the data, Recurrent Neural Networks (RNNs) were chosen. These networks are well-suited for tasks involving sequential and time-dependent data. Specifically, Long Short-Term Memory (LSTM) networks were applied due to their ability to capture temporal dependencies and patterns in complex datasets. The proposed method was tested in Shabestar aquifer, in semi-arid Iran, a region where agriculture relies heavily on groundwater resources. Data from monitoring wells located in a confined aquifer was used to validate the approach. Various validation techniques, including Leave-One-Station-Out (LOSO), Leave-One-Time-Period-Out (LOTPO), and 5-fold cross-validation, were employed to ensure the robustness and generalizability of the proposed methodology. The results of the study revealed that integrating InSAR time-series data with LSTM networks provided accurate GWL estimates. This success is attributed to the method's ability to exploit the temporal information contained within the InSAR data. Moreover, the LSTM-based approach outperformed traditional machine learning models like Random Forests. Overall, the proposed methodology offers a promising pathway for providing more accurate estimations of GWL by harnessing the power of satellite data and state-of-the-art deep learning techniques.