Deformation Prediction of Ground Radar Based on GRUU-Net With Attention Mechanism Model

Real-time monitoring of ground deformation is crucial for disaster mitigation, enabling timely responses to landslides, earthquakes, and structural instabilities. Conventional methods, such as the interferometric synthetic aperture radar (InSAR), provide accurate deformation measurements but suffer several limitations, including high computational costs, atmospheric disturbances effect, and reliance on satellite re-visit cycles. To address these challenges, we propose a deep learning (DL)-based approach using gated recurrent units (GRUs) to estimate deformation time series directly from ground-based radar complex data. The proposed model is a U-Net architecture in which the encoder and decoder consist of GRU layers, and an attention mechanism is presented between the encoder and decoder layers. This approach is motivated by the need for faster deformation monitoring, particularly in early warning applications. Using DL, we aim to reduce the computational burden associated with traditional deformation monitoring frameworks, simplifying the deformation estimation process. We applied our proposed model to a real outdoor dataset collected using the recently developed multiple-input multiple-output (MIMO) radar, NanoRadar. The results show that our proposed DL model outperforms benchmark methods for estimating deformation time series.