InSAR-Based Prediction of Time-Series Displacements Using a New Physics-Informed Neural Network with Prior Parameter Inversion

Deep learning algorithms have become useful tools for predicting time-series displacements from historical displacements measured using interferometric synthetic aperture radar (InSAR) techniques. However, nearly all existing InSAR-related studies are based on data-driven deep learning algorithms, causing poor robustness, especially for long-term prediction with small-scale training samples. In this study, we propose a new algorithm, named physics-informed neural network with prior parameter inversion (PINNPI), for InSAR-based prediction of time-series displacements. PINNPI is a hybrid data-driven and knowledge-guided deep learning network, where two coupled deep neural networks are first constructed for network training and parameter inversion of prior knowledge. The outputs of these two deep neural networks are coupled by an automatic differentiation module. By minimizing a hybrid physics-informed and data-driven loss function, the proposed network simultaneously models time-series displacement and estimates prior parameters. Subsequently, time-series displacements are predicted based on the trained networks and inverted parameters. The incorporation of physical knowledge into PINNPI enhances the capability of long-term displacement prediction with respect to data-driven learning algorithms. In addition, PINNPI effectively improves the poor robustness of classical PINNs, when prior parameters are unknown. Simulations and two real-world tests suggest that the accuracy of displacement prediction by PINNPI is, on average, 85% and 88% higher than that of classical data-driven deep learning and PINN algorithms, respectively. This work offers a new insight for predicting InSAR-based displacements associated with anthropogenic and geophysical activities.