A Transfer Learning Approach for Deformation Pattern Recognition in InSAR Time Series

The multiepoch interferometry synthetic aperture radar (InSAR) technique is a widely applied geodetic tool for measuring surface displacement. Yet, traditional interpretations of InSAR results have primarily centered on linear displacement velocities, often neglecting the rich insights that InSAR displacement time-series data can offer. This study innovatively addresses this gap by proposing a deep learning (DL)-based method for interpreting InSAR deformation time series in urban environments. We first introduce six canonical deformation patterns: stable, linear, stepwise, piecewise linear, power-law, and undefined. A novel postprocessing approach integrates DL models—bidirectional long short-term memory (BiLSTM), temporal convolutional network (TCN), and Transformer—with transfer learning techniques. The strategy involves pretraining models on simulated data and fine-tuning with real-world data, significantly reducing dependence on extensive labeled datasets. This study demonstrates the effectiveness of these DL models in processing InSAR deformation sequences, illustrating how transfer learning can tackle the challenge of limited labeled InSAR datasets. The experimental results reveal that the TCN model achieves the best performance, with an accuracy of 91%. Tested on InSAR data from Kunming City, the proposed approach effectively classified deformation sequences into predefined categories. The findings demonstrate that time-series analysis reveals more detailed deformation insights—particularly in regions with low deformation rates—than traditional velocity-based methods. Furthermore, incorporating transfer learning significantly reduces the dependency on extensive real-world datasets, enhancing overall model performance and facilitating future advancements in the field.