Recurrent neural networks for atmospheric noise removal from InSAR time series with missing values

Atmospheric noise is one of the primary challenges for improving the accuracy of deformation estimation by InSAR technologies. Temporal filtering methods, like Gaussian filtering, are commonly used to remove the atmospheric noise from the InSAR time series. Such low-pass filters can effectively suppress stochastic noise. Yet, its performance heavily depends on the parameter settings and can easily be affected by the seasonal variations and missing values presented in the InSAR time series. Recurrent neural networks (RNNs) have been successfully adapted in many time series or sequential data applications. Still, there is little work on exploiting the ability of RNNs for modeling InSAR time series. This paper proposes a bidirectional RNN with gated recurrent units (GRU) for removing the atmospheric noise from the InSAR time series. A physical-based method of synthesizing InSAR time series is developed to tackle the lack of training data problem. The proposed GRU model integrates a GRU-D layer for handling the missing values, and all the model components are jointly trained to produce the denoised time series. Besides, we introduce the seasonal factor (SF) signal as an auxiliary input to help the model better capture the seasonality of the deformation and improve the denoising results. Experiments on synthetic datasets and HKIA real-world datasets demonstrate that our proposed GRU model achieves better denoising performance than Gaussian filtering and other RNN baseline models.