Improving time-series InSAR deformation estimation for city clusters by deep learning-based atmospheric delay correction

Atmospheric delay (AD) is the main source of error in time-series interferometric synthetic aperture radar (InSAR) deformation estimation over large areas. In this study, we propose a bidirectional gated recurrent unit (BiGRU) model to correct random and seasonal ADs adaptively. The BiGRU model decomposes InSAR time-series measurements into non-seasonal and seasonal components by adopting a branched network structure and extracting component-wise features separately. To remove seasonal ADs and meanwhile preserve true seasonal deformation, a dense and fully connected layer with weighted feature learning was designed. Five typical time-series deformation patterns were simulated for model training, and its robustness was evaluated using synthetic data. We applied the trained model to two city clusters in China (Guangdong and Jiangxi-Hunan) using 178 Sentinle-1 images. The results showed that BiGRU with moderate Generic Atmospheric Correction Online Service (GACOS) and spatiotemporal filtering (pGA_Fi_BiGRU) reduced the standard deviation of InSAR time-series measurements by 64.3% in the Guangdong region and by 53.5% in the Jiangxi-Hunan region compared with the raw data. Compared with the traditional combined GACOS and spatiotemporal filtering processing methods, the pGA_Fi_BiGRU improved the AD reduction performance by 4.7% and by 8.5% in Guangdong and Jiangxi-Hunan, respectively. The InSAR time-series deformation after pGA_Fi_BiGRU processing removed residual ADs and preserved true deformation, which agreed well with the geodetic leveling and Global Navigation Satellite System data. The first overall subsidence velocity of the Irrawaddy Delta city cluster in Myanmar was then mapped, followed by time-series deformation estimation using pGA_Fi_BiGRU. Representative time-series deformation due to groundwater extraction, coastal erosion, and accretion were properly derived, suggesting that the proposed model can be generalized to other city clusters with different atmospheric noise and geophysical dynamics.