A machine learning-based method for multi-satellite SAR data integration

Abstract

Large- and small-scale subsidence coexist in the world's coastal cities due to extensive land reclamation and fast urbanization. Synthetic aperture radar (SAR) images are typically limited by either low resolution or small coverage, making them ineffective for fully monitoring displacement in coastal areas. In this research, a machine learning-based method is developed to investigate the reclaimed land subsidence based on multi-satellite SAR data integration. The proposed method requires at least a pair of SAR images from complementary tracks. First, the line-of-sight (LOS) displacements are recovered in connection to a series of extremely coherent points based on the differential interferometry synthetic aperture radar (DInSAR). These LOS displacements are then converted into their vertical component, geocoded to a common grid, and simultaneously integrated (i.e., pixel-by-pixel) based on Support Vector Regression (SVR). The proposed methodology does not necessitate the simultaneous processing of huge DInSAR interferogram sequences. The experiments include high-resolution COSMO-SkyMed (CSK) and TerraSAR-X (TSX) images, as well as a small monitoring cycle Sentinel-1 (S1) images of reclaimed territories near Hong Kong Kowloon City. The overall average annual displacement (AAD) ranges from -12.86 to 11.63 mm/year derived from 2008 to 2019. The evaluation metrics including RMSE, MAE, correlation coefficient, and R-squared are used to investigate the impact of SVR in the integration of SAR datasets. Based on these evaluation metrics, SVR is superior in terms of integration performance, accuracy, and generalization ability. Thus, the proposed method has potentially performed multi-satellite SAR data integration.