Machine learning-based integration of multi-satellite InSAR to retrieve long-term time series displacement

Abstract

In the world's fast-growing mega-cities, ground deformation is one of the crucial issues threatening many cities in both societal and economic aspects. Traditional field procedures (e.g., leveling and global navigation satellite system (GNSS)) have been utilized for deformation monitoring. Notwithstanding these methods' high reliability, their main drawbacks are low spatial resolution in large-scale projects, relatively high cost, and lack of manpower. In contrast, remote sensing techniques, particularly, differential InSAR (DInSAR) provide high-resolution deformation maps in large-spatial coverage at high levels of accuracy. Recently, synthetic aperture radar (SAR) data are available from diverse bands including C-band (e.g., SIR-C, European remote-sensing satellite (ERS), ENVISAT, RADARSAT-1/2, and Sentinel-1), X-band (e.g., TerraSAR-X and COSMO-SkyMed), L-band (e.g., JERS, ALOS-1/2, TerraSAR-L, and DESDynl), and P-band (e.g., BIOMASS). A key challenge of integrating multi-band SAR datasets is that they have diverse maximum detection gradients, degrees of decorrelation, noise rejection capability, etc. The integration of multiple operational bands, polarimetric channels, and orbit orientations is anticipated to enrich the gained information thus enabling depth interpretation of the surface deformation. In this research, the problem of integrating multi-satellite SAR data is addressed based on two aspects (1) adaptation of the traditional small baseline subset (SBAS) time series and (2) utilization of machine learning machine learning (ML) to perform integration. The proposed methodology exploits complementary information from different SAR data to generate integrated long-term ground displacement time series. Part I of this thesis focuses on the integration of multi-satellite SAR data by adaptation of traditional SBAS. The proposed method is employed to generate the vertical displacement maps of Almokattam City in Egypt from 2000 to 2020. The experiments have shown promising results based on ERS, ENVISAT, and Sentinel-1A interferograms. Significantertical deformation has been recorded along the west of the city with a mean value of - 2.32 mm/year and a standard deviation of 0.21 mm/year. Moreover, the research findings are in line with those from previous studies in the area. Accordingly, the proposed integration approach has great potential in retrieving long-term vertical displacement based on multi-satellite SAR data. In part II of the thesis, machine learning is used to integrate multi-satellite SAR data. At least a pair of SAR images from complementary tracks is the input of the proposed method. The line-of-sight (LOS) displacements are computed based on DInSAR at a series of high-coherence points. The vertical components of displacement are then computed from the recovered LOS displacement. After that, the vertical displacement maps are geocoded to the ground coordinate system. Finally, the support vector regression (SVR) is used to integrate the displacement on a pixel-by-pixel basis. The proposed method does not employ simultaneous processing of huge DInSAR interferogram sequences, which is a key advantage compared to other methods. The SVR integration is tested using COSMO-SkyMed (CSK),TerraSAR-X (TSX) images, and a small monitoring cycle Sentinel-1 (S1) images to monitor the deformation of the reclaimed territories near Hong Kong Kowloon City. The results show that the average annual displacement (AAD) ranges from -12.86 to 11.63 mm/year from 2008 to 2020 with a Standard Deviation (STD) of 0.69 mm/year. Moreover, the root mean square error (RMSE), MAE, correlation coefficient, and R-squared are computed. Accordingly, a potential performance of the proposed method in multi-satellite SAR data integration has been recorded.