Localized deformation detection and prediction in wide-range urban environments

Recent increases in acquisition and availability of SAR images have significantly advanced deformation monitoring. An existing challenge includes wide-area time-series InSAR (TSInSAR) analyses that demand high-performance computing resources and extensive storage capacity. Moreover, a clear definition of localized deformations and established methods for their detection are missing, and TSInSAR results are time-intensive to deliver. To address these challenges, this paper proposes to first generate wide-area TSInSAR results using PS-based concatenation. This is followed by a hierarchical classification to preliminarily identify regions of uplift or subsidence, combined with the Getis-Ord (Formula presented.) statistic. High-resolution network (HRNet) models are then applied for deformation prediction. We tested this process on area covering parts of the Netherlands, Germany, and Belgium. Three Sentinel-1 SAR frames were used, i.e. Frame 415 and 420 on Path 37 and Frame 421 on Path 110, each comprising 60 SAR acquisitions from June 2017 to June 2019. Compared to leveling data, the concatenated TSInSAR results demonstrated higher precision, with a mean value difference (MVD) of 0.2 mm, root mean square error (RMSE) of 0.1 (Formula presented.), while the non-concatenated results were 1.8 mm and 1.7 (Formula presented.), respectively. The hierarchical classification combined with (Formula presented.) successfully identified localized deformation regions. Additionally, the HRNet model, trained on benchmark TSInSAR datasets, accurately predicted deformations using newly wrapped interferograms (IFGs). In conclusion, this study presents an efficient solution for generating wide-ranging TSInSAR results and offers robust methods for detecting and analyzing localized deformation.