Deep learning facilitated local deformation monitoring with large-scale SAR interferometry

SAR interferometry has stepped in the big-data era, particularly with the acquisition capability and open-data policy of ESA’s Sentinel-1 SAR mission. Large amount of Sentinel-1 SAR images has been acquired and archived, allowing for generating thousands of interferograms, covering millions of square kilometers. In such a large-scale interferometry scenario, many applications still focus on monitoring kilometer-scale local deformation, sparsely distributed in a large area. It is thus not effective to apply the time-series InSAR analysis to the whole image stack, but to focus on areas with deformation. Aiming at this target, we present our recent work built upon deep neural networks to firstly detect localized deformation and then carry on the time-series analysis on small interferogram patches with deformation signals.

Here, we first introduce our burst-based Sentinel-1 processor, which has been fully paralleled for large-scale InSAR processing. From these interferograms, we adapt and train several deep neural networks for masking decorrelation areas, detecting local deformation, and unwrapping high-gradient phases. We apply our networks for mining subsidence and landslides monitoring. Comparing with traditional time-series InSAR analysis, the presented strategy not only reduces the computation time, but also avoids the influence of large-scale tropospheric delays and the propagation of possible unwrapping errors.

The presented methods introduce artificial intelligence to the time-series InSAR processing chain and make the mission of regularly monitoring localized deformation sparsely distributed in large scale feasible and more efficient. As future work, we can further improve the temporal resolution of InSAR based local deformation monitoring by training networks combining interferograms from C-band and L-band SAR images, which will be available soon from future SAR missions such as NiSAR and LuTan-1.