A Neural Network-Based Block Region SAR Image Registration Algorithm

The LuTan-1 (LT-1) mission is China’s first civil L-band differential interferometric SAR (D-InSAR) satellite system, comprising the 01 Group A and B satellites, which were successfully launched in 2022. LT-1 has been extensively utilized for large-scale topographic mapping, geohazard risk identification, and natural resource management. Since June 2023, the LT-1 satellites have entered the strip1 mode to acquire repeat-pass observation data for long-term ground deformation monitoring. However, the initial orbit position vectors lacked sufficient precision, and without external orbit correction data, accurate initial offset estimation between image pairs could not be achieved. This limitation rendered conventional cross-correlation-based region registration algorithms ineffective, posing significant challenges for automated SAR image registration and geocoding. Moreover, long-baseline data introduced registration noise errors, further reducing observation accuracy. To address these challenges, we implemented a neural network-based feature point matching technique to estimate the initial offset between SAR image pairs. Additionally, a block-based registration approach was adopted to suppress registration noise. These methods were applied to the D-InSAR data  processing of the Ji Shishan, Gansu earthquake (Mw 6.2) on December 18, 2023. The results demonstrate that our approach successfully achieved accurate and automated region registration and geocoding while improving interferometric coherence and phase quality.