Large-area urban TomoSAR method with limited a priori knowledge and a complex deep learning model

Buildings are crucial to cities, and tomographic synthetic aperture radar (TomoSAR) is an important tool for monitoring the heights, linear deformations and thermal amplitudes of buildings. However, existing TomoSAR height inversion methods do not fully leverage a priori knowledge, compromising the accuracy of deformation estimation; deep learning-based methods involve the integration of multiple steps, complicating the process. Additionally, the computational inefficiency of existing algorithms significantly hinders the large-scale practical deployment of TomoSAR. To address the above issues, this study proposes a novel large-area urban TomoSAR method integrating limited a priori knowledge constraints with a complex-valued (CV) deep learning model. By refining scatterer types and Permanent Scatterer (PS) height sample sets under limited a priori height data constraints, the proposed CV-TomoPS-Net establishes an end-to-end framework for scatterer classification and PS height regression. Additionally, the proposed fast beamforming method, paired with an adaptive spatial search mechanism, enables rapid large-area inversion of deformation and thermal amplitude parameters. Experiments were conducted in Shenzhen city using COSMO-SkyMed SAR data from 2020 to 2023 and limited a priori data. Results show that the proposed method improves the accuracy of scatterer type classification by 16 %, reduces the height calculation error by 30 %, and improves the monitoring efficiency by 80 % compared with the traditional beamforming method. Validation via corner reflectors deformation monitoring confirmed reliability, with a 1.5 mm average error. These results highlight the practical applicability of the proposed method for large-scale urban monitoring and its potential to provide technical support for sustainable urban development.