From repeat-pass to single-pass: City-scale 3D point-cloud reconstruction via InSAR-TomoSAR fusion with the Hongtu-1 multistatic SAR constellation

Spaceborne synthetic aperture radar (SAR) provides an all-weather capability for reconstructing 3D point clouds of radar-visible urban scatterers; this is valuable for urban mapping and post-disaster assessment. However, conventional spaceborne tomographic SAR (TomoSAR) for urban point-cloud reconstruction relies on repeat-pass acquisitions over months to years and typically yields sparse point clouds, thereby limiting rapid response and interpretation of radar-visible structural details. Hongtu-1 (HT-1, also known as PIESAT-01), the first commercial four-satellite formation multibaseline SAR system, acquires four SAR images in a single pass and is designed to generate digital surface model (DSM) products at a 1:50,000 scale (25-m posting). To further exploit the capability of HT-1, we propose the sparse-view 3D multibaseline InSAR-TomoSAR fusion (SV-MBTF) framework, a single-pass SAR point-cloud reconstruction chain compatible with sparse-view stripmap data. With only four images and a maximum bandwidth of 150 MHz, SV-MBTF generates dense urban point clouds corresponding to structures illuminated along the radar line of sight at a challenging single-look grid spacing of approximately 0.7m×1.7m (slant range × azimuth), thereby extending HT-1 beyond its nominal DSM products toward high-resolution SAR point-cloud mapping. SV-MBTF is centered on a noise-resilient adaptive 3D multibaseline InSAR (MB-InSAR) algorithm, which integrates (i) two-stage intensity-guided statistically homogeneous point extraction, (ii) residual-phase estimation based on multibaseline phase-gradient linking, and (iii) model-based residual-phase compensation, enabling robust elevation retrieval under heterogeneous phase quality. We further fuse 3D MB-InSAR and TomoSAR reconstructions to improve point-cloud density and 3D geometric accuracy, and propose the energy-entropy dispersion index, together with point-cloud filtering methods, for point selection and quality control. Using HT-1 data collected during a 3.6-s single-pass observation, the proposed method reconstructs more than 200,000,000 points covering more than 600km2, thereby substantially improving the mapping efficiency compared with that of conventional repeat-pass acquisitions. Experiments over complex high-rise areas in Las Vegas and representative urban scenes in Beijing validate the reconstruction performance through comparisons with high-resolution airborne LiDAR point clouds and optical 3D models, achieving a relative 3D localization error (Chamfer distance) below 2.5 m. Real-data comparisons with recent state-of-the-art deep-learning-based and representative conventional noise-suppression methods show that the proposed method reduces the 3D MB-InSAR height-reconstruction error by 20.9% relative to the best competing result. Additionally, we demonstrate extended applications enabled by the reconstructed HT-1 point clouds, including 3D geocoding and coastal construction change monitoring in Fujian, China. The demo code and data-access information are available at: https://github.com/yu-cas/SV-MBTF