ISAR-GS: Gaussian Splatting for ISAR Space Target 3-D Reconstruction and Viewpoint Rendering

Reconstructing the 3-D structure of noncooperative space targets from inverse synthetic aperture radar (ISAR) image sequences is critically challenging in the field of space situational awareness. The complex electromagnetic scattering involved in ISAR imaging makes reliable 3-D reconstruction difficult and complicates ISAR image rendering from limited-view sequences. To address these challenges, this study proposes ISAR Gaussian splatting (ISAR-GS), a method for the 3-D reconstruction and viewpoint rendering of space targets from ISAR image sequences. A set of Gaussian primitives is used to represent the geometry and scattering properties of space targets. To account for ISAR-specific imaging physics, a differentiable ISAR Gaussian renderer is introduced to jointly model energy attenuation and scattering intensity accumulation. A gradient propagation scheme is further proposed to propagate supervision from rendered images to the Gaussian representation, enabling the joint optimization of geometric parameters and scattering attributes. On the three evaluated simulated datasets, ISAR-GS achieves better 3-D reconstruction and viewpoint-rendering results than ISAR-neural radiance fields (NeRF), a representative implicit neural rendering method, while requiring substantially less training time and enabling faster inference. It also exhibits greater robustness than ISAR-NeRF under the evaluated projection-vector perturbations and echo signal-to-noise ratio levels. Furthermore, the practical applicability of the proposed method is verified on a measured ISAR image sequence of the Tiangong-1 space station.