SAR-NeuS: Neural Surface Inverse Rendering for 3D Reconstruction and Novel view Synthesis from SAR Imagery
Year: 2026
Type: conference-paper
Abstract: We present SAR-NeuS, a neural surface inverse rendering framework for reconstructing high-fidelity 3D geometry and synthesizing novel views from SAR imagery. SAR-NeuS leverages implicit neural surface representations together with a physics-informed differentiable SAR image-formation model to jointly estimate scene geometry and backscatter characteristics. The proposed method learns a continuous signed distance function and a backscatter intensity field optimized to reproduce observed SAR returns under varying acquisition geometries, enabling accurate recovery of complex structures even in single-pass, sparse-view scenarios. We demonstrate the capability of SAR-NeuS on realistic electromagnetic-simulated multi-view SAR datasets, achieving accurate geometry reconstruction and high-fidelity novel view synthesis across diverse object geometries, scattering conditions, and sampling densities. Quantitative evaluation reports Chamfer Distance values in the range of 0.005-0.024 m2and average MS-SSIM of 0.85 in single-bounce, with only moderate degradation under multibounce conditions. These results highlight the efficacy of SAR-NeuS for next-generation SAR interpretation and 3D reconstruction.
Institutions: Arquimea (Spain), Universität der Bundeswehr München
Language: English
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202340694