3D EO/SAR Reconstructions from Single and Limited Viewing Perspectives

Toyon Research proposes to develop a novel 3D SAR image reconstruction algorithm by leveraging concurrent or historical multi-view EO imagery within a physics-based machine learning-accelerated reconstruction algorithm. Specifically, we propose utilizing single and multi-view EO-based reconstruction to improve 3D SAR imaging via (1) 2D/3D point-scattering model initialization, (2) EO-based fiducial tracking for 3D k-space coordinate refinement, and (3) geometric multi-bounce and ground-plane estimation models. The reconstruction algorithm is based on an objective function combining data consistency, total-variation, and L1-sparsity terms, together with a multi-scale low-rank (MSLR) factorization of the target-scene, enabling data- and time-efficient imaging of potentially large scenes. To further improve computational efficiency, we propose to use real-time stochastic gradient descent (SGD) with momentum, enabling the potential of on-board 3D target model refinement, scene understanding, automatic target recognition (ATR) using fast 2D/3D neural networks we have previously developed for airborne platforms.