S-Band SAR Target Classification via 2D and 3D Deep Learning Methods

Tyler Rust, Michael William Pergeorelis, Chandra Kambhamettu, Colin Kelly

Abstract: S-band Synthetic Aperture Radar (SAR) offers advantages in foliage, ground penetration, and weather tolerance. However, it is comparatively underutilized for object classification due to the preference for higher band frequencies. This paper presents the first application of deep learning to S-band Synthetic Aperture Radar (SAR) data for local object classification. Our method of extrapolating the 2D SAR image to a 3D Radar Cross Section (RCS) response differs from previous work that uses targets physical 3D point clouds. We also present a novel preliminary 2D-3D fusion method for S-band SAR to demonstrate the effectiveness of ensemble methods on this type of data. We combine a lightweight custom convolutional neural network (CNN) with a PointNet-based network, enhancing feature extraction from image and point cloud domains. Our method is more precise and robust to clutter compared to single-modality techniques.