SAR target recognition network based on frequency domain covariance matrix and Riemannian manifold

Synthetic Aperture Radar (SAR) image target recognition is a critical problem in remote sensing. Traditional deep learning-based methods often overlook the rich information in the frequency domain of SAR images. Existing frequency domain based approaches typically rely on prior knowledge to select different frequency components or guide network learning, failing to fully exploit the information contained in the frequency domain. We believe that knowledge from the macroscopic high-dimensional space needs to be harnessed to learn about the frequency domain. Therefore, in this paper, we propose an innovative SAR target recognition framework known as FDCM-Net (Frequency Domain Covariance Matrix and Riemannian Manifold Network). Our method integrates the frequency domain information of SAR images with the covariance matrix, revealing the complex relationships between different frequency components and delving into the latent features within the frequency domain. Additionally, we explore the geometric properties of the covariance matrix and design an innovative network architecture based on Riemannian manifold theory. This network effectively captures important information from the frequency domain of SAR images in high-dimensional space while preserving the geometric characteristics of the original data and performing classification. We conducted comprehensive evaluations of FDCM-Net using the MSTAR, OpenSARShip 2.0 and FUSAR-Ship 1.0 datasets. Our results demonstrate that FDCM-Net outperforms previous methods based on spatial and frequency domain baselines in terms of classification performance. • FDCM-Net integrates SAR frequency info and covariance matrices for latent features. • Innovative network captures frequency info and preserves geometric properties. • Combines texture, SPD manifold, and multi-manifold fusion with classification. • Experiments show FDCM-Net outperforms baselines on three datasets.