LINEAR ARRAY SAR IMAGING VIA COMPRESSED SENSING

• Published in: Electromagnetic waves (Cambridge, Mass.) Vol. 117; pp. 299 - 319
• Main Authors: Wei, S.-J, Zhang, X.-L, Shi, J
• Format: Journal Article
• Language: English
• Published: Cambridge Electromagnetics Academy 01.01.2011
• Subjects: Data compression; Equipment and supplies; Imaging systems; Methods; Synthetic aperture radar; China
• ISSN: 1559-8985; 1070-4698; 1559-8985
• DOI: 10.2528/PIER11033105

In recent years, various attempts have been undertaken to obtain three-dimensional (3-D) reflectivity of observed scene from synthetic aperture radar (SAR) technique. Linear array SAR (LASAR) has been demonstrated as a promising technique to achieve 3-D imaging of earth surface. The common methods used for LASAR imaging are usually based on matched filter (MF) which obeys the traditional Nyquist sampling theory. However, due to limitation in the length of linear array and the "Rayleigh" resolution, the standard MF-based methods suffer from low resolution and high sidelobes. Hence, high resolution imaging algorithms are desired. In LASAR images, dominating scatterers are always sparse compared with the total 3-D illuminated space cells. Combined with this prior knowledge of sparsity property, this paper presents a novel algorithm for LASAR imaging via compressed sensing (CS). The theory of CS indicates that sparse signal can be exactly reconstructed in high Signal-Noise-Ratio (SNR) level by solving a convex optimization problem with a very small number of samples. To overcome strong noise and clutter interference in LASAR raw echo, the new method firstly achieves range focussing by a pulse compression technique, which can greatly improve SNR level of signal in both azimuth and cross-track directions. Then, the resolution enhancement images of sparse targets are reconstructed by L1 norm regularization. High resolution properties and point localization accuracies are tested and verified by simulation and real experimental data. The results show that the CS method outperforms the conventional MF-based methods, even if very small random selected samples are used.