Quantitative and Qualitative Comparison of SAR Images from Incomplete Measurements Using Compressed Sensing and Nonuniform FFT Hamed Kajbaf, Joseph T. Case, Yahong Rosa Zheng, Sergey Kharkovsky, and Reza Zoughi Electrical and Computer Engineering Department Missouri University of Science and Technology Rolla, Missouri 65409 Email: {hamed.kajbaf, j.t.case, zhengyr, sergiy, zoughir}@mst.edu Abstract— In this paper the performance of two wideband synthetic aperture radar (SAR) imaging methods from incom- plete data sets are compared quantitatively and qualitatively. The first approach uses nonuniform fast Fourier transform (NUFFT) SAR to form images from nonuniform spatial and frequency data points. The second approach benefits from the emerging compressed sensing (CS) methodology to recover raw data from undersampled measurements. The results of our experimental tests show that CS has a better performance in terms of error and image contrast while NUFFT SAR has lower computational complexity. I. I NTRODUCTION Microwave synthetic aperture radar (SAR) imaging is a high resolution nondestructive testing and evaluation (NDT&E) technique which can be exploited to detect discontinuities in critical structures by raster scanning using a single antenna probe (e.g. an open-ended waveguide) [1]–[3]. Wideband SAR is capable of determining depth of discontinuities and providing 3D images of specimens under test (SUT) such as spacecraft tiles, airplane coating, bonding of adhesive or composite materials. However, the drawback of microwave SAR imaging for NDT&E applications is the time needed for scanning the region of interest, which for a relatively large SUT might be hours of data acquisition. Reducing the number of spatial samples significantly helps in decreasing the acquisition time. The emerging compressed sensing (CS) theory have in- troduced one method of reducing the number of samples by sampling below the conventional Nyquist rate [4], [5]. Nonuniform fast Fourier transform (NUFFT) SAR is another method for forming SAR images from incomplete data with low computational requirements. In this paper the performance of sample reduction in SAR imaging using CS and NUFFT SAR are compared in terms of metrics indicating the quality of the SAR images. II. NUFFT SAR Suppose that an antenna probe located at ( , , 0 ) illu- minates a target and a general point on the target, (, , ), reflects back the pulse. The same probe receives the backscat- tered coherent signal, ( , ,), which is the superposition of reflection from all points in the illuminated area ( , ,)= ∫∫∫ (, , )  ddd(1) where is the range between the probe and the target point = ( ) 2 +( ) 2 +( 0 ) 2 , = is the wavenumber, is the propagation speed, and (, , ) is the reflectivity function of the target, which is the ratio of the reflected field to the incident field. Decomposing the spherical wave propagation into a superposition of plane wave components, we can rewrite (1) in 3D Fourier transform form [6] ( , ,) = ∫∫ [∫∫∫ (, , ) ( + +) ddd ] × ( + +0) d d (2) where and are Fourier transform variables correspond- ing to and , respectively, and the one corresponding to is = 4 2 2 2 . The triple integral in (2) is the 3D Fourier transform of (, , ). Solving this equation and dropping the distinction between primed and unprimed coordinates, the 3D image is formed by [6] (, , )= 1 3 { 2 {(, , )} 4 2 2 2 0 } (3) where ℱ{.} and 1 {.} are the Fourier and inverse Fourier transform operators, respectively. Since in (3) the frequencies are uniformly sampled in the frequency band and the probe scans at uniform step size along X-Y dimensions, ’s are nonuniformly distributed and Stolt interpolation is normally used to interpolate the uniform points in [7]. It has been shown that NUFFT can be exploited to improve the accuracy of SAR imaging [8] since it is a good approximation of the nonuniform discrete Fourier transform (NDFT) [9], [10]. Using NUFFT, (3) becomes (, , )= 1 2 { 1  { 2 {(, , )} 4 2 2 2 0 }} (4) 978-1-4244-8902-2/11/$26.00 ©2011 IEEE 592