Nonlinearity and the effect of detection on single-channel synthetic aperture radar imagery

Abstract

When signals exhibit non-Gaussian statistics, nonlinear signal processing techniques offer advantages over their linear counterparts. Nonlinearity in high-resolution synthetic aperture radar (SAR) imagery is an intrinsic phenomenon often overlooked in the radar literature. In this paper, we study the nonlinear dynamics, and the effect of detection, in SAR imagery. To this end, two complementary methods for exposing the nonlinear statistics are presented. The first method utilizes histogram fitting with relevant statistical models. The second method is based on hypothesis testing. Our results are demonstrated on real-world Radarsat-2 target chips. It is found that in the presence of extended targets (e.g., ships), the nonlinear effect in the SAR chip is predominant. Nonlinearity is observed to be negligible in the absence of extended targets. As the SAR chip is detected, the nonlinear dynamics are either diminished/altered (i.e., for power-detection) or obliterated (i.e., for magnitudedetection). To take full advantage of nonlinear statistics, it is recommended to utilize the complex-valued SAR image rather than the detected one. Furthermore, the Student's T locationscale distribution is seen to offer an excellent model for the SAR chip.

On circularity/noncircularity in single-channel synthetic aperture radar imagery

“…Furthermore, there seems to be a proportional relationship between noncircularity and nonlinearity (see H 1,NC and H 1,NL in Table I). Nonlinearity in SAR imagery was investigated in [15] through utilizing a hypothesis test for nonlinearity (i.e., H 1,NL ). The basic idea involves the use of a GLRT test to calculate the shape parameter (ĉ) of the CGGD distribution, fitted to the SAR data, at a specific PFA.…”

“…The test result is given in terms ofĉ. Note that 0 <ĉ < 1 implies nonlinearity,ĉ = 1 implies Gaussianity, andĉ > 1 implies sub-Gaussianity [7,15]. The relationship between noncircularity and nonlinearity is not surprising, as the latter effect is the reason for violation of conventional radar resolution theory [3][4][5].…”

SAR2HEIGHT: Height Estimation from a Single SAR Image in Mountain Areas via Sparse Height and Proxyless Depth-Aware Penalty Neural Architecture Search for Unet

“…Details about the algorithm are shown in Algorithm 1. At first, the mask function M S×S is established as described in Equation (7). Then, H S is generated by the Hadamard product (Hadamard product: For X =…”

“…Costante [6] attempts to use single SAR imagery and a phase map for height estimation. Since the phase map from a single SAR image is uniformly distributed [7] and meaningless, a more efficient postprocessing approach is required to be performed on the phase map for height estimation. Amirkolaee [8] proposes an improved decoder network for Unet.…”

Complex-Valued Vs. Real-Valued Neural Networks for Classification Perspectives: An Example on Non-Circular Data

“…Notably, CVNNs are more adapted to extract phase information [2], which could be helpful, e.g., for retrieving Doppler frequency in radar signals, classifying polarimetric Synthetic Aperture Radar (SAR) data [3,4], etc. Furthermore, the hypothesis of circularity is not always satisfied as shown in [5,6] for SAR images which depends on the region of interest. Therefore, non-circularity parameters are key factors to improve performance in radar estimation and classification tasks, as proposed in [7,8].…”