Enhanced LRR-Based RFI Suppression for SAR Imaging Using the Common Sparsity of Range Profiles for Accurate Signal Recovery

RFI Suppression for SAR via a Dictionary-Based Nonconvex Low-Rank Minimization Framework and Its Adaptive Implementation

“…Recently, for narrowband RFI suppression, the previous literature has put forward plenty of effective methods, which can be divided into three categories, namely nonparametric methods [7][8][9][10][11][12][13], parametric methods [14][15][16], and semiparametric methods [17][18][19][20][21][22][23]. The common nonparametric methods include the notched filtering (NF) method [7,8], the eigen-subspace projection (ESP) method [9,10], and the LMS filtering method [11][12][13].…”

“…The newly proposed semiparametric methods especially consider the low-rank property of a two-dimensional RFI matrix, which is generated by stacking multiple pulses along the azimuth [17][18][19][20][21][22][23]. The low-rank property benefits from the fact that RFI usually occupies constant frequencies within the SAR working band for a period of time, which ensures its high degree of correlation in azimuth.…”

“…If the LFM-based useful signal, which is noise-like and thus non-sparse, is directly used as the sparse regularization without pulse compression, the performance of the optimization model is inevitably limited. Secondly, the low-rank models proposed in most of the literature [17][18][19]21] approximate the rank function of the two-dimension RFI matrix with the sum of all the singular values, namely the nuclear norm. The corresponding low-rank minimization problem can be solved by the singular value thresholding (SVT) algorithm, whereas for the nuclear norm, the SVT algorithm punishes all the singular values with the same penalty [27,28].…”

“…For the former sparsity-related problem, one possible solution is to construct an overcompleted dictionary from the discrete time-shifted versions of the transmitted signal [17,21,29]. Therefore, with the corresponding dictionary, the sparse coefficient is used for the sparse regularization, while the improper punishment is still ignored in the researches mentioned above.…”

“…Therefore, the corresponding sparse coefficient, rather than the useful signal itself, is used as the sparse regularizer for better modeling and protection. The constructed dictionary is inspired by [17,21,29], but is also combined with nonconvex functions for the first time, then further applied to our new framework. Benefitting from the discrimination for different singular values, the obtained low-rank matrix can better represent RFIs, and the useful signal protected by the sparser regularization has better fidelity with less residual RFIs.…”