Ferkans — Interactive Telecom Tutor

Compressed Sensing Meets SAR

The matched-filter image $\mathbf{A}^{H} \mathbf{y}$ is fast and simple, but it suffers from sidelobes, limited dynamic range, and resolution bounded by the Rayleigh limit. Compressed sensing SAR (CS-SAR) exploits the sparsity of many SAR scenes to achieve three goals simultaneously: super-resolution beyond the diffraction limit, sub-Nyquist acquisition with fewer pulses, and artifact suppression without windowing.

The theoretical foundations come from Chapter 11 (RIP, sparse recovery guarantees); the optimization algorithms from Chapter 4 (ISTA, FISTA, ADMM). Here we focus on the SAR-specific applications and the role of the structured sensing matrix.

Definition:
CS-SAR Formulation

The CS-SAR reconstruction solves:

$\hat{\mathbf{c}} = \arg\min_{\mathbf{c}} \frac{1}{2}\|\mathbf{y} - \mathbf{A}\mathbf{c}\|_2^2 + \lambda\|\boldsymbol{\Psi}\mathbf{c}\|_1,$

where $\boldsymbol{\Psi}$ is a sparsifying transform:

Identity: For scenes with isolated point targets.
Wavelet transform: For natural scenes with multi-scale structure.
Learned dictionary: For domain-specific priors (urban, vegetation).

The Kronecker structure $\mathbf{A} = \mathbf{A}_{\text{az}} \otimes \mathbf{A}_{\text{rg}}$ enables efficient matrix-vector products in all iterative solvers without ever forming the full $\mathbf{A}$ matrix.

Definition:
TV-Regularized SAR

For scenes with extended regions and sharp edges (e.g., urban areas, sea-land boundaries), total variation (TV) regularization preserves edges while suppressing noise:

$\hat{\mathbf{c}} = \arg\min_{\mathbf{c}} \frac{1}{2}\|\mathbf{y} - \mathbf{A}\mathbf{c}\|_2^2 + \lambda\,\text{TV}(\mathbf{c}),$

where $\text{TV}(\mathbf{c}) = \sum_{i,j}\sqrt{|\Delta_x \mathbf{c}_{i,j}|^2 + |\Delta_y \mathbf{c}_{i,j}|^2}$ is the isotropic total variation.

TV-SAR avoids the staircase artifacts of $\ell_1$ on extended targets while retaining the super-resolution capability for isolated scatterers.

Theorem: CS-SAR Recovery Guarantee

Consider a stripmap SAR with $N$ azimuth positions, sub-sampled to $M$ randomly selected positions. If the scene reflectivity $\mathbf{c}$ is $S$ -sparse in some orthonormal basis $\boldsymbol{\Psi}$ , and

$M \geq C \cdot S \cdot \log^4 N,$

then $\ell_1$ minimization recovers $\mathbf{c}$ with high probability.

The constant $C$ depends on the coherence between the sensing matrix rows and the sparsity basis. For random sub-sampled Fourier matrices (the SAR case), $C$ is small.

The SAR sensing matrix is a partial Fourier matrix — precisely the class of structured matrices for which compressed sensing theory provides the strongest guarantees. This is not a coincidence: SAR data lives in the Fourier domain (spatial frequencies), and sparsity in the image domain is a natural physical property.

Proof

RIP of sub-sampled Fourier

The restricted isometry property (RIP) of random sub-sampled Fourier matrices was established by Candes and Tao (2006) and Rudelson and Vershynin (2008). Sub-sampling $M$ rows of the $N \times N$ DFT matrix yields a matrix satisfying the RIP of order $S$ with high probability when $M = O(S \log^4 N)$ .

Application to SAR

The azimuth sensing matrix $\mathbf{A}_{\text{az}}$ is (up to phase factors) a partial DFT. Random pulse selection is equivalent to random row sub-sampling. Combined with the range Kronecker factor, the full SAR CS guarantee follows. $\blacksquare$

CS-SAR vs Matched Filter Reconstruction

Compares matched-filter and compressed-sensing reconstruction for a sparse SAR scene.

Top-left: True scene (sparse point targets). Top-right: Matched filter with full data. Bottom-left: Matched filter with sub-sampled data (note aliasing). Bottom-right: CS-SAR (ISTA) with the same sub-sampled data.

Reduce the sub-sampling percentage to see how CS-SAR gracefully degrades while the matched filter develops severe artifacts.

Parameters

Sub-sampling (% of full data)25

SNR (dB)20

Number of targets8

Example: CS-SAR with 75% Data Reduction

A SAR system collects $N = 1024$ azimuth pulses for a scene with $S = 20$ point targets. Using only $M = 256$ randomly selected pulses (75% reduction):

(a) Verify the CS recovery condition.

(b) Compare matched-filter and $\ell_1$ reconstruction.

Solution

Recovery condition

$M = 256 > C \cdot S \cdot \log^4 N \approx C \cdot 20 \cdot 10^4 \approx 200$ for $C \approx 1$ . The condition is satisfied.

Matched filter from sub-sampled data

Zero-filling the missing pulses and applying the DFT produces an image with severe azimuth ambiguities (sidelobes from the non-uniform sampling pattern).

$\ell_1$ reconstruction

ISTA/FISTA with the sub-sampled sensing matrix recovers all 20 targets at full resolution, with sidelobes suppressed below the noise floor. This is the power of CS: the sparsity prior replaces the missing data.

Definition:
Joint Autofocus and Sparse Reconstruction

Combining autofocus with sparse recovery yields the joint optimization:

$\min_{\mathbf{c}, \phi_e} \frac{1}{2}\|\mathbf{y} - \mathbf{D}(\phi_e)\mathbf{A}\mathbf{c}\|_2^2 + \lambda\|\mathbf{c}\|_1,$

where $\mathbf{D}(\phi_e) = \text{diag}(e^{j\phi_e(\eta_1)}, \ldots, e^{j\phi_e(\eta_{N_a})}) \otimes \mathbf{I}_{N_f}$ .

Alternating minimization:

Fix $\phi_e$ , solve for $\mathbf{c}$ via ISTA.
Fix $\mathbf{c}$ , solve for $\phi_e$ via gradient descent.

The sparsity prior regularizes both the image and the phase errors, avoiding the error propagation of sequential PGA-then-reconstruct pipelines.

,

Definition:
Polarimetric SAR (PolSAR) Imaging

Polarimetric SAR transmits and receives on multiple polarizations (HH, HV, VH, VV), forming the scattering matrix at each pixel:

$\mathbf{S}(\tgtpos) = \begin{bmatrix} S_{HH} & S_{HV} \\ S_{VH} & S_{VV} \end{bmatrix}.$

The Pauli decomposition extracts physical scattering mechanisms:

Surface scattering ( $|S_{HH} + S_{VV}|$ ): Smooth surfaces.
Double-bounce ( $|S_{HH} - S_{VV}|$ ): Building-ground interaction.
Volume ( $|S_{HV}|$ ): Vegetation canopy.

Polarimetric data can be incorporated into the CS framework via group sparsity ( $\ell_{2,1}$ norm): the four polarization channels of each scatterer share the same support.

🎓CommIT Contribution(2026)

Unified Illumination and Sensing Model for RF Imaging

G. Caire — TU Berlin Technical Report (internal)

Caire's unified model (2026) shows that SAR, ISAR, and multi-static RF imaging all derive from the same forward model with different parameterizations of the sensing matrix $\mathbf{A}$ . In SAR, the rows of $\mathbf{A}$ are indexed by (pulse position, frequency); in ISAR, by (slow-time, frequency) with the target rotation implicitly embedded; in multi-static imaging, by (Tx, Rx, frequency) triples. This unification enables a single reconstruction framework — from matched filtering through compressed sensing to learned regularizers — to be applied across all modalities.

RF imagingforward modelsensing matrix

Connecting CS-SAR to the Book's Framework

CS-SAR brings together several threads from the book:

Chapter 1: SVD and operator analysis of $\mathbf{A}$ .
Chapter 2: Regularization theory (Tikhonov, truncated SVD).
Chapter 3: Convex analysis and proximal operators ( $\ell_1$ , TV).
Chapter 4: ISTA/FISTA/ADMM for efficient computation.
Chapter 8: Kronecker structure of the SAR sensing matrix.
Chapter 11: RIP and CS recovery guarantees.

The techniques developed in Part IV — ISTA (§Fast Algorithms for Structured Operators), ADMM, group sparsity — are all directly applicable to SAR with the structured $\mathbf{A}$ from this chapter.

SAR Reconstruction Methods Comparison

Method	Formula	Advantage	Limitation
Matched filter (RDA)	$\mathbf{A}^{H} \mathbf{y}$	Fast, simple	Sidelobes, Rayleigh-limited
Tikhonov	$(\mathbf{A}^{H}\mathbf{A} + \lambda\mathbf{I})^{-1}\mathbf{A}^{H}\mathbf{y}$	Reduced sidelobes	Blurs point targets
$\ell_1$ (CS-SAR)	$\min \\|\mathbf{y} - \mathbf{A}\mathbf{c}\\|^2 + \lambda\\|\mathbf{c}\\|_1$	Super-resolution, sub-Nyquist	Assumes sparse scene
TV-SAR	$\min \\|\mathbf{y} - \mathbf{A}\mathbf{c}\\|^2 + \lambda\text{TV}(\mathbf{c})$	Edge-preserving	Staircase artifacts

Historical Note: Compressed Sensing and SAR — A Natural Marriage

2008--2014

The application of compressed sensing to SAR was proposed almost simultaneously by several groups around 2008--2010, shortly after the foundational CS papers of Candes, Romberg, and Tao (2006). The SAR community was quick to recognize that SAR data naturally lives in the Fourier domain and that many SAR scenes are sparse in the image domain — precisely the conditions under which CS theory provides the strongest guarantees. The survey by Cetin et al. (2014) consolidated a decade of work into a unified framework that connects SAR to the broader inverse problems literature.

Quick Check

Which of the following is NOT an advantage of CS-SAR over matched-filter SAR?

Super-resolution beyond the Rayleigh limit

Reduced data acquisition (fewer pulses)

Faster computation time

Suppressed sidelobes without windowing

Correction:

Faster computation time

CS-SAR requires iterative optimization (e.g., hundreds of ISTA iterations), which is much slower than the single-pass FFT of the matched filter. The advantage of CS-SAR is better image quality, not faster computation.

Key Takeaway

CS-SAR applies compressed sensing to exploit scene sparsity for super-resolution, sub-Nyquist acquisition, and sidelobe suppression. The SAR sensing matrix is a partial Fourier matrix — the ideal structure for CS guarantees. TV regularization extends the approach to non-sparse (but piecewise-smooth) scenes. Joint autofocus + sparse recovery avoids error propagation from sequential processing. Polarimetric SAR adds a group-sparsity dimension. All methods reduce to instances of the universal model $\mathbf{y} = \mathbf{A}\mathbf{c} + \mathbf{w}$ .

Sparse SAR and Compressed Sensing SAR

Compressed Sensing Meets SAR

Definition: CS-SAR Formulation

Definition: TV-Regularized SAR

Theorem: CS-SAR Recovery Guarantee

RIP of sub-sampled Fourier

Application to SAR

CS-SAR vs Matched Filter Reconstruction

Parameters

Example: CS-SAR with 75% Data Reduction

Recovery condition

Matched filter from sub-sampled data

$\ell_1$ reconstruction

Definition: Joint Autofocus and Sparse Reconstruction

Definition: Polarimetric SAR (PolSAR) Imaging

Unified Illumination and Sensing Model for RF Imaging

Connecting CS-SAR to the Book's Framework

SAR Reconstruction Methods Comparison

Historical Note: Compressed Sensing and SAR — A Natural Marriage

Quick Check

Key Takeaway

Definition:
CS-SAR Formulation

Definition:
TV-Regularized SAR

Definition:
Joint Autofocus and Sparse Reconstruction

Definition:
Polarimetric SAR (PolSAR) Imaging