Ferkans — Interactive Telecom Tutor

Why Joint Angle-Doppler Processing is Necessary

In airborne radar, ground clutter is not confined to zero Doppler. Platform motion at velocity $v_p$ gives clutter at azimuth angle $\phi$ a Doppler shift $f_d = 2v_p \sin\phi / \lambda$ . Clutter occupies a ridge in the angle-Doppler plane, cutting across the Doppler dimension at every angle. Neither spatial filtering (beamforming) alone nor temporal filtering (Doppler processing) alone can suppress this ridge — joint processing in both dimensions simultaneously is required.

STAP achieves this by forming the optimal space-time filter.

Definition:
Space-Time Snapshot Vector

Consider an array of $N_a$ elements collecting $N_p$ pulses. The space-time snapshot for a given range bin stacks all spatial-temporal samples into a single vector:

$\mathbf{x} = \begin{bmatrix} \mathbf{x}_1 \\ \mathbf{x}_2 \\ \vdots \\ \mathbf{x}_{N_p} \end{bmatrix} \in \mathbb{C}^{N_a N_p},$

where $\mathbf{x}_n \in \mathbb{C}^{N_a}$ is the spatial snapshot from pulse $n$ .

The space-time steering vector for a target at angle $\theta$ and Doppler $f_d$ has Kronecker structure:

$\mathbf{v}(\theta, f_d) = \mathbf{b}(f_d) \otimes \mathbf{a}(\theta),$

where $\mathbf{a}(\theta) \in \mathbb{C}^{N_a}$ is the spatial steering vector and $\mathbf{b}(f_d) = [1, e^{j2\pi f_d T_{\text{PRI}}}, \ldots, e^{j2\pi f_d (N_p - 1)T_{\text{PRI}}}]^T$ is the temporal steering vector.

Theorem: Optimal STAP Weight Vector

The filter that maximizes the output SINR for a target with space-time steering vector $\mathbf{v}$ in the presence of clutter-plus-noise with covariance $\mathbf{R}_{cn} = \mathbb{E}[\mathbf{x}\mathbf{x}^H]$ is:

$\mathbf{w}_{\text{opt}} = \mathbf{R}_{cn}^{-1}\mathbf{v},$

and the maximum output SINR is:

$\text{SNR}_{\text{out}} = |\alpha|^2 \mathbf{v}^H \mathbf{R}_{cn}^{-1} \mathbf{v},$

where $\alpha$ is the target complex amplitude.

This is the minimum-variance distortionless response (MVDR) beamformer generalized to the joint space-time domain. $\mathbf{R}_{cn}^{-1}$ whitens the clutter-plus-noise, and then $\mathbf{v}$ steers toward the target. The Kronecker structure $\mathbf{v} = \mathbf{b} \otimes \mathbf{a}$ connects STAP to the sensing operator factorization in Ch 07.

Proof

Output SINR

The filter output is $y = \mathbf{w}^H \mathbf{x}$ . Signal power: $|\alpha|^2 |\mathbf{w}^H\mathbf{v}|^2$ . Interference-plus-noise power: $\mathbf{w}^H \mathbf{R}_{cn} \mathbf{w}$ . SINR $= |\alpha|^2 |\mathbf{w}^H\mathbf{v}|^2 / (\mathbf{w}^H \mathbf{R}_{cn} \mathbf{w})$ .

Apply Cauchy-Schwarz with metric $\mathbf{R}_{cn}$

Let $\tilde{\mathbf{w}} = \mathbf{R}_{cn}^{1/2}\mathbf{w}$ and $\tilde{\mathbf{v}} = \mathbf{R}_{cn}^{-1/2}\mathbf{v}$ . Then SINR $= |\alpha|^2 |\tilde{\mathbf{w}}^H \tilde{\mathbf{v}}|^2 / \|\tilde{\mathbf{w}}\|^2 \leq |\alpha|^2 \|\tilde{\mathbf{v}}\|^2 = |\alpha|^2 \mathbf{v}^H\mathbf{R}_{cn}^{-1}\mathbf{v}$ , with equality when $\tilde{\mathbf{w}} \propto \tilde{\mathbf{v}}$ , i.e., $\mathbf{w} \propto \mathbf{R}_{cn}^{-1}\mathbf{v}$ .

Definition:
The Clutter Ridge

For an airborne radar with platform velocity $v_p$ , the ground clutter at azimuth angle $\phi$ has Doppler frequency:

$f_d(\phi) = \frac{2v_p \sin\phi}{\lambda}.$

In the angle-Doppler plane, this traces a sinusoidal ridge: the clutter occupies a narrow band around the curve $f_d = (2v_p/\lambda)\sin\phi$ , not a single Doppler bin.

A target moving at velocity $v_t$ at angle $\theta_t$ must be distinguished from the clutter at the same angle, which has Doppler $f_{d,c}(\theta_t) = 2v_p\sin\theta_t/\lambda$ . The target's Doppler is $f_{d,t} = 2(v_t + v_p\sin\theta_t)/\lambda$ . STAP exploits the Doppler difference $2v_t/\lambda$ to separate target from clutter.

Definition:
Brennan Rule for Sample Support

The optimal STAP filter requires the covariance $\mathbf{R}_{cn} \in \mathbb{C}^{N_a N_p \times N_a N_p}$ to be known. In practice, it is estimated from $K$ i.i.d. training snapshots (secondary data from neighboring range bins):

$\hat{\mathbf{R}}_{cn} = \frac{1}{K}\sum_{k=1}^{K}\mathbf{x}_k\mathbf{x}_k^H.$

Brennan's rule: For the sample-matrix-inversion (SMI) STAP filter to achieve within 3 dB of the optimal SINR, the number of training snapshots must satisfy:

$K \geq 2 N_a N_p.$

For a modest system with $N_a = 16$ , $N_p = 32$ , this requires $K \geq 1024$ i.i.d. range bins — often infeasible. This motivates reduced-rank and reduced-dimension STAP methods.

Definition:
Reduced-Rank STAP Methods

To reduce the sample support requirement, several approaches restrict the dimensionality of the STAP filter:

Joint Domain Localized (JDL): Process small angle-Doppler patches centered on the look direction, reducing dimensionality from $N_a N_p$ to $\bar{N}_a \bar{N}_p$ where $\bar{N}_a, \bar{N}_p \ll N_a, N_p$ .
Displaced Phase Center Antenna (DPCA): For a sidelooking array, clutter cancellation by subtracting pulses from elements displaced by exactly one inter-pulse distance ( $v_p T_{\text{PRI}} = d$ , element spacing).
Principal Components (PC): Project onto the dominant eigenvectors of $\hat{\mathbf{R}}_{cn}$ , suppressing clutter in the low-rank subspace.

These methods reduce the training requirement from $2N_aN_p$ to $2\bar{N}_a\bar{N}_p$ or the rank of the clutter subspace.

Example: STAP SINR Improvement for Airborne Radar

An airborne radar has $N_a = 8$ elements, $N_p = 16$ pulses, $d = \lambda/2$ , $v_p = 150$ m/s, $\lambda = 3$ cm. Compute (a) the clutter Doppler at broadside ( $\phi = 0°$ ) and at $\phi = 30°$ , (b) the minimum detectable velocity (MDV), and (c) the required training snapshots for full-rank STAP.

Solution

Clutter Doppler

$f_d(0°) = 0$ Hz (broadside clutter is at zero Doppler). $f_d(30°) = 2 \times 150 \times \sin(30°)/0.03 = 5000$ Hz.

MDV

The MDV is approximately the Doppler resolution of the clutter-notch filter: $\Delta f_d = f_{\text{PRF}} / N_p$ . If PRF $= 5000$ Hz, $\Delta f_d = 312.5$ Hz, giving $v_{\text{MDV}} = \Delta f_d \cdot \lambda/2 = 4.7$ m/s.

Training requirement

Brennan rule: $K \geq 2 N_a N_p = 2 \times 8 \times 16 = 256$ i.i.d. range bins. For JDL with $\bar{N}_a = 3, \bar{N}_p = 3$ : $K \geq 18$ — a dramatic reduction.

STAP Improvement Factor

Visualize the SINR improvement factor as a function of target Doppler for full-rank STAP vs. factored processing (spatial beamforming followed by Doppler filtering). The clutter notch at the platform Doppler is visible.

Parameters

N_a

8

N_p

16

v_p

(m/s)150

CNR (dB)30

⚠️Engineering Note

Practical STAP Limitations

Implementing STAP in real radar systems faces several challenges:

Non-stationarity: The clutter covariance changes with range (due to geometry) and terrain type. Training data from neighboring range bins may not be i.i.d.
Computational cost: Full-rank STAP requires inverting an $N_aN_p \times N_aN_p$ matrix — $O(N_a^3 N_p^3)$ operations. JDL and other reduced-dimension methods bring this to manageable levels.
Heterogeneous clutter: Urban terrain, mountains, and coastlines create discrete clutter returns that violate the Gaussian clutter assumption.
Jamming: Active jammers create additional interference that must be suppressed jointly with clutter.

Knowledge-aided STAP (KA-STAP) uses terrain databases and prior information to improve covariance estimation when training data is limited.

Why This Matters: STAP and the Imaging Forward Model

The STAP weight vector $\mathbf{w} = \mathbf{R}_{cn}^{-1}\mathbf{v}$ is the one-voxel version of the imaging problem. For a full image, we apply this filter for every look direction $\theta$ and every Doppler $f_d$ — sweeping $\mathbf{v}(\theta, f_d)$ over the angle-Doppler plane.

The connection to the forward model is:

The data vector $\mathbf{x}$ is one column of $\mathbf{y}$ .
The steering vector $\mathbf{v}$ is one column of $\mathbf{A}$ .
The STAP weight $\mathbf{R}_{cn}^{-1}\mathbf{v}$ is the MVDR beamformer, which is a column of the Capon imaging matrix (Ch 13.4).

Full-scene imaging is STAP applied to all voxels simultaneously.

See full treatment in Chapter 13

Common Mistake: Insufficient Training Data for STAP

Mistake:

Applying full-rank SMI-STAP when the number of available training snapshots $K < 2N_aN_p$ , expecting near-optimal performance.

Correction:

With $K < 2N_aN_p$ , the estimated covariance $\hat{\mathbf{R}}_{cn}$ is poorly conditioned, and the STAP filter amplifies noise instead of suppressing clutter. The SINR loss can exceed 10 dB. Use reduced-rank methods (JDL, PC-STAP) that require only $K \geq 2\bar{N}$ training snapshots, where $\bar{N}$ is the reduced dimension.

STAP (Space-Time Adaptive Processing)

Joint spatial and temporal adaptive filtering for suppressing clutter in airborne or spaceborne radar. The optimal STAP weight is $\mathbf{w} = \mathbf{R}_{cn}^{-1}\mathbf{v}$ where $\mathbf{R}_{cn}$ is the clutter-plus-noise covariance and $\mathbf{v}$ is the space-time steering vector.

Key Takeaway

STAP jointly filters in angle and Doppler via $\mathbf{w} = \mathbf{R}_{cn}^{-1}\mathbf{v}$ , suppressing clutter that occupies a sinusoidal ridge in the angle-Doppler plane. The Brennan rule ( $K \geq 2N_aN_p$ ) governs sample support; reduced-rank methods (JDL, DPCA) make STAP practical. STAP is the single-voxel version of imaging — full-scene reconstruction applies it to every voxel simultaneously.

STAP: Space-Time Adaptive Processing