Cognitive ISAR: Adaptive Spectrum-Aware Radar

• Cognitive ISAR is an adaptive radar system that closes the perception–action loop via spectrum sensing, tailored waveform synthesis, and robust data recovery.
• It employs convex QCQP and recovery techniques like compressed sensing and rank minimization to mitigate the effects of spectral notching and slow-time gaps.
• This approach enables spectral coexistence in congested RF environments, offering resilience under low SNR and interference in diverse operational settings.

Searching arXiv for papers on cognitive ISAR and closely related ISAR methods. Cognitive inverse synthetic aperture radar (ISAR) denotes an ISAR architecture that closes a perception–action loop in congested radio-frequency environments, alternating between environmental perception and adaptive transmission so that imaging objectives are met while spectral coexistence with overlaid emitters is preserved. In the formulation introduced for spectral compatibility applications, the radar first senses the spectrum to obtain the true relevant spectral parameters of active emitters, then synthesizes and transmits a tailored waveform with bespoke spectral notches, and finally recovers the missing data induced by those notches and by cognitive scheduling before forming the final image (Rosamilia et al., 11 Jul 2025). A related line of work extends the cognitive perspective beyond spectral compatibility by embedding learned scene priors and a differentiable radar forward model into ISAR reconstruction, thereby supporting adaptive sensing under sparse, noisy, and even Non-Line-of-Sight measurement conditions (Oshim et al., 2024).

1. Conceptual definition and architectural scope

Cognitive ISAR is defined by the closure of the perception–action loop. In the spectral-compatibility formulation, the radar perceives the RF environment using a spectrum sensing module that provides the true relevant spectral parameters of active emitters, including center frequencies, bandwidths, occupancy, and time variation; decides a coexistence strategy in terms of spectral masks, stop-bands, and coexistence constraints; acts by synthesizing a spectrally shaped waveform with tailored notches; and applies a recovery stage that compensates for missing data in the frequency domain and in slow time. Updated spectral awareness is then fed back to refine subsequent waveforms and processing (Rosamilia et al., 11 Jul 2025).

The resulting system is not merely a waveform-design procedure. It is an end-to-end sensing-and-imaging chain comprising multi-snapshot sensing outputs, spectral mask determination, QCQP-based waveform synthesis, transmission and reception, standard ISAR processing, and recovery through either compressed sensing or rank minimization. This suggests that “cognition” in this context is operationalized as adaptive spectral coexistence coupled to adaptive reconstruction, rather than as a generic label for intelligent processing (Rosamilia et al., 11 Jul 2025).

A broader interpretation appears in NeRF-enabled Analysis-through-Synthesis for ISAR, where cognition is associated with the incorporation of scene priors, differentiable physical modeling, and the potential for adaptive acquisition. There, a learned radar forward model is used to reconstruct small objects from sparse and noisy UWB measurements, and the differentiability of the model is linked to next-best-view selection, waveform adaptation, and bandwidth allocation as prospective closed-loop controls (Oshim et al., 2024). This suggests that cognitive ISAR can encompass both spectrum-aware transmission and model-aware acquisition.

2. Signal model, imaging formation, and the effect of missing data

The multi-scatterer baseband model used for cognitive ISAR considers a target with $M$ point scatterers indexed by $m$, complex reflectivities $a_m$, instantaneous ranges $R_m(t)$, carrier $f_c$, and wideband baseband transmit signal $s(t)$. After dechirp or matched filtering and range compression, the received baseband over fast time $\tau$ and slow time $p$ is modeled as

$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$

where $w(\cdot)$ is the compressed pulse shape, $m$0 captures residual motion-induced phase after translational motion compensation, and $m$1 is noise or interference. After FFT along $m$2,

$m$3

When spectral notches are present, $m$4 is observed only for $m$5; when slow-time pulses are skipped, observations are available only for $m$6 (Rosamilia et al., 11 Jul 2025).

Range–Doppler formation proceeds through FFT along slow time:

$m$7

with $m$8 the PRI. Missing data in frequency and slow time appear as holes in $m$9 and produce elevated sidelobes and artifacts unless explicitly recovered (Rosamilia et al., 11 Jul 2025).

The paper states the range-resolution relationship under spectral notching as

$a_m$0

where $a_m$1 is the usable non-notched bandwidth. For the two-notch example inside a nominal $a_m$2 band, notches of $a_m$3 and $a_m$4 reduce $a_m$5 from $a_m$6 to approximately $a_m$7, so $a_m$8 versus approximately $a_m$9 without notches (Rosamilia et al., 11 Jul 2025). Cross-range and Doppler resolutions are described by

$R_m(t)$0

with the caveat that missing slow-time pulses can degrade effective aperture and Doppler resolution unless recovery reconstructs the gaps (Rosamilia et al., 11 Jul 2025).

The ATS-NeRF paper uses the classical ISAR relations

$R_m(t)$1

and emphasizes that sparse angular coverage, indoor clutter, and low-RCS targets intensify the limitations of backprojection, range-Doppler, and Polar Format Algorithms in practical settings (Oshim et al., 2024). In that framing, cognitive capability is tied less to coexistence constraints and more to extracting coherent images from scarce measurements through a physics-grounded prior.

3. Spectral perception and coexistence constraints

In the spectral-compatibility architecture, perception is performed by multi-snapshot spectrum sensing with block-sparsity exploitation over multichannel coherent receivers or SDR, with extensions to off-grid sensing, and it outputs the occupied bands $R_m(t)$2 for $R_m(t)$3 emitters, their notch depth requirements, and time activity in temporal slots (Rosamilia et al., 11 Jul 2025). The role of sensing is therefore not merely detection; it parameterizes waveform synthesis directly.

Spectral compatibility is enforced through per-band interference constraints. If $R_m(t)$4 denotes the discrete-time transmit sequence, and $R_m(t)$5 is built from the Fourier submatrix columns corresponding to $R_m(t)$6, then the average signal energy injected into each coexistence band is limited by

$R_m(t)$7

The associated spectral representation is written as

$R_m(t)$8

These constraints define per-band notch depths and compatibility with licensed or coexisting users (Rosamilia et al., 11 Jul 2025).

An important limitation is stated explicitly: no explicit $R_m(t)$9 or sensing-latency figures are reported. The perception module is assumed to provide reliable spectral parameters for downstream design (Rosamilia et al., 11 Jul 2025). This suggests that, in the reported system, coexistence performance is evaluated chiefly through waveform compliance and imaging quality rather than through a complete end-to-end joint characterization of sensing reliability and adaptation latency.

4. Tailored waveform synthesis and QCQP design

The waveform-design objective is to maximize usable bandwidth while carving stop-bands that satisfy coexistence requirements, preserve favorable ambiguity or autocorrelation behavior, and respect power constraints and block-energy bounds for temporal smoothness (Rosamilia et al., 11 Jul 2025). The reference signal is a chirp $f_c$0 of $f_c$1 samples with $f_c$2 bandwidth and $f_c$3 duration.

The global design is cast as a convex QCQP:

$f_c$4

and is implemented through a tractable block-iterative QCQP for large $f_c$5. For the first block,

$f_c$6

while subsequent blocks exploit an overlap $f_c$7 with the previous block and enforce the same block-energy and spectral-interference conditions on the concatenated vector $f_c$8 (Rosamilia et al., 11 Jul 2025).

For the reported design, $f_c$9 and $s(t)$0. Two licensed emitters occupy $s(t)$1 and $s(t)$2, corresponding to $s(t)$3 and $s(t)$4 notches. The synthesized waveform exhibits notch depths of approximately $s(t)$5 and $s(t)$6, with small spurious components near band edges. The normalized autocorrelation peak sidelobe level degrades from $s(t)$7 for the reference chirp to $s(t)$8 for the notched waveform, while the $s(t)$9 mainlobe remains essentially unchanged (Rosamilia et al., 11 Jul 2025).

Several trade-offs are identified explicitly. Reducing $\tau$0 increases $\tau$1; notches distort the matched filter and can raise range sidelobes and produce image artifacts; the block-energy constraint maintains temporal smoothness; and PAPR can rise with spectral shaping. Doppler robustness remains chiefly determined by CPI and motion compensation, although frequency notches can still affect range–Doppler artifacts (Rosamilia et al., 11 Jul 2025). A common misconception is therefore that spectral notching alone delivers coexistence “for free.” The reported architecture shows that notching must be paired with recovery to prevent quality loss from the induced missing samples.

5. Recovery of missing frequency bins and slow-time gaps

Missing-data recovery is the central compensatory mechanism that makes cognitive ISAR viable under spectral coexistence constraints. In the frequency domain, with full-band range profile $\tau$2, observed samples $\tau$3 on index set $\tau$4, and partial Fourier or matched-filter operator $\tau$5, the model is

$\tau$6

Under sparsity assumptions, the paper lists the noiseless and noisy formulations

$\tau$7

and

$\tau$8

In the implemented image-domain recovery, the incomplete slow-time/frequency matrix $\tau$9 is related to the image $p$0 through undercomplete Fourier dictionaries $p$1 and $p$2, giving

$p$3

The 2D-SL0 algorithm approximates $p$4 minimization via smoothed $p$5 with parameters $p$6, $p$7, $p$8, and $p$9 (Rosamilia et al., 11 Jul 2025).

The sparsity argument is that ISAR images are sparse in a 2D Fourier basis at high frequencies because only a few dominant scatterers contribute strongly. The paper notes that incoherence is promoted by randomized missing patterns, whereas deterministic notches are mitigated by the structure and sparsity of the data (Rosamilia et al., 11 Jul 2025). This is a limited but precise claim: the system does not assert arbitrary recoverability for any notch pattern.

Slow-time gaps introduced by cognitive scheduling or MPAR are treated through rank minimization. With image $$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$0, the convex relaxation is

$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$1

and the MM-based update is

$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$2

followed by SVD and singular-value soft-thresholding:

$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$3

with

$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$4

Termination occurs when the Shannon effective rank of $$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$5 drops below that of the incomplete RD image (Rosamilia et al., 11 Jul 2025).

The paper gives a clear preference criterion. Compressed sensing is preferred when a few bright scatterers dominate and SNR is moderate to high; rank minimization is preferred when noise or interference dominates or sparsity is weak, because correlated slow-time evolution is well modeled by low rank (Rosamilia et al., 11 Jul 2025). It also quantifies the computational trade-off: 2D-SL0 is computationally lightweight with rapid convergence, whereas RM-MM is dominated by an SVD with per-iteration complexity $$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$6, although partial or randomized SVD is suggested for real-time constraints (Rosamilia et al., 11 Jul 2025).

6. Processing chain, experimental evidence, and performance regimes

The complete processing chain consists of dechirp or matched filtering and range compression, motion compensation or autofocus, recovery through CS or RM, Doppler or cross-range processing, and feedback from image quality to the next cognitive cycle (Rosamilia et al., 11 Jul 2025). The interaction with cognitive scheduling is explicit: pulse-skipping for coexistence with other RF activities in MPAR induces slow-time gaps, and motion compensation must remain robust to nonuniform sampling, with precompensation integrated before recovery when needed (Rosamilia et al., 11 Jul 2025).

The reported experiments use a drone measurement dataset in the $$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$7–$$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$8 band with HH polarization, $$r(\tau,p) \approx \sum_{m=1}^{M} a_m \, w\!\left(\tau - \tfrac{2 R_m(p)}{c}\right) \, e^{-j \tfrac{4\pi f_c}{c} \Delta R_m(p)} + n(\tau,p),$$9 bandwidth, and $w(\cdot)$0 frequency step. The target is a DJI Matrice 100 over an aspect angle span of approximately $w(\cdot)$1 with $w(\cdot)$2 steps. Spectral analysis uses Welch PSD with a Blackman–Harris window, segments of $w(\cdot)$3 samples, and $w(\cdot)$4 overlap (Rosamilia et al., 11 Jul 2025).

The principal quantitative results are summarized below.

Scenario	Method	IC / COH / NMSE
Two interferers	GT	0.0944 / 1.0000 / 0.0000
Two interferers	Standard chirp with interference	0.0839 / 0.9986 / 0.0600
Two interferers	Notched (no recovery)	0.0855 / 0.9987 / 0.0533
Two interferers	Notched + CS/RM	0.0933 / 0.9994 / 0.0342
Multiple emitters active in different temporal slots	GT	0.0944 / 1.0000 / 0.0000
Multiple emitters active in different temporal slots	Standard	0.0776 / 0.9980 / 0.0826
Multiple emitters active in different temporal slots	Notched	0.0824 / 0.9988 / 0.0540
Multiple emitters active in different temporal slots	Notched + CS/RM	0.0898 / 0.9994 / 0.0348
MPAR with two interferers (50% slow-time gap)	GT	0.0944 / 1.0000 / 0.0000
MPAR with two interferers (50% slow-time gap)	Standard	0.0672 / 0.9976 / 0.0864
MPAR with two interferers (50% slow-time gap)	Notched	0.0731 / 0.9982 / 0.0647
MPAR with two interferers (50% slow-time gap)	Notched + CS/RM	0.0900 / 0.9990 / 0.0451

These data support a precise interpretation. Conventional non-cognitive ISAR suffers from strong interference artifacts; notching alone avoids interference but degrades the image because of missing bins; notching combined with CS or RM recovers near-ground-truth images with limited compromise in resolution and sidelobes while preserving spectral compatibility (Rosamilia et al., 11 Jul 2025).

The low-SNR regime exposes an important distinction between the two recovery paradigms. At single-pulse SNR $w(\cdot)$5 after compression with two interferers, notched plus CS struggles and produces spurious scatterers, whereas Notched + RM suppresses interference and reconstructs a faithful ISAR image. Across SNR from $w(\cdot)$6 to $w(\cdot)$7, Notched + RM achieves higher COH and lower NMSE than alternatives, and IC is highest for RM above approximately $w(\cdot)$8, while below that threshold IC does not correlate with fidelity (Rosamilia et al., 11 Jul 2025). This addresses a common misconception that image sharpness metrics alone are sufficient to assess cognitive ISAR quality under heavy interference.

A distinct but related evidence base appears in the ATS-NeRF study, which evaluates small-object coherent ISAR using a monostatic PulsON P440 UWB impulse radar over $w(\cdot)$9–$m$00 with center frequency $m$01 and classical range resolution $m$02. With synthetic additive Gaussian noise $m$03, ATS outperforms BP across PSNR, LPIPS, and MSE for one to four reflectors. For one reflector, ATS achieves PSNR $m$04, LPIPS $m$05, and MSE $m$06, versus BP PSNR $m$07, LPIPS $m$08, and MSE $m$09; analogous advantages are reported for two, three, and four reflectors (Oshim et al., 2024). The same work reports sparse-view and NLOS robustness, including reconstruction of a soda can inside a cardboard box, while also quantifying latency as ATS approximately $m$10 versus BP approximately $m$11 on an NVIDIA RTX 3080 Ti (Oshim et al., 2024). This suggests that cognitive ISAR can also be understood as robust, prior-informed reconstruction under measurement scarcity, albeit with present computational costs.

7. Applications, limitations, and emerging directions

The reported application domains for cognitive ISAR with spectral compatibility include spectrum-sharing environments with legacy communications, MPAR platforms performing search, track, imaging, and communications concurrently, urban RF crowding, and EW-contested scenarios. Rank minimization is stated to be particularly effective under jamming and low SNR (Rosamilia et al., 11 Jul 2025). These are direct application settings for the coexistence-oriented formulation.

The limitations are equally explicit. Rapidly time-varying interferers challenge perception latency and notch agility; deep or wide notches reduce $m$12 and increase $m$13; residual edge spurious may increase sidelobes without robust recovery; CS depends on sparsity and adequate SNR; RM is more robust but computationally heavier; motion-compensation errors can reduce recovery fidelity; and perception errors degrade coexistence (Rosamilia et al., 11 Jul 2025). The architecture therefore depends on the joint adequacy of sensing accuracy, waveform agility, motion compensation, and inverse reconstruction.

Several extensions are named directly: multi-band agility, MIMO or multistatic cognitive ISAR, joint sensing–communication with intra-pulse embeds and adaptive notch control, learned priors and deep unfolding for CS or RM, integration with regulatory frameworks and dynamic QoS-driven notch shaping, and Schatten-$m$14 norm regularization generalizing the nuclear norm in RM (Rosamilia et al., 11 Jul 2025). These extensions preserve the underlying structure of perception, constrained action, and recovery.

The ATS-NeRF line adds a different set of prospective directions: next-best-view selection through expected information gain or Fisher information, waveform adaptation via pulse width, center frequency, or time-gating, bandwidth allocation to maximize sensitivity to uncertain structures, and on-the-fly acquisition tuning based on residuals and uncertainty (Oshim et al., 2024). A plausible implication is that the future of cognitive ISAR may involve the convergence of two historically distinct mechanisms of cognition: spectrum-aware coexistence control and model-aware adaptive measurement selection.

Taken together, the current literature represented here frames cognitive ISAR as a technically specific class of adaptive radar imaging systems. In one form, it senses occupied spectrum, synthesizes notched waveforms under convex interference constraints, and reconstructs missing data to preserve image fidelity in crowded RF environments (Rosamilia et al., 11 Jul 2025). In another, it embeds differentiable radar physics and learned priors into the reconstruction itself, enabling principled adaptation under sparse, noisy, and NLOS measurements (Oshim et al., 2024). The common denominator is not a single algorithm, but the systematic use of feedback, constraints, and prior structure to optimize ISAR imaging under operational limits.