arXiv CS open access

Multistatic ISAC with Deep Learning-Based Clutter Suppression for 5G and Beyond

Yalin E. Sagduyu et al. · arxiv_cs

arXiv CS · Papers

distributed-systems internet networking protocols

Integrated sensing and communications (ISAC) systems using 5G and beyond signals of opportunity enable joint wireless sensing and communication using shared infrastructure and spectrum resources. While multistatic ISAC can increase sensing coverage with spatially distributed receivers, dense and dyn

Conceptio Open Archive

Analyze & explore related research in Conceptio

Connect this record directly into your research workflow or AI agent. Explore citations across open repositories, query full text, and export structured proof bundles.

Explore in Search →Connect via MCP / API ↗

Multistatic ISAC with Deep Learning-Based Clutter Suppression for 5G and Beyond Yalin E. Sagduyu, Tugba Erpek, and Sastry Kompella

arXiv:2610.40586v1 [cs.NI] 29 Oct 2026

Nexcepta, Gaithersburg, MD, USA

Abstract—Integrated sensing and communications (ISAC) systems using 5G and beyond signals of opportunity enable joint wireless sensing and communication using shared infrastructure and spectrum resources. While multistatic ISAC can increase sensing coverage with spatially distributed receivers, dense and dynamic clutter can significantly degrade bistatic range and Doppler estimation, leading to errors in target localization and velocity estimation. This paper presents a multistatic ISAC framework with clutter suppression for high-fidelity estimation of target position and velocity. Distributed clutter is modeled with randomly located scatterers with varying Doppler spreads and radar cross sections to capture dense and dynamic clutter effects. A broad range of conventional signal processing baselines are evaluated for clutter suppression, showing limited robustness under strong clutter conditions. To address this limitation, deep learning-based clutter suppression methods are introduced to improve robustness against dense and dynamic clutter by employing residual and attention-enhanced deep neural network architectures that directly enhance target responses from cluttercontaminated observations. These methods substantially outperform conventional approaches in target position and velocity estimation across diverse clutter conditions in multistatic ISAC. Index Terms—ISAC, multistatic, deep learning, 5G, 6G, localization, velocity estimation, clutter suppression.

I. I NTRODUCTION Integrated sensing and communications (ISAC) is a key technology for next-generation wireless networks that repurposes communication signals for sensing using shared spectrum and infrastructure [1]. An important advantage of ISAC is the ability to exploit signals of opportunity, such as commercial 5G NR signals, for target sensing and detection [2]–[4]. In particular, 5G NR positioning reference signals (PRS) provide structured pilot patterns that can be effectively repurposed for passive sensing without dedicated radar transmitters and support persistent low-cost sensing over large areas [5]. The ISAC capabilities are particularly important for mission-critical applications such as intelligence, surveillance, and reconnaissance (ISR), border monitoring, and counter-unmanned aerial systems (counter-UAS), which require robust sensing, localization, and tracking in complex environments. Monostatic ISAC systems employ co-located transmitters and receivers, which simplifies synchronization and signal processing but limits sensing coverage and geometric diversity [6]–[10]. In contrast, multistatic ISAC systems use spatially distributed receivers observing reflections from a common transmitter, enabling improved spatial diversity, wider coverage, enhanced target observability, and more accurate

localization [11], [12]. Despite these advantages, multistatic ISAC sensing remains challenging in dense clutter environments, where delayed and Doppler-shifted clutter reflections can overlap with target returns in bistatic range-Doppler space. Since clutter characteristics can vary across spatially distributed receivers, clutter can significantly degrade target localization and velocity estimation performance [13], [14]. In this paper, we study multistatic ISAC with 5G NR PRS in clutter environments for bistatic range-Doppler estimation, two-dimensional (2D) target localization, and velocity estimation using spatially distributed receivers. A multistatic sensing framework based on cross-ambiguity-function (CAF) processing and nonlinear multilateration is adopted. The clutter is introduced using distributed scatterers with randomized locations, radar cross sections (RCSs), and Doppler spreads to capture diverse clutter conditions including highly dynamic clutter with significant Doppler overlap with target returns. We show that conventional signal processing-based clutter suppression methods can only provide partial improvement and their effectiveness drops under dense or dynamic clutter conditions that degrade multistatic ISAC performance. To improve robustness, we introduce deep learning (DL)based clutter suppression methods to directly enhance target responses from clutter-contaminated observations. The first method employs a residual U-Net architecture for targetenhanced residual estimation, while the second method introduces directional convolutional processing and attention-gated feature fusion to further improve clutter suppression, target enhancement, and multistatic sensing performance. We show that the proposed residual and attention-aware DL methods substantially outperform conventional clutter suppression approaches across a wide range of clutter conditions, achieving significantly improved bistatic range-Doppler estimation, target localization, and velocity estimation performance. The remainder of the paper is organized as follows. Sec. II presents signal and clutter models. Sec. III describes position and velocity estimation. Sec. IV discusses conventional clutter suppression. Sec. V presents DL-based clutter suppression. Section V evaluates performance. Sec. VII concludes the paper. II. M ULTISTATIC ISAC S IGNAL AND C LUTTER M ODEL We consider an illuminator (base station) that transmits 5G NR PRS and M = 3 distributed receivers that cooperatively process the reflected signals for 2D sensing, as shown in

(c)

is the sampling interval, τm and τm,q denote the target and clutter propagation delays, respectively, αm and βm,q are their reflection coefficients, and wm [n] ∼ CN (0, σn2 ) is additive white Gaussian noise. The reflection coefficients account for bistatic propagation loss, antenna gains, carrier wavelength, and target and clutter RCS. With a narrowband approximation, the echo amplitude for a scatterer at location x is modeled as √ √ PT GT GR λ σ , (6) γ(x) = (4π)3/2 Rtx Rrx

ISAC Sensor

Clutter

Fig. 1: Multistatic ISAC in clutter environment.

Fig. 1. 5G numerology 0 is assumed as in [11] with subcarrier spacing ∆f = 15 kHz, FFT size N = 1024, carrier frequency fc = 2.5 GHz, and sampling rate fs = 15.36 MHz. The PRS spans KRB = 76 resource blocks and follows the 3GPP TS 38.211 generation procedure using QPSK Gold sequences [15]. Resource allocation uses a comb structure with spacing Kcomb =2 and offset koffset = 0, excluding the DC subcarrier. Each NR slot contains 14 OFDM symbols, including LPRS = 2 PRS symbols. Consecutive PRS transmissions across multiple slots produce K=128 coherent slow-time sensing observations. For receiver m ∈ {1, 2, 3}, the bistatic propagation length is Bm (p) = ∥p − t∥ + ∥p − rm ∥,

(1)

where t is the transmitter location, rm is the receiver m position, p and v are target position and velocity (all in R2 ). The associated bistatic radial velocity at receiver m is p − rm p−t ⊤ + , (2) vm (p, v) = v ∥p − t∥ ∥p − rm ∥ where (·)⊤ denotes transpose. This represents the projection of the target velocity vector v onto the bistatic observation direction associated with receiver m. The resulting Doppler shift at receiver m is fc fd,m = vm (p, v), (3) c where c is the speed of light. The sensing region contains Nc randomly distributed clutter scatterers. The q-th clutter scatterer (c) with RCS σq is located at cq ∈ R2 . The clutter RCS values are modeled as independent log-uniform random variables whose scale is controlled by the clutter-strength parameter αc , with larger αc corresponding to stronger clutter returns. Each scatterer contributes a receiver-dependent Doppler component (c) 2 fm,q ∼ N (0, σf,c ),

(4)

(c)

where fm,q is a zero-mean Gaussian random variable with 2 variance σf,c and σf,c controls the clutter Doppler spread. Larger σf,c yields more dynamic clutter, while larger Nc yields denser clutter. The received signal at receiver m is the superposition of the target echo, clutter reflections, and noise: ym [n] = αm s[n − τm ]ej2πfd,m nTs +

Nc X

+ wm [n],

(5)

where PT is transmit power, GT and GR are transmit and receive antenna gains, λ = c/fc is the carrier wavelength, Rtx and Rrx are the transmitter-to-scatterer and scatterer-to-receiver distances, and σ is the RCS. Before CAF processing, the direct path is estimated and removed from the received signal. III. TARGET L OCALIZATION AND V ELOCITY E STIMATION Each sensing receiver correlates the received samples with the known PRS reference sequence. Let 0 ≤ k ≤ K −1 denote the slow-time PRS observation index, tk denote the midpoint time of the k-th PRS symbol, and sk [n], 0 ≤ n ≤ N −1, denote the associated cyclic-prefix (CP)-free PRS reference symbol. For delay bin 0 ≤ d ≤ D − 1, the correlation output is Cm (k, d) =

N −1 X

∗ ym [ncp k + d + n]sk [n]

at receiver m, where ncp k is the starting sample index of the CP-free PRS region for observation k and (·)∗ is the complex conjugate. The resulting matrix Cm ∈ CK×D represents the slow-time delay-response for clutter mitigation and Doppler analysis. Following clutter suppression (see Secs. IV-V), coherent integration is applied over the slow-time dimension using a Doppler grid with Nf = 401 frequency bins over [−400, 400] Hz. Using a Hann window W [k], the CAF is CAFm (d, f ) =

K−1 X

W [k]Cm (k, d)e−j2πf tk

(8)

k=0

for receiver m at Doppler frequency f . The magnitude response |CAFm (d, f )| produces the bistatic range-Doppler map. Target detection uses two-dimensional cell-averaging constant false alarm rate (CA-CFAR) on the CAF magnitude response. Threepoint parabolic interpolation around the detected cell is applied in delay and Doppler dimensions to refine bistatic range and radial velocity estimates B̂m and v̂m . Target localization jointly fits the bistatic range measurements from all receivers. Each bistatic measurement defines an ellipse with foci located at the transmitter and the corresponding receiver position. When ϵm (p) = ∥p − t∥ + ∥p − rm ∥ − B̂m

(9)

is the bistatic residual associated with receiver m, the estimated target position is determined from solving

q=1

where s[n] denotes the transmitted PRS waveform, Ts = 1/fs

(7)

n=0

p̂ = arg min p

M X m=1

(ϵm (p)) .

(10)

Because the bistatic constraints are nonlinear, a trust-region reflective least-squares solver with multiple randomized initialization points is employed. The solution yielding the minimum aggregate residual error is selected as the final estimate. After estimating target position, the bistatic radial velocity relations become linear in the Cartesian velocity vector. Define um =

p̂ − rm p̂ − t + , ∥p̂ − t∥ ∥p̂ − rm ∥

(11)

as the bistatic observation direction for receiver m. Stacking these vectors forms the observation matrix U ∈ RM ×2 , while the estimated radial velocities are collected into v̂r = [v̂1 , . . . , v̂M ]⊤ .

(12)

The Cartesian velocity is estimated using ridge-regularized least squares (with a small regularization parameter ε > 0) −1 ⊤ v̂ = U⊤ U + εI U v̂r , (13) where I is the identity matrix. The target speed is given by ∥v̂∥ and the heading angle is computed using atan2(v̂y , v̂x ). IV. BASELINES FOR C LUTTER S UPPRESSION We next compare a broad set of conventional clutter suppression methods operating on the slow-time delay correlation matrix Cm . They suppress clutter using techniques like temporal differencing, subspace projection, Doppler-domain filtering, adaptive prediction, and smooth-component estimation. After clutter suppression, the resulting correlation matrices are used for CAF computation and bistatic range-Doppler estimation. 1) No Clutter Suppression (NCS): NCS skips clutter mitigation and computes the CAF from the correlation matrix Cm . 2) Current-Reference Subtraction (CUR) : CUR serves as a coherent reference-based clutter suppression benchmark. Let Ccur m denote the current clutter-contaminated correlation matrix and Cref m denote the clutter-only reference correlation matrix for receiver m. A global clutter phase offset ϕm is estimated from dominant clutter components in the reference and current observations. The current observation is phase-aligned prior to subtraction, yielding the clutter-suppressed correlation matrix ref CCUR = e−jϕm Ccur m m − Cm .

(14)

The clutter-suppressed matrix CCUR is used for CAF compum tation according to (8) to estimate bistatic range and Doppler. 3) Space-Projection Adaptive Difference (SP-AD): SP-AD suppresses slowly varying clutter by differencing adjacent slow-time snapshots prior to CAF processing. 4) Rank-One Slow-Time Projection (SP-R1): SP-R1 models clutter as a dominant rank-one slow-time component and removes the strongest singular mode before Doppler processing. 5) DPSS Band-Stop Filtering (DPSS): Discrete prolate spheroidal sequences (DPSS) model clutter as a low-Doppler process and aim to suppress clutter by projecting the signal onto the orthogonal complement of a DPSS clutter subspace. 6) Clutter Map Temporal Doppler Filtering (CMTD): CMTD suppresses near-zero-Doppler clutter using FFT-based slow-time notch filtering centered near zero frequency.

7) Total Variation High-Pass Filtering (TVHP): TVHP models clutter as a temporally smooth component and removes smooth clutter responses using total-variation regularization. 8) Rolling-Median High-Pass Filtering (RMHP): RMHP suppresses clutter using a sliding-window median filter along the slow-time dimension to remove slowly varying clutter responses while preserving localized target peaks. 9) Wavelet High-Pass Filtering (WHP): WHP applies Haar wavelet decomposition along slow time and reconstructs the signal using high-frequency detail components. 10) Adaptive Least Mean Squares (LMS) Residual Filtering (ALMS): ALMS applies adaptive LMS prediction along slow time and uses the prediction residual for clutter suppression. 11) Polynomial High-Pass Filtering (POLYHP): POLYHP models clutter as a slowly varying low-order polynomial along slow time and removes the estimated polynomial component. V. D EEP L EARNING -BASED C LUTTER S UPPRESSION In addition to signal processing-based clutter suppression methods, we introduce two supervised DL methods that operate directly on the slow-time delay correlation matrix C ∈ CK×D obtained from matched filtering. The objective is to learn a nonlinear mapping that suppresses clutter while preserving target responses required for accurate bistatic range-Doppler estimation. Training data are generated from the multistatic ISAC signal and clutter model described previously. Each training realization contains random receiver geometries, target positions, target velocities, clutter densities, clutter RCSs, clutter Doppler spreads, and noise realizations. The training input is the clutter-contaminated matched-filter matrix Craw = Ctarget + Cclutter + N,

(15)

where Ctarget and Cclutter denote the target-only and clutteronly correlation matrices, respectively, and N denotes the additive noise. The desired output is Ctarget . Next, we present two DL-based clutter suppression (DLCS) methods that estimate Ctarget from Craw . The estimated targetonly correlation matrix is subsequently used in (8) in place of Cm for range-Doppler estimation. A. Residual U-Net-based DLCS The first DL method, DLCS-1, employs a supervised residual U-Net architecture that directly estimates the target-only matched-filter response from clutter-corrupted observations. 1) Input Representation: The input is given by (Q) X = [C(I) raw , Craw ], (I)

(Q)

(16)

where Craw and Craw are the in-phase and quadrature components of the complex-valued correlation matrix, respectively, and are normalized using the root-mean-square (RMS) magnitude of the input tensor before being passed to the network. 2) Network Architecture: The network follows an encoderdecoder U-Net structure with skip connections. The encoder consists of repeated convolutional blocks that progressively extract hierarchical clutter and target features while reducing

spatial resolution through max-pooling operations. Each convolutional block contains 3 × 3 convolutional layers, batch normalization, and rectified linear unit (ReLU) activations. The feature tensor F(ℓ) at layer ℓ is encoded as F(ℓ+1) = ReLU BN Conv(F(ℓ) ) , (17) where Conv(·) denotes convolution and BN(·) denotes batch normalization. The decoder progressively reconstructs a targetenhanced residual map using bilinear upsampling and skipconnected encoder features. Skip connections preserve fine spatial structures corresponding to localized target peaks in the range-Doppler domain. The final output layer is a 1 × 1 convolution that produces a two-channel estimate (I)

(Q)

Ŷ = [Ĉtarget , Ĉtarget ], (I)

(18)

(Q)

where Ĉtarget and Ĉtarget denote the predicted in-phase and quadrature components of the target-only correlation matrix. 3) Training Objective: The network is trained using a residual reconstruction loss composed of mean-squared error and sparsity regularization: 2

LDLCS = Ĉtarget − Ctarget

+ λ1 Ĉtarget

, 1

(19)

where λ1 > 0 controls sparsity regularization. The ℓ2 term encourages accurate target reconstruction. The ℓ1 term promotes sparse target responses and suppresses residual clutter artifacts. 4) Optimization: The network parameters are optimized using the AdamW optimizer with a learning rate ηDLCS = 10−3 and a weight-decay regularization coefficient λwd = 10−5 . Mini-batch training is used with random clutter and target realizations. Gradient clipping is applied to stabilize optimization: ∥∇θ L∥2 ≤ γmax ,

(20)

where θ denotes the trainable network parameters and the clipping threshold γmax is 5. After training, the output Ĉtarget is processed in the CAF and range-Doppler estimation stages.

lutions, elongated (1 × 5) convolutions, batch normalization, and ReLU nonlinearities. The (5 × 1) kernels learn slowtime Doppler structures and clutter trajectories, whereas the (1 × 5) kernels learn range-domain correlations and delayspread clutter structures. Let F denote an intermediate feature map. An anisotropic block computes Fst = Conv5×1 (F) and Frg = Conv1×5 (F), followed by nonlinear feature fusion. This directional processing better distinguishes between target signatures and clutter structures that exhibit different geometric behavior across delay and slow-time dimensions. 3) Attention-Gated Skip Connections: The decoder incorporates attention-gated skip connections to suppress clutterdominated encoder activations before feature fusion. Let S denote the encoder skip feature map and G denote the decoder gating feature map. The attention mask is computed as A = σ (ψ (ϕ(S) + θ(G))), where ϕ(·), θ(·), and ψ(·) are learned convolutional transformations and σ(·) is the sigmoid activation function. The refined skip feature becomes Satt = A ⊙ S, where ⊙ denotes elementwise multiplication. This mechanism suppresses clutter-dominated feature activations while preserving target-dominant spatial structures by emphasizing features consistent with the decoder’s target-relevant context. 4) Decoder Structure: The decoder reconstructs the targetenhanced residual map through bilinear upsampling and feature concatenation with the attention-refined skip features. Each decoder stage employs anisotropic convolutional blocks similar to those used in the encoder. Anisotropic convolutions employ direction-dependent convolution kernels with unequal spatial support across dimensions. Kernels of size 5 × 1 and 1 × 5 are used to separately model slow-time Doppler structures and delay-domain correlations, respectively. The final output layer generates a two-channel estimate of the target residual: (I)

(Q)

Ŷ = [Ĉtarget , Ĉtarget ].

(22)

5) Loss Function: The network is trained using a combination of reconstruction and sparsity losses: 2

B. Attention-Gated Anisotropic DLCS

LAttn = Ĉtarget − Ctarget

The second DL method, DLCS-2, extends the residual UNet architecture through anisotropic convolutional processing and attention-gated skip connections designed specifically for clutter-dominated range-Doppler maps. 1) Multi-Channel Input Representation: Instead of using only real and imaginary channels, the network exploits four complementary input representations: (Q) X = [C(I) raw , Craw , |Craw |, log(1 + |Craw |)].

(21)

The magnitude channel explicitly captures target energy distributions, while the logarithmic channel compresses dynamic range and improves representation of weak target responses in strong clutter environments. 2) Anisotropic Convolutional Encoder: The encoder uses anisotropic convolutional blocks that separately model slowtime and delay-domain structures. Each anisotropic block contains standard 3 × 3 convolutions, elongated (5 × 1) convo-

+ λs Ĉtarget

+ λTV TV(Ĉtarget ),

(23)

where λs controls sparsity regularization and λTV controls total-variation regularization. The total-variation term promotes spatial smoothness while preserving localized target peaks. 6) Optimization and Training: Training is performed using the AdamW optimizer with mini-batch gradient descent. Gradient clipping and weight decay are applied to improve optimization stability and generalization performance. After training, the network output is directly processed using CAFbased range-Doppler estimation and bistatic localization. VI. P ERFORMANCE E VALUATION For the numerical results, we assume a transmit power of 40 dBm, transmit and receive antenna gains of 10 dBi, a target RCS of 4 m2 , and a noise variance of 10−3 . All DL models were trained using NVIDIA RTX PRO 6000 Blackwell GPUs.

90°

TABLE I: Per-receiver bistatic range and radial-velocity estimation results. Btrue 331.16 422.90 533.39 331.16 422.90 533.39 331.16 422.90 533.39

Best 293.11 292.80 312.32 331.75 429.64 312.53 331.85 429.51 527.51

Berr -38.05 -130.09 -221.07 0.58 6.74 -220.85 0.68 6.61 -5.88

vtrue -25.61 -13.86 -49.05 -25.61 -13.86 -49.05 -25.61 -13.86 -49.05

vest -2.17 5.82 6.85 -25.61 -13.88 7.04 -25.61 -13.86 -46.70

verr 23.44 19.69 55.91 -0.00 -0.01 56.10 -0.00 0.00 2.36

NCS RMHP DLCS-1

Speed Error (m/s) 22.184 9.509 0.445

Angle Error (deg) 114.065 156.948 0.045

45°

30 0°

180°

315°

90°

True velocity Estimated velocity

135°

225°

45°

30 0°

180°

225°

30 0°

315°

270°

(a) NCS

True velocity Estimated velocity

135°

315°

270°

(b) RMHP

Fig. 2: Velocity estimation results. 600

Transmitter Receivers

500

600

True target Estimated target

Transmitter Receivers

500

600

True target Estimated target

400

300

200

100

200 300 X (m)

(a) NCS

400

500

600

Transmitter Receivers

500

400

−100 −100

The DLCS-1 and DLCS-2 architectures contain 1,885,186 and 12,398,789 trainable parameters, respectively. Let Ntrain and Ntest denote the numbers of training and test samples, respectively, and let Nepoch denote the number of training epochs. For each sample, the transmitter and receiver geometry, target position and velocity, clutter scatterer locations, and additive noise realizations are randomly generated. Clutter effects are controlled by the values of Nc , αc , and σf,c . The training and test data are generated independently, with clutter density, strength, and Doppler spread randomized during training. We first consider a moderate training setting of Ntrain = 500 and Nepoch = 200 under a strong clutter environment with Nc = 40, αc = 1, and σf,c = 30 Hz. For a randomly selected test sample (Ntest = 1), Tables I and II compare DLCS-1 with NCS and RMHP. NCS produces severe bistatic-range and radialvelocity estimation errors, leading to large localization and velocity estimation errors, as shown in Figs. 2 and 3. RMHP, the strongest conventional baseline, improves the estimates by suppressing slowly varying clutter, but fails to accurately recover the target response for the other receivers, causing large localization and velocity-estimation errors. DLCS-1 accurately estimates bistatic range and radial velocity for all receivers, achieving a position error of 4.462 m and speed and angle estimation errors of only 0.445 m/s and 0.045◦ , respectively. The range-Doppler maps in Fig. 4 further illustrate these performance differences. For NCS and RMHP, clutter obscures the target response, causing large estimation errors. In contrast, DLCS-1 produces cleaner range-Doppler maps with localized target peaks closely aligned with the true target locations. As a result, DLCS-1 achieves highly accurate velocity estimation (Fig. 2) and target localization (Fig. 3). Next, we increase the number of test samples to Ntest = 2000 and statistically compare DLCS-1 with conventional clutter-suppression baselines under different clutter settings specified by Nc , αc , and σf,c . For each clutter case, the success rate is defined as the percentage of trials with position error below T = 10 m, and ηsucc is computed as the normalized success rate relative to the reference method CUR of Sec. IV-2. As shown in Table III, DLCS-1 consistently achieves a higher normalized success rate ηsucc than all non-reference

225°

Y (m)

Method

Position Error (m) 114.985 158.673 4.462

45°

180°

TABLE II: Localization and velocity estimation errors. RMS B Error (m) 149.715 127.570 5.122

90°

True velocity Estimated velocity

135°

Y (m)

RX 1 2 3 1 2 3 1 2 3

Y (m)

Method NCS NCS NCS RMHP RMHP RMHP DLCS-1 DLCS-1 DLCS-1

True target Estimated target

200 100 0

−100 −100

100

200 300 X (m)

400

500

600

−100 −100

(b) RMHP

100

200 300 X (m)

400

500

600

Fig. 3: Target localization results.

clutter-suppression methods across all clutter environments. Conventional methods degrade significantly under dense and dynamic clutter. RMHP is the strongest conventional baseline, yet remains substantially outperformed by DLCS-1. Table IV compares ηsucc of DLCS-1 and DLCS-2 for different clutter and training configurations. As Ntrain and Nepoch increase to 5000 and 500, respectively, DLCS-2 approaches the performance of the reference method while consistently outperforming DLCS-1 across all clutter settings. Fig. 5 shows ηsucc as a function of position error threshold T . Across all T and clutter settings, the DLCS methods, trained with Ntrain = 5000 and Nepoch = 500, consistently outperform NCS and RMHP. DLCS-2 achieves the best performance, followed closely by DLCS-1, demonstrating the effectiveness of DL-based clutter suppression in multistatic ISAC. VII. C ONCLUSION In this paper, we studied clutter suppression in multistatic ISAC using 5G NR PRS for bistatic range-Doppler estimation, target localization, and velocity estimation in dense and dynamic clutter environments. We showed that clutter can severely degrade sensing accuracy of multistatic ISAC and that conventional signal processing methods for clutter suppression provide only limited gains under challenging clutter conditions. To improve robustness of multistatic ISAC against clutter, we introduced DL-based clutter suppression methods to directly enhance target responses from clutter-corrupted observations. These methods substantially improve clutter suppression and reduce sensing errors relative to conventional approaches across diverse clutter conditions in multistatic ISAC. R EFERENCES [1] N. González-Prelcic, M. F. Keskin, O. Kaltiokallio, M. Valkama et al., “The integrated sensing and communication revolution for 6G: Vision, techniques, and applications,” Proceedings of the IEEE, 2024. [2] N. K. Nataraja, S. Sharma, K. Ali, F. Bai et al., “Integrated sensing and communication (ISAC) for vehicles: Bistatic radar with 5G-NR signals,” IEEE Transactions on Vehicular Technology, 2024.

TABLE III: Normalized success rate ηsucc for different clutter-suppression methods.

Receiver 3

true target estimated target

600

500

400

300

CMTD 1.72% 3.52% 3.26% 6.03% 4.09% 9.77%

true target estimated target

700

600

300

DPSS 2.62% 5.16% 5.62% 10.35% 29.84% 46.42%

−10

−20

−30

300 −40

200

100 0

200

100

−40

−20

−40

−20

Radial velocity (m/s)

Receiver 1

Receiver 2

−40

−20

−60

true target estimated target

700

−10

600

500

400

300

−20

−30

0.5 0.4 0.3 0.2 0.1

0.9

200

−20

−40

Radial velocity (m/s)

−20

Receiver 3

600

true target estimated target

700

−10

600

500

400

300

200

100

−20

NCS RMHP

−20

Radial velocity (m/s)

−40

−20

Radial velocity (m/s)

−30

−50

−40

−20

TABLE IV: Normalized success rate comparison of DLCS-1 and DLCS-2. αc

σf,c

30 Hz

10 Hz

0.1 5

0.5 0.4 0.3 0.2 0.1 5

0.9

DLCS-1 DLCS-2

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0

NCS RMHP

0.8

Position Error Threshold T (m)

Ntrain 500 5000 500 5000 500 5000 500 5000 500 5000 500 5000

Nepoch 200 500 200 500 200 500 200 500 200 500 200 500

0.9

NCS RMHP

1.0

DLCS-1 DLCS-2

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 5

Position Error Threshold T (m)

0.9

NCS RMHP

DLCS-1 DLCS-2

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.0

Position Error Threshold T (m)

(e) Nc = 40, αc = 1, σf,c = 10 Hz. (f) Nc = 20, αc = 1, σf,c = 10 Hz. Fig. 5: Normalized success rate ηsucc versus position error threshold T (m).

−60

0.2

Radial velocity (m/s)

Fig. 4: Bistatic range-Doppler maps.

0.3

1.0

DLCS-1 DLCS-2

0.6

0.0

−40

0.4

Position Error Threshold T (m)

0.7

1.0

CAF (dB)

true target estimated target

700

500

0.5

Radial velocity (m/s)

Receiver 2

600

0.6

−60

(b) RMHP Receiver 1 true target estimated target

DLCS-1 DLCS-2

0.7

0.0

NCS RMHP

0.8

Position Error Threshold T (m)

−50

100

−40

Radial velocity (m/s)

700

0.9

DLCS-1 62.77% 74.39% 72.21% 80.93% 90.27% 94.22%

200

100

−40

0.8

−40 200 100

Bistatic range (m)

0.6

0.0

Normalized Success Rate ηsucc

Bistatic range (m)

600

500

0.7

POLYHP 2.62% 5.65% 5.62% 10.51% 2.62% 43.08%

1.0

0.8

1.0

Receiver 3

NCS RMHP

ALMS 10.23% 17.92% 19.89% 29.99% 46.85% 62.87%

(a) Nc = 40, αc = 2, σf,c = 30 Hz. (b) Nc = 40, αc = 1, σf,c = 30 Hz.

Radial velocity (m/s)

true target estimated target

700

DLCS-1 DLCS-2

0.9

WHP 10.56% 18.74% 22.00% 32.19% 54.62% 68.08%

Position Error Threshold T (m)

CAF (dB)

true target estimated target

600

1.0

−50

(a) NCS 700

RMHP 25.70% 35.11% 40.18% 48.17% 85.61% 89.98%

0.0

100

Radial velocity (m/s)

TVHP 18.41% 29.46% 28.61% 38.30% 59.93% 65.55%

Normalized Success Rate ηsucc

Receiver 2

SP-R1 1.72% 3.03% 3.83% 6.68% 5.40% 11.64%

Normalized Success Rate ηsucc

Bistatic range (m)

SP-AD 24.30% 35.27% 36.92% 46.70% 65.90% 73.94%

Normalized Success Rate ηsucc

700

Receiver 1 700

NCS 1.72% 2.86% 2.85% 5.46% 2.29% 4.97%

Normalized Success Rate ηsucc

true target estimated target

σf,c 30 Hz 30 Hz 30 Hz 30 Hz 10 Hz 10 Hz

Normalized Success Rate ηsucc

αc 2 1 2 1 1 1

CAF (dB)

Nc 40 40 20 20 40 20

DLCS-1 62.77% 71.19% 74.39% 80.44% 72.21% 76.69% 80.93% 87.61% 90.27% 95.59% 94.22% 98.70%

DLCS-2 66.61% 72.42% 78.48% 84.29% 74.16% 78.24% 88.18% 92.42% 90.52% 95.67% 98.05% 99.41%

[3] M. Wypich, R. Maksymiuk, and T. P. Zielinski, “5G-based passive radar utilizing channel response estimated via reference signals,” IEEE Transactions on Radar Systems, 2025. [4] S. Blandino, N. Golmie, A. Sahoo, T. Nguyen et al., “Detecting airborne objects with 5G NR radars,” in IEEE Military Communications Conference (MILCOM), 2025. [5] Z. Wei, Y. Wang, L. Ma, S. Yang et al., “5G PRS-based sensing: A sensing reference signal approach for joint sensing and communication system,” IEEE Transactions on Vehicular Technology, 2022. [6] S. Lu, F. Liu, and L. Hanzo, “The degrees-of-freedom in monostatic

ISAC channels: NLoS exploitation vs. reduction,” IEEE Transactions on Vehicular Technology, 2022. [7] Y. E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Joint sensing and semantic communications with multi-task deep learning,” IEEE Communications Magazine, 2024. [8] Y. E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Will 6G be semantic communications? Opportunities and challenges from task oriented and secure communications to integrated sensing,” IEEE Network, 2024. [9] U. Demir, K. Davaslioglu, Y. E. Sagduyu, T. Erpek et al., “I-SCOUT: Integrated sensing and communications to uncover moving targets in NextG networks,” in IEEE MILCOM, 2024. [10] M. F. Keskin, M. M. Mojahedian, J. O. Lacruz, C. Marcus et al., “Fundamental trade-offs in monostatic ISAC: A holistic investigation towards 6G,” IEEE Transactions on Wireless Communications, 2025. [11] Y. E. Sagduyu, K. Davaslioglu, T. Erpek, S. Kompella et al., “MULTISCOUT: Multistatic integrated sensing and communications in 5G and beyond for moving target detection, positioning, and tracking,” in IEEE MILCOM, 2025. [12] Y. Yang, Z. Li, J. Zhao, Q. Wu et al., “Cooperative multi-static ISAC networks: A unified design framework for active and passive sensing,” IEEE Transactions on Wireless Communications, 2025. [13] H. Luo, Y. Wang, D. Luo, J. Zhao et al., “Integrated sensing and communications in clutter environment,” IEEE Transactions on Wireless Communications, 2024. [14] R. Liu, P. Li, M. Li, and A. L. Swindlehurst, “Clutter-aware integrated sensing and communication: Models, methods, and future directions,” Proceedings of the IEEE, 2026. [15] 3GPP, “NR: physical channels and modulation,” 3rd Generation Partnership Project, Technical Specification (TS) 38.211, 2024, ver. 18.2.0.

Multistatic ISAC with Deep Learning-Based Clutter Suppression for 5G and Beyond

Analyze & explore related research in Conceptio

Related documents