Deep Learning Clutter Suppression: Multistatic ISAC Achieves High-Precision Localization and Velocity Estimation

The Opportunity and Clutter Challenge of Multistatic ISAC

Integrated sensing and communication (ISAC) systems that use 5G and future signals as signals of opportunity make it possible for wireless sensing and communication to share infrastructure and spectrum resources. Multistatic ISAC, through spatially dispersed receivers, can significantly expand sensing coverage—an advantage that monostatic systems find hard to match.

But expanded coverage brings new troubles. In multistatic scenarios, dense and dynamic clutter severely degrades bistatic range and Doppler estimates, which in turn leads to errors in target localization and velocity estimation. Simply put, the sensing range is extended, but the environment is full of reflectors, true and false target signals are mixed together, and traditional methods begin to struggle.

Bistatic range and Doppler estimation are core intermediate quantities in multistatic (bistatic/multistatic) radar and ISAC. Bistatic range refers to the sum of the path lengths from the target to the transmitter and to the receiver, obtained by multiplying the delay of the echo relative to the direct signal by the speed of light; Doppler frequency reflects the rate of change of this path length over time and is related to the target's radial velocity component. A single transmit-receive pair gives only one elliptical iso-range contour and one Doppler curve, and multiple receivers must be jointly used to solve for the target's three-dimensional position and velocity vector. Precisely because final localization and velocity estimation depend entirely on these intermediate estimates, any unsuppressed clutter will first cause false peaks or spectral broadening in the range-Doppler domain, and then be amplified through geometric solving into target position offset and velocity error. This is why the article emphasizes that clutter causes errors to "propagate stage by stage."

Clutter Modeling: Making the Problem Quantifiable

This study does not avoid the complexity of clutter but chooses to model it directly. The authors use randomly distributed scatterers to characterize distributed clutter and give these scatterers different Doppler spreads and radar cross sections (RCS) to capture dense, dynamic clutter effects.

The value of this modeling approach is that it allows different suppression methods to be compared fairly under unified and realistic conditions. The closer the clutter is to the dense and dynamic characteristics of real scenarios, the more meaningful the performance gaps between methods become.

Radar cross section (RCS) is the equivalent area that measures a scatterer's ability to intercept and reradiate incident electromagnetic waves, in square meters, often expressed in dBsm. It is not the geometric projected area of the scatterer but depends on shape, material, polarization, and incidence angle. For example, a flat plate can reach tens of dBsm in the specular direction, while a small drone is often only on the order of −20 dBsm. Doppler spread describes the broadening of the echo from the same scatterer in the frequency domain: stationary clutter is theoretically a zero-Doppler spectral line, but vegetation swaying in the wind, undulating sea surface, or moving cloud and rain particles spread its energy over a range of frequencies, and the spread width corresponds to the dispersion of radial velocities within the scatterer. In modeling, superimposing randomly distributed scatterers with different RCS and Doppler spreads is precisely intended to reproduce the spatially non-uniform and temporally non-stationary characteristics of clutter in real environments, thereby avoiding method evaluation being misled by overly idealized Gaussian clutter assumptions.

Traditional Signal Processing Baselines: General Fragility Under Strong Clutter

The paper evaluates a wide range of traditional signal processing baseline methods for clutter suppression. The conclusion is not optimistic: under strong clutter conditions, the robustness of these conventional methods is generally limited.

This is not surprising. Traditional methods often rely on some assumption about the statistical characteristics of clutter, while clutter in multistatic scenarios is spatially dispersed and temporally dynamic, making fixed assumptions easy to fail. Once clutter suppression is incomplete, residual clutter directly contaminates the two key intermediate quantities—bistatic range and Doppler estimation—and the error propagates stage by stage to the final position and velocity results.

Traditional clutter suppression methods can be roughly divided into several categories: moving target indication (MTI) uses the Doppler difference between clutter and target for high-pass filtering; constant false alarm rate (CFAR) detection adaptively sets a threshold based on local background estimation; subspace methods (such as eigenvalue decomposition and MUSIC) decompose the received covariance matrix into clutter and target subspaces and project out clutter. The common implicit premise of these methods is that clutter is temporally stationary and spatially statistically homogeneous, so a reliable covariance or background level can be estimated from a small number of samples. Multistatic scenarios precisely break this premise: the clutter seen by each receiver comes from different viewing angles and different scatterers, with mutually different statistical characteristics that change rapidly with platform or environmental motion, leading to insufficient covariance estimation samples, threshold mismatch, and subspace leakage, so residual clutter disguises itself as weak targets. This explains why the paper observes that traditional baselines generally have limited robustness under strong clutter.

Deep Learning Solution: Directly Enhancing Target Response with Residuals and Attention

To address the limitations of traditional methods, the authors introduce a deep learning-based clutter suppression method. The core idea is to use a deep neural network architecture enhanced by residual networks and attention to directly enhance the target response from observations contaminated by clutter.

The role of the attention mechanism here is noteworthy: it enables the network to learn to distinguish which components are more likely to be target echoes and which are more likely to be clutter, rather than treating all signals indiscriminately. The residual structure helps the network preserve useful information from the original observations during deep transformations. Together, the two give the network stronger robustness in the face of dense and dynamic clutter.

Residual networks (ResNet) add the input directly to the outputs of several layers through skip connections, so that the network only needs to learn a "residual" mapping relative to the input. This design alleviates the vanishing gradient problem in deep networks and also allows the network to preserve detailed information from the original observations while suppressing clutter, rather than regenerating everything from scratch. The attention mechanism assigns learnable weights to each channel or position of the feature map, essentially implementing soft feature selection: patterns similar to target echoes are amplified, patterns similar to clutter are suppressed, and the weights are data-driven rather than manually set. In clutter suppression tasks, the combination of the two usually adopts a structure similar to "observation minus predicted clutter" or "mask multiplied by observation," so that the network outputs a signal with enhanced target response and attenuated clutter response, which is then passed to subsequent range-Doppler estimation and localization solving.

Results: Comprehensive Superiority Across Clutter Conditions

The paper reports that these deep learning methods substantially outperform traditional methods in target position and velocity estimation under diverse clutter conditions. This improvement is of practical significance in multistatic ISAC scenarios—position and velocity estimation are precisely the core outputs of the sensing function.

From a methodological perspective, this work sends a clear signal: when traditional signal processing reaches its robustness ceiling in complex clutter environments, data-driven learning methods provide a considerable complementary path. Of course, the practical deployment of deep learning methods also depends on how well the training data matches the real clutter distribution, which is also a direction worthy of continued attention.

One-line Summary

Multistatic ISAC expands sensing coverage but suffers from dense dynamic clutter; a residual attention deep network substantially outperforms traditional methods in position and velocity estimation.

Paragraph Summary

The article focuses on the clutter suppression problem in multistatic integrated sensing and communication (ISAC). Multistatic systems significantly expand sensing coverage through dispersed receivers, but dense and dynamic clutter severely degrades bistatic range and Doppler estimation, leading to target localization and velocity errors. The study models distributed clutter using randomly distributed scatterers with different Doppler spreads and radar cross sections, allowing different methods to be compared fairly under unified and realistic conditions. Evaluation shows that traditional signal processing baselines generally have limited robustness under strong clutter because their statistical assumptions struggle to adapt to spatially dispersed and temporally dynamic clutter. The authors therefore introduce a deep neural network enhanced by residual networks and attention mechanisms to directly enhance the target response from contaminated observations. Results show that deep learning methods substantially outperform traditional methods in position and velocity estimation under diverse clutter conditions, providing a data-driven complementary path for sensing in complex clutter environments.