ABSTRACT This paper investigates the joint beamforming design for ISAC‐enabled UAV systems, aiming to optimise the trade‐off between communication sum rate (SR) and sensing Cramér–Rao lower bound (CRLB) under stringent size, weight and power (SWaP) constraints. Departing from traditional relaxation methods, we first address the nonconvexity and variable coupling by constructing a tractable quadratic surrogate objective, which transforms the intractable original objective into an unconstrained optimisation framework strictly defined on the complex sphere manifold (CSM). We then propose a Riemannian trust‐region algorithm on the complex sphere manifold (RTR‐CSM). The problem is reformulated as a tractable quadratic surrogate, with a finite‐difference approximated Hessian used to construct the trust‐region subproblem, which is solved via truncated conjugate gradient to avoid expensive exact second‐order derivations and matrix inversions. Simulations show that RTR‐CSM achieves near‐optimal performance and converges significantly faster than state‐of‐the‐art benchmarks in terms of iteration count (e.g., reducing total runtime by 5.2×, 3.9×, and 2.9× compared with SCA, FP‐SGDA, and IRO‐CSM, respectively), making it highly suitable for resource‐constrained UAV‐enabled ISAC beamforming.
ABSTRACT Frequency diverse arrays (FDAs) generate range‐angle‐dependent beampatterns, which provide significant advantages for modern radar and communication applications. Synthesising a flat‐top beampattern is highly desirable for enhancing spatial coverage and improving scanning efficiency. However, achieving this objective with phase‐only excitation, which maximises transmit power efficiency, remains a challenging non‐convex optimization problem. Existing continuous optimization algorithms are often limited by strict phase‐matching constraints, high computational complexity and poor convergence performance. In this paper, we propose a novel phase‐only flat‐top beampattern synthesis method for FDA systems. To increase the optimization degrees of freedom, the objective function is formulated to minimise mainlobe ripple by evaluating the squared difference between the synthesised and desired beam energies. This formulation eliminates the need for strict phase matching and provides additional degrees of freedom for optimization. Furthermore, the number of antenna elements is explicitly incorporated into the objective function to enhance sidelobe suppression and increase the optimization degrees of freedom. To efficiently solve the formulated problem, we employ an ADMM‐based framework, where the Adam optimization algorithm is incorporated into the inner loop to handle the nonconvex phase‐weight updates. Numerical simulations demonstrate that the proposed method achieves a stable flat‐top mainlobe with effective sidelobe suppression.
ABSTRACT Frequency modulated continuous wave (FMCW) inverse synthetic aperture radar (ISAR) imaging has the advantages of high integration, low cost, low power consumption and light weight. However, due to the limited transmission power of single channel millimetre wave radar, the system relies on narrow beam gain compensation to ensure the detection range; yet, this narrow beam significantly compresses the azimuth coverage of a single illumination. In the absence of prior constraints on motion parameters such as range, velocity and azimuth, it is difficult for narrow beams to achieve rapid target search and stare‐mode ISAR imaging. To address these issues, this paper proposes a motion parameter‐based ISAR imaging method for target regions, tailored for narrow beam phased array millimetre wave (mmWave) radar. The proposed method first extracts target echoes based on the region of interest (ROI); then performs coarse translational compensation using prior radial velocity information, followed by low‐order trend estimation and fine correction of residual range cell migration (RCM), which significantly enhances both compensation efficiency and robustness. Finally, a phased array mmWave radar platform is developed for experimental validation, encompassing target motion parameter accuracy testing, end‐to‐end verification of the ‘detect‐first, then stare‐mode ISAR imaging’ workflow for single targets, as well as comparative imaging evaluation under prior‐informed and prior‐free conditions in multi‐target scenarios. Quantitative experimental results show that the proposed hierarchical coarse‐to‐fine compensation architecture outperforms the traditional envelope correlation method, reducing image entropy by 12.1%, improving image contrast by 22.5% and optimising peak sidelobe ratio (PSLR) by 1.03 dB. These results validate the rationality and superiority of the proposed framework in practical mmWave ISAR imaging applications.
ABSTRACT As a typical and widely used passive jamming method, chaff has an extremely strong interference effect on radar. Current research has mainly focused on the identification of chaff jamming. However, for mixed echo signals of the target and chaff, accurately extracting target information from the mixed echo signal is a key problem that needs to be solved. To address this problem, an effective method based on variational mode decomposition (VMD) parameter adaptive optimization and a two‐stage filtering algorithm (TSFA) is proposed for chaff jamming suppression and target echo signal reconstruction. In order to achieve optimal decomposition of mixed echo signal, a VMD parameter adaptive optimization method based on the coati optimization algorithm is developed to determine the optimal mode number and penalty factor . Then, the optimised VMD is utilised to decompose the mixed echo signal. To suppress chaff jamming, a TSFA based on centre frequency and spectral entropy is designed to filter out intrinsic mode functions (IMFs) containing chaff jamming and select those associated with target information. The target echo signal is reconstructed by adding the selected IMFs. The experimental results show that the proposed method achieves higher signal to noise ratio, signal to noise ratio improvement and correlation coefficient, while yielding lower root mean square error and mean absolute error, indicating effective chaff jamming suppression and accurate target echo signal reconstruction.
ABSTRACT The Equatorial Ionospheric Anomaly (EIA) occurs within the ionosphere on both sides of the Earth's magnetic equator, typically at magnetic latitudes of approximately ± 10° to ± 20°. In this region, the spatiotemporal evolution of ionospheric total electron content (TEC) is highly complex, leading to anomalous enhancements in electron density. These ionospheric disturbances become particularly pronounced during geomagnetic storms, posing significant challenges for high‐precision prediction. To address this prediction challenge, we introduce a deep learning model termed an Encoder‐Decoder with Self‐Attention Convolutional Gated Recurrent Unit (ED‐SA‐ConvGRU), where multiple physical indices were incorporated, including solar activity indices (F10.7, Solar Radio Flux at 10.7 cm; SSN, Sunspot Number) and geomagnetic indices (Kp, K‐index; Dst, Disturbance Storm Time Index), with three distinct input combinations constructed. This study utilised GNSS data from 70 stations of the Australian Regional GNSS Network (ARGN) from 2023 to 2025. Test results indicate that, compared to the GRU, ConvGRU, and ED‐ConvGRU models, the proposed ED‐SA‐ConvGRU model reduced the Root Mean Square Error (RMSE) by 16.8%, 2.4%, and 6.8%, respectively. Moreover, the prediction accuracy of the ED‐SA‐ConvGRU model was further enhanced through the incorporation of physical indices. Specifically, Input Combination III (historical TEC + geomagnetic indices Kp and Dst + solar activity indices F10.7 and SSN) yielded the lowest RMSE. These findings indicate that the ED‐SA‐ConvGRU model has competitive performance in ionospheric TEC prediction over the low‐to mid‐latitude EIA region.
ABSTRACT Multiple‐input multiple‐output (MIMO)‐space‐time adaptive processing (STAP) can effectively reject clutter in airborne radar with superior spatial resolution. Nonetheless, MIMO‐STAP suffers from severe performance decrease due to the large system degrees‐of‐freedom according to Reed, Mallet and Brennan (RMB) rule. To this end, this paper proposes a MIMO‐STAP with scaling modification (SM), named SM‐MIMO‐STAP. The developed MIMO‐STAP first maps the radar data into another domain with a linear transformation, and the modification is implemented in the transformed domain. Then, a scaling parameter is designed to apply to the MIMO‐STAP weight vector, which is optimised to minimise the radar output power. Ultimately, full‐dimension MIMO‐STAP is effectively implemented with the modified weight vector. The proposed method is in closed form and easy to conduct. Numerical experiments show that the proposed MIMO‐STAP performs better than the conventional SM‐MIMO‐STAP under conditions of limited snapshots and exhibits robust performance against array gain and phase errors.
ABSTRACT Sparse vertical arrays can reduce the measurement distance required for cylindrical‐wave near‐field radar cross‐section (RCS) measurement, but their grating lobes may illuminate the target region directly or through conducting‐ground reflection. Therefore, the target distance should be screened before excitation synthesis when the array geometry and measurement site are already fixed. This paper presents a ground‐aware target‐range selection and field‐synthesis workflow for wideband cylindrical‐wave near‐field RCS measurement. Given the element spacing, array‐centre height, operating band, simulated element pattern and conducting‐ground model, direct and mirror‐ground grating‐lobe intervals are mapped into the target‐range domain. To avoid underestimating ground‐reflected risk, the present design uses the geometric exclusion interval conservatively while retaining element‐pattern and ground‐reflection weighting for more general site models. For the considered 21‐element X‐band geometry, the interval construction gives a recommended target distance of approximately 3.0 m. A multiconstraint wideband alternating projection method (MC‐WAPM) is then applied as the postselection shared‐excitation synthesis stage at the selected distance. Multilevel fast multipole method (MLFMM) simulations showed that MC‐WAPM produced a flatter and more phase‐consistent target‐zone field than Baseline APM and Baseline Taylor. The concrete‐ground response remained close to the no‐ground profile and did not change the PEC‐based range recommendation. In the direct sphere‐calibrated RCS comparison, MC‐WAPM yielded pooled five‐frequency MAEs of 1.60 dB for PEC ground and 1.49 dB for concrete ground.
ABSTRACT Significant non‐Gaussian characteristics are observed in echo signal spectra during severe convective weather. Current weather radar simulators, typically based on Gaussian spectrum assumptions, are difficult to validate non‐Gaussian spectra signal processing algorithms. This paper proposes a simulator capable of generating arbitrarily shaped echo spectra by reconstructing the power spectra from meteorological particle microphysical information. Within each radar resolution volume (RRV), sub‐grids are created, and WRF model outputs are interpolated onto these grids. The T‐matrix method computes microphysical properties—including aspect ratios and orientation distributions—for rain, snow, ice, and graupel particles. A particle terminal velocity model maps particle diameter to radial velocity, enabling calculation of each velocity bin's contribution to the Doppler spectrum per sub‐grid. Sub‐grid spectra are then integrated using antenna pattern weighting to form the overall Doppler spectrum and power spectral density for the RRV, from which IQ data are generated. The simulator is validated using stratospheric typhoon observations, demonstrating its effectiveness. It offers a physically consistent simulation framework for developing and testing non‐Gaussian spectra processors in severe convective weather and supports flexible radar scenario configuration.
ABSTRACT In modern electromagnetic reconnaissance, the ability to recognise intra‐pulse modulation patterns in complex environments has become increasingly critical. Traditional recognition algorithms are typically designed under ideal conditions and handcrafted features, and the presence of substantial noise interference in complex electromagnetic conditions can significantly degrade the performance of these algorithms. Although transformers have achieved great success in computer vision tasks, existing vision transformers lack the ability to capture long‐range dependencies. To remedy these flaws, this paper proposes an accurate automatic modulation recognition algorithm based on the vision transformer. First, we utilise traditional image processing techniques to provide the position of signals in the time frequency image and introduce a radial basis function to generate the positional embeddings. This procedure enhances the spatial awareness of the model. Second, dynamic scales of 2D patches are embedded as inputs to the encoders. Furthermore, inspired by dilated convolutions, vanilla attention is reformed in the cascade encoders to promote the training and inference procedures. Experiments and ablation studies exhibit the proposed algorithm that performs well across various signal‐to‐noise ratio conditions, demonstrating its robustness and accuracy in modulation recognition.
ABSTRACT Automated radar‐based tracking offers a scalable solution for avian monitoring, particularly in remote or data‐intensive settings where manual annotation is impractical. This study evaluates the performance of the GlobAl Nearest Neighbour targEt Tracker (GANNET), a low‐cost and customisable algorithm for detecting and tracking bird‐like targets in X‐band marine radar imagery. Using data from the Fall of Warness tidal test site in Orkney, Scotland, we compared manual and automated outputs across more than 55,000 filtered scan‐level detections and over 34,000 consolidated trajectories derived from good‐quality imagery. Trajectories were defined using a minimum track‐length threshold of six detections, with a mean of 22 detections per track. GANNET processed data over 10 times faster than manual annotation and showed higher sensitivity, particularly in medium‐ and low‐quality imagery where manual performance deteriorated. These results highlight GANNET's value for retrospective analysis of archived radar datasets and its potential application in environmental impact monitoring for birds in marine renewable energy contexts. Targeted improvements in clutter suppression, motion modelling and validation against a cooperative target would further strengthen its suitability for operational deployment. GANNET therefore represents a promising foundation for next‐generation radar ornithology and offshore biodiversity assessment.
ABSTRACT Multi‐object tracking (MOT) is a fundamental task in radar technology. Leveraging the visual attributes in radar video data provides an effective approach to enhance tracking robustness, particularly in long‐term scenarios where maintaining track continuity is challenging. However, effectively utilising the visual features of targets in radar videos also faces many limitations, primarily due to inter‐frame jitter in target positions and limited unstable nature of target appearances. To overcome these limitations, we propose a novel association framework for multi‐object tracking in radar video sequences. This framework achieves robust tracking through an implicit‐to‐explicit association strategy: An implicit association neural network (IANN) first establishes implicit associations between detections across consecutive frames end‐to‐end while performing detection, which are subsequently transformed into explicit associations by an explicit association module. Experimental results on real radar video data demonstrate the framework's effectiveness and superior performance. The proposed method stablishes a new benchmark for future research in multi‐object tracking for radar video sequences.
ABSTRACT Modern airborne intercept radars typically operate in the medium pulse repetition frequency (MPRF) mode. Measurements in this mode suffer from inherent ambiguity in both range and Doppler dimensions. Traditionally, methods based on frame‐to‐frame staggering rely on local decision mechanisms, idealised numerical rules and preset kinematic models. In complex scenarios, performance drops rapidly. To tackle this, a joint target discrimination and ambiguity resolution algorithm based on the physics‐guided spatio‐temporal transformer (PG‐STT) is proposed. Firstly, physically meaningful disambiguation hypotheses and high‐dimensional heuristic features are extracted from the ambiguous radar plots. Secondly, a spatial transformer is utilised to learn the spatial distribution of the hypotheses. Also, a temporal transformer is applied to learn global motion patterns and counteract measurement noise. Finally, target discrimination, range disambiguation and velocity disambiguation are accomplished by three linear prediction heads, respectively. Simulation results validate the effectiveness of the proposed method. The algorithm demonstrates strong capabilities in target discrimination and ambiguity resolution, even in complex scenarios with low SNR, strong clutter and manoeuvring targets. Compared with traditional methods and monolithic baseline methods, our algorithm shows significant advantages in overall performance.
ABSTRACT The performance of the adaptive beamformer is not only related to the array weight but also influenced by the array structure. Different sparse array configurations exhibit varying sensitivities to uncertainties in the signal direction of arrival (DOA). To address the degradation of array gain caused by desired signal DOA mismatch, this paper proposes an effective robust sparse array design method. Because DOA uncertainty leads to deviations in the steering vector (SV), we first introduce a SV uncertainty set into the sparse array design model and impose a reweighted ‐norm penalty to promote sparsity in the weight vector. The joint optimisation model is efficiently solved using the alternating direction method of multipliers (ADMM), yielding a sparse array structure with inherent resistance to DOA mismatch. Subsequently, the beamforming weight vector is optimised in two aspects. On the one hand, an improved integral reconstruction method is adopted to reconstruct the interference‐plus‐noise covariance matrix (INCM), avoiding the influence of desired signal components in the covariance matrix. On the other hand, based on subspace projection and power constraint principles, we have proposed a sparse array SV estimation algorithm to further enhance the beamforming performance. Simulation results demonstrate that the proposed approach can effectively solve the problem of serious degradation of the output signal‐to‐interference‐plus‐noise ratio (SINR) caused by DOA mismatch. It enhances beamforming robustness when prior information such as ideal INCM and accurate interference signal DOAs are unavailable.
ABSTRACT Distributed array radar (DAR) expands the aperture of the array radar by adding multiple synchronized auxiliary arrays with the main array, thereby enhancing the ability to counter mainlobe jamming. However, the long baseline of DAR causes signal sources to fall into the near‐field region. The coupling of range and angle parameters in the near‐field signal model poses challenges to anti‐jamming methods based on jamming parameter estimation and cancelation. To address this issue, this paper proposes a mainlobe jamming suppression method for DAR based on variational sparse Bayesian learning (SBL) jamming estimation and range‐angle two‐dimensional null broadening beamforming. To decouple the range and angle parameters in the near‐field steering vector model, a variational grid optimisation nonuniform sparse recovery dictionary is designed. Afterwards, iterative‐optimised variational SBL using prior information is performed to estimate the range‐angle parameters of jammers accurately. Given potential estimation errors, two‐dimensional null broadening beamforming based on steering vector perturbation is proposed to suppress jamming. Simulation and experimental results verify that the proposed method can effectively reduce the computational complexity of sparse recovery, obtain more accurate jamming parameters, and achieve higher output signal‐to‐interference‐plus‐noise ratio (SINR) for jamming suppression.
ABSTRACT This paper presents an in‐depth study of mixer metasurfaces comprising a frequency‐selective surface (FSS) mixer layer and a secondary transmission line local oscillator (LO) layer. Through both simulation and experimental measurement, we demonstrate how an upconverted signal is maximised by optimising metamaterial design and the separation of mixer components. Local oscillator signals at 2.1, 3, 4.2, and 4.45GHz were applied to the LO layer to drive the switching diodes on the mixer layer. The resulting metamaterial reflectivity, transmission, surface currents, scattering patterns, and 1D probe data were analysed, with results from a large‐scale fabricated model verifying the simulations. Finally, the structure is augmented with a metamaterial absorber to suppress generated harmonics, demonstrating a novel multi‐layered approach to incident signal manipulation and scattering control.
ABSTRACT In the application of group‐target overall tracking, it is necessary to simultaneously track the positions, measurement rates and extended shape states of the group targets. Among existing studies, the gamma Gaussian inverse Wishart (GGIW) distribution is the prevailing model for describing the state of group targets. The PMBM filter, with its rigorous mathematical framework, high‐precision tracking performance and ease of implementation, has been widely applied to group‐target overall tracking. However, in actual group target tracking, the PMBM filter requires the assumption of new group targets. Inaccurate prior information about the group targets may make it impossible to track them. To address this issue, this paper proposes a measurement‐driven adaptive birth PMBM filter. In this filtering framework, the group target parameter estimation based on single‐frame measurements is achieved through the proposed two‐layer nested expectation‐maximisation (EM) GGIW parameter estimation method. The inner EM algorithm realises the maximum posterior of GGIW parameters under the given group GGIW parameter prior, whereas the outer EM algorithm estimates the optimal GGIW parameter prior. During implementation, both serial and parallel methods are presented. The effectiveness of the proposed PMBM filter is verified through simulation experiments.
ABSTRACT In this paper, we address the problem of joint adaptive detection and range‐Doppler estimation for MIMO radar systems in heterogeneous environments. Specifically, we extend the Two‐Step Generalised Likelihood Ratio Test Detector designed for homogeneous Gaussian background to work in compound‐Gaussian scenarios. To this end, we replace the conventional sample covariance matrix of the secondary data with the estimators of clutter statistics based on the normalised sample covariance matrix and the recursive estimate. In this context, three detectors are proposed that can inherently provide estimates of target position in the delay‐Doppler domain. Numerical examples demonstrate that the proposed detectors significantly outperform the homogeneous counterparts in heterogeneous environments whilst maintaining satisfying performance in homogeneous environments. Remarkably, the analysis of the false alarm rate reveals that the proposed detectors are insensitive to changes in the scaling factor and/or covariance structure of the background.
ABSTRACT Passive localisation with narrowband external illuminators is attractive for low‐cost and flexible deployment, yet accurate positioning becomes difficult when time‐delay (range) information is unreliable or unavailable. To address this limitation and leverage signals from multi‐illuminator, this paper investigates range‐free 3‐D localisation for a monostatic passive radar. Doppler and DOA measurements (and, when available, Doppler‐rate) are incorporated into a maximum‐likelihood framework for joint position–velocity estimation. To mitigate the resulting high‐dimensional optimisation, we propose an iteration based velocity estimation that expresses the velocity estimate as a function of a candidate position, reducing the original 6‐D problem to a 3‐D position search. A Gauss–Newton (GN) guided dimension‐reduced Particle Swarm Optimisation (PSO) is then employed to accelerate convergence by steering elite particles along local GN directions while preserving global exploration. Simulation and measured‐data results demonstrate that the proposed method achieves close to the Cramér‐Rao Lower Bound (CRLB) accuracy with significantly improved efficiency, enabling stable localisation with fewer observations in both single‐illuminator and multi‐illuminator configurations.
ABSTRACT In the context of electronic warfare (EW), we focus on the detection of unknown radar pulses with a superheterodyne receiver. The main objective is to temporally locate the radar pulses by a detection step at each time sample. To optimise the thresholding step, we introduce and assess preprocessing methods of the received signals. We aim to demonstrate the pertinence of deep‐learning methods, such as spiking neural networks (SNNs) for this task, compared to recurrent neural networks (RNNs) and long‐short term memory (LSTM). To do so, we first perform a theoretical comparison between the SNN, RNN and LSTM, especially in terms of energy consumption, which is crucial for embedded systems. To apply the SNN, we propose appropriate pre and postprocessing steps. We also develop suitable performance metrics. Then, we compare the three methods on a simulated database of radar pulses. After simulating a real environment, we transform the database by including different clutters (rural and maritime). We compare the different networks in similar situations, considering the distribution and the level of clutter, or the characteristics of the pulses. Although RNN fails to learn the detection task, other deep learning approaches provide better results than the standard methods. LSTM offers slightly better performance than SNN. However, we also conclude that SNN has a lower power consumption than RNN and LSTM networks. Therefore, the size of the SNN architecture can be increased to improve the results.
ABSTRACT This paper presents an in‐depth study of a simulation framework that generates inverse synthetic aperture radar images and interferometric point clouds from computer‐aided design models of resident space objects in Keplerian orbits with varying radar and orbital parameters, modelling orbital radar backscatter via a physical optics approximation using multistatic radar configurations with two setups: one featuring a single transmitter and three receivers, and another with one combined transmitter/receiver and two receivers. We detail the creation of an extensive database of 42 resident space objects encompassing both inverse synthetic aperture radar images and interferometric point clouds, with a focus on training deep learning architectures to evaluate their performance across various signal‐to‐noise ratio conditions. The inverse synthetic aperture radar images dataset, comprising both complete and reduced training sets, is leveraged for deep learning model training, while a parallel methodology is applied to interferometric point cloud data, utilising computer‐aided design models for replicating inverse synthetic aperture radar interferometric processing. Our findings reveal that the proposed modifications to the point cloud based deep learning architectures, specifically PointNet, DGCNN, and Point Transformer, outperform their original versions and image‐based convolutional neural networks (ResNet, DenseNet, MobileNet, EfficientNet) at lower SNR levels (below 20 dB). Conversely, noise‐augmented training enhances the robustness of both modalities, with image‐based convolutional neural networks exhibiting superior adaptability to noise‐corrupted inverse synthetic aperture radar images. When trained with reduced data, point cloud‐based architectures perform best with noise‐free training, while noise‐augmented approaches deliver comparable results across both ISAR images and point clouds based classification. The results underscore the necessity for advanced simulation environments to produce high‐fidelity training data as described in this article, emphasising that computer‐aided design model augmentation alone is insufficient for effective training in radar point cloud applications.
