The masked autoencoder (MAE) is a self-supervised learning model that reconstructs masked patches of an input image. Its pretrained encoder significantly enhances downstream tasks, even surpassing supervised methods. This is particularly important for synthetic aperture radar (SAR) target recognition because labeled SAR images are scarce. Leveraging large amounts of unlabeled SAR data for network pretraining mitigates this challenge. Therefore, by combining the scale-fusion spatial-channel attention (SF-SCA) and multihead self-attention (MHSA) modules, the multiscale perceptive convolutional transformer (MSP-CT) is proposed as the backbone of the multiscale perceptive MAE (MSP-MAE) for SAR target recognition tasks. The MHSA module exhibits locality and sparsity in shallow features, thereby introducing redundancy in global dependence modeling. In contrast, the SF-SCA module facilitates the interaction of localized and sparse patches through a sliding window. Therefore, the MSP-CT architecture employs alternating stacks of SF-SCA and MHSA blocks to balance the ability of the model to extract local and global features. The experimental results show that MSP-MAE pretraining can effectively improve the performance of the downstream task and can greatly alleviate the problem of insufficient labeled samples.

Micro-motions of low-slow-small (LSS) targets induce time-varying micro-Doppler modulations in radar echoes and provide informative signatures for target analysis. Reliable time-frequency representation (TFR) remains challenging under noisy and multi-component conditions, especially when weak periodic components, local ridge crossings, and asymmetric Doppler branches coexist. This paper proposes an adaptive dynamic complex convolution network, termed ADCC-TFNet, for radar micro-Doppler TFR generation. The network jointly processes the in-phase and quadrature (I/Q) components of complex radar echoes, generates input-conditioned complex analysis kernels, and applies direction-aware refinement to improve ridge continuity and local concentration. On simulated signals with analytically generated reference TFRs, ADCC-TFNet reduces RMSE and R & eacute;nyi entropy by 88.83% and 20.02%, respectively, relative to the average of all compared non-ADCC-TFNet entries, and increases the structural similarity index measure (SSIM) by 0.32 in absolute terms. Qualitative evaluation on measured FMCW and S-band radar echoes further shows compact and clearly structured micro-Doppler patterns under the evaluated measurement conditions.

Radar jamming recognition is a critical element in the field of Electronic Counter-Measures (ECM). However, in realistic electromagnetic environments, the extreme scarcity of labeled samples from non-cooperative targets causes traditional deep learning models to suffer from severe feature collapse and prototype drift. To address the challenge of insufficient robustness under few-shot conditions, this paper proposes a Text-Anchored Generative Prototypical Network (TAGPNet). First, we design a prior-guided representation alignment mechanism based on a Semantic-Attribute Dual-Anchor backbone. By fusing semantic anchors encoding expert knowledge with data-driven attribute anchors under the regulation of a learnable confidence gating, this mechanism leverages semantic priors to constrain the feature manifold, effectively mitigating feature bias in extreme low-shot regimes. Second, a confidence-driven multi-view collaborative decision-making framework is proposed to establish a cognitive closed-loop of “Feature Projection – Geometric Metric – Prior Rectification”. This framework utilizes an adaptive branch weighting loss to dynamically integrate decision advantages from different views, endowing the model with self-calibration capabilities in uncertain environments. Furthermore, to fundamentally overcome the statistical instability caused by data scarcity, we construct a generative evolution paradigm incorporating Progressive Manifold Curriculum Learning and uncertainty calibration. This paradigm employs a stage-wise strategy to build a robust initial feature space, utilizes a feature diffusion model to achieve intra-class manifold densification, and reconstructs prototypes via uncertainty-aware inverse variance weighting, effectively mitigating prototype drift. Extensive experiments on a complex dataset containing 19 types of radar jamming demonstrate that TAGPNet achieves notable recognition accuracies of 53.7%, 70.2%, 74.9%, and 79.1% under 1-shot, 3-shot, 5-shot, and 10-shot settings, respectively, validating the superiority of the proposed method in data-limited scenarios.

In the above article [1], after publication, the authors identified several inadvertent figure-related errors. The corrections are as follows. The scope of this correction is limited to replacing two incorrectly placed qualitative result figures and correcting two interchanged figure captions. These changes do not affect the experimental data, numerical results, comparative analysis, or conclusions of the article.

Sea clutter significantly impacts the radar detection of maritime targets. Existing sea clutter suppression methods often face challenges in complex, dynamic marine environments, and their generalization capabilities may be limited. This article proposes a network architecture named triplet diffusion attention multiscale Res-KAN Net (TD-AMRKNet), based on a diffusion model and triplet attention. By introducing lightweight multiscale generalized spatial convolutions (multiscale-GSConvs) and several small model networks, TD-AMRKNet effectively reduces model parameters, making it a compact and efficient network. The AMRK module, designed with a gating mechanism, captures long-range dependencies in images. It also integrates multisource knowledge through cross-resolution image fusion, thereby enhancing semantic understanding and improving the representation of details and local features. TD-AMRKNet effectively suppresses sea clutter across different radar data types, including time-frequency spectrograms from staring radar and PPI images from scanning radar. Experimental results show that the model contains only 3.46-M parameters and requires approximately 0.0305 s for overall average (OA) processing. It achieves competitive performance on six real-world sea clutter datasets, with clutter suppression effectiveness evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), P-S, and clutter suppression ratio (CSR) metrics.

Under the sea-air background, the measurements of low-altitude, slow-speed, and small-sized (LSS) targets by digital ubiquitous radar are affected by factors such as target motion mode, sea clutter interference, and detection distance, which may produce the time-varying complex non-Gaussian noise and result in false tracking or track interruption. The adaptive q-Renyi kernel-based extended Kalman filter (AQEKF) is proposed to address this issue. First, the extended Kalman filter (EKF) for radar target tracking based on q-Renyi kernel is developed to process noise of non-Gaussian distribution. Then, projection statistics (PS)-based parameter mapping is established to adjust the weights of abnormal measurement vectors. Furthermore, the residual-based kernel width function is designed to enhance the algorithm's stability in complex noise environments. In the digital ubiquitous radar tracking experiment under sea-air background, the root-mean-square error (RMSE) of the AQEKF is reduced by 4.79%similar to 40.31% compared with the traditional EKF.

In radar maritime target detection tasks, most radars need to cover a wide observation area by scanning. The beam can not stay in one direction for a long time to obtain multiple accumulated pulses. Therefore, the detection capability of weak targets is limited. In addition, affected by the complex characteristics of strong sea clutter, the detection method based on statistical theory and model features are insufficient to distinguish between target and clutter signals. By transmitting omnidirectional radar signals, digital array ubiquitous radar can realize simultaneous continuous observation in a wide area. Long-term pulse accumulation can be achieved, improving the detection performance of weak targets. In this paper, we propose a Spatial-Temporal Graph Fourier Transform (ST-GFT) target detection method, which extract Eigenvalue-Doppler (E-D) features from high-dimensional signals of digital array radar. The proposed method is not affected by the direction of target echo. Through separating target and sea clutter signals, target features are enhanced, and target detection performance is improved. Comparing with the conventional Digital Beam Forming and Moving Target Detection (DBF-MTD) method, the proposed method increases the target Signal to Clutter Ratio (SCR) by 21.59dB in the Quadcopter detection experiment in sea clutter background.

Accurate identification of radar modulation patterns is crucial for advanced electronic countermeasures and intelligent sensing systems. However, many existing deep learning architectures do not fully exploit the synergy between fine-grained intra-pulse characteristics and global spectral information. This shortcoming thereby limits recognition performance. This work introduces GLF-Net (Global-Local Fusion Network), an attention-driven framework for radar modulation recognition. First, an enhanced Cross-Channel Global Attention (GCGA) module is incorporated to model long-range dependencies in radar echo sequences. Next, Large-Kernel Attention (LKA) is instantiated as a localized branch to extract discriminative intra-pulse features. These components are integrated into a multi-scale feature-fusion scheme, which significantly enhances representational capacity. The scheme also facilitates effective classification of complex modulation types. Extensive experiments demonstrate consistent gains over representative baselines in recognition accuracy. These results confirm the effectiveness of the introduced attention mechanisms and the overall architecture.

The recognition accuracy in radar signal modulation recognition (RSMR) is severely impacted by the lack of training data. Conventional deep learning approaches typically depend on extensive labeled datasets, which are scarce in real-world scenarios. To mitigate this limitation, a convolutional-transformer hybrid network based on self-supervised contrastive learning (CTNet-SSCL) is proposed for RSMR. In self-supervised contrastive pre-training, an amplitude distortion data augmentation technique is proposed, which enables the model to effectively utilize unlabeled data, allowing it to learn meaningful feature representations. Subsequently, the pre-trained multi-scale perceptual transformer (MSPFormer) encoder, combined with a randomly initialized classifier, is fine-tuned using labeled samples. The encoder combines multi-scale feature fusion and time-frequency attention mechanisms to further enhance the robustness and recognition accuracy of the model in complex environments. The excellent performance of the proposed method is verified in experiments on a dataset with 10 different waveforms. The recognition accuracy of the proposed method reaches 99.98% at a signal-to-noise ratio (SNR) of 2 dB.

Maritime radar systems play a critical role in ocean surveillance and navigation; however, the presence of sea clutter significantly degrades measurement performance. Traditional sea clutter suppression methods often struggle with the complexity and nonstationary characteristics of sea clutter, especially under harsh maritime conditions. Therefore, a framework based on multipath decoder network (MPDNet) is proposed to achieve efficient sea clutter suppression. MPDNet integrates a multipath feature extraction and fusion strategy. It incorporates innovative intermediate modules, such as Mamba2Block and k-nearest neighbors (KNNs) Layer, as well as decoder modules, including Pinwheel-shaped convolution and TiedBlockConv2d. Furthermore, a multidecoder architecture is designed, together with a multiobjective loss joint optimization strategy to balance the contributions of each path and improve clutter suppression performance. This architecture allows the model to capture both global and local features while maintaining computational efficiency. Experimental results on six real-world radar datasets show that the lightweight MPDNet outperforms existing networks in terms of peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the PSNR-SSIM (P-S) metric. Ablation studies further validate the effectiveness of using three encoder branches over other configurations. In the target detection stage, You Only Look Once version 8 achieves an inference time reduced to approximately one-fifth of that under the original clutter background after processing with the proposed method, along with improved detection accuracy. These results demonstrate good generalization performance and further verify the effectiveness of MPDNet in sea clutter suppression.

For solving the detection challenge of distributed targets in compound-Gaussian clutter under steering vector uncertainty, we propose two efficient detectors based on the generalized likelihood ratio test (GLRT) and Wald test under the assumptions that the clutter texture is deterministic but unknown and the distributed target steering vector is confined to a specified subspace while its coordinates remain undetermined. In the parameter estimation phase, we use maximum likelihood estimation to estimate target amplitude and texture, followed by target coordinate estimation via the projected gradient descent method. In the detector design phase, we adopt a two-step strategy. First, we derive detectors under known clutter covariance matrix (CM). Subsequently, the CM is substituted with its approximate maximum likelihood estimate. Simulation and real-data experiments confirm that the developed detectors exhibit superior detection performance compared to existing methods, with the GLRT-based detector achieving better results than the Wald-based detector. Moreover, both detectors' detection performance improves with increased training data, reduced target distributed dimension, and signal subspace dimension. In addition, the proposed detectors exhibit constant false alarm rate properties for texture and CM structure.

Highlights What are the main findings? What are the implications of the main findings?Highlights What are the main findings? What are the implications of the main findings?Abstract Low-slow-small (LSS) target recognition using multi-dimensional radar remains challenging due to weak signatures, similar kinematics, and overlapping short-term Doppler patterns. Digital-array radar provides continuous, complementary Doppler-spectrum and kinematic measurements; however, their heterogeneity in dimension, distribution, and physical meaning often makes direct fusion under-exploit discriminative complementarity and inadequately model temporal track evolution. To address this, we propose a Doppler-Kinematic Spatio-Temporal Graph Learning framework named Dual-Stream Spatio-Temporal Cross-Attention Graph Convolutional Network (DS-STCAGCN) for LSS target recognition using multi-dimensional radar observations. The method separately encodes Doppler-spectrum and kinematic features to preserve their modality-specific characteristics, fuses them through bidirectional cross-attention, captures long-range temporal dependencies via self-attention, and aggregates local frame-to-frame correlations through graph convolution on a time-ordered observation graph. On the public L-band digital-array dataset LSS-DAUR-1.0, DS-STCAGCN achieves 99.73% mean accuracy and maintains 98.64% at 5 dB signal-to-noise ratio (SNR). On the passive-radar dataset LSS-PR-1.0, it reaches 99.86% mean accuracy, demonstrating strong cross-modal generalization. This work provides an effective spatio-temporal modelling framework for multi-dimensional radar sensing and robust LSS target recognition.

In synthetic aperture radar (SAR) target recognition, the interpretation of SAR images and the accurate recognition of targets are significantly affected by speckle noise. Traditional recognition methods often fail to meet the high accuracy requirements. Therefore, this letter proposes a dual-ended fusion network (DEFNet) for SAR target recognition. The network consists of three main components: the local refinement feature extraction (LRFE) network, the large-scale feature extraction (LSFE) network, and the adaptive feature interaction fusion (AFIF) module. It employs a parallel structure, utilizing the LRFE and LSFE branch modules to extract local and large-scale features, respectively. These two features are then adaptively fused through the AFIF module to further enhance feature representation. Experimental results on the moving and stationary target acquisition and recognition dataset indicate that DEFNet demonstrates significant performance improvement compared to traditional methods, showcasing its effectiveness and adaptability in SAR target recognition tasks.

Sea clutter significantly impacts radar maritime target detection. Existing methods still show limited clutter suppression performance, and their generalization ability re quires further improvement. To address this issue, a sea clutter suppression method based on the pyramid multi-scale residual attention generative adversarial network (PMS-RA GAN) is proposed. In the generator, an integrated residual attention mechanism (IRAM) is designed to highlight the target and effectively suppress the background clutter. Furthermore, a multi-scale pyramid adaptive feature extraction (MPA) module based on sea clutter characteristics is designed to generate high quality radar time-frequency images. Additionally, a composite loss function is proposed to balance between image quality and detail performance. Experimental results on real world datasets demonstrate that PMS-RA GAN outperforms the existing methods in sea clutter suppression, significantly improving target detection performance. The complete code for the proposed PMS-RA GAN model is publicly available at https://github.com/202400358077/PMS-RA-GAN and archived on Zenodo at https://zenodo.org/records/19805662.

In this paper, we investigate the adaptive detection of range-distributed targets in compound-Gaussian clutter, where the texture component follows a Weighted Generalized Inverse Gaussian (WGIG) distribution. We propose adaptive detectors for WGIG-distributed clutter based on two-step Rao, Wald, Durbin, and Gradient tests. The unknown covariance matrix is estimated using Approximate Maximum Likelihood (AML) and the Normalized Sample Covariance Matrix (NSCM). To address the analytical intractability of Maximum A Posteriori (MAP) estimation for the texture component, we adopt an alternative approach: The MAP estimator of the reciprocal expectation of the texture component, which is used in designing adaptive detectors based on the Rao, Wald, and Durbin tests. For the Gradient test-based detector, the test statistic is derived directly from the posterior probability density function. Our theoretical analysis confirms the consistency of the detectors derived from the Rao, Durbin, and Gradient tests. Extensive evaluations on both simulated and real data yield three key findings: (1) The proposed AML-based detectors maintain the constant false alarm rate property; (2) Under matched signal conditions, the detectors based on the Rao and Wald tests achieve the best performance on both the IPIX radar dataset and the Journal of Radar’s maritime surveillance dataset—specifically, they outperform the two-step generalized likelihood ratio test-based detector, requiring 0.1～0.5 dB and 0.7～0.8 dB lower Signal-to-Clutter Ratio (SCR) to achieve the same detection probability, respectively; and (3) Under mismatched signal conditions, the Rao test-based detector with AML estimation exhibits superior robustness, while the Wald test-based detector demonstrates the strongest suppression capability against mismatched signals.

In the context of neural network-based radar feature extraction and detection methods, single-feature detection approaches exhibit limited capability in distinguishing targets from background in complex environments such as sea clutter. To address this, a Multi-Feature Extraction Network and Graph Fusion Detection Network (MFEn-GFDn) method is proposed, leveraging feature complementarity and enhanced information utilization. MFEn extracts features from various time-frequency maps of radar signals to construct Multi-Feature Graph Data (MFG) for multi-feature graphical representation. Subsequently, GFDn performs fusion detection on MFG containing multi-feature information. By expanding the feature dimension, detection performance is further improved. Experimental results on dataset composed of real measured IPIX data demonstrate that MFEn-GFDn detection probability is approximately 8% higher than that of the Dual-Channel Convolutional Neural Network (DCCNN). Additionally, MFEn-GFDn enhances detection performance by expanding the feature dimension, particularly in environments lacking corresponding training samples.

Micro-Doppler (m-D) signals are susceptible to interference from a large number of Doppler signals and ambient noise, and the single use of m-D signatures (MDSs) for the classification of small, slow, and low-speed targets poses certain limitations. In this article, a dynamic multifeature data fusion neural network (DMFFNNet) classification method is proposed. First, K-band frequency-modulated continuous-wave (FMCW) radar is used to collect echo data from five types of rotor drones and bionic bird. After preprocessing the data, 2-D range-period graphic and 2-D time-frequency (TF) spectrograms are obtained. We investigate the construction of new data representations in the range-periodic domain, designing networks to extract dynamic time-varying features of the data. To be able to obtain accurate localized features, a local feature extraction module is proposed to extract local features from the range-period graph, while a global feature extraction module is used to extract global features from the TF spectrograms. To be able to extract dynamic information about the data, a 3-D network is used to capture dynamic change feature in the 3-D range-period data. Finally, a feature fusion module is designed to integrate the extracted features, and to be able to better extract the features of the target, an attention mechanism is added to the fusion network to extract the temporal and spatial features in the spectrogram and fuse them to further improve the overall performance of the model. Experimental results show that compared with single-channel CNN classification methods, incorporating dynamic feature data enables the network to achieve better classification accuracy.

This paper proposes a deep learning-based time-frequency representation approach that employs the enhanced time-frequency micro-Doppler network (ETF-MDNet) model to improve the characterization of micro-Doppler features for radar targets, particularly “low, slow, and small” ones. The ETF-MDNet model consists of four key components: the micro-Doppler target signal input module, the basis function selection module, the feature aggregation module, and the energy concentration module. A notable characteristic of this method is its utilization of the inherent adaptive learning capabilities of deep learning, which are combined with an attention mechanism to enhance the aggregation of time-frequency energy. This integration optimizes the method's capacity to represent micro-motion features across both channel and spatial dimensions. Consequently, this approach effectively captures the micro-motion information of the target while suppressing extraneous noise. In comparison to traditional short time Fourier transform, generalized warblet and reassigned spectrogram analysis methods, the proposed method achieves an average enhancement of 31.5% in time-frequency energy concentration, higher time-frequency energy aggregation, and the ability to reveal micro-motion feature details not captured by traditional methods.

Conventional sparse uniform arrays (SUAs) is composed of multiple identical and rigorously collinear uniform linear arrays. By adjusting the baseline length between the subarrays, the array aperture can be arbitrarily large, thus substantially improving the accuracy of the direction-of-arrival (DOA) estimation. However, in practical applications, it is challenging to meet the strict collinearity requirement due to geographical constraints. In this letter, to address this problem, we propose the non-collinear sparse uniform array (NCSUA) model to mitigate the influence of the non-ideal terrain and enhance the practicality of the SUA. A novel estimation algorithm is then proposed to resolve the angle ambiguity in NCSUA and effectively achieve high-accuracy DOA estimation. Compared with the conventional SUA, numerical simulation results demonstrate the superiority of NCSUA employing the new de-ambiguity algorithm in DOA estimation performance and practical applications.