The absence of publicly available, large-scale, high-quality datasets for Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) has significantly hindered the application of rapidly advancing deep learning techniques, which hold huge potential to unlock new capabilities in this field. This is primarily because collecting large volumes of diverse target samples from SAR images is prohibitively expensive, largely due to privacy concerns, the characteristics of microwave radar imagery perception, and the need for specialized expertise in data annotation. Throughout the history of SAR ATR research, there have been only a number of small datasets, mainly including targets like ships, airplanes, buildings, etc. There is only one vehicle dataset MSTAR collected in the 1990 s, which has been a valuable source for SAR ATR. To fill this gap, this paper introduces a large-scale, new dataset named ATRNet-STAR with 40 different vehicle categories collected under various realistic imaging conditions and scenes. It marks a substantial advancement in dataset scale and diversity, comprising over 190,000 well-annotated samples-$10\times$ larger than its predecessor, the famous MSTAR. Building such a large dataset is a challenging task, and the data collection scheme will be detailed. Secondly, we illustrate the value of ATRNet-STAR via extensively evaluating the performance of 15 representative methods with 7 different experimental settings on challenging classification and detection benchmarks derived from the dataset. Finally, based on our extensive experiments, we identify valuable insights for SAR ATR and discuss potential future research directions in this field. We hope that the scale, diversity, and benchmark of ATRNet-STAR can significantly facilitate the advancement of SAR ATR.

The growing density of communication and radar devices renders multiple-input multiple-output (MIMO) radar systems susceptible to severe detection performance degradation caused by even slight steering vector mismatches. To ensure robust detection under such mismatches, this paper presents a robust waveform design approach based on a steering vector uncertainty-constrained Max-Min signal-to-interference-plus noise ratio (SINR) formulation. Compared to conventional Max SINR designs, the proposed method optimizes waveforms that maintain high SINR even in the presence of steering vector errors. The problem incorporates constraints for spectral compatibility and peak-to-average power ratio (PAPR). To solve this non convex problem, we develop an efficient iterative algorithm that employs successive convex approximation (SCA) to transform the original problem into a sequence of convex subproblems, which are then solved in parallel via the alternating direction method of multipliers (ADMM). Numerical simulations show a reduction in convergence time of up to 30% compared to existing techniques.

In order to address the low radar-cross-section target detection challenge under high sea states and remove the dependence of existing methods on idealized prior knowledge, a two-stage closed-loop adaptive pulse-agile radar framework is proposed in this letter. First, the clutter and target parameters are estimated with the expectation-maximization scheme, followed by the processing of principal component analysis-based clutter suppression and generalized likelihood ratio test detection. Then, the multi-pulse polyphase-coded frequency modulate waveform is designed to minimize the complementary integrated sidelobe level in strong clutter regions with the iterative method. Numerical experiments are conducted to demonstrate the advantage of the proposed method in clutter suppression. To further confirm the robustness of our method, we also present the Monte Carlo experiments in challenging sea condition.

Information-theoretic principles provide a rigorous foundation for adaptive radar waveform design in contested and dynamically varying environments. This paper addresses the joint optimization of constant modulus waveforms to enhance both target detection and parameter estimation concurrently. A unified design framework is developed by maximizing a mutual information upper bound (MIUB), which intrinsically reconciles the tradeoff between detection sensitivity and estimation accuracy without heuristic weighting. Realistic, potentially non-Gaussian statistics of target and clutter returns are modeled using Gaussian mixture distributions (GMDs), enabling tractable closed-form approximations of the MIUB’s Kullback–Leibler divergence and mutual information components. To tackle the ensuing non-convex optimization, a tailored metaheuristic phase-coded dream optimization algorithm (PC-DOA) is proposed, incorporating hybrid initialization and adaptive exploration–exploitation mechanisms for efficient phase-space search. Numerical results substantiate the proposed approach’s superiority in achieving favorable detection estimation trade-offs over existing benchmarks.

Instead of maximizing the detection probability, an information-theoretic waveform design method using Cauchy-Schwarz Divergence (CSD) between the probability density functions of the two hypotheses (viz., the target is present/absent) is proposed for improving the target detection performance. A Minorization-Maximization (MM) approach is introduced to deal with the non-convex waveform optimization problem under the energy constraint. Numerical results demonstrate the feasibility and effectiveness of the proposed algorithm.

Clutter suppression remains a fundamental challenge in multi-channel SAR moving target detection. Among various approaches, Space-Time Adaptive Processing (STAP) has been widely employed due to its excellent clutter suppression capabilities. However, STAP intrinsically assumes complex Gaussian clutter, which often significantly deviates from actual heterogeneous clutter environments encountered in practical scenarios. Such deviations can markedly impair the clutter suppression performance of STAP. To overcome this issue, this paper proposes a Kullback-Leibler Divergence (KLD)-based Hermitian Positive Definite (HPD) manifold Dimensionality Reduction (DR) technique, aimed at enhancing the feature discrimination of moving target on HPD manifold. Initially, the airborne multi-channel SAR image domain data are mapped into the HPD manifold to form a feature manifold. Next, an optimal dimensionality reduction mapping is constructed by formulating an optimization problem on the HPD manifold that maximizes the KLD, and an analytical solution to this problem is derived. Based on this solution, an optimal reduction matrix is designed, projecting the data into a low-dimensional HPD manifold. Finally, on the reduced HPD manifold, a Matrix Information Geometry (MIG) detector based on KLD is employed to achieve enhanced detection of moving targets. The proposed multi-channel SAR moving target detection framework incorporating KLD-based dimensionality reduction does not rely on specific parametric statistical model of the clutter background, only utilizes geometric properties derived from second-order statistic. This effectively addresses the performance degradation caused by statistical distribution model mismatch in complex environments. Moreover, deriving an analytical solution to the optimization problem avoids the high computational costs associated with Riemannian conjugate gradient algorithms, thus reducing complexity substantially and facilitating practical implementation in SAR systems. Experimental results, including simulations and measurements with real clutter backgrounds, demonstrate the superior effectiveness of the proposed approach.

Heterogeneous clutter suppression is an important challenge for multi-channel synthetic aperture radar (SAR) moving target detection. The knowledge-aided space time adaptive processing (STAP) method is widely used due to its superior clutter suppression performance. In this process, it is necessary to use independent and identically distributed (IID) training data to estimate the clutter covariance matrix (CCM) of the cell under test (CUT). However, heterogeneous environments such as clutter power fluctuations and moving target contamination in training samples will significantly reduce the estimation accuracy of CCM, which is not conducive to subsequent clutter suppression processing. To solve this problem, this paper introduces the covariance matrix estimation method based on the geometric mean of Hermitian positive-definite (HPD) manifold into the multi-channel SAR moving target detection. Firstly, the airborne multi-channel SAR echo is characterized on the HPD manifold, which better reveals the nonlinear geometric structure of the multi-channel SAR echo data than in Euclidean space. Based on the echo characterization on the HPD manifold, the CCM estimation problem is transformed into the geometric mean estimation problem on the HPD manifold. At the same time, the generalized inner product (GIP) criterion is used in training sample selection to reduce the negative impact of contaminated samples. Finally, the CCM estimation based on geometric mean is applied to the knowledge-aided SAR-STAP, which effectively improves the clutter suppression performance. Given the large volume of multi-channel SAR echo data, this paper comprehensively evaluates algorithm performance and computational complexity, and selects the geometric metric criterion for multi-channel SAR moving target detection. Based on this criterion, a multi-channel SAR moving target detection framework driven by HPD manifold is proposed. The detection performance of the proposed framework is verified on simulated and measured clutter backgrounds.

Despite the remarkable progress in synthetic aperture radar automatic target recognition (SAR ATR), recent efforts have concentrated on detecting and classifying a specific category, e.g., vehicles, ships, airplanes, or buildings. One of the fundamental limitations of the top-performing SAR ATR methods is that the learning paradigm is supervised, task-specific, limited-category, closed-world learning, which depends on massive amounts of accurately annotated samples that are expensively labeled by expert SAR analysts and have limited generalization capability and scalability. In this work, we make the first attempt towards building a foundation model for SAR ATR, termed SARATR-X. SARATR-X learns generalizable representations via self-supervised learning (SSL) and provides a cornerstone for label-efficient model adaptation to generic SAR target detection and classification tasks. Specifically, SARATR-X is trained on 0.18 M unlabelled SAR target samples, which are curated by combining contemporary benchmarks and constitute the largest publicly available dataset till now. Considering the characteristics of SAR images, a backbone tailored for SAR ATR is carefully designed, and a two-step SSL method endowed with multi-scale gradient features was applied to ensure the feature diversity and model scalability of SARATR-X. The capabilities of SARATR-X are evaluated on classification under few-shot and robustness settings and detection across various categories and scenes, and impressive performance is achieved, often competitive with or even superior to prior fully supervised, semi-supervised, or self-supervised algorithms. Our SARATR-X and the curated dataset are released at https://github.com/waterdisappear/SARATR-X to foster research into foundation models for SAR image interpretation.

In recent years, deep learning has been widely used in SAR ATR and achieved excellent performance on the MSTAR dataset. However, due to constrained imaging conditions, MSTAR has data biases such as background correlation,i.e., background clutter properties have a spurious correlation with target classes. Deep learning can overfit clutter to reduce training errors. Therefore, the degree of overfitting for clutter reflects the non-causality of deep learning in SAR ATR. Existing methods only qualitatively analyze this phenomenon. In this paper, we quantify the contributions of different regions to target recognition based on the Shapley value. The Shapley value of clutter measures the degree of overfitting. Moreover, we explain how data bias and model bias contribute to non-causality. Concisely, data bias leads to comparable signal-to-clutter ratios and clutter textures in training and test sets. And various model structures have different degrees of overfitting for these biases. The experimental results of various models under standard operating conditions on the MSTAR dataset support our conclusions. Our code is available at https://github.com/waterdisappear/Data-Bias-in-MSTAR.

Video synthetic aperture radar (ViSAR) can produce continuous images with a high frame rate and contain the moving target’s shadow, which provides numerous advantages for detecting moving targets. In this letter, a novel moving target detection method based on convolutional neural network (CNN) is proposed. The proposed network has a 3D convolutional encoding path, a 2D convolutional decoding path, and a bridge path to efficiently capture the target’s shadow information and summarize the spatiotemporal features from raw continuous images to high-level semantics. Furthermore, based on coordinate attention (CA), a temporal tri-coordinate attention (TTCA) module is proposed to obtain key spatiotemporal features in ViSAR data. The validity of the proposed method has been confirmed through experiments with actual ViSAR data and shows superior performance in the suppression of false alarms and missing alarms.

Micro‐range (m‐R) signatures which are induced by micro‐motion dynamics can be observed from range profiles, providing that the range resolution of radar is high enough. For real scenarios, micro‐motion is often mixed with macro‐motion (translation). To extract the micro‐motion signatures, it is required to remove the macro‐motion component. The widely employed range alignment technique fails for rigid‐body targets with micro‐motion, since the relative distances between different scattering centres on a rigid‐body target are varying and it is unable to obtain a stable reference range profile. Thus, the extracted m‐R signatures will be accompanied with residual macro‐motion, which may lead to the degradation. However, this issue is often ignored in the research of m‐R signatures extraction. In this work, by modelling the motions of scattering centres as the superimposition of a polynomial signal (represents macro‐motion) and a sinusoidal signal (represents micro‐motion), a micro‐motion period estimation method based on high‐order difference sequence is proposed. The property that the difference operation can decrease the order of polynomial signals while preserve sinusoidal signals with the same frequency enables the proposed method to extract m‐R signatures in the presence of macro‐motion. The effectiveness of the proposed method is validated by synthetic and measured radar data.

Target detection, tracking and classification are three essential and closely coupled subjects for most surveillance systems. In the finite set statistics (FISST) framework, this paper presents a Bayesian and recursive solution to joint detection, tracking and classification (JDTC) of a manoeuvring target in a cluttered environment, which is inspired by previous work on joint target tracking and classification in the classical Bayesian filter framework. The derived JDTC algorithm exploits the dependence of target state on target class by using class‐dependent dynamical model sets. The relative merits of this JDTC algorithm are demonstrated via a two‐dimensional example using a sequential Monte Carlo implementation. It is shown that handling those three closely coupled subjects jointly can achieve comparable detection and tracking performance to that of the exact filter in the FISST framework with a prior known class. The classification results are consistent with the previous work.

This study considers the problem of jointly detecting whether a target is present in a scene and estimating its state, if it is there. This joint detection and estimation problem can be solved using a special case of the multi-target Bayes filter (referred to as the joint target detection and tracking (JoTT) filter). However, if the model used by the JoTT filter does not match the actual dynamics, the filter will tend to miss-detection directly or diverge such that the actual errors fall outside the range predicted by the filter's estimate of the error covariance. A similar difficulty arises, if the target behaviour can switch between different modes of operation, since the filter may then be accurate for only one particular mode. This study proposes a novel joint detection and tracking filter, which is the multiple model extension of the JoTT filter to accommodate the possible target manoeuvring behaviour. In addition, a sequential Monte Carlo implementation (for generic models) and a Gaussian mixture implementation (for linear Gaussian models) are proposed. The simulation results are presented to show the effectiveness of the proposed filter over the original JoTT filter.