Yuanpeng Zhang

Space micromotion target recognition (SMMTR) is a key technology in space defense systems. Unlike monostatic radars, which provide only a single observation dimension, the networked radar systems offer richer information and significantly enhance target recognition performance. Traditional graph neural networks can only model binary relations, whereas hypergraph networks can effectively capture high-order complex interaction relations by allowing multiple nodes to share the same hyperedge, thereby extracting more accurate features in complex systems. In this article, a multiview adaptive hypergraph convolutional network (MVA-HGCN) architecture is proposed. The MVA-HGCN adaptively learns the hypergraph structure from two perspectives: time steps and radar entities. First, the multiscale temporal adaptive hypergraph convolutional network (MSTA-HGCN) treats time steps as hypergraph nodes at different scales. This module captures more comprehensive time-varying micromotion features across multiple temporal resolutions, reflecting the dynamic evolution of the target. Second, the radar entity adaptive hypergraph convolutional network (REA-HGCN) treats each radar as a hypergraph node. This module adaptively captures the potential complex spatial dependencies from the perspective of multisensor collaborative observation. Extensive experiments demonstrate that the proposed MVA-HGCN achieves reliable SMMTR performance even under harsh conditions of low signal-to-noise ratio (SNR) and low pulse repetition frequency (PRF). Specifically, it achieves a recognition accuracy of 93.89% at 5-dB SNR and 100-Hz PRF, representing a 4.78% improvement over the best existing method.

A networked radar system composed of radars deployed in different locations and working in different frequency bands can provide multi-band and multi-view target information, which can effectively improve the recognition accuracy of space micro-motion targets. The feature acquisition ability of each part of the networked radar systems is different. To fully tap the potential information of networked radar systems, a multi-domain feature-level fusion method based on networked radar systems for space micro-motion targets recognition is proposed in this paper. Firstly, the multi-domain features dataset of the networked radar systems is constructed. Secondly, the feature extraction and adaptive fusion subnetwork is designed. Two linear attention modules, cross-covariance atttention (XCA) module and linear angular attention (LA-Att) module, are introduced to extract fine-grained features in parallel and efficiently. The cross-attention fusion module (CAFM) is introduced for adaptive feature fusion. Then, a spatial-temporal feature extraction subnetwork is designed. The temporal features are extracted by bidirectional gated recurrent unit (BiGRU). The spatial dependencies between nodes are captured by the dual graph fusion network (DGFN) which integrates the adaptive graph and the predefined graph. Finally, the spatial-temporal feature vector is input into the prediction layer to obtain robust recognition results. Extensive experiments have proved the effectiveness and robustness of the proposed method.

The radar network system can provide multi-band and multi-view target information, which helps improve target recognition ability. A space micro-motion targets recognition method based on radar network systems (RNS) with a dual-channel bidirectional gated recurrent unit (BiGRU)-Transformer-graph fusion (DC-BiGT-GF) network is proposed in this paper. Firstly, a temporal feature extraction subnetwork based on BiGRU-Transformer is utilized to process the real and imaginary parts of complex-valued radar cross section (RCS) in parallel to capture the local and global temporal dependencies. Secondly, a spatial feature extraction subnetwork is designed to extract the potential spatial dependencies, which integrates a predefined graph and an adaptive graph. In the real part channel, the Euclidian distance between radars is used to construct the adjacency matrix to represent the predefined graph structure, and the adaptive adjacency matrix is designed to learn the potential graph structure from end to end. To represent the frequency domain features, the phase difference is applied in the imaginary part channel to build a predefined adjacency matrix. Meanwhile, the adaptive adjacency matrix is calculated using cosine similarity to obtain the geometric features. Finally, extensive experiments show that the DC-BiGT-GF network can reliably recognize the space micro-motion targets under low SNR and low radar pulse repetition frequency conditions. Recognition accuracy is greatly improved compared to the baseline methods.

Heterogeneous radar network systems can provide multiband and multiangle information about targets, enhancing the ability to recognize space targets. This article proposes a space target recognition method based on a bidirectional gated recurrent unit (BiGRU)-transformer and dual graph fusion (BiGT-DGF) network. Through a temporal information extraction subnetwork, the BiGRU and Transformer are used to dynamically model a radar cross-section (RCS) time series under multiple bands and angles, effectively exploiting both the local and global temporal dependencies. Through a spatial information extraction subnetwork, which integrates predefined graphs with self-adaptive graphs, the spatial dependencies between various radars are dynamically and adaptively captured. On this basis, the prediction output layer utilizes the spatiotemporal information extracted by the above two subnetworks to effectively recognize space targets. The experimental results show that the proposed method can reliably recognize space targets even under low signal-to-noise ratios (SNRs) and low pulse repetition frequencies.

This letter proposes a novel open set domain adaptation method for heterogeneous synthetic aperture radar (SAR) image classification. In this problem, the test data often differ from the distribution of training data and may contain unseen classes that are not present in the training data. Considering the inaccuracy of existing studies in classifying samples of unknown class in target domain, we propose a dynamic self-training network structure to distinguish known classes from unknown class by dynamically adjusting the discriminant threshold. Then in order to further improve the accuracy, a self-training domain-specific weak alignment module is designed, which abandons the previous idea of hard aligning source and target features, and achieves the weak alignment of target domain by the idea of prototype normalization and domain-specific knowledge extraction. Finally, experiments on two heterogenous SAR image benchmark datasets show that the proposed method improves the classification accuracy by more than 3% over the state-of-the-art methods.

The micro-Doppler effect of radar targets reflects the target motion attribute, which can provide important information for target recognition. Based on the micro-Doppler effect, the estimation of micro-motion parameters of the cone is of great significance in the field of anti-missile. However, the existing methods based on monostatic radar can only extract some micro-motion parameters, and the research on micro-Doppler properties is not in-depth enough. In the article, the micro-Doppler properties of the precession cone with fins are studied. Firstly, the motion model of the precession cone with fins is given. Subsequently, the frequency shift equation of the fin is studied, and a set of equations containing the relationship between the micro-motion parameters is constructed. The micro-motion parameters (precession angle, bottom radius, angle between conning vector and radar line of sight) are estimated by solving the set of equations. The effectiveness of the algorithm is verified by simulated data and measured data.

For space targets with the same shapes and composite relationships in terms of micromotion forms (STSSCRMFs), the accuracy of deep learning approaches with single-domain data is seriously degraded. To address this issue, this article proposes a hybrid neural network based on feature-level fusion that utilizes multidomain radar information to recognize STSSCRMFs. The proposed network simultaneously processes radar cross-section (RCS) time series, high-resolution range profile (HRRP) sequences, and time-frequency (TF) spectrograms through three branches, leveraging their complementary characteristics. Additionally, an attention mechanism is incorporated into the feature-level fusion module to selectively enhance the features of important domains, thereby achieving improved space target recognition. Consequently, the temporal-spatial features of multidomain data are adaptively fused, capturing discriminative and complementary representa-tions, which effectively enhances the accuracy of space target recognition. Experimental results demonstrate the superiority and effectiveness of the proposed method.

Existing deep learning-based space target recognition methods rely on abundantly labeled samples and are not capable of recognizing samples from unseen classes without training. In this article, based on generalized zero-shot learning (GZSL), we propose a space target recognition framework to simultaneously recognize space targets from both seen and unseen classes. First, we defined semantic attributes to describe the characteristics of different categories of space targets. Second, we constructed a dual-branch neural network, termed the global-local visual feature embedding network (GLVFENet), which jointly learns global and local visual features to obtain discriminative feature representations, thereby achieving GZSL for space targets with higher accuracy. Specifically, the global visual feature embedding subnetwork (GVFE-Subnet) calculates the compatibility score by measuring the cosine similarity between the projection of global visual features in the semantic space and various semantic vectors, thereby obtaining global visual embeddings. The local visual feature embedding subnetwork (LVFE-Subnet) introduces soft space attention, and an encoder discovers the semantic-guided local regions in the image to then generate local visual embeddings. Finally, the visual embeddings from both branches were combined and matched with semantics. The calibrated stacking method is introduced to achieve GZSL recognition of space targets. Extensive experiments were conducted on an electromagnetic simulation dataset of nine categories of space targets, and the effectiveness of our GLVFENet is confirmed.

Range instantaneous Doppler (RID) methods show superior advantages in imaging performance over range-Doppler (RD) algorithm for maneuvering targets. However, traditional RID methods are susceptible to noise and sparse aperture (SA), which makes the resolution of imaging results low and the quality degraded. To improve the imaging performance, we propose a RID high-resolution (HR) imaging framework for maneuvering targets, in which a Hybrid Transformer-based HR network (HTHRNet) is introduced into the RID method. HTHRNet is trained to map low-resolution (LR) time-frequency distribution (TFD) images to HR TFD images under different conditions. Our deep neural network fuses the convolutional neural network (CNN) and Transformer to make full use of their advantages. In addition, we modify the network structure to make HTHRNet training easier and more efficient. The experimental results of TFD data, simulated data, microwave chamber data and measured data show that the proposed imaging method has superior performance, which further indicates that the complex deep neural network can be well applied to ISAR HR imaging.

High-resolution range profile (HRRP) sequences have great potential for space target classification because they can provide both scattering information and micro-motion information. However, many factors cause an obtained HRRP sequence for a space target to be corrupted in real cases due to noise interference, limited radar resources, and the requirement of multi-target observations. Many space target classification methods cease to be effective when HRRP sequences are corrupted, so classifying space targets with corrupted HRRP sequences is still a challenging problem. To solve this problem, a novel space target classification method based on a temporal-spatial feature aggregation network (TSFA-Net) is proposed by using the corrupted HRRP sequences directly. First, a sequence-to-token module (S2T-module) is designed to extract low-level and fine-grained features from the raw inputs. Second, to effectively model the long-range dependencies among corrupted HRRP sequences and capture global representations without losing target local features, we propose a parallel and dual-branch block, i.e., a temporal-spatial feature aggregation block (TSFA-block), by combining a Transformer network and a convolutional neural network (CNN). Then, via progressively hierarchically stacking TSFA blocks, a hierarchical temporal-spatial feature aggregation subnetwork (H-TSFA-subnetwork) is constructed to obtain the final temporal-spatial features. Finally, a token-to-label module (T2L-module) is adopted to obtain the classification results. Extensive experiments demonstrate that the proposed method achieves state-of-the-art classification accuracy for space target classification with HRRP sequences, especially under the conditions of a low signal-to-noise ratio and a high missing rate.