Naive retrieval-augmented generation (RAG) methods enhance large language models (LLMs) by retrieving relevant textual information, improving the accuracy of responses. However, they are limited in capturing the complex relationships and structures in textual graphs, where both textual and topological information are crucial for graph reasoning. To address this issue, we propose a multi-level pruning graph RAG framework, called MGRAG. MGRAG consists of three stages: sub-graph retrieval, multi-level graph pruning, and answer generation. We first index and rank the subgraphs to improve retrieval efficiency. Then, we apply dynamic pruning at both the subgraph and node levels to extract the most relevant graph structures. Finally, we integrate the query, graph description, and optimized graph structure as inputs to the LLM for response generation. Experimental results on multi-hop reasoning benchmarks demonstrate that MGRAG effectively eliminates irrelevant structures, significantly improving response accuracy and overall model performance.

In complex maritime scenarios, multiple targets within the same radar beam often degrade the quality of inverse synthetic aperture radar (ISAR) imaging. Most existing methods typically leverage a single domain for target separation, hardly considering the relationships across multiple domains. To fully exploit the inter-domain interactions, we propose an end-to-end multi-domain interaction deep unrolling network based on a block-aware optimization model, termed MDIB-Net, to simultaneously achieve multi-target separation and echo reconstruction. Combining target structural features and deep unrolling techniques, this model-driven network effectively explores the correlations of the same target across different domains to achieve target separation. The MDIB-Net comprises cascaded iteration blocks, where each iteration block consists of an optimization block and a domain interaction (DIR) block. The optimization block solves the proposed block-aware multi-target separation function, leveraging low-rank properties, local similarity, and structural priors to capture the physical characteristics of targets. Moreover, a learnable module is incorporated in this block to discover the optimal transformation domain, thereby enhancing the structural prior of targets. The DIR block introduces a lightweight module to extract the semantic maps from the high-range resolution profiles (HRRPs) domain. The DIR block further incorporates a spatial-adaptive semantic guidance module, which takes the semantic maps as guidance, to refine the transformation domain features, effectively promoting cross-domain feature interactions. Additionally, the MDIB-Net achieves multi-target separation by dynamically adjusting its iterative strategy according to the predefined target number, demonstrating both robustness and flexibility. Simulated and measured experiments validate the effectiveness of the proposed method.

Icebergs threaten the navigation safety of ships in the Arctic, small icebergs are difficult to detect in harsh sea conditions. Timely detection of these targets is crucial for aiding ships in navigational decisions. Infrared detection technology is an effective solution for detecting icebergs. Due to the low contrast of iceberg infrared images and severe sea wave clutter, current methods can not suit iceberg detection tasks well. In this article, we propose an Adaptive Multi-scale Enhanced Interaction Network (AMEINet) and construct the Single-Frame Infrared Iceberg Target Dataset (SFIIT). First, the Global Contrast Enhance Module (GCE) is designed, which adaptively enhances the image with the target as the center, and reconstructs the potential features submerged by low contrast. Then, to obtain rich features, the Boundary Feature Extraction Module (BFE) and Gated Context Feature Extraction Module (GCFE) are designed to extract target information at different levels. BEF adopts a new computation method to obtain the edge contour structures without edge GroundTruth supervision. GCFE effectively controls redundant information transmission through the gate mechanism and jump connection, filtering out the false alarms and miss detection caused by strong wave clutter. Finally, experiments on the publicized SIRST dataset and proposed SFIIT dataset show that the proposed method can achieve superior performances in infrared iceberg detection.

With the capacity to carry the orbital angular momentum (OAM), the vortex electromagnetic wave (VEMW) radar observes targets from a unique dimension. Since the VEMW radar introduces the azimuthal angle information, it certainly extends the potential of 3-D inverse synthetic aperture radar (3-D ISAR) imaging. Nevertheless, the large rotational angle and OAM mode number limit the application of reconstructing noncooperative targets’ spatial construction with high quality. To address it, this letter proposes a novel 3-D ISAR imaging method based on the migration through the azimuthal angle cells (MTAAC) compensation. First, we construct the MTAAC phase by forming and analyzing the rotational geometric model. Following that, the related MTAAC compensation phase is established. Then, we perform the CLEAN algorithm on the conventional ISAR image to estimate the range and cross-range of different scatterers. Finally, we iteratively extract the 3-D structure of the target by compensating the MTAAC within the VEMW ISAR image. The MTAAC domain is also provided to further confirm the condition and the coverage of the proposed method. The simulated experiments show that the proposed method achieves a 10.6% reconstruction error under 0-dB signal-to-noise ratio (SNR) scenarios, verifying the effectiveness of the proposed method.

Ground-penetrating radar (GPR) has become indispensable for subsurface reconstruction, especially in complex, multilayered infrastructure environments. Efficient and accurate estimation of permittivity and layer thickness, crucial for understanding these scenes, presents significant challenges. This article explores the integration and interaction between the back projection (BP) algorithm and full-wave inversion (FWI) to optimize reconstruction outcomes. We propose a ray-based refraction method and a high-order moment (HOM)-oriented local BP estimation, each aimed at improving efficiency from the critical perspectives of “delay” and “summation” in the BP algorithm, respectively, with a focus on the target layer. Crucially, the BP estimation results provide a reliable initial model for FWI, reducing the risk of local minima and decreasing the iteration count for more precise parameter estimation. FWI also compensates for the BP algorithm’s limitations in estimating parameters in nontarget layers. Furthermore, full-wave modeling (FWM) mitigates antenna effects prior to parameter estimation, thereby enhancing accuracy. During reconstruction, the BP algorithm primarily generates target imaging, while FWI provides detailed information about layer interfaces. This synergistic approach leverages the complementary strengths of both algorithms: the BP algorithm captures spatial information depicted by the targets observed in B-scans, while FWI incorporates detailed electromagnetic wave propagation along layer interfaces presented in A-scans via radar equations. Experiments systematically analyze the accuracy and effectiveness of the proposed approach using ideal simulation models, sandbox laboratory data, and road data from the Belgian Road Research Centre (BRRC) facilities. This comprehensive evaluation underscores the approach’s substantial potential in advanced geophysical surveys and environmental research.

In real-world situations, multiple targets may locate in the same radar antenna beam, resulting in degraded readability of inverse synthetic aperture radar (ISAR) imaging. Due to the different motions of multiple targets, the existing monotarget imaging methods generally lead to tight overlap and severe defocusing. To tackle this problem, this article proposes a two-step method based on dual-precision optimization for multitarget ISAR imaging. First, the Radon transform estimates the range walks of different targets, subsequently generating the proposed coarse motion compensation transforms (MCTs). Moreover, the short imaging interval commonly brings strong similarity within adjacent range profiles. This internal property motivates us to model the multitarget coarse separation as a problem jointly constrained by local similarity and low-rank property. Such a double-constrained problem is then solved by incorporating the alternating direction method of multipliers with linearized alternative direction method with adaptive penalty (ADMM–LADMAP) framework. Through these procedures, range profiles of different targets are partially separated, allowing the correlation-based range alignment methods to estimate the range walk with higher accuracy. With the help of these accurate estimations, MCTs are modified, and the sparsity of the ISAR image is simultaneously enhanced. The sparse prior is thus integrated into the previous optimization to convert coarse separation to a more precise one so as to achieve entire separation. Finally, phase autofocus and cross-range compression are carried out on the completely isolated range profiles to yield well-focused ISAR images. Experiments based on both simulated and measured data demonstrate the effectiveness of the proposed method.

Supervised learning for semantic segmentation has achieved impressive success in remote sensing, while this normally has a high demand on pixel-level ground truth from the testing images (target domain). Labeling data for semantic segmentation is labor-intensive and time-consuming. To reduce the workload of manual labeling, domain adaptation (DA) utilizes preexisting labeled images from other sources (source domain) to classify the images in the target domain. In this article, we propose a bispace alignment network for DA named BSANet. BSANet is designed to have a dual-branch structure which is able to extract features in the image domain and the wavelet domain simultaneously. To minimize the discrepancy between the source and target domains, we propose a bispace adversarial learning strategy. Specifically, BSANet employs two discriminators in different spaces, one aligning the source and target feature distributions, and the other helping the classification outputs render reasonable spatial layouts. The proposed method shows the ability to train an end-to-end network for semantic segmentation without using any label in the target domain. Extensive experiments and ablation studies are conducted in cross-city scenarios. Comparative experiments with several state-of-the-art DA methods show that our method achieves the best performance.

High-squint synthetic aperture radar (HS-SAR) imaging technology greatly enhances the flexibility of SAR. However, the existing HS-SAR imaging algorithms are generally based on the assumption that the illuminated target is static during the synthetic aperture formation. As a consequence, the state-of-the-art algorithms are unable to focus on moving targets well, thus leading to displaced and blurred images of the targets. A maritime ship is a typical kind of moving target with complex noncooperative motion, which has received much attention. Thus, it is a difficult but valuable issue to study HS-SAR imaging of ship targets. In this work, the influence of ship translation and fluctuation is theoretically analyzed. On the basis of these analyses, the new conception of “third range compression (TRC)” caused by ship translation is proposed, and the squint minimization (SM) operation is implemented for the compensation of TRC. Moreover, considering that the target Doppler parameters induced by ship fluctuation are related to its position, a novel method called WASH-CLEAN (watershed and CLEAN) is proposed to automatically focus ship scattering points with different swings. Finally, we propose an integrated modified range-Doppler (RD) imaging algorithm by combining the advantages of HS-SAR and inverse synthetic aperture radar (ISAR). The simulation results show the validity and effectiveness of the presented method.

With the rapid development of deep learning technology, semantic segmentation methods have been widely used in remote sensing data. A pretrained semantic segmentation model usually cannot perform well when the testing images (target domain) have an obvious difference from the training data set (source domain), while a large enough labeled data set is almost impossible to be acquired for each scenario. Unsupervised domain adaptation (DA) techniques aim to transfer knowledge learned from the source domain to a totally unlabeled target domain. By reducing the domain shift, DA methods have shown the ability to improve the classification accuracy for the target domain. Hence, in this letter, we propose an unsupervised adversarial DA network that converts deep features into 2-D feature curves and reduces the discrepancy between curves from the source domain and curves from the target domain based on a conditional generative adversarial networks (cGANs) model. Our proposed DA network is able to improve the semantic labeling accuracy when we apply a pretrained semantic segmentation model to the target domain. To test the effectiveness of the proposed method, experiments are conducted on the International Society for Photogrammetry and Remote Sensing (ISPRS) 2-D Semantic Labeling data set. Results show that our proposed network is able to stably improve overall accuracy not only when the source and target domains are from the same city but with different building styles but also when the source and target domains are from different cities and acquired by different sensors. By comparing with a few state-of-the-art DA methods, we demonstrate that our proposed method achieves the best cross-domain semantic segmentation performance.

More recently, due to the enormous potential of long short‐term memory (LSTM) neural network in various fields, some efficient training algorithms have been developed, including the extended Kalman filter (EKF)‐based training algorithm and particle filter (PF)‐based training algorithm. However, it should be noted that if the system is highly non‐linear, the linearisation employed in the EKF may cause instability. Moreover, the PF usually suffers from the particle degeneracy. Therefore, the PF‐based training algorithm may only find a poor local optimum. To solve these problems, an unscented Kalman filter (UKF)‐based training algorithm is proposed. The UKF employs a deterministic sampling method; hence, there is no linearisation in it and it does not have the degeneracy problem. Moreover, the computational complexity of the UKF is the same order as that of the EKF. To further reduce the computational complexity, the authors propose a minimum norm UKF (MN‐UKF) to obtain a good trade‐off between performance and complexity. To the best of the authors’ knowledge, this is the first reported solution to this problem. Simulations using both benchmark synthetic signal and real‐world signal illustrate the potential of the algorithms developed.

Recently, the quaternion-valued feedforward neural network (QFNN) has been developed to process three dimensional (3-D) and 4-D signals in quaternion domain, and the weight matrices and bias vectors of the QFNN were obtained based on the quaternion backward propagation (QBP) method. However, it should be noted that the QBP is a first-order quaternion gradient descent algorithm. The convergence speed of the QBP is usually slow and may not be very suitable to process nonstationary quaternion-valued signals. To address this problem, a widely linear quaternion unscented Kalman filter (WLQUKF) algorithm is proposed to train the QFNN. This is derived by utilizing some recent studies in the augmented quaternion statistics and the $\mathbb {HR}$ -calculus. With the augmented quaternion statistics, the WLQUKF is able to process general quaternion-valued noncircular, nonlinear, and nonstationary signals, effectively. Simulations on both benchmark circular and noncircular quaternion-valued signals, and on real-world quaternion-valued signals support the analysis.

The very high resolution (VHR) images can be seen as multiview data. For better organizing and highlighting similarities and differences between the multiple views of data, a semisupervised multiview feature selection (SemiMFS) method is proposed in this paper, based on consensus and complementary principles. In SemiMFS, feature views are generated by decomposing features into multiple disjoint and meaningful groups. Each feature group represents a view, and each view describes a data characteristic. Then features are evaluated and selected within each view. The experiments on a Worldview-2 VHR satellite image verify the effectiveness and practicability of the method, compared with traditional single-view algorithms.