Complex battlefield environments are characterised by diverse target types, intricate task constraints, and highly dynamic environmental conditions, thereby imposing requirements on intelligent target recognition that go beyond conventional optimisation of perceptual accuracy. In these environments, recognition results are not only used to describe target attributes but also directly affect the reliability of task planning and decision-making. However, most current target recognition research mainly concentrates on static scenarios and perception-based metrics, which do not adequately capture the practical significance of recognition results in task execution. To address this gap, a task-driven paradigm for target recognition has gradually emerged in recent years, in which task-related information is explicitly incorporated into model design, training, and evaluation, thereby enabling recognition results to support task deployment and system-level decision-making better. Following this research trend, this paper presents a systematic survey of task-driven intelligent target recognition methods from a methodological perspective. Firstly, the fundamental concepts of task-driven target recognition were analysed, and its key differences from traditional perception-driven approaches were clarified with respect to output representations, optimisation objectives, and system role positioning. Then, from the perspective of task-related information modelling, existing methods were systematically reviewed with respect to semantic and attribute representations, target state and behaviour modelling, and uncertainty and risk representation. After that, task-constraint modelling during training and optimisation, as well as the collaborative interfaces between recognition outputs and task-execution and decision modules, were further discussed. Finally, using the typical demands of complex battlefield environments as a key context, the paper summarized the major challenges in task-driven target recognition, including adapting to dynamic environments, managing unknown targets, ensuring trustworthy uncertainty representation, and coordinating at the system level. It also outlines potential directions for future research.

During multi-source information fusion, efficiently combining features from High Resolution Range Profiles (HRRP) and Inverse Synthetic Aperture Radar (ISAR) images is challenging due to the difficulty in fusion, limited sample availability, and the open-set recognition problem. To address these issues, a spatial target recognition method based on feature alignment and data augmentation was proposed. Firstly, Principal Component Analysis (PCA) was adopted to reduce the dimensionality of HRRP and extract time-frequency features. The Pauli decomposition was then utilised to expand the data of polarised ISAR images. After that, an improved ResNet18 network and a Transformer fusion module were constructed to align and fuse HRRP and ISAR features. Finally, the OpenMax open-set recognition framework was introduced, and the Weibull distribution was used to model class boundaries to achieve discrimination of unknown classes. Experimental results show that the proposed method achieves 90.43% accuracy in closed-set recognition and 91.39% rejection rate for unknown classes in open-set recognition, verifying its effective recognition and generalisation abilities in complex scenarios.

This paper proposes a recognition framework based on a hierarchical Bayesian model to address the challenges associated with fragmented prior knowledge and the absence of uncertainty quantification in decision-making processes for aerial target recognition within complex electromagnetic environments. By developing a three-tiered hierarchical structure encompassing "measurement noise-individual characteristics-class commonality", the intra-class physical variability of target Radar Cross Section (RCS) and sensor random noise were explicitly modelled as probability distributions, representing a novel contribution. Posterior inference was performed using Markov Chain Monte Carlo (MCMC) methods, simultaneously outputting target-class probabilities with confidence intervals. Simulation results show that under harsh observation conditions at 5dB SNR, the recognition accuracy reaches 78%, improving by 6% to 10% over Support Vector Machine (SVM) and Naive Bayes classifiers. In small-sample scenarios (5 training samples per class), the accuracy advantage increases to approximately 13%. The 95% confidence interval coverage rate exceeds 88%, validating the effectiveness of uncertainty quantification. The proposed method provides a practical pathway to robust target recognition within complex battlefield environments characterized by "small-sample + high-noise" conditions.

To address the challenges of information warfare in complex electromagnetic environments, this paper proposes a Reconnaissance-Jamming-Communication (R-J-C) integrated system and constructs a dynamic intelligent management architecture and strategy. Based on the theory of Electromagnetic Manoeuvre Warfare (EMW), an integrated management framework comprising the command-control, perception, countermeasure, and communication subsystems was established. Moreover, management strategies with a global architecture, intelligent decision-making, and rapid reconfiguration capabilities were designed, and the comprehensive management of the electromagnetic spectrum and the cooperative countermeasures process were validated through a game-theoretic integrated cognition system. In summary, the R-J-C management strategies provide a significant academic reference for enhancing the dynamic adaptability and multi-domain coordination capabilities of integrated electromagnetic technologies.

To overcome the limitations of traditional approaches in suppressing sea clutter which displays complex, non-stationary, non-Gaussian distributions and strong temporal correlation, this paper introduces RTFEN (Radar Temporal Feature Enhancement Network), a deep learning method that improves temporal feature extraction. The technique employs a temporal compression module to suppress clutter via space-time filtering, a bidirectional ConvLSTM to model motion differences between targets and clutter, and a temporal attention mechanism to adaptively enhance target information in key frames. Experimental results show that the proposed method can effectively enhance sea-clutter suppression performance.

Addressing the problem of insufficient robustness of the intelligent recognition model of radar High Resolution Range Profile (HRRP) under adversarial sample attacks, a lightweight enhancement method was proposed in this study. Firstly, a comprehensive analysis was conducted using the Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), and a black-box migration attack to assess the vulnerability of the lightweight Convolutional Neural Network (CNN). Secondly, a cascaded defense strategy of “fast adversarial training + input denoising auto encoder + post-anomaly detection”was established. Finally, countermeasure experiments were carried out using three types of air targets and 6,000 sets of measured samples. The results show that this strategy can reduce the attack success rate to 9.2%, sacrificing only 2.1 percentage point of the cleaning accuracy, and increasing inference delay by less than 20%. It achieves a stable balance among model size, real-time performance, and robustness, providing a practical solution for the anti-interference design in radar-intelligent recognition systems.

Although deep neural networks have achieved significant advancements across a range of visual tasks, they continue to be vulnerable to adversarial attacks. Compared to digital-domain attacks, physical-world adversarial attacks pose greater threats. In the context of adversarial attacks on UAV remote-sensing image object detection, it’s essential to maintain stable effectiveness under complex conditions, such as varying viewpoints, distances, and lighting conditions. Optimising attack methods must be fully considered in light of the dynamics and diversity of real-world imaging environments. Although existing physical-domain adversarial attack methods can degrade the performance of object detection models, they often rely solely on pixel-level local texture optimisation, resulting in monotonous adversarial texture patterns and limited adaptability. To address the aforementioned issues, this paper proposed a diffusion model-based physical adversarial attack method. The proposed approach employed a pre-trained diffusion model as the generator, leveraging both image and text priors to guide the generation of adversarial textures. Within a comprehensive physical attack framework, it enabled vehicle camouflage in UAV remote-sensing object-detection tasks. Experimental results demonstrate that the proposed method achieves high attack success rates and strong cross-model transferability across multiple object detection models, outperforming comparative methods in attack effectiveness and texture pattern diversity.

This paper introduces an improved adaptive DBSCAN clustering algorithm to tackle the issues of point cloud clutter removal and sparse target classification in radar imaging systems, which are traditionally handled by signal processing methods. The proposed method constructed a Euclidean distance matrix for classification, rapidly identified central sampleswithin categories, detected and eliminated anomalous stray points, and adaptively adjusted the neighbourhood density and radius parameters for future frames based on the Euclidean distances and mutation indices of the central points. Initially, the engineering advantages of the improved algorithm were validated through simulation experiments, followed by further verification using real-world road scene data to confirm its practical effectiveness. Experimental results show that the proposed algorithm effectively eliminates clutter from target point clouds and dynamically adjusts clustering parameters to reduce sparse classification errors in targets.

This study addresses the issues of low efficiency and accuracy in target matching within remote sensing ship tracking scenarios, which are attributed to significant modal differences between optical images and synthetic aperture radar (SAR) images, as well as the inadequacy of cross-modal matching technology. In response, a novel optical-SAR cross-modal target matching method employing a window attention mechanism is proposed. The method designed a cross-modal dual-branch embedding module to process the two image types separately and extracted modality-agnostic features via a hierarchical window attention mechanism. It fused the modal information embedding and ship-size embedding to supplement the semantic and physical attribute information of ships and to enhance the learning of cross-modal-aligned features. Experimental results show that the proposed method achieves an overall mean Average Precision (mAP) of 46.0%, Top-1 (R1) matching accuracy of 60.8%, Top-5 (R5) matching accuracy of 74.4%, and Top-10 (R10) matching accuracy of 79.5% on the HOSS dataset. Compared with the state-of-the-art TransOSS model, R5 and R10 achieve improvements of 3.4% and 1.1% respectively. The key matching indicators in both the optical-to-SAR and SAR-to-optical directions are superior to those of the current optimal model. The research indicates that the proposed method outperforms the SOTA(state-of-the-art) model and provides technical support for continuous ship tracking across scenarios such as maritime search and rescue and shipping supervision.

This paper proposes a deep learning-based infrared ship-and-wake detection algorithm to address missed and false detections of low-emission, small targets in complex sea-sky backgrounds. The algorithm enhanced the YOLO network by incorporating a dual attention mechanism that suppresses feature maps within the YOLOv8 architecture. In addition, a ship-wake matching module was developed, leveraging more prominent wake features to assist ship detection, effectively reducing false alarms and missed detections of weak and small ships in complex backgrounds. Finally, a ship dataset was constructed for testing and analysis. Results show that the proposed algorithm achieves 98% precision and a high detection speed, demonstrating strong robustness in detecting weak infrared ship targets and significantly enhancing detection performance.

This paper introduces a data augmentation technique using a conditional diffusion model to tackle issues related to limited samples and cross-domain distribution shifts in High-Resolution Range Profile (HRRP) data. The proposed approach enhanced the traditional diffusion model by incorporating an angle modulation mechanism that processes azimuth and elevation angles. These angular values were mapped to a high-dimensional continuous space using sine-cosine encoding and then used to modulate category embeddings via a linear transformation, thereby strengthening the model's capacity to capture dependencies between viewing angles and target classes. Additionally, a one-dimensional Fréchet Inception Distance (FID) evaluation metric, leveraging a Temporal Convolutional Network (TCN) for feature extraction, was employed to quantitatively assess the distributional similarity between generated and real HRRP data. Experimental results show that the HRRP data generated by the proposed method achieves a significantly lower one-dimensional FID score than produced by conventional conditional diffusion models. Adding generated samples to the actual training dataset increases the average classification accuracy by 8.55 percent point, demonstrating the effectiveness and practical value of the proposed HRRP data augmentation method.

This paper introduces a voice command-and-control system based on Large Language Models (LLMs), designed to handle the high dynamics, intense conflict, and multi-source, heterogeneous information in modern combat environments. The limitations of traditional 'point-and-click' interfaces and standard voice systems were addressed, which often struggle with noise robustness and limited semantic adaptability. By integrating advanced audio denoising and tactical hot word enhancement, speech recognition accuracy and domain adaptability in noisy conditions were improved. Domain-specific Prompt engineering, data augmentation, and LoRA fine-tuning further enhanced the LLMs’ understanding of non-standard expressions and tactical semantics, enabling end-to-end voice-to-command conversion. Experimental results show that the proposed approach outperforms baseline methods in battlefield noise conditions, giving a dependable, natural, and efficient framework for human-machine collaborative command.

A target guidance algorithm is proposed to solve the two-dimensional target information transfer problem of detector in passive detection scenarios. Through target azimuth and angular altitude detected by a single station, the algorithm is effective to guide the other detector to search or track without target distance, which is used to achieve cooperative detection and triangulation localization. Simulation verification demonstrates that the algorithm enables detector to guide other detectors for cooperative detection and complete target distance calculation in passive tracking scenarios, with a guidance success rate exceeding 85% in the simulation experiment. In physical environment validation, all three tests were successful. Besides, the Kalman filter algorithm is used to deal with the detected data, which is valid to remove outlier, smooth the data, and improve success rate of target guidance.

In the issue of evaluating threats in air combat scenarios, current assessment techniques frequently encounter difficulties in delivering precise results. To mitigate these limitations, this paper concentrates on multi-aircraft group confrontation combat scenarios. It proposed a novel air combat target threat assessment method based on combined weighting, Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), and Grey Relational Analysis (GRA) to calculate the threat level of the target group. Firstly, an improved objective weighting method based on Criteria Importance Through Intercriteria Correlation (CRITIC) and the order relation analysis method (G1) was employed to determine the objective and subjective weights of threat assessment indicators, respectively. Based on the principle of minimum discriminative information, a combined weighting scheme was used to calculate the comprehensive indicator weights. Secondly, the Improved TOPSIS (ITOPSIS) and GRA were integrated to comprehensively utilise the distance metric and grey relational degree between threat indicators and the positive/negative ideal solutions. Then, the relative closeness of target groups was calculated as the threat assessment result. Finally, a typical simulation case study was designed to validate the effectiveness of the assessment method. Simulation results show that the proposed threat assessment method, based on combined weighting-ITOPSIS-GRA, considers a broader set of factors and yields more reasonable and accurate assessment outcomes.

In extended-range air-target strike scenarios, conventional kill chain models encounter difficulties in managing the challenges presented by highly dynamic and intensively contested environments. To address this concern, this paper proposes a dynamic kill-web construction methodology grounded in collaborative tactical knowledge. Firstly, the kill web was modelled as a temporally constrained quadruple task-resource allocation graph (S-D-I-T graph), in which the three categories of nodes—S (sensing), D (decision-making), and I (engagement)—must act on the T (target) nodes in a tactically logical sequence. A collaborative tactical knowledge base was then introduced to systematically define task priorities, resource-capability matching rules, and temporal dependencies in common air-target strike scenarios. Finally, a dynamic construction algorithm based on constraint satisfaction and multi-agent negotiation was designed to generate and configure the kill web in realtime, achieving real-time generation and reconfiguration under time-window, resource-capacity, and tactical-logic constraints. Simulation results show that the proposed method significantly outperforms traditional static kill chains and unguided random-allocation strategies in terms of target damage rate, task completion speed, and resilience. This study provides efficient and robust decision support for long-range air-target strike operations.

To address issues such as insufficient consideration of interrelationships among operational links, imperfect construction of multi-simulation-node evaluation systems, limited sources of assessment data, and significant interference from subjective data, this study introduces an integrated system-of-systems effectiveness evaluation framework that combines complex network theory and multi-agent simulation technology, facilitating a comprehensive analysis and quantitative assessment of dynamic interactions and overall effectiveness in complex systems. Firstly, integrating simulation platforms with physical models produced indicator data for system-level evaluation under specified scenarios and schemes. Secondly, simulation nodes and their corresponding data were mapped into nodes and relationships within a complex network, where corresponding capability matrices were constructed. Finally, within the defined combat scenarios and simulation nodes, effectiveness metrics for each scheme were quantified by incorporating link information derived from complex network analysis and capability matrices for different combat loops, followed by a systematic comparison and analysis of their effectiveness. Experimental results show that the proposed method effectively assesses each scheme's strengths and weaknesses, while also intuitively illustrating how effectiveness evolves over time and highlighting overall trends across various combat loops.

The modular combat system, characterised by its inherent flexibility, reconfigurability, and adaptability, has emerged as a critical enabler for supporting new warfare concepts such as Mosaic Warfare and for addressing dynamic combat environments. Developing a dedicated metamodel is vital for creating a standardized architectural description method for modular combat systems. Based on domain meta-modelling methodologies and following an analysis of the conceptual framework and key elements of modular combat systems, this study tailored and extended the service and resource viewpoints of the Unified Architecture Framework meta-model while establishing a data viewpoint. This results in a meta-model that supported a consistent multi-viewpoint description. Through an air and missile defence case study, the complete architectural modelling process, from capability decomposition to service clustering and finally to resource grouping, was demonstrated. The analysis emphasizes its flexibility, offering a robust meta-model and methodological basis for designing modular combat systems.