Accurate radio map construction is essential for 6 G wireless network optimization, yet faces significant challenges due to sparse real-world measurements and dynamic environmental obstacles. This paper presents RMF, a novel one-step generative model based on mean flow matching that enables direct mapping from noise to radio map distribution in a single forward pass. Our approach integrates a multi-feature U-Net architecture with specialized branches for processing building layouts, base station configurations, sparse measurements, and dynamic obstacles through cross-attention fusion. Extensive evaluations on the RadioMapSeer dataset demonstrate that RMF achieves state-of-the-art performance, reducing RMSE by 7.5–12.2% compared to diffusion-based methods while maintaining competitive SSIM scores of 0.9557–0.9674. In challenging zero-measurement scenarios, RMF attains PSNR improvements of 1.45–1.65 dB over existing approaches, showcasing robust performance in both static and dynamic environments. The model's balance of accuracy and efficiency makes it particularly suitable for real-time 6 G applications including coverage optimization and dynamic resource management.

Deep learning-based radio frequency fingerprint identification (RFFI) shows significant promise for Internet of Things device authentication. However, current mainstream RFFI methods often assume training and testing data originate from the same receiver, leading to performance degradation in cross-receiver scenarios. Direct application of a model trained for one receiver to a new one results in substantially reduced identification accuracy. While domain adaptation (DA) techniques can mitigate this domain shift, most existing DA-RFFI methods are limited to single-source single-target settings, hindering their scalability to complex multi-target domain environments. Specifically, a new identification model needs to be retrained for each new target receiver. To address this challenge, this article proposes a novel adversarial multi-target domain adaptation (AMTDA) method. The core of this approach is a decomposed domain discriminator, which simultaneously aligns the source domain with all target domains (source-target alignment) and the intrinsic distributions among different target domains (target-target alignment) through adversarial training. This enables the model to learn more generalizable features. Furthermore, we incorporate virtual adversarial training to further enhance the model’s robustness and generalization capabilities. Extensive experiments conducted on two subsets of the WiSig dataset demonstrate that the proposed AMTDA significantly outperforms current state-of-the-art algorithms in identification performance, while maintaining high efficiency.

In the field of spectrum situation monitoring, space-based spectrum mapping can overcome the limitations of ground-based monitoring, such as restricted coverage, environmental constraints, and difficulties in equipment deployment. However, current research on radio map reconstruction methods based on space-based satellite data is insufficient. Traditional methods mostly set missing patterns based on the type of data collected on the ground, such as random sampling or strip sampling. These methods can meet the needs of generating spectrum maps for local spectrum monitoring on the ground, but they are hardly adaptable to the task of space-based global spectrum mapping. To address this challenge, we have designed a tensor reconstruction method for radio maps based on local low-rankness. Before conducting global completion, we first identify local regions in the map with more prominent low-rank characteristics for completion. This enables the method to handle radio map reconstruction in the context of space-based wide-area coverage. During the reconstruction phase, targeting the two-dimensional spatial completion problem of space-based spectrum maps, we extend the autoregressive process from the temporal dimension in classical spatiotemporal models to the spatial domain, proposing a low-rank tensor completion model based on dual autoregression (LRTCDAR). By incorporating a dual autoregressive structure to capture spatial correlations and combining it with the inherent low-rank property of radio maps, the model achieves high-precision reconstruction of incomplete radio maps. Extensive experimental results demonstrate that, compared to state-of-the-art methods, the proposed approach effectively reconstructs radio maps under various conditional backgrounds. Notably, LRTCDAR exhibits greater advantages when applied to satellite data acquisition scenarios.

Accurately reconstructing Radio Environment Maps (REMs) under sparse measurements is a critical challenge for spectrum management and wireless resource optimization. Conventional interpolation and model-based methods often degrade in urban environments due to multipath scattering, shadowing, and the lack of reliable environmental priors. In this study, we propose CMCTNet, a multimodal deep learning framework that integrates sparse Reference Signal Received Power (RSRP) measurements with Synthetic Aperture Radar (SAR) imagery and building footprint maps to reconstruct high-resolution REMs. CMCTNet is a cross-modal completion network designed with wavelet-based encoders, dense connectivity, and an attention-driven fusion mechanism. Unlike existing methods that rely on detailed geographic information, our framework leverages SAR as a structural prior, capturing obstruction layouts and propagation-relevant textures. Extensive experiments on a synthetic urban dataset show that CMCTNet consistently outperforms strong baselines, including DeepREM, RMTransformer, and RobUNet, especially under extreme sparsity. Quantitative results demonstrate substantial improvements in MAE, RMSE, SSIM, and PSNR across both uniform and Gaussian-distributed sampling. Ablation studies further validate the contributions of SAR input, wavelet decomposition, and cross-modal attention, while a polarization sensitivity analysis confirms robustness under mismatched SAR polarizations. These findings highlight the feasibility of using SAR data as a globally accessible structural prior and demonstrate the potential of multimodal learning to capture propagation-relevant spatial features for accurate and scalable REM reconstruction.

The escalating disparity between the supply and demand of maritime radio spectrum resources necessitates the construction of high-fidelity Radio Environment Maps (REM) for effective spectrum situational awareness and dynamic management. However, this task is severely hampered by distinctive maritime challenges, including extreme data sparsity, highly dynamic electromagnetic propagation characteristics, and complex spatio-temporal correlations, which significantly degrade conventional terrestrial REM reconstruction methods. To overcome these limitations, this paper proposes a hierarchical REM reconstruction framework that synergistically integrates Spiking Neural Networks with physical constraints. Our contributions are threefold. First, we devise an adaptive Unmanned Aerial Vehicle sampling strategy based on a refined Ant Colony Optimization algorithm, incorporating a hierarchical priority decision mechanism and joint heuristic function to improve data collection efficiency under sparse sampling. Second, we architect a Frequency-Spatio-Temporal Attention (FSTA) -enhanced Spiking Neural Network (SNN) model that captures spatio-temporal dynamics from sparse observations for high-precision 2D REM completion. Third, we introduce a physics-guided knowledge distillation paradigm that embeds maritime electromagnetic propagation models as multi-stage soft constraints through three coordinated mechanisms, which direct input correction via height-weighted physical deviation terms, and loss-level supervision penalizing physically inconsistent 3D reconstructions. Extensive simulations conducted in a high-fidelity maritime scenario demonstrate, which is constructed using real geographic environments, GMTED digital elevation data, and representative meteorological conditions. Our framework outperforms tensor completion U-Net and PINN baseline across various sampling rates. Notably, at sampling rates of 20%, 50%, and 80%, the proposed framework consistently attains superior Root Mean Square Error (RMSE) and Normalized Mean Square Error (NMSE), with the Physical Residual Metric (PRM) further serving as a diagnostic indicator confirming internalization of physical priors.

Radio Frequency Fingerprint Identification (RFFI) utilizes non-ideal hardware features present in the signal to identify different transmitters. However, existing RFFI models have poor generalization capabilities. When a model trained on one receiver is deployed on a new receiver, the identification performance of the model degrades due to the effect of different receiver characteristics, which can cause the signal distribution to be shifted. To address this problem, we propose a crossreceiver RFFI based on domain adaptation with dynamic distribution alignment. First, deep features are extracted using the ResNet18 network and the global distribution of the features is aligned using Maximum Mean Discrepancy (MMD). Then multilevel features are extracted using a designed multiscale feature extraction module and the subdomain distribution is aligned using Local Maximum Mean Discrepancy (LMMD). Finally, a dynamic parameter is introduced to adaptively adjust the relative importance between the global and subdomain distributions. Twelve sets of cross-receiver experiments are conducted on the WiSig dataset, and the algorithm in this paper achieves an average identification rate of 92.52 domain. Meanwhile, the experimental results under different signal-to-noise ratios (SNR) also verify the algorithm has strong robust performance. It shows that the algorithm can effectively alleviate the model performance degradation problem in the cross-receiver scenarios.

The comprehensive acquisition of electromagnetic spectrum data is crucial for applications such as cognitive radio and electromagnetic mapping in electromagnetic environment monitoring. However, the inherent limitations of spectrum sensor sampling rates and environmental unreliability often result in data loss. Existing spectrum data recovery methods predominantly adopt generic algorithms from other domains, failing to ade-quately exploit the intrinsic characteristics of spectrum signals. This paper introduces a novel low-rank autoregressive tensor com-pletion algorithm leveraging the Shift-Invariant Discrete Wavelet Transform (SIDWT) to recover and reconstruct incomplete elec-tromagnetic spectrum data. The approach focuses on minimizing the truncated nuclear norm of the SIDWT coefficient tensor, which exhibits stronger low-rank properties. Additionally, the AR(p) model is incorporated to impose temporal constraints, en-suring the reconstructed data closely approximates the original in both global and local characteristics. Additionally, to address the issue of high computational complexity, we have developed a fast algorithm. Experimental validation using real-world data demon-strates the efficacy of the proposed algorithm. It outperforms state-of-the-art algorithms on 78.1% of electromagnetic spectrum datasets and all audio signal spectrum datasets, achieving at least a 2% improvement in Relative Squared Error (RSE) compared to the best-performing baseline. Notably, under high missing rates and on TM datasets, the improvement exceeds 30%. The proposed method provides a novel solution for the completion of incomplete electromagnetic spectrum data. The proposed fast algorithm achieves approximately 40% improvement in computational speed without significant performance degradation.

In urban area, 2-D radio map is no longer sufficient for spectrum management, highlighting the growing significance of 3-D radio map. However, traditional methods for constructing 3-D radio map face challenges, including high storage requirements and computational complexity, limiting their effectiveness. To address these issues, this article introduces a systematic approach for constructing 3-D radio maps. Initially, a portion of the real-world 3-D city map is captured and discretized into grids of uniform size. Subsequently, spectrum data is collected using an unmanned-aerial-vehicle (UAV) platform. To enhance sampling efficiency, iterative path planning algorithms refine sampling locations to capture the spectrum conditions of the target space. Finally, the trajectory tensor of the entire spectrum space is constructed through adaptive embedding, and the global spectrum is compressed and reconstructed by using the enhanced tensor singular value decomposition (t-SVD) algorithm to construct the radio map of the city scenario. Comparing our algorithm with the existing methods, we analyze the influence factors of frequency-spatial compression sampling on the spectrum situation and strength recovery. Numerical results show that, compared with the existing methods, the proposed method achieves more accurate spectrum mapping.