Radio Frequency Fingerprint (RFF), a physical-layer authentication technique that requires no cryptographic pairing and exhibits inherent resistance to cloning or spoofing, has emerged as a promising solution for device authentication in the Internet of Things (IoT) and unmanned systems. However, practical RFF authentication depends on extracting device-specific hardware signatures from received waveforms in a physically explainable manner. Existing deep-learning-based methods often provide limited interpretability, whereas feature-specific methods usually capture only a small subset of handcrafted impairments. This motivates a more general, hardware-oriented RFF representation. In this paper, we construct such a representation by using the received signal together with its corresponding ideal reference waveform to estimate the transmitter’s linear and nonlinear behavioral features. The estimated behavior features are then used as the radio frequency fingerprint for device classification and identification. Under a fixed receiver and a standard preprocessing pipeline (timing alignment, CFO compensation, and amplitude normalization), we represent this transmitter-dependent behavior with a Volterra series: the zeroth-order term accounts for the constant offset, the first-order kernel describes the equivalent linear response with memory, and the higher-order kernels describe nonlinear interactions among delayed signal samples. To obtain a compact representation, we approximate the kernels using wavelet bases and estimate the corresponding coefficients via regularized least-squares fitting. The magnitudes of the fitted coefficients are then classified using a lightweight convolutional neural network. Experiments on a public LoRa dataset achieve 99.49% mean accuracy in the static-channel setting and above 90% mean accuracy under combined multipath and Doppler impairments. Compared with recent RFF methods, the proposed approach provides competitive accuracy and substantially lower complexity than most deep-learning baselines, while offering a more physically interpretable representation of transmitter behavior. We further validate dataset-level generalization on a self-constructed 16-device dataset. The implementation is publicly available at: https://github.com/thomas-smith123/RFFI.

With the rapid development of Internet of Things (IoT) technologies, efficient spectrum management becomes crucial. Wideband monitoring helps detect abnormal channels, ensuring reliable operation of IoT devices. Software-defined radio (SDR), with its wideband acquisition capability, offers flexibility for spectrum monitoring. We propose a two-stage monitoring method for signal modulation classification, combining coarse and fine detection techniques. Initially, the signal is transformed into a time-frequency (TF) matrix using time-frequency transformation (TFT). We design a complex network to perform approximate modulation recognition on TF domain. Based on the coarse detection results, we separate the signals in the TF domain and apply inverse TFT to obtain time-domain signals. To overcome TF matrix resolution limitations and improve signal center frequency (CF) estimation, we combine the Fast Fourier Transform (FFT) and Chirp Z-transform (CZT), performing cyclic convolution to refine CF estimation. This enables a more accurate frequency offset estimation. Next, frequency down-conversion is applied to the time-domain signal based on the estimated CF. For classification, we use complex Gated Recurrent Unit (GRU) and complex residual networks to extract and fuse features from time-domain signals, enabling precise modulation type classification for non-downsampled MQAM/MPSK signals. Experimental results validate that our method, when tested on the complex YOLOv7, improves the [email protected]:0.95 by 4.3% compared to the original YOLOv7 framework, with faster network convergence. Our approach also achieves significantly lower variance in CF estimation and outperforms existing methods in MQAM/MPSK classification accuracy under various signal-to-noise ratios (SNR). This study offers a promising approach for future wireless IoT communication systems management.

With the growing demand for large-scale interconnection of electronic devices, efficient utilization of limited spectrum resources has become increasingly important. Cognitive radio, which communicates by sensing the electromagnetic environment, significantly enhances spectrum utilization and offers a promising solution. Dynamic spectrum access and automatic modulation classification (AMC) are two key technologies in cognitive radio, attracting more attention. However, most existing AMC methods face limitations under flexible time-frequency resource allocation, particularly in identifying high-order modulation signals at low signal-to-noise ratios. To address this issue, we propose a multitask learning network called downlink control information network (DCI-Net), and apply it to the existing fifth-generation (5G) new radio (NR) communication system. The network aims to detect the effective time-frequency resources of the signal while identifying its modulation type. Furthermore, to further reduce the dependence of noncooperative communication receivers on pilot signals, we introduce a data-assisted channel estimation (CE) algorithm. This algorithm combines the identified modulation type with the equalized signal to generate pseudo-pilots, which are then used for signal demodulation in subsequent time slots. Notably, this method only requires the transmission of pilot signals in the first time slot, thereby reducing the receiver’s dependence on pilot signals. Simulation and over-the-air (OTA) test results demonstrate that the proposed multitask learning network significantly improves the accuracy of signal modulation recognition, while the data-assisted CE algorithm outperforms the least squares method in static environments.

Dual-functional radar-communication (DFRC) systems have gained attention for their advantages in spectrum sharing and hardware integration. Frequency-hopping (FH) MIMO radar waveforms enable communication via index modulation but introduce challenges such as increased range sidelobes, suboptimal communication performance, and incompatibility with conventional MIMO detection. Based on these considerations, we propose a new DFRC waveform design that combines frequency-hopping code selection (FHCS), binary frequency-shift keying (BFSK), and permutation selection (PS), termed FHCS-BFSK-PS. This design utilizes FH code vectors, frequency codes, and carrier-to-antenna mapping matrices as modulation parameters, embedding communication information across three dimensions: FH code index, frequency, and spatial phase. For communications, leveraging the weak coupling among these parameters, a simple and efficient demodulation algorithm based on the matched filtering criterion is proposed. For radar detection, we analyze the impact of fast-time domain FH on conventional MIMO detection and propose a modified detection method. This approach calculates and sums the log-likelihood ratios across multiple subpulses, thereby avoiding the incoherent accumulation of steering vectors in the subpulse domain. Finally, simulation results demonstrate the feasibility of the proposed scheme in target detection and highlight its superior performance in terms of symbol error rate (SER), data rate, and ambiguity function characteristics.

The main difficulty of synthetic aperture radar (SAR)-optical image matching or registration lies in the significant heterogeneous characteristics introduced by the different imaging mechanisms between SAR and optical images. Instead of directly using the raw image pair, transforming the pair into a feature domain, where they have homogeneous feature representation, is believed more effective. Inspired by image segmentation, we develop an end-to-end deep learning model for the SAR-optical matching, based on a siamese U-net with a fast Fourier transform (FFT) correlation layer. First, the siamese U-net with sharing weights extracts the feature maps of the SAR and optical images and projects the heterogeneous images into a homogeneous space. Then, the two feature maps are cross-correlated or normalized cross-correlated by the FFT layer and a similarity heatmap is obtained. Finally, the heatmap is send into a softmax2d classifier to determine the best matching, and thus matching is converted into classification. The nonlinear mapping capability of deep learning can well tackle the intensity variation across the different imaging modals; the encoder–decoder architecture with skip connections in the U-net can take full advantage of the global information and simultaneously preserve the local resolution and position information and thus guarantees high accuracy and robustness; besides, the FFT correlation is helpful for the efficiency improvement and training with large image pairs. Experiments show that the proposed method can achieve a pixel-level matching error.

Video synthetic aperture radar (VideoSAR) provides the potential of high-resolution target images with also high frame rate. Different from the target signal in VideoSAR, target shadow straightforwardly indicates its location and track without suffering from the Doppler shift and smearing induced by the motion modulation. The target shadow can be exploited to realize the real-time tracking of the moving target with VideoSAR. In this letter, we propose a robust single target shadow tracker, which integrates the discriminative correlation filter(DCF) and interacting multiple model probabilistic data association filter(IMM-PDAF) to simultaneously overcome the interference of around clutter and target maneuvers. Considering the dependence between target shadow appearance and its dynamic motion, a filtering template updating strategy based on the estimated mode probability was proposed. VideoSAR tracking experiment shows that the proposed algorithm outperforms conventional visual tracking techniques in the robustness aspects.

Limited by the total length, the total number of the antenna units as well as their topology, the radar images always suffered from the sidelobe/grating lobe which severely impacts the quality of the radar images. In this article, a convolutional neural network (CNN)-based radar image-enhancing method is proposed. Using the original radar images as the input samples and using their corresponding ideal radar images with no sidelobe/grating lobe as the label to train the CNN. A well-trained CNN can suppress the sidelobe/grating lobe in the radar images. The structure of the specific CNN, the generation methods of the samples and the labels, the training procedure of the CNN, as well as some other detailed implementation strategies are specifically illustrated in this article. The proposed method is utilized to suppress the sidelobe/grating lobe in both the simulated and real recorded radar images. Compared to other existing methods, the proposed method is with better sidelobe/grating lobe suppressing performance and better robustness.

The ultrahigh frequency ultra-wideband synthetic aperture radar (UHF UWB SAR) has the well foliage penetrating and high-resolution imaging, which can be used to detect the concealed area under the foliage in forests. This paper presents an airborne UHF UWB SAR experiment and imaging results. During the winter, an airborne campaign has been carried out in Shanxi Province in China, and the raw data was collected. In this experiment, the SAR system was integrated onboard a CESSNA-172 airplane. The antenna was fixed on the suspension arm of the right wing of the airplane, while the other part of the SAR system was placed on the back seat of this airplane. The experimental results have been obtained from the collected raw data, which proved the imaging performance of the airborne UHF UWB SAR system as well as the validity of the imaging method.

This letter is devoted to the layout reconstruction via the ultra-wideband (UWB) through-wall imaging radar under one single observation and simultaneously takes account of the real-time human indication. In the proposed framework, layout reconstruction is taken as the preprocessing, where a coherent processing interval consisting of several successive received echoes in the initial stage is first employed to construct a range-Doppler (RD) spectrum. Then, in the RD spectrum, a series of selected discrete Doppler frequency signals is used to form Doppler back projection (BP) images. Finally, in the Doppler BP image stack, we design a 3-D constant false alarm rate detector to extract the building layout. Once completed, the achieved layout as auxiliary information is fused with the simultaneous human indication. Through-wall experiments show that the proposed method can effectively extract the covered layout of multiple walls under one single view and accordingly provide strong support for the concealed human sensing.

The work presented in this letter focuses on the time-variant radio frequency interference (RFI) issue in the low-frequency ultra-wideband (UWB) radar. Different from many previous studies, we first analyze the characteristics of RFIs and scattered echoes in the slow-time dimension and take advantage of overlapped short-time Fourier transform to adapt to the time-variant RFIs and update the frequency Doppler spectrum. Then, in the frequency Doppler spectrum, we adopt the minimum statistic combined with 1-D cell-averaging constant false alarm rate to estimate and separate the RFI power spectrum from the scattered echoes based on their differences. Finally, to mitigate the estimated RFIs, a suboptimal filter controlled by the defined entropy of range profiles after math filtering is designed. Employing a UWB radar, different experiments were conducted, and results verify the proposed method.

In through-wall indication of a moving human target in enclosed structures, a shadow effect because of the human target blocking parts from illumination on the back wall will emerge, referred to as a “ghost” in indication results. The shadow ghost moves as the human target does, which makes causal change detection (CD) invalid to separate them. To mitigate the shadow ghost, we analyze its differences from the moving human target. Based on the difference that the illumination is only blocked in partial channels of the multiple-input–multiple-output (MIMO) array while target echoes exist in most channels and the fact that shadow ghosts overlap more between successive indication results than the imaged targets as a result of their larger size, we proposed a mitigation method including a coherence factor and noncausal CD processing. Through-wall experiments via a MIMO through-wall imaging radar validate the proposed method.

Through-wall indication of human targets is highly desired in many applications. Generally, human targets behind wall are noncooperative, and rare prior knowledge about the circumstance behind wall could be available. Thus, it requires the ability to indicate human targets with different motions from clutters. To investigate this problem, we first examine the conventional time-domain indication methods, and find that their performances are controlled by the historical pulse number adopted to estimate background, which corresponds to the tap-length from the angle of filter. Then, based on an intermittent mode of human target echoes, we define the optimum tap-length as the shortest tap-length that makes the filter output signal-to-clutter-and-noise ratio reach maximum and develop an adaptive indication method with a gradient tap-length control scheme to search the optimum tap-length. Finally, through-wall experiments with an impulse through-wall radar demonstrate that the proposed method can obtain a good adaptive indication performance on human target with different motions.