Owing to complex environments and small radar cross section of target, traditional methods based on statistical models and manually designed features struggle to achieve effective detection of weak targets with low signal-to-clutter ratios. Deep learning-based radar target detection algorithms, by virtue of their data-driven working paradigm, can achieve favorable detection performance. However, radar target echoes often exhibit dynamic variations, making it difficult for existing methods to effectively capture target information. To address this issue, we propose an attention-based adaptive multi-domain joint radar target detection algorithm for adaptively extracting dynamic target information. The proposed method adjusts the network’s attention to target in the Doppler domain through a channel attention mechanism, and captures the temporal dynamic features of target in the pulse domain via a self-attention mechanism. Meanwhile, a local peak module is adopted to enhance the contrast between local target and clutter, thereby further improving the network performance. The effectiveness of the proposed method is validated using two real-recorded datasets.
To enhance the performance of downstream tasks such as land-cover classification and target monitoring, robust fusion of Synthetic Aperture Radar (SAR) and optical imagery has become increasingly important. However, traditional fusion frameworks primarily emphasize spectral fidelity and structural preservation, while offering limited capability in suppressing the inherent speckle noise in SAR images. They also lack effective mechanisms to handle degradation factors such as cloud occlusion commonly present in optical remote sensing. Recent advances in deep learning have demonstrated strong feature representation capabilities in other multi-modality fusion tasks, offering useful insights for remote sensing fusion. Nevertheless, existing methods are generally designed to maximize information preservation, making it difficult for their feature extraction modules to distinguish high-frequency noise from true structural details. Moreover, these methods fail to account for cloud-induced degradation, leading to severe performance deterioration under realistic conditions where SAR speckle and optical occlusion coexist. To address these challenges, we propose an end-to-end SAR–optical fusion network tailored for remote sensing applications. The core contributions of this work are twofold. First, we design a Sparse Feature Extraction Module (SFEM) as the core building block of the encoder, which integrates Sparse Self-Attention (SSA) with a Hybrid-Scale Feed-forward Network (HSFN) to extract stable and discriminative multi-scale features from degraded inputs. Second, we introduce a degradation-aware data augmentation strategy that simulates realistic cloud occlusion, forcing the model to learn stronger cross-modal complementarity and significantly improving gener-alization under complex conditions. Comprehensive comparisons against four state-of-the-art fusion algorithms on the public WHU-OPT-SAR dataset demonstrate that our method achieves superior qualitative and quantitative performance. Even under severe degradation, the proposed network effectively maintains spectral consistency and structural integrity, yielding visually clear and information-rich fused images.
This work presents an anomaly detection method based on the Rosenblatt Transformation (RT) for identifying nonhomogeneous range bins in radar data cubes. The algorithm estimates the empirical probability density function from training data and uses the corresponding cumulative distribution function to map a test data vector to independent and identically distributed samples with a uniform distribution. Deviations from uniformity in the transformed variables indicate statistical dissimilarity, allowing the test sample to be flagged as anomalous. We evaluate the method using high-fidelity radar data generated by RFView®, applying the Kullback-Leibler Divergence (KLD) to measure departures from uniformity. A sharp increase in KLD values indicates the onset of nonhomogeneity, consistent with visual observations from range-Doppler and range-angle maps. To validate the method, we estimate the sample covariance matrix using the predicted homogeneous bins and compare it to a ground-truth covariance matrix. Results show that the RT-based approach reliably identifies homogeneous regions and improves covariance estimation in heterogeneous radar environments.
Detecting a target signal in noise is a central problem in radar signal processing. The Normalized Matched Filter (NMF) is widely used, typically evaluated on a discrete grid of target parameters. In practice, targets rarely align with grid points, reducing detection performance. The off-grid NMF addresses this by maximizing over all possible parameters, improving sensitivity to off-grid targets. Building on recent results establishing the relationship between the threshold of the off-grid test and the false alarm probability, we investigate properties of the off-grid NMF that does not seem to be well known in the literature and illustrate through simulations that it exhibits an almost constant false alarm rate in realistic radar scenarios.
Selection of a sequence of pulse repetition frequencies (PRFs) is core to the design of pulse-Doppler radars seeking to detect and decode targets that are ambiguous in range or Doppler. By varying the PRF over multiple coherent processing intervals (CPIs), the radar can decode ambiguous targets and enhance visibility. Pulse sequence selection is constrained by radar system hardware and operating conditions. This paper draws a connection between pulse repetition sequence selection and the problem of finding cliques in a graph. The Bron-Kerbosch algorithm is then shown to efficiently identify mutually compatible sets of PRFs.
