The integration of Low-Earth Orbit (LEO) satellites with hypersonic aircraft presents a critical frontier for next-generation non-terrestrial networks (NTN). However, the ultra-high dynamics inherent in these scenarios introduce severe Doppler frequency shifts and time-varying delays, significantly degrading link reliability. To address these challenges, this article proposes a robust Orthogonal Time Frequency Space (OTFS) communication framework tailored for hypersonic-LEO links. First, we establish a three-dimensional (3D) kinematic model based on the Earth-Centered Inertial (ECI) coordinate system. This model accurately characterizes the dynamic delay-Doppler (DD) channel features across the aircraft’s ascent, cruise, and descent phases. Guided by these channel insights, we develop a phase-adaptive dynamic pilot design scheme that jointly optimizes pilot position, pilot power, pilot quantity, and guard intervals to maximize spectral efficiency and estimation accuracy. Furthermore, leveraging the temporal correlation of the OTFS channel in the DD domain, we propose a Multi-Frame Joint Cross-Correlation Matching (MF-JCM) channel estimation algorithm. This algorithm aggregates multi-frame channel responses to suppress noise and enhance detection probability. Simulation results demonstrate that the proposed scheme achieves a Bit Error Rate (BER) of 10−4at an signal-to-noise ratio (SNR) of 7.5 dB in the ascent phase and 11 dB in the descent phase. Hardware experiments based on Software-Defined Radio (SDR) further validate the feasibility and superiority of the proposed approach in practical ultra-high dynamic scenarios.

Data-assisted channel estimation (DA-CE) faces unique challenges in Low Earth Orbit (LEO) scenarios with large Doppler, where phase distortion and pilot sparsity degrade the utility of data symbols for refinement. Moreover, myopic, context-agnostic reliability metrics often fail to identify data symbols that are truly useful for improving estimation accuracy. To resolve these challenges, we innovatively reformulate the data selection as a pixel-level utility masking problem, similar to semantic segmentation tasks, where each resource element (RE) is evaluated for its utility in channel refinement. We introduce an integrated framework rooted in physics-informed reinforcement learning (PIRL), addressed by two symbiotic components. First, a Physics-informed Subspace Basis Expansion Models-LMMSE (PiSBEM-LMMSE) algorithm acts as the perception layer of our framework, which yields a high-fidelity, physics-consistent state representation by expressing the dominant Doppler-induced dynamics via a low-rank complex-exponential basis and decoupling them from residual stochastic fading. Second, a lightweight U-Net-based deep reinforcement learning (DRL) agent, acting as the cognitive decision core, learns an optimal, context-aware masking policy upon this structured representation. The U-Net architecture, with its encoder-decoder structure and skip connections, is specifically chosen to process the multi-channel, image-like state representation, capturing both global channel dynamics and local perturbations. Extensive simulations demonstrate that our framework achieves significant performance improvements over state-of-the-art methods, exhibiting remarkable robustness in LEO scenarios where conventional approaches fail.

Non-terrestrial networks (NTNs), connecting components such as low-earth orbit (LEO) satellites, exhibit highly dynamic behaviors in both their topological configurations and radio channels. Traditional pilot-aided channel estimation (PACE) methods are susceptible to significant Doppler shifts and rapid channel variations, limiting further enhancements in data transmission rates or incurring unaffordable pilot overheads. Data-aided channel estimation (DACE), as a potential improvement path, leverages data vectors to supplement pilot signals. Nevertheless, conventional DACE faces difficulties in effectively selecting data vectors under frequency offsets and rapid channel variations. To address this, this paper introduces a dynamic threshold-based scheme for selecting reliable data vectors, which leverages an adaptive threshold to precisely identify reliable data vectors within a frame. Furthermore, we propose a data vector deviation estimation method as a preprocessing step to correct the overall offset of data vectors, thereby ensuring the efficacy of data selection even amidst significant channel variations. Our approach significantly elevates the performance of LEO links within current protocols and is compatible with higher modulation orders and frequency bands. Simulation results validate the effectiveness of the proposed scheme.

Abstract Anomaly detection is a challenging task, especially detecting and segmenting tiny defect regions in images without anomaly priors. Although deep encoder‐decoder‐based convolutional neural networks have achieved good anomaly detection results, existing methods operate uniformly on all extracted image features without considering disentangling these features. To fully explore the texture and semantic information of images, A novel unsupervised anomaly detection method is proposed. Specifically, discriminative features are extracted from images by using a deep pre‐trained network, where shallow and deep features are aggregated into texture and semantic modules, respectively. Then, a feature fusion module is developed to interactively enable feature information in two different modules. The texture and semantic segmentation results are obtained by comparing the texture features and semantic features before and after reconstruction, respectively. Finally, an anomaly segmentation module is designed to generate anomaly detection results by integrating the results of the texture and semantic modules by setting a threshold. Experimental results on benchmark datasets for anomaly detection demonstrate that our proposed method can efficiently and effectively detect anomalies, outperforming some state‐of‐the‐art methods by 2.7% and 0.6% in classification and segmentation.