This paper presents an integrated genetic algorithm-based method for detecting, focusing, imaging, and parameter estimation of small, weak airborne targets (e.g., drones, stealth aircraft) using synthetic aperture radar (SAR). The method improves the signal-to-clutter ratio through preprocessing (range compression, DPCA clutter suppression, downsampling) and narrows the optimization space via a fixed-point search for target region and velocity. A genetic algorithm is then used to optimize motion speed estimation, minimizing Shannon entropy of BP imaging results. Simulation results show a 75% improvement in search efficiency, with comparable accuracy and better error distribution and stability compared to traditional methods. This method offers an effective solution for real-time detection of airborne targets.

Small object detection in unmanned aerial vehicle (UAV) imagery remains a critical challenge in computer vision. To address the low pixel coverage of small objects and the strong coupling between foreground and background in UAV imagery, we propose MCF-DETR, a lightweight model leveraging multi-scale feature fusion. First, we introduce DCH-ResNet18, a backbone network based on the improved Dual-Channel Heterogeneous Convolution, which decouples foreground–background information while extracting multi-level gradient features. Second, we develop a composite encoder with three novel modules: (1) Partial Single-Head Attention, enhancing feature representation by modeling object-context relationships; (2) Spatial-Pooling Down-sampling, reducing computation via spatial decoupling and enhancing high-frequency features with max pooling; and (3) Regional Dynamic Calibration, reconstructing spatial features to improve small-object perception. On the VisDrone2019, MCF-DETR achieves 53.2% fewer parameters and 37.3% fewer GFLOPs than RT-DETR, alongside a 1.2% AP50 gain. Additionally, its generalization capability on UAV imagery is further validated via the CODrone. Ultimately, by constructing a lightweight detection framework capable of decoupling foreground and background, this work achieves an optimal balance between localization accuracy and deployment efficiency, offering a practical and effective solution for resource-constrained UAV applications.

This article proposes a novel detection framework based on pixel-level temporal classification and spatio-temporal association filtering. Addressing the challenge of distinguishing target shadows from background speckle under low signal-to-clutter ratio (SCR) conditions in video SAR, we introduce a low-rank sparse decomposition technique utilizing weighted nuclear norm minimization. This effectively enhances the SCR of faint moving targets. Building on this foundation, we design an integrated scheme combining a dilated-attention bidirectional long short-term memory network with a physics-informed spatio-temporal constrained graph neural network. This approach precisely captures pixel amplitude transition patterns in the temporal dimension, enabling robust pixel-level classification. Notably, by incorporating physical rule constraints, the proposed method automatically eliminates isolated false alarms in non-Euclidean space, overcoming the limitations of traditional Euclidean distance-based association algorithms in dense clutter. Experiments on real-world airborne video SAR data validate the effectiveness and robustness of the proposed method.

Although context windows have expanded significantly in recent years, hallucinations in long-context summarization remain a challenge. Long novels are better suited than news or papers for researching these hallucinations, due to their intrinsic information and detailed descriptions of events and dialogues. However, current research lacks a multi-scale benchmark for hallucination detection in long-context novel summarization and does not fully explore how hallucinations change as the context grows longer. In this study, we propose LongNovel, a multi-scale long-context bilingual (Chinese and English) novel benchmark for hallucination detection. This benchmark is constructed from 29 Chinese novels (ranging from 16k to 100k tokens) and chapter-level data from the BookSum dataset. We design 8 hallucination types and employ a combination of Multi-Model Arbitration and Entity-Referenced Hallucination Generation to ensure both data authenticity and a balanced distribution of hallucination categories. Furthermore, we manually revise the content in the test set to guarantee data reliability. Extensive experimental results demonstrate that LongNovel is a challenging benchmark. We release LongNovel for future research. https://github.com/BDML-lab/LongNovel

To enhance bearing fault diagnosis performance under various operating conditions, this paper proposes a hybrid approach based on generative adversarial networks (GANs), transfer learning, wavelet transform time-frequency representations, asymmetric convolutional networks, and the multi-head attention mechanism (MAC-MHA). Firstly, GANs are utilized to generate new bearing fault data to meet the model’s training requirements. Then, wavelet transform is applied to convert the bearing vibration signals into time-frequency representations, capturing the temporal evolution of frequency components. Next, an improved asymmetric convolutional network (MAC-MHA), combined with the multi-head attention mechanism, is employed to enhance the focus on key time-frequency features, further improving fault diagnosis accuracy. Considering the differences in operating conditions, transfer learning techniques are applied to facilitate knowledge transfer from the source domain to the target domain, thereby enhancing the model’s generalization ability. Experimental results demonstrate the effectiveness and robustness of the proposed method under various operating conditions. Finally, the proposed hybrid fault diagnosis approach is validated using the PADERBORN and CWRU datasets.

The properties of wood change after heat treatment, affecting its applications. Glossiness, a key aesthetic property, is of great significance in fields like furniture. Precise prediction can optimize the process and improve product quality. Although the traditional back propagation neural network (BPNN) has been applied in the field of wood properties, it still has issues such as poor prediction accuracy. This study proposes an improved whale optimization algorithm (IWOA) to optimize BPNN, constructing an IWOA-BPNN model for predicting the glossiness of heat-treated wood. IWOA uses chaos theory and tent chaos mapping to accelerate convergence, combines with the sine cosine algorithm to enhance optimization, and adopts an adaptive inertia weight to balance search and exploitation. A dataset containing 216 data entries from four different wood species was collected. Through model comparison, the IWOA-BPNN model showed significant advantages. Compared with the traditional BPNN model, the mean absolute error (MAE) value decreased by 66.02%, the mean absolute percentage error (MAPE) value decreased by 64.21%, the root mean square error (RMSE) value decreased by 69.60%, and the R2 value increased by 12.87%. This model provides an efficient method for optimizing wood heat treatment processes and promotes the development of the wood industry.

Autoencoders (AEs) have received extensive attention in hyperspectral anomaly detection (HAD) due to their capability to separate the background from the anomaly based on the reconstruction error. However, the existing AE methods routinely fail to adequately exploit spatial information and may precisely reconstruct anomalies, thereby affecting the detection accuracy. To address these issues, this study proposes a novel Multi-scale Autoencoder Suppression Strategy (MASS). The underlying principle of MASS is to prioritize the reconstruction of background information over anomalies. In the encoding stage, the Local Feature Extractor, which integrates Convolution and Omni-Dimensional Dynamic Convolution (ODConv), is combined with the Global Feature Extractor based on Transformer to effectively extract multi-scale features. Furthermore, a Self-Attention Suppression module (SAS) is devised to diminish the influence of anomalous pixels, enabling the network to focus more intently on the precise reconstruction of the background. During the process of network learning, a mask derived from the test outcomes of each iteration is integrated into the loss function computation, encompassing only the positions with low anomaly scores from the preceding detection round. Experiments on eight datasets demonstrate that the proposed method is significantly superior to several traditional methods and deep learning methods in terms of performance.

To address the challenging problem of multi-scale inshore–offshore ship detection in synthetic aperture radar (SAR) remote sensing images, we propose a novel deep learning-based automatic ship detection method within the framework of compositional learning. The proposed method is supported by three pillars: context-guided region proposal, prototype-based model-pretraining, and multi-model ensemble learning. To reduce the false alarms induced by the discrete ground clutters, the prior knowledge of the harbour’s layout is exploited to generate land masks for terrain delimitation. To prepare the model for the diverse ship targets of different sizes and orientations it might encounter in the test environment, a novel cross-dataset model pretraining strategy is devised, where the SAR images of several key ship target prototypes from the auxiliary dataset are used to support class-incremental learning. To combine the advantages of diverse model architectures, an adaptive decision-level fusion framework is proposed, which consists of three components: a dynamic confidence threshold assignment strategy based on the sizes of targets, a weighted fusion mechanism based on president-senate check–balance, and Soft-NMS-based Dense Group Target Bounding Box Fusion (Soft-NMS-DGT-BBF). The performance enhancement brought by contextual knowledge-aided terrain delimitation, cross-dataset prototype-based model pretraining and check–balance-based adaptive decision-level fusion are validated with a series of ingeniously devised experiments based on the FAIR-CSAR-Ship dataset.

This study integrates the Backpropagation (BP) Neural Network with several optimization algorithms, namely Hippopotamus Optimization (HO), Parrot Optimization (PO), Osprey Optimization Algorithm (OOA), and Goose Optimization (GO), to develop four predictive models for the adhesive strength of heat-treated wood: HO-BP, PO-BP, OOA-BP, and GO-BP. These models were used to predict the adhesive strength of the wood that was heat-treated under multiple variables such as treatment temperature, time, feed rate, cutting speed, and abrasive particle size. The efficacy of the BP neural network models was assessed utilizing the coefficient of determination (R2), error rate, and CEC test dataset. The outcomes demonstrate that, relative to the other algorithms, the Hippopotamus Optimization (HO) method shows better search efficacy and convergence velocity. Furthermore, XGBoost was used to statistically evaluate and rank input variables, revealing that cutting speed (m/s) and treatment time (hours) had the most significant impact on model predictions. Taken together, these four predictive models demonstrated effective applicability in assessing adhesive strength under various processing conditions in practical experiments. The MAE, RMSE, MAPE, and R2 values of the HO-BP model reached 0.0822, 0.1024, 1.1317, and 0.9358, respectively, demonstrating superior predictive accuracy compared to other models. These findings support industrial process optimization for enhanced wood utilization.

With the development of infrared technology, infrared small target detection (IRSTD) is widely applied in fields such as environmental monitoring, marine rescue, and forest fire prevention. Existing IRSTD methods are often based on spatial domain approaches, which preserve target features in the spatial domain but overlook the characteristics of infrared small targets in the frequency domain. In frequency domain methods, infrared small targets are typically considered as high-frequency components, while the continuous background is regarded as low-frequency components. However, infrared small targets often have complex backgrounds, strong edges, and noise generated during imaging, all of which are also reflected as high-frequency components, leading to false detections. To overcome this issue and fully explore the potential of IRSTD in the frequency domain, we propose a novel network, SFDTNet, which integrates frequency-domain attention and U-Structure for IRSTD. In the encoding phase, spatial feature extraction is applied to the infrared small target. In the decoding stage, global-scale spatial features are modeled in the frequency domain to achieve more precise reconstruction of small targets while reducing the interference of background high-frequency clutter. Frequency domain self-attention (FDSA) introduces an attention mechanism to model global information in the frequency domain and capture the importance of different frequency components. Adaptive frequency selection network (AFSN) incorporates learnable masks to adaptively modulate high- and low-frequency components in the frequency domain. Finally, a deep supervision strategy is employed to help the network learn features more effectively. Experimental results demonstrate that it effectively retains the shape and contours of small targets while achieving a very low false detection rate. Compared with existing state-of-the-art methods, our approach shows superior performance and better robustness.

This paper addresses limitations of traditional ocean current retrieval by adapting the M4S model for dual-antenna squinted SAR. A virtual front-and-side-looking system processes ATI-SAR data for simulation, validated by error analysis between retrieved and standard flow fields.

Aiming at the current problems of small-target defect detection difficulty, weak resistance to complex background interference, and low detection precision in insulator defect detection, an insulator defect detection model, SnakeNet, is proposed based on the YOLOv8 detection framework. Firstly, the Dynamic Snake Convolutional Neck (DSConv-Neck) is proposed to extract rich fine-grained features and improve the detection ability of small-target defects. Then, the background interference is suppressed by the C2f Efficient Multi-Scale Attention (C2MA) module to enhance the anti-background resistance; finally, the GhostConv VanillaNet (GCVNet) module is proposed to minimize the number of model parameters while ensuring the detection precision. Many experiments are carried out on two public datasets, the China power line insulator dataset (CPLID) and the Insulator Defect Image Dataset (IDID). The results show that the optimal detection precision and recall of SnakeNet are 98.7% and 98.8%, respectively, and the number of model parameters and computation volume are only 2.9M and 6.7G. Compared with the original YOLOv8, the detection precision and recall are improved by 2.4% and 3.5%, respectively, and the number of model parameters and computation volume are reduced by 0.3M and 1.3M, respectively. Compared with other algorithms, such as SSD and RT-DETR, SnakeNet performs well in detecting small targets and adapting to complex backgrounds with lower computational costs.

Synthetic aperture radar (SAR) can detect moving targets on the ground/sea, and high-resolution imaging on the ground/sea has critical applications in both military and civilian fields. This paper attempts to use a spaceborne SAR system to detect and image moving targets in the air for the first time. Due to the high velocity of aerial targets, they usually appear as two-dimensional range and azimuth direction defocus in SAR images, and clutter will also have a profound impact on target detection. To solve the above problems, a method of detecting and focusing on a spaceborne SAR target based on a two-dimensional velocity search is proposed by combining the BP algorithm. According to the current environment of the aerial target and the number of system channels, the clutter suppression methods are set and combined with two-dimensional velocity search with different precision, the Shannon entropy under different search velocity groups is used to obtain the search velocity group closest to the actual velocity and realize the integrated processing of moving target detection–focused imaging parameter estimation. Combined with simulation data, the effectiveness of the proposed method is verified.

In recent years, the detection performance of SAR-GMTI (synthetic aperture radar-ground moving target indication) algorithm based on deep learning has always been limited by insufficient measured data due to the heavy operation complexity and high cost of real SAR systems. To solve this problem, this paper proposes an overall DT-based implementation framework for SAR ground moving target intelligent detection tasks. In particular, by virtue of a SAR imaging algorithm, a high-fidelity twin replica of SAR moving targets is established in digital space through parameter traversal based on the prior target characteristics of the obtained measured datasets. Then, the constructed SAR twin datasets is fed into the neural network model to train an intelligent detector by fully learning features of the moving targets and preset the SAR scene in the twin space, which can realize the robust detection of ground moving targets in related practical scenarios with no need for multiple and complex field experiments. Moreover, the effectiveness of the proposed framework is verified on the MiniSAR measured system, and a comparison with traditional CFAR detection method is given simultaneously.

Combined with the synthetic aperture radar (SAR) imaging algorithm, diverse simulation sample sets of ground moving targets are constructed to tackle the problem of insufficient measured data in the SAR ground moving target indication algorithm based on deep learning. In view of this, a overall scheme of realizing robust detection of ground moving targets under varying backgrounds is conducted on the integration of the adaptive spatial location extraction network based on deformable module and the multichannel clutter suppression technology. In particular, a spatial deformable module is incorporated into the network to enhance its modeling capacity of the input targets with different shapes. Furthermore, the multichannel clutter suppression technology of airborne SAR is adopted to significantly mitigate the interference of complex background clutter. The effectiveness of the proposed method is verified on the simulation sample sets, and comparison with other detection methods is given simultaneously.

In this paper, an end-to-end system framework is proposed for the Digital Twin study of spaceborne ATI-SAR ocean current velocity inversion. Within this framework, a fitting inversion approach is proposed to enhance the conventional spaceborne ATI-SAR ocean current velocity inversion algorithm. Consequently, the issue of possible local inversion errors stemming from the mismatch between the traditional spaceborne ATI-SAR inversion algorithm and various dual-antenna configurations is resolved to a certain extent. A simulated spaceborne ATI-SAR system, featuring a dual-antenna configuration comprising a baseline direction perpendicular to the track and a squint angle, is presented to validate the efficacy of the Digital Twin methodology. Under the specified simulation parameters, the average inversion error for the final ocean current velocity is recorded at 0.0084 m/s, showcasing a reduction of 0.0401 m/s compared with the average inversion error prior to optimization.

A high precision estimation algorithm for ground moving targets in multi-channel wide-area surveillance ground moving target indication systems is proposed based on maximum likelihood method. The main concept of this novel algorithm is to estimate the azimuth angle of the detected targets using maximum likelihood method with the space steering vector formed by the estimated interferometric phase extracted from the mainlobe clutter region of the real data. Through this novel algorithm, the effect of channel errors among the multi-channels can be well reduced. Simulation experiments demonstrate the effectiveness of the proposed algorithm.

Low probability of intercept (LPI) radar has played an important role in in modern radar systems due to its difficulty in being intercepted by noncooperative interceptors. The modulation methods of LPI radar signals are becoming increasingly complex, causing some difficulties in accurately identifying radar signals in complex electromagnetic environments. To address this problem, we propose an automatic recognition framework for LPI radar modulated signals in complex multipath electromagnetic environments. Specifically, the radar time-domain signal is converted into a time-frequency distribution image through time-frequency analysis technology, and then, adaptive filtering is performed using an adaptive network-based fuzzy inference system in the preprocessing stage to enhance the time-frequency characteristics of the signal under low signal-to-noise ratios (SNRs). In the automatic recognition stage of radar signal, the inherent characteristics of time-frequency images of signals are extracted by exploiting deep learning, and a split-attention networks combined with a joint loss function is designed, namely ResNeSt. The experimental results show that compared with the existing automatic modulation recognition methods for radar signals, this framework has higher recognition accuracy in complex electromagnetic environments and exhibits robustness against superimposed multipath effects. When the SNR is as low as -12 dB, the average probability of accurately identifying 15 typical LPI radar signals is 94.93%.

In the field of multi-view SAR target location, the greater the difference in viewing angles, the higher the target location accuracy. However, this makes it difficult to match the same target between the SAR images with different viewing angles. The performance of traditional SAR image-matching algorithms will deteriorate or even fail to match the images correctly when the viewing angle is gradually increased. To solve this problem, a fast SAR matching method for the SAR images with large viewing angles based on inertial navigation information and neighborhood structure consensus (ININSC) is proposed in this paper. In this algorithm, the key targets are detected in the SAR images by using the maximum connected domain algorithm and the K-means clustering algorithm, and the connected domain centroid of the target is taken as a feature point. The approximate position of the key targets after the viewing angle change is found through inertial navigation information, and then accurate and fast matching is achieved by using the consensus of the neighborhood topological structure of the key targets. The measured data sets demonstrate that compared with traditional SAR image-matching algorithms, the proposed ININSC algorithm solves such a problem that SAR images cannot be accurately matched under the differences at large viewing angles and has better robustness and timeliness.

Short-term photovoltaic power forecasting (PVPF) is crucial in the scheduling and functioning of contemporary electrical systems. A short-term PVPF model based on correlation analysis, similar day clustering, mode decomposition and hybrid deep learning is proposed to address the volatility and stochasticity of PV output. Screening of major meteorological factors affecting photovoltaic output using the Kendall rank correlation coefficient method. Considering the influence of seasonal factors, the historical PV power data were divided by season as the initial dataset for clustering. Subsequently, based on the main meteorological factors screened, the K-means++ clustering method was used for similar day clustering, and the initial dataset was classified into sunny, cloudy, and rainy days, and historical data with the same weather type as the predicted day were used as training samples to train the model. The PV power sequence is then decomposed into intrinsic mode functions (IMFs) at different frequencies by the improved variational mode decomposition (IVMD) to reduce its non-stationarity and complexity. Introducing the improved whale optimization algorithm (IWOA) for optimal hyperparameters of BiLSTM, and an Attention mechanism is incorporated for effective key information capture. Finally, the IVMD-IWOA-BiLSTM-Attention prediction model is constructed, and the prediction results of each IMF are summed and reconstructed to obtain the final PV power prediction results. Simulation results show that the proposed deep learning-based hybrid prediction model can obtain satisfactory prediction accuracy under different weather conditions and is significantly better than other models, which is of great significance for the decision-making scheduling of PV grid-connection as well as for enhancing the new energy consumption.