Millimeter-wave (mmWave) radar is a key non-contact sensor for vibration measurement. Under low signal-to-noise ratio (SNR) conditions, the coupled effects of noise and static clutter significantly distort the amplitude and phase of complex echoes, thereby hindering reliable micro-vibration detection. This study formulates low-SNR mmWave micro-vibration enhancement as a supervised in-phase and quadrature (IQ)-domain restoration problem and develops a Cascaded U-Net based restoration framework for recovering clean vibration-related IQ sequences from noise- and static-clutter-corrupted observations. The network learns an implicit IQ-domain enhancement mapping from noise- and static-clutter-corrupted observations to reference vibration-related IQ sequences, thereby reducing noise-induced fluctuations and mitigating the influence of static-clutter-induced phasor offsets under the training distribution. Validation is conducted using both numerical simulations and precision linear-stage experiments. The simulation results show that the Cascaded U-Net achieves better IQ-domain signal enhancement and vibration recovery performance than the compared baseline models, including the traditional U-Net, residual U-Net (ResUNet), Transformer, and long short-term memory (LSTM). In particular, the proposed method provides lower reconstruction errors under very low-SNR and small-amplitude vibration conditions, indicating its stronger capability in recovering weak vibration-related phase variations from noise- and static-clutter-corrupted observations. Experiments with real measurements further support the simulation results. In three linear-stage experiments, the root mean square error (RMSE) is reduced by 11.54%, 10.94%, and 9.09%, respectively, compared with the traditional U-Net. These results indicate that the cascaded architecture provides a more effective IQ-domain restoration framework for weak micro-vibration signal enhancement in mmWave radar measurements.

Three-dimensional reconstruction with compact millimeter-wave (mmWave) radar is challenging due to sparse measurements and limited angular resolution. DART (Doppler Aided Radar Tomography) addresses this problem by exploiting ego-motion-induced Doppler diversity to learn a continuous neural radar field and synthesize range-Doppler (RD) observations from novel viewpoints. However, when applying DART to selfcollected radar data, we observe geometric artifacts whose origins are not directly revealed by the implicit reconstruction pipeline or the final RD reconstruction loss. To investigate this issue, we develop a simulation-based analysis method tailored to DARTlike implicit Doppler tomography. Using fixed RD observations, we independently perturb the position, velocity, and orientation priors and further analyze their coupled effect under odometry drift. The controlled results show distinct degradation patterns and, under the tested perturbations, reveal a pronounced Doppler misalignment when the velocity prior is erroneous. We further show that this error affects not only the Doppler projection but also the Doppler-constrained 3-D sampling support, causing reconstruction gradients to be assigned to incorrect spatial locations. Based on these findings, we propose DART-VBE, a frame-wise velocity-bias estimation framework that introduces a learnable speed correction into the differentiable DART rendering pipeline before RD sampling. To make the correction identifiable and better decouple motion-induced errors from scene-field parameters, we combine the original RD reconstruction loss with a Doppler centroid alignment loss and a motion regularization term. Experiments on public and self-collected datasets show that DART-VBE generally improves RD synthesis quality, reduces Doppler centroid error and Chamfer Distance, and produces more consistent 3-D radar reconstructions than the original DART baseline. Speed-bias simulation studies further show that the learned correction provides a physically meaningful estimate of the injected velocity bias.

Systematically characterizing the time series deformation evolution of open-pit mine slopes is key to revealing their potential instability development and supporting subsequent deformation-level classification. Interferometric Synthetic Aperture Radar (InSAR), by enabling measurement of ground deformation at a global scale approximately every ten days, may hold the key to those interactions. However, atmospheric propagation delays still have a significant impact on deformation calculations, and open-pit mine slopes monitored by InSAR often suffer from low coherence. This noise can obscure nonlinear and transient precursory signatures in deformation time series, reducing the identifiability of key temporal patterns required for automated interpretation. Here, we present a Coherence-conditioned Encoder–Decoder Long Short-Term Memory (CED-LSTM) denoising network for deformation time series. We generate a physics-aware synthetic dataset by modeling coherence-dependent measurement noise and temporally correlated atmospheric delays. The network jointly models deformation time series and coherence, using residual learning and adaptive gated composite loss to preserve deformation trends. It is designed to autonomously extract ground deformation signals from noise in InSAR time series without prior knowledge of where deformation occurs or how it evolves. On the synthetic validation set, the network achieved a root mean square error (RMSE) of 2.2 mm across the validation sequences. Applied to three InSAR datasets over an open-pit mine from March 2019 to March 2022, denoising suppresses noise and stabilizes deformation boundaries, enabling extraction of trend and transient indicators and a data-driven deformation-level score. Using quantile-based thresholds, these scores are then used to produce multi-year deformation-level classification maps.

In the monitoring of discontinuous ground-based synthetic aperture radar (GB-SAR), challenges such as repositioning error and atmospheric phase screen (APS) can significantly impact the accuracy of deformation inversion. Existing compensation methods are limited to specific scanning modes (linear-scanning or arc-scanning) and lack a unified framework, leading to suboptimal performance in complex scenarios. We propose a novel joint compensation model applicable to both linear-scanning and arc-scanning GB-SAR. By formulating repositioning error as ternary functions of positional shifts and linearizing them through first-order approximation, the method establishes a unified phase error model. A high-order range error component is integrated to characterize APS effects. The combined model parameters are optimized by gradient descent. Experimental validation using near-field and far-field datasets demonstrates significant improvements: in linear-scanning mode, the residual phase RMSE is reduced by 40.4%, while in arc-scanning mode, it decreased by 6.8%. The proposed framework effectively compensates for two errors, outperforming conventional approaches by unifying compensation across scanning geometries. This study enables high-precision deformation monitoring in diverse GB-SAR applications, advancing the reliability of geological hazard early warning and infrastructure assessment.

To improve the fusion modeling capability and prediction accuracy of multimodal marine meteorological data in complex environments, this paper proposes a cross-scale fusion method for multimodal marine meteorological data (MMAF). First, to address the problem of inconsistent resolution of data from different modalities, a cross-scale feature consistency modeling mechanism is proposed. Through multi-resolution feature extraction and scale normalization processing, the modeling capability of heterogeneous modalities in a unified feature space is achieved. Secondly, to address the inconsistency in feature distribution of multimodal data, a multimodal alignment method combining maximum mean difference and adversarial learning is constructed. Through global distribution matching and modal discrimination confusion, it guides the efficient alignment of multimodal data in a shared embedding space. Finally, to address the possible redundancy, conflict and missing problems between modal information, a multimodal data anomaly perception and processing mechanism is designed to achieve robust multimodal fusion expression. Experimental results demonstrate that the proposed method can effectively improve alignment and fusion accuracy, and reduces the MSE index by approximately 15.35

Tianjin has long been plagued by land subsidence, which is primarily caused by groundwater over-exploitation and poses major threats to urban construction, economic development, and infrastructure safety. This study aims to explore the relationship between groundwater storage (GWS) changes and land subsidence in Tianjin by combining GRACE and InSAR technologies. To clarify this relationship macroscopically, GRACE and GLDAS data were used to analyze GWS changes in the study area from January 2023 to February 2025, while PS-InSAR was applied to obtain concurrent land subsidence information. Time-lagged cross-correlation analysis was introduced to quantify the correlation and lag time between GWS changes and land subsidence time series. Results showed: (1) GWS was positive in most areas (maximum annual rate 59.13 mm/yr, concentrated in east-central Jinghai, south-central Xiqing, Jinnan and central Binhai New Area), decreasing from the center to north and south, with a significant decline (-9.22 mm/yr) in northern Binhai New Area; severe land subsidence showed a "peripheral aggregation" pattern, with 4 contiguous zones, 7 funnels and a max rate of 67.62 mm/yr (near Yangjiapo Town, Binhai New Area); (2) The cross-correlation coefficient between GWS changes and land subsidence ranged from 0.257 to 0.882 (0.542-0.853, 0.431-0.826, 0.257-0.882, 0.538-0.846 for Zones I-IV respectively), and land subsidence lagged GWS changes by 41-263 days overall. This study confirms the link and lag effect between GWS and land subsidence in Tianjin, providing scientific guidance for subsidence prevention and early warning.

Millimeter-wave traffic surveillance radars can detect moving vehicles and measure their speed, playing a crucial role in traffic supervision applications. However, these radars cannot image moving vehicles. Integrating radar imaging technology enables the acquisition of vehicle shape information for classification and recognition, thereby boosting their functional capability. Inverse synthetic aperture radar (ISAR) achieves two-dimensional high-resolution imaging of moving targets, yet traditional ISAR methods suffer from high algorithm complexity, heavy computational load and poor real-time performance, making them unsuitable for time-sensitive traffic scenarios. Given that existing traffic surveillance radars operate in high-frequency bands, sub-aperture-based processing maintains high resolution while reducing imaging algorithm complexity. Thus, this paper proposes a fast imaging method for moving vehicles in traffic surveillance radars based on the second-order Keystone transform. It first analyzes and constructs the radar’s observation geometry and sub-aperture signal model, then defines sub-aperture segmentation parameters with moving target imaging focus quality as the criterion. The method processes received echo data frame by frame to extract the target azimuth angle corresponding to each pulse, implements sub-aperture segmentation, adopts the second-order Keystone transform for range cell migration (RCM) correction within each sub-aperture, and performs range Inverse fast fourier transform (IFFT), azimuth fast fourier transform (FFT), and polar-to-Cartesian coordinate conversion to obtain 2D high-resolution vehicle images. The effectiveness of the proposed method is verified via moving vehicle point-target simulations and real data experiments.

Arc scanning synthetic aperture radar (ArcSAR) can achieve high-resolution panoramic imaging and retrieve submillimeter-level deformation information. To monitor buildings in a city scenario, ArcSAR must be lightweight; have a high resolution, a mid-range (around a hundred meters), and low power consumption; and be cost-effective. In this study, a novel high-resolution wide-beam single-chip millimeter-wave (mmwave) ArcSAR system, together with an imaging algorithm, is presented. First, to handle the non-uniform azimuth sampling caused by motor motion, a high-accuracy angular coder is used in the system design. The coder can send the radar a hardware trigger signal when rotated to a specific angle so that uniform angular sampling can be achieved under the unstable rotation of the motor. Second, the ArcSAR’s maximum azimuth sampling angle that can avoid aliasing is deducted based on the Nyquist theorem. The mathematical relation supports the proposed ArcSAR system in acquiring data by setting the sampling angle interval. Third, the range cell migration (RCM) phenomenon is severe because mmwave radar has a wide azimuth beamwidth and a high frequency, and ArcSAR has a curved synthetic aperture. Therefore, the fourth-order RCM model based on the range-Doppler (RD) algorithm is interpreted with a uniform azimuth angle to suit the system and implemented. The proposed system uses the TI 6843 module as the radar sensor, and its azimuth beamwidth is 64∘. The performance of the system and the corresponding imaging algorithm are thoroughly analyzed and validated via simulations and real data experiments. The output image covers a 360∘ and 180 m area at an azimuth resolution of 0.2∘. The results show that the proposed system has good application prospects, and the design principles can support the improvement of current ArcSARs.

Objective Image transmission through multimode fiber (MMF) is now widely used in medical imaging, biological tissue detection, communication technology, and other fields. In multimode fiber imaging, the light pulse carrying the spatial information of the object enters the multimode fiber, and thousands of transmission modes excited in the fiber form encoded spatial information. Due to the complex mechanisms of interference, coupling, self-phase modulation, and group velocity dispersion among the fiber modes, the exit end of the fiber eventually forms a speckle image. With the development of optical modulators and computational optics, the advantages of deep learning methods in image reconstruction have become increasingly prominent. The high operational efficiency and strong resistance to fiber disturbances have pushed MMF image transmission into practical applications. Most existing studies use the MNIST handwritten digit set (28x28 resolution) for both training and testing, which is insufficient to train the generalization ability of network models. This reliance on limited data reduces the practical performance of the models. To enhance the practical application of multimode fiber imaging, we propose a hybrid model--TMnn (Transmission Matrix and Neural Network), based on complex value operations and a neural network that incorporates the physical processes of multimode fiber light field modulation. The model is applied to train and verify different natural scene image datasets, and the results show that the model training speed is significantly improved while maintaining the quality of image reconstruction. At the same time, the generalization ability of the neural network is also enhanced in the image restoration task. Methods Combining the physical mechanisms of optical fiber and neural networks, we propose a multimode optical fiber speckle reconstruction algorithm, TMnn, based on complex value operations, which is trained on a natural scene image dataset. According to the response relationship between the input and output optical fields of multimode fiber, the inverse transmission matrix of the fiber is fitted using an iterative algorithm. The reconstruction optimization is then performed through a convolutional neural network to complete the speckle image reconstruction. The model is mainly divided into two modules. The first is the reconstruction module, which constructs the complex value deep neural network to fit the transmission matrix and initially reconstructs the images. The network consists of an input layer, complex convolution layers, complex batch normalization layers, and complex dense connection layers. The second part is the optimization module, which optimizes the initially reconstructed image by constructing a 3x3 convolutional neural network. The initial reconstructed image is taken as input, and the image features are extracted deeply. Image details are then reconstructed through the convolution layer, pooling layer, and fully connected layer in sequence. Results and Discussions By comparing with traditional neural networks (SCNN, DCNN), CANN (Complex Artificial Neural Network), and USINET, we confirm the advantages of the model in terms of reconstruction effect and training speed. In terms of model training, we make a comparison with CANN on the ImageNet dataset. Compared with CANN, the SSIM index shows a significant improvement, and the number of iterations is reduced by 200. However, the addition of two complex convolution layers increases the number of parameters, which has a certain effect on the training time cost. We also compare the model with USINET, and the training information is shown in Table 2. The results show that the average SSIM index of this algorithm improved by about 0.5%, and the training time is reduced by 6.46 hours. The TMnn model outperforms USINET in the first 4 hours of training and tends to converge after about 6 hours of training, with the SSIM value stabilizing at around 0.8. This indicates that the model constructed in this paper does not compromise training speed or model performance despite the complex operations. Conclusions We integrate the physical mechanism of optical fiber transmission with deep learning technology to construct a deep neural network based on complex-valued operations, which achieves high-quality reconstruction with an SSIM index above 0.7. Through the reconstruction of various datasets, the validity and generalization of the model are demonstrated. By comparing it with traditional neural networks, fully connected complex networks, and USINET, the advantages of the model are confirmed in terms of reconstruction quality and training speed. However, the network still has limitations in reconstructing more complex, detailed images. The network structure and model parameters for feature extraction need optimization to better capture detailed features and further enhance the quality of natural scene image reconstruction.

Coal mines play an important role in the global energy supply. Monitoring the displacement of open-pit mines is crucial to preventing geological disasters, such as landslides and surface displacement, caused by high-intensity mining activities. In recent years, multi-temporal Synthetic Aperture Radar Interferometry (InSAR) technology has advanced and become widely used for monitoring the displacement of open-pit mines. However, the scattering characteristics of surfaces in open-pit mining areas are unstable, resulting in few coherence points with uneven distribution. Small BAseline Subset InSAR (SABS-InSAR) technology struggles to extract high-density points and fails to capture the overall displacement trend of the monitoring area. To address these challenges, this study focused on the Shengli West No. 2 open-pit coal mine in eastern Inner Mongolia, China, using 201 Sentinel-1 images collected from 20 May 2017 to 13 April 2024. We applied both SBAS-InSAR and distributed scatterer InSAR (DS-InSAR) methods to investigate the surface displacement and long-term behavior of the open-pit coal mine over the past seven years. The relationship between this displacement and mining activities was analyzed. The results indicate significant land subsidence was observed in reclaimed areas, with rates exceeding 281.2 mm/y. The compaction process of waste materials was the main contributor to land subsidence. Land uplift or horizontal displacement was observed over the areas near the active working parts of the mines. Compared to SBAS-InSAR, DS-InSAR was shown to more effectively capture the spatiotemporal distribution of surface displacement in open-pit coal mines, offering more intuitive, comprehensive, and high-precision monitoring of open-pit coal mines.

Nowadays, since millimeter-wave (MMW) radar can stably perform simultaneous localization and mapping (SLAM) in bad weather, it plays an important role in application scenarios such as autonomous driving and disaster rescue. Existing research considered the observation of MMW radars is conformed to 0-mean Gaussian distribution, so the localization based on MMW radar is conformed to the distribution, too. However, the change of observation's distribution because of coordinate transform was ignored. Hence, the conclusion is drawn out that the estimated position will cumulate error with fixed direction when the robot is located by MMW radar after the analysis of this article. This article utilizes the correlative scan matching (CSM) method in radar SLAM to analyze the error form with different coordinate transforms. The error of one landmark, the effect of localization due to a set of landmarks, and the trend of accumulation of errors are analytic, respectively. A simulation method based on Monte Carlo sampling is used to prove the conclusion in theory. In order to solve this problem, a CSM-based cumulative error expectation compensation (CEEC-CSM) method is proposed. Four scene experiments were used to prove that the position error of the CEEC-CSM method is reduced by 49%.

In a complex electromagnetic environment, the tracking of jamming source by passive radar network is of great significance for enhancing anti-jamming capability, military combat safety, and strategic decision-making. However, traditional jamming source tracking algorithms suffer from low tracking accuracy and convergence speed, primarily due to the high nonlinearity and the unknown noise characteristics of the passive radar system. In order to improve the capability of jamming source tracking for passive radar network, a maximum correntropy cubature Kalman filter based on improved grey wolf optimization algorithm is proposed. Firstly, the grey wolf optimization mechanism improved by Gaussian random walk and Gaussian mutation strategies is proposed to accurately estimate the characteristics of unknown process and measurement noise, providing more accurate model parameters for the cubature Kalman filter algorithm. Then, an adaptive maximum correntropy criterion is designed, which optimizes the filter gain by adaptively adjusting the kernel size to suppress the influence of outliers on the filtering estimation and enhances the robustness of the algorithm. Finally, experiment of jamming source tracking indicates that the proposed algorithm significantly outperforms traditional algorithms in terms of tracking accuracy and convergence speed under diverse unknown noise environments.

As a crucial component of the transportation infrastructure, the health of bridge plays a direct role in the traffic safety. Over time, gradual structural deformation can compromise a bridge's stability and safety. Therefore, accurately predicting bridge deformation is essential for analyzing its causes and detecting potential safety hazards in a timely manner. Satellite-based synthetic aperture radar interferometry (InSAR) technology, which detects deformation at millimeter-scale precision over large areas, offers significant advantages in monitoring bridge deformation. However, most existing time-series deformation prediction methods based on InSAR data primarily focus on land subsidence. Given that bridge is complex, singular structures with unique spatial-temporal characteristics, existing methods designed for land subsidence are not directly applicable to bridge deformation prediction. To address this challenge, we propose a novel K-shape and complete linkage hierarchical cluster long short-term memory (KCC-LSTM) approach for predicting bridge deformation based on time-series InSAR data. The approach initially combines two machine learning based clustering algorithms, K-Shape for better capturing shape features of time series and complete linkage hierarchical clustering combined with spatial geographic location captures the spatial characteristics of time series to derive clusters with unique spatiotemporal deformation behavior, improving clustering accuracy and spatiotemporal correlation. Clustering results generated from this unsupervised machine learning approach are later used as training labels to develop long short-term memory (LSTM) networks. We validate the proposed approach using time-series data from 100 X-band TerraSAR-X images, acquired from 13 April 2010 to 13 December 2019. Our results demonstrate that compared to standard LSTM, the proposed approach reduces root mean square error of Bridge 1 from 3.6 to 0.5 mm and Bridge 2 from 3.6 to 1.3 mm, improving prediction accuracy. The results underscore the effectiveness of the KCC-LSTM model in predicting deformation in complex infrastructure, such as bridge.

Canopy chlorophyll content (CCC) is a key indicator for assessing the carbon sequestration capacity and material cycling efficiency of ecosystems, and its accurate retrieval holds significant importance for analyzing ecosystem functioning. Although numerous destructive and remote sensing methods have been developed to estimate CCC, the accurate estimation of CCC remains a significant challenge in mountainous regions with complex terrain and heterogeneous vegetation types. Through the synergistic analysis of ground hyperspectral and Sentinel-2 data, this study employed Pearson correlation analysis and spectral resampling techniques to identify Sentinel-2 blue band B1 (443 nm) and red band B4 (665 nm) as chlorophyll-sensitive bands through spectral matching with the hyperspectral reflectance of typical grassland vegetation. Based on this, we developed a new four-band vegetation index (VI), the Dual Red-edge and Coastal Aerosol Vegetation Index (DRECAVI), for estimating the CCC of heterogeneous grasslands in the middle section of the Tianshan Mountains. DRECAVI incorporates red-edge anti-saturation modules (bands B4 and B7) and aerosol correction modules (bands B1 and B8). In order to test the performance of the new index, we compared it with eight commonly used indices and a hybrid model, the Sentinel-2 Biophysical Processor (S2BP). The results indicated the following: (1) DRECAVI demonstrated the highest accuracy in CCC retrieval for mountainous vegetation (R2 = 0.74, RMSE = 16.79, MAE = 12.50) compared to other VIs and hybrid methods, effectively mitigating saturation effects in high biomass areas and capturing a weak bimodal distribution pattern of CCC in the montane meadow. (2) The blue band B1 enhances atmospheric correction robustness by suppressing aerosol scattering, and the red-edge band B7 overcomes the sensitivity limitations of conventional red-edge indices (such as NDVI705, CIred-edge, and NDRE), demonstrating the potential application of the synergy mechanism between the blue band and the red-edge band. (3) Although the S2BP achieved high accuracy (R2 = 0.73, RMSE = 19.83, MAE = 14.71) without saturation effects and detected a bimodal distribution of CCC in the montane meadow of the study area, its algorithmic complexity hindered large-scale operational applications. In contrast, DRECAVI maintained similar precision while reducing algorithmic complexity, making it more suitable for regional-scale grassland dynamic monitoring. This study confirms that the synergistic use of multi-source data effectively overcomes the limitations of the spectral–spatial resolution of a single data source, providing a novel methodology for the precision monitoring of mountain ecosystems.

Synthetic Aperture Radar (SAR) ship classification is crucial for maritime surveillance. Most existing methods primarily focus on visual or polarimetric features, often constrained by a limited feature set and facing challenges in data diversity and multimodal information integration. This study introduces a text-enhanced multimodal framework for SAR ship classification (TeMSC), an extensible and unified approach that integrates multimodal information related to SAR ships. It consists of textform geometry information embedding, polarization and visual information embedding, and a multimodal prediction module. By incorporating ship geometry information in text format, TeMSC leverages text representation to enhance feature expressiveness, compensating for the limited discriminative power of traditional visual and polarization features, especially in low-resolution scenarios. TeMSC effectively processes complementary multimodal information through a multimodal prediction module, while avoiding the complexity associated with traditional decisionlevel feature fusion strategies. Additionally, a classification token mechanism is introduced to streamline the classification process. Through a two-stage training strategy, TeMSC captures information across multiple SAR datasets, enhancing its generalization and adaptability. Extensive experiments on the FUSAR-Ship and OpenSARShip datasets demonstrate the superior performance of TeMSC and highlight the benefits of multimodal integration for SAR ship classification. TeMSC provides a foundation for future research on SAR-focused multimodal learning applications.

Monitoring the structural deformation of bridge with high precision during the operation process is crucial for assessing its health. This study proposes a practical strategy for jointly measuring multi-scale periodic dynamic deformation in bridges using both spaceborne and ground-based Interferometric Synthetic Aperture Radar (InSAR) technologies. The proposed strategy involves extracting seasonal periodic deformation by applying thermal expansion components with spaceborne Persistent Scatterer InSAR (PS-InSAR) and capturing daily periodic deformation using a two-stage atmospheric phase screen compensation ground-based InSAR method. This study focuses on a double-tower cable-stayed and rigid frame system bridge to investigate the spatiotemporal evolution of bridge multi-scale periodic dynamic deformation patterns. The monitoring results indicate that the geometric state and deformation pattern of the bridge remained stable, exhibiting significant seasonal and daily dynamic deformations that were either positively or negatively correlated with temperature changes. Seasonal periodic deformation captured by spaceborne InSAR showed maximum displacements near expansion joints, while tower deformation remained constrained within +/- 5 mm. Daily periodic deformation captured by ground-based InSAR revealed significant displacements at the bridge tower top, contrasting with minimal deformation of +/- 2 mm near fixed bearings. These deformations exhibited significant correlations with temperature changes. Both the deformation trend and magnitude confirmed to the computational results of the bridge structure design.

In view of the complex and harsh working conditions at the coal production site of the auxiliary vertical shaft in mines, the traditional management model has significant limitations in timeliness, accuracy, and predictability due to its high reliance on human labor, and also shows poor economic efficiency. To address the above issues, this paper proposes an AI image recognition management system integrating cutting-edge technologies. The system deeply integrates advanced image acquisition, preprocessing, analysis, and intelligent decisionmaking technologies. Additionally, it innovatively applies AIBOX and edge computing technologies in the mine image recognition system, adopting a hierarchical distributed architecture design to construct an AI image recognition management system comprising a data sensing layer, transmission layer, edge computing layer, decision control layer, and application display layer. This aims to achieve comprehensive, real-time, and high-precision monitoring and control of equipment, personnel, and the environment, while reducing the input of human and material resources. Practical tests have proven that the system has achieved remarkable results in key aspects such as the timeliness of fault warning, personnel management efficiency, and environmental monitoring accuracy. It can effectively guarantee the safe production of coal mines, reduce investment in safety supervision, and demonstrate significant economic benefits.

To address the challenges of traditional marine meteorological prediction methods, which struggle to effectively capture intervariable correlations in multivariate time series data and suffer from insufficient prediction accuracy, this article proposes a multivariate short-term marine meteorological prediction model. First, an intelligent marine prediction fusion architecture is constructed, which is well-suited to artificial intelligence (AI) technologies. This architecture optimizes the process of marine meteorological data collection, processing, and analysis, providing a flexible and efficient infrastructure for short-term marine prediction. Second, an influence-based importance attention mechanism for meteorological variables is designed. By exploiting the differences in interactions among meteorological variables, it selects significant attention heads for computation, effectively reducing the model's computational complexity and enhancing its response speed. Finally, a multivariate dimension prediction method for marine meteorology is proposed. By independently processing the time series of each variable, it enhances the capability to capture interactions among meteorological variables, thus improving the model's understanding of and predictions for dynamic changes in marine meteorology. The experimental results show that the model can fully capture and analyze the complex relationship between variables in a multivariable marine meteorological environment, effectively improve the accuracy and efficiency of the prediction, and verify its application potential in marine meteorological prediction.

The three-dimensional (3D) imaging of targets in enclosed spaces using Synthetic Aperture Radar (SAR) is currently a hot research topic in the field of radar technology. At present, there are various operational modes for SAR-based 3D imaging, but they involve high data acquisition costs and complex data processing. Some studies have utilized rotating array SAR to obtain SAR images from different angles, and then reconstruct the three-dimensional structure of the target based on its geometric variations. However, this method can introduce errors in the height estimation. To address this issue, this study proposes a Bayesian estimation-based method for rotating SAR 3D imaging, which aims to mitigate the problem of unreliable height measurements. A hybrid distribution model based on stereo matching is established in this research, and the Bayesian estimation method is employed to minimize the error distribution of the target across multiple angles, thereby obtaining the optimal height value. The proposed method is validated using real-world data, and 3D point cloud reconstruction of the target is performed to verify the effectiveness of the approach.

With the rapid development of unmanned aerial vehicles (UAVs) and their applications in various fields, accurate image registration, especially for small UAV targets, has become crucial. Existing feature point matching methods face challenges in small UAV target registration due to scarce feature points, small size, and complex backgrounds, leading to unstable outcomes. To overcome this issue, this paper proposes a cross-modal image registration (CMIR) model, which consists of a multi-scale feature point extraction network (MSFEN) and a feature refinement matching module (FRMM). In this method, optical images are first transformed into pseudo-infrared images using a modality transformation network (MTN) to reduce modality disparities. The MSFEN then extracts robust feature points for preliminary matching, while the FRMM refines the matches by analyzing the relative positions and local structures of the points. Experimental results show the method excels in small UAV target image registration, effectively addressing challenges like scarce feature points, small target size, and complex backgrounds.