Search Results (636)

In continuous sign language recognition (CSLR), existing methods predominantly adopt frame-wise feature extraction, neglecting temporal continuity and motion trajectory modeling, thereby struggling to capture complete spatio-temporal dynamics. Meanwhile, global cross-modal attention approaches typically directly model interactions between text and the entire video sequence but lack structural constraints, making them susceptible to interference from redundant frames, which leads to attention distribution dilution and undermines fine-grained motion alignment capability. To address these issues, this paper proposes CMST-Net, a Cross-modal Interaction and Spatio-temporal Feature Enhancement Method for Continuous Sign Language Recognition. CMST-Net comprises two principal modules: the Local–Global Cross-modal Fusion Module (LGCFM) and the Spatio-Temporal Feature Enhancement Module (STFEM). LGCFM introduces a synergistic modeling mechanism that combines local sliding-window attention with global attention, achieving a unified fusion of structured local alignment and global semantic modeling. STFEM incorporates multi-scale spatial dilated convolution and coordinate attention to extract fine-grained spatial features while leveraging channel partitioning and a hierarchical residual structure to enhance long-range temporal modeling capability; the two modules collaboratively yield high-quality spatio-temporal feature representations. Experiments on three public benchmark datasets (PHOENIX2014, PHOENIX2014-T, and CSL-Daily) demonstrate that CMST-Net can effectively improve continuous sign language recognition performance, achieving state-of-the-art performance on the PHOENIX2014 and CSL-Daily datasets and competitive results on the PHOENIX2014-T dataset. Full article

Advances in artificial intelligence models offer promising approaches for intelligent convection initiation (CI) identification—a critical step in severe weather nowcasting—thereby driving demand for high-quality, long-term labeled datasets. Accordingly, this study develops a CI identification technique using quality-controlled, gridded composite reflectivity data from the weather radar and CMA global atmospheric reanalysis wind data over central-eastern China (2018–2023) to construct a labeled dataset. The proposed CI identification method integrates a “forward-time search and backward-time verification” strategy, which involves three key steps: screening grid points with absent or weak convection; monitoring these points for convective development within 30 min; and finally, confirming the first occurrence of convection. Additionally, quality control is applied to eliminate the influence of outliers and anomalous radar data. The resulting dataset constructed from 829 severe convective processes comprises ~25.6 million grid points labeled for CI occurrences at one or more lead times of 10, 20, or 30 min to resolve spatiotemporal evolution. Of these, 71.90% are accompanied by surface weather phenomena. This study provides a reliable dataset to support the learning of intelligent identification and nowcasting models for CI events. Furthermore, based on this dataset, the spatiotemporal distribution characteristics of warm-season CI over central-eastern China are delineated. Full article

The Jianghan Plain is one of the important freshwater aquaculture regions in China. Accurate information on the spatial distribution and spatiotemporal dynamics of aquaculture ponds is essential for regional ecological management. However, large-scale and accurate identification remains challenging for aquaculture ponds because they have spectral and seasonal hydrological characteristics similar to those of rice fields, rivers, canals, and lakes. In this study, we developed a framework for mapping inland aquaculture ponds using multi-source remote sensing data from Sentinel-2, Sentinel-1, and PlanetScope. The framework integrated elevation-zoned Otsu thresholding for candidate water extraction, phenological features for rice field removal, and object classification based on a Gradient Boosting Decision Tree (GBDT) model using 12 shape and spatial-context features. It was applied to identify aquaculture ponds in the Jianghan Plain from 2016 to 2025. Overall accuracy exceeded 93%, and the F1-score exceeded 0.93 in all validations. Results from different sensors also showed high spatial consistency. In 2025, aquaculture ponds covered 2365.17 km² in the Jianghan Plain and were mainly concentrated in the central and eastern parts of the plain, especially near Honghu Lake and along the Yangtze and Hanjiang river meanders. Over the ten-year period, the aquaculture pond area fluctuated between 2033.58 and 2383.52 km², showing an initial decline followed by recovery. Lost aquaculture ponds were mainly located around lakes, while newly added ponds were mostly distributed along the margins of existing clusters. The decade dataset generated in this study can support freshwater aquaculture management and wetland conservation. Full article

Power cable tunnels are typical enclosed cable-routing spaces in which cable overheating, insulation aging, and partial discharge may gradually develop into fire hazards. During the early stage of cable fires, abnormal sensor responses are often weak, localized, and continuously evolving, which increases the difficulty of early warning based on a single sensor or a single temporal feature. Motivated by cable fire early warning in power cable tunnels, this study uses cable-fire records from a publicly available indoor EN 54 fire-test-room dataset with distributed multi-sensor nodes to evaluate the proposed model under controlled laboratory conditions. In monitoring scenarios with fixed sensor nodes, temporal-only modeling methods often struggle to simultaneously characterize short-term variations, temporal evolution, and spatial differences in node responses. To address this limitation, this study proposes a multi-sensor spatiotemporal feature fusion model that integrates a gated recurrent unit (GRU), a Modern Temporal Convolutional Network (ModernTCN), and an enhanced graph convolutional network (GCN+). The proposed model adopts the ModernTCN as the temporal modeling backbone. A GRU module is introduced at the front end to encode local fluctuations and short-term continuous changes between consecutive time steps, while GCN+ is embedded at the intermediate feature stage of the backbone to model spatial correlations and cross-node coordinated responses among fixed sensor nodes. Experimental results show that the proposed model achieves strong classification performance in the cable fire early warning discrimination task, with a test accuracy of 0.9884 and a false negative rate (FNR) reduced to 0.0150. The comparative experimental results indicate that the proposed model achieves better overall performance than typical temporal baseline models. The ablation study further shows that, under the experimental settings of this study, the introduction of GRU and GCN+ leads to overall improvements in the main evaluation metrics, suggesting that both modules provide a certain enhancement to the cable fire early warning discrimination performance of the ModernTCN backbone. Full article

To address the challenges of copyright protection for marine environmental datasets in open sharing environments, this paper proposes ST-Mark, a robust watermarking framework that synergistically integrates established geometric invariants and Quantization Index Modulation (QIM) techniques to meet the strict physical constraints of marine environmental datasets. The proposed method first extracts feature points by analyzing the spatiotemporal distribution of the data. It then constructs a local reference frame from the convex hull vertices and computes the geometrically invariant angles and distance ratios of the feature points relative to this reference pair to achieve robust partitioning. Finally, the watermark is embedded into the attribute domain of the grouped data through the Quantization Index Modulation (QIM) strategy while constraining perturbations within observational uncertainty bounds. Extensive experiments demonstrate that ST-Mark exhibits strong robustness: under extreme conditions such as temporal deletion attacks with an intensity of 0.9, the average normalized correlation (NC) remains above the robustness threshold of 0.75, with peak values reaching 0.99 on high-resolution datasets, although performance may fluctuate or fall near this threshold under severe spatial restrictions and numerical quantization, while still supporting reliable copyright verification under typical operational conditions. Full article

Dynamic gesture recognition under small-sample conditions remains challenging due to the large variations caused by users, viewpoints, motion patterns, and hand configurations. This study proposes a lightweight dynamic gesture recognition framework that integrates meta-learning, Neural Architecture Search (NAS), and knowledge distillation (KD) to achieve rapid adaptation with limited training samples. Different from conventional recognition methods that mainly focus on feature extraction and model optimization, this study further considers the intrinsic symmetry characteristics of hand gestures. A mirror-aware skeleton representation strategy is introduced by modeling the structural correspondence between left- and right-hand keypoints, which reduces the distribution differences caused by hand-side variations and improves the generalization ability under few-shot conditions. The proposed framework adopts a lightweight spatio-temporal feature extraction module, an optimization-based meta-learning strategy, and a knowledge distillation mechanism to balance recognition accuracy and computational efficiency. Experiments are conducted on DHG-14, SHREC2017, FPHA, LMDHG, and 20BN-Jester datasets under different few-shot settings, including cross-user variation, viewpoint variation, motion speed variation, partial occlusion, and background interference. The experimental results demonstrate that the proposed method achieves competitive recognition performance while significantly reducing model complexity and computational cost. The proposed framework provides an efficient solution for dynamic gesture recognition in resource-constrained scenarios. Full article

Accurate traffic flow prediction is a core task in intelligent transportation systems because urban traffic observations are spatially distributed, temporally dynamic, and commonly held by different regional management entities. Existing centralized and local models remain limited when traffic data are non-independent and identically distributed across regions, and when raw data cannot be directly exchanged because of privacy, ownership, and communication constraints. To address these challenges, this study proposes a personalized similarity-aware federated spatiotemporal learning framework for multi-regional traffic flow prediction. The framework integrates three mechanisms: client-specific adaptation for regional distributional heterogeneity, adaptive delayed graph learning for dynamic congestion propagation, and similarity-aware federated aggregation for information-quality-based cross-client collaboration. Spatial dependency, temporal evolution, traffic-flow-theory-informed variables, road attributes, and temporal contextual features are jointly modeled without sharing raw client data. Experiments on controlled synthetic data and the Q-Traffic real-world dataset demonstrate that the proposed method consistently outperforms independent training, FedAvg, FedProx, FedSTN-inspired, and FedAGCN-inspired baselines. On the Q-Traffic grid-level setting, the proposed adaptive graph version reduces MSE by 35.3% compared with FedAvg, while the CNN version reduces MSE by 27.5%. Under the cluster-level setting, the adaptive graph version reduces MSE by 26.3% compared with FedAvg. Ablation, sensitivity, communication-cost, differential-privacy, and client-dropout analyses further show that the proposed framework improves predictive accuracy, cross-client stability, and robustness under heterogeneous federated traffic scenarios. Full article

Accurate regional lightning-occurrence prediction is important for operational weather-risk management, but its development is challenged by the severe class imbalance of grid-hour lightning samples. This study proposes a repeated random undersampling (RUS) stacking ensemble that combines heterogeneous machine-learning models and produces probabilistic lightning-occurrence forecasts using atmospheric variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5), together with spatial and temporal predictors. The effects of the undersampling ratio, ensemble size, and positive-class weighting were systematically evaluated, and a configuration with a RUS ratio of 1:15, 20 ensemble members, and a positive-class weight of 2 was selected. Using an operational threshold of 0.591 selected exclusively on the 2024 validation set, the final model achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.945, an area under the precision–recall curve (PR-AUC) of 0.235, a probability of detection (POD) of 0.521, and an F1-score of 0.302 on the independent 2025 test set. Physical-variable-group ablation and aggregated TreeSHAP (tree-based Shapley additive explanations) analyses were further conducted to interpret the predictions. Removing the spatiotemporal predictors produced the largest reduction in performance, followed by removing cloud and microphysical variables. Both the Light Gradient Boosting Machine (LightGBM) and categorical boosting (CatBoost) models consistently identified longitude, latitude, total-column cloud ice water, and convective available potential energy as the leading predictors, while higher cloud-ice-water content and stronger convective-instability indices generally shifted model outputs towards lightning occurrence. Direct transfer of the ERA5-trained ensemble to corresponding ECMWF forecast fields without retraining retained useful predictive skill and substantially outperformed the operational ECMWF lightning product over the collocated evaluation samples, although performance degradation indicated a cross-dataset distribution shift. These results demonstrate the value of combining imbalance-aware ensemble learning with physically interpretable predictors for regional lightning-risk forecasting, while the strong influence of geographic variables indicates that external validation and local recalibration or retraining are required before application to other regions. Full article

Seawater intrusion (SWI) represents an increasingly critical challenge for coastal groundwater systems, with particularly pronounced impacts observed along the Indian coastline. Coastal aquifers constitute a vital freshwater resource supporting domestic, agricultural, and industrial demands. This study evaluates SWI vulnerability along the entire Indian coast, extending from Gujarat to West Bengal, covering approximately 7517 km of shoreline and up to 100 km inland. The assessment applies the GALDIT vulnerability framework that combines several hydrogeological and hydrochemical criteria such as groundwater occurrence, aquifer hydraulic conductivity, depth to groundwater, distance from shoreline, hydrochemical data, and groundwater quality data. We also assessed the intrusion of existing seawater, shoreline location, and aquifer thickness. However, conventional vulnerability assessments are inherently static and often fail to capture anthropogenic influences. To address this limitation, the present study integrates multi-temporal land use and land cover (LULC) datasets derived from ESA WorldCover remote sensing data for the period 2017–2024. Incorporating LULC dynamics enables a more comprehensive evaluation of the impacts of urban expansion and agricultural intensification on coastal susceptibility to SWI. Accordingly, a modified GALDIT-LU framework is developed to assess the spatiotemporal evolution of coastal vulnerability. The outcomes suggest that huge parts of the Indian coastline are vulnerable to moderate or very high classes, with the very high vulnerability class growing from 13,295 km² in 2017 to 38,257 km² in 2024, a 188% increase in vulnerability over the course of seven years. Groundwater chloride concentrations from Central Ground Water Board (CGWB) monitoring well locations have been used for validation over the proposed assessment, and show good spatial agreement between areas identified as high vulnerability and the spatial distribution of groundwater salinity for all three assessment periods, lending support to the robustness and predictive power of the proposed groundwater salinity assessment. The findings carry direct implications for the United Nations 2030 Agenda, demonstrating that the identified vulnerability patterns intersect with critical targets related to clean water and sanitation, food security, public health, climate action, and poverty reduction along one of the world’s most densely populated coastlines. Full article

Existing microseismic prediction methods commonly incorporate domain prior knowledge as ordinary input features, which makes it difficult to fully exploit its structural constraints and physical implications. To address this limitation, this study proposes a geological-structure-prior-guided method (SPG) for predicting the distribution of microseismic energy fields. In the study, a dataset was constructed from 11,081 original microseismic events, yielding 476 day-indexed microseismic energy maps. In SPG, historical microseismic energy maps are used to characterize the recent dynamic evolution of the energy field, whereas geological spatial fields, including coal-seam depth, coal-seam thickness, fault distance, fold distance, and goaf distance, are used to represent static structural priors. A decoder-side prior modulation mechanism is designed to introduce structural constraints into the reconstruction of the target-day energy field. The experimental results show that SPG achieves the best overall continuous-field prediction performance compared with a baseline U-Net and representative spatiotemporal prediction models, including PredRNN-V2, SwinLSTM, and Earthformer. In particular, SPG reduces the mean squared error (MSE) by 11.77% and improves the structural similarity index measure (SSIM) by 2.20% relative to the best-performing comparison model, while also decreasing the average structural gap by 21.29% compared with the baseline U-Net. Furthermore, supplementary high-energy-mask evaluation indicates that SPG maintains competitive capability in identifying high-energy regions. These results indicate that SPG can more effectively reconstruct continuous energy distributions, thereby providing a feasible methodological reference for spatially resolved rockburst early warning and domain-knowledge-informed predictive modeling. Full article

Sheep- and goatpox are highly contagious, transboundary viral diseases caused by capripoxviruses (CaPV), severely impacting small ruminant production and resulting in significant economic losses. Our study aimed to analyze the spatiotemporal SGP distribution using the available genome sequence data and to evaluate the recombination occurrence. For this, the WAHIS, WOAH and FAO databases were utilized for the epidemiological analysis of SGP outbreaks. The cross-correlation was calculated to assess the impact of massive animal movements associated with the Islamic holiday of Eid al-Adha on the SGP epizootic situation, and phylodynamic and phylogeographic analysis, as well as a recombination analysis were performed. A total of 1629 SGP outbreaks were reported to WAHIS during 2010–2024, with the majority occurring in Mongolia, the Balkan countries, Russia and the Middle East. Over 60% of all SGP outbreaks were associated with croplands or grasslands, with the highest proportion corresponding to animal densities of 10–50 head/km². Statistically significant positive cross-correlation (p < 0.05) was identified between the month of the Eid al-Adha celebration and the number of SGP outbreaks in Russia, Mongolia, Bulgaria and Tajikistan, while in Greece and China no significant correlation was found. The inferred goat pox virus (GTPV) transmission pathways from China to Vietnam and from India to Bangladesh; for the sheeppox virus SPPV, the routes between Kazakhstan and Russia, Kazakhstan and India, as well as between Russia and China, had the strongest Bayes factor support. Intra-specific recombination events were not detected for the SSPV and GTPV datasets. However, inter-specific CaPV recombination analysis identified a single recombination event in GTPV. Therefore, the use of molecular epidemiological tools, along with the time-calibrated phylodynamic and phylogeographic analyses, has significant applications in the local and international surveillance of the occurrence of SGP outbreaks and for identification of potential recombination events. Full article

Federated learning provides a feasible way to support collaborative vehicle part recognition without sharing raw data across distributed vehicle service nodes. However, real vehicle scenarios exhibit spatiotemporal heterogeneity, where client distributions vary and new vehicle models, parts, and viewpoints emerge over time. Such heterogeneity causes catastrophic forgetting and degrades knowledge retention in federated continual learning. To address this problem, this paper proposes FedACA (Federated Continual Learning with Adaptive Class Anchors), a class-anchor-based federated continual learning method for vehicle part image classification. The method uses a frozen Vision Transformer backbone and introduces scene-aware prompt adaptation, adaptive class anchors, non-parametric global prototype selection, and class-aware prompt fusion. These modules stabilize class semantics, reduce feature drift, and improve cross-client knowledge alignment in a data-local federated setting with raw-data non-sharing; however, FedACA does not provide formal privacy guarantees. Experiments on a self-constructed 67-class multi-view vehicle-part dataset and CIFAR-100 show that, under the task-aware class-masking protocol, FedACA achieves final-stage average accuracies of 97.34% ± 0.72% and 95.71% ± 0.63%, respectively, over three independent random seeds, outperforming the evaluated federated continual learning baselines. These results demonstrate that FedACA improves recognition performance and historical knowledge retention under spatiotemporally heterogeneous federated continual learning settings. Full article

Regional multi-isotope mapping provides an effective means of investigating deep crustal composition and crustal reworking processes. The western margin of the Yangtze Craton has experienced multiple episodes of magmatism, crustal reworking, and mineralization since its formation, making it an ideal region for exploring the relationship between deep crustal evolution and metallogenic systems. In this study, Sr-Nd-Hf isotopic and whole-rock geochemical mapping was conducted based on compiled geochronological, geochemical, and isotopic datasets, integrated with deep geophysical data, to reconstruct the Precambrian crustal composition and reworking processes of the region. The results reveal significant crustal heterogeneity in the western margin of the Yangtze Craton. West of the Anninghe Fault Zone, rocks exhibit depleted Nd-Hf isotopic signatures (ε_Hf(t) = 0 to +17; ε_Nd(t) = 0 to +6), low I_Sr values (0.702–0.710), and high Nb/Ta (14–26) and Zr/Hf (36–46) ratios, indicating derivation from the remelting of juvenile mafic lower crust and providing direct evidence for Proterozoic underplating of mantle-derived magmas. In contrast, rocks east of the Anninghe Fault Zone show enriched Nd-Hf isotopic signatures (ε_Nd(t) = −9 to 0; ε_Hf(t) = −7.5 to 0), higher I_Sr values (0.711–0.726), and lower Nb/Ta and Zr/Hf ratios (8–13 and 21–35), suggesting remelting of ancient continental crust. Mineralization is closely associated with crustal architecture. IOCG deposits mainly occur in juvenile crustal domains, whereas Sn polymetallic and phosphate deposits are associated with ancient reworked crust. Under the Precambrian subduction setting, continuous arc magmatism promoted juvenile crust growth, whereas asthenospheric upwelling facilitated ancient crustal reworking, jointly controlling the spatiotemporal distribution of metallogenic systems along the western margin of the Yangtze Craton. Full article

Motor Imagery (MI) decoding based on electroencephalogram (EEG) is promising for Brain–Computer Interface (BCI) applications, yet existing methods generally suffer from two major limitations. (1) Global EEG signal analysis overlooks region-specific neural activity patterns, causing biased feature extraction and poor inter-subject generalization. (2) Few MI-EEG decoding studies adopt frequency decomposition for multi-rhythm feature extraction. Even when adopted, conventional methods rely on predefined frequency bands and suffer from mode mixing, failing to preserve the inherent fractal self-similarity and nonlinear characteristics of EEG signals, restricting the extraction of fine-grained specific features. To address these issues, we propose RMF-Net, a novel model integrating brain region division and Multi-variable Variational Mode Decomposition (MVMD). The model partitions EEG into functional brain regions based on MI neural mechanisms, performs dynamic modal feature extraction for each region via MVMD, and enables efficient cross-regional spatiotemporal feature interaction through an adaptive fusion. On the BCI Competition IV 2a open EEG MI dataset, our model achieves 80.06% accuracy in cross-session tasks and 63.05% in cross-subject tasks, outperforming other mainstream methods. Further analysis verifies that the cross-regional feature weight distribution of RMF-Net conforms to neuroanatomical principles. This work demonstrates that the spatiotemporal feature fusion framework combining brain region segmentation and fractal-aware multimodal signal decomposition is effective for EEG MI decoding tasks. Full article

Accurate crime spatiotemporal prediction is crucial for crime prevention. However, crime occurrences are influenced by diverse and interacting social factors, resulting in dynamically evolving distributions with non-stationarity and spatial heterogeneity. Most existing methods focus on data preprocessing or architectural enhancements and remain offline models, which limits their generalization capability. To address these challenges, we propose a novel spatial semantic-guided online learning framework. Specifically, we first compute the spatial semantic similarity between urban regions using points of interest. Based on this, we then introduce a contrastive learning objective guided by this similarity during training. This design aims to enhance the model’s ability to capture both the similarities and discrepancies among regions. During the prediction process, an iterative online learning strategy is employed to adapt to dynamically changing crime patterns. By continuously fine-tuning the model with streaming data, the proposed framework improves robustness and generalization under non-stationary crime spatiotemporal distributions. Finally, extensive experiments on real-world crime datasets indicate the effectiveness and stability of our proposed approach. Full article