Search Results (676)

The abrupt change of canopy structure at the flowering stage causes a long-standing monitoring blind zone for SAR-based inversion of winter rapeseed. This study proposed a feature-based inversion method for the Total Photosynthetic Area Index (TPAI) of winter rapeseed, in which a multi-dimensional spatiotemporal feature framework was constructed by integrating radar-derived features such as polarization interactions and spatial texture with acquisition-date-derived phenological priors including day-of-year harmonics and normalized z-score features, and a regularized linear Ridge model was selected from six machine-learning candidates. Evaluated under a spatial five-fold cross-validation protocol with geographically disjoint validation folds, the model achieved an overall R² of 0.615 and an RMSE of 1.033 m² m⁻², against an R² of 0.584 for a baseline encoding the growth calendar alone, indicating only a modest overall gain from the radar-derived features; the radar-derived features nevertheless delivered their largest within-stage improvement over this baseline at the flowering stage, suggesting that they can partly alleviate—though not resolve—the sensitivity loss of conventional single-polarization features during flowering. Error magnitudes were insensitive to systematic bias of up to ±40% in the SAI scaling assumption, and full-coverage TPAI mapping across five phenological stages, produced from seam-corrected Sentinel-1 mosaics, matched the observed growth trajectory. Given that the SAI was estimated indirectly and the analysis relied on a single growing season at a single study area, these results should be regarded as site- and season-specific. The framework therefore shows promise as a spatially validated, feature-based pathway toward TPAI monitoring of winter rapeseed in cloudy-rainy regions, with further validation across multiple seasons and regions required before operational application. Full article

Real-world spatiotemporal data are frequently imbalanced across spatial regions and collected from heterogeneous sources with different structures, sampling frequencies, and semantics. Sparse regions provide limited supervision, whereas modality redundancy and structural heterogeneity make it difficult to exploit complementary information during fusion. To address these challenges, we propose MSTDP, a Multimodal Spatio-Temporal Data Prediction framework. A causal target-clock alignment procedure first maps high-frequency, low-frequency, and irregular observations to the forecasting timeline without using future information. MSTDP then employs modality-specific encoders to obtain source-dependent representations and constructs first-, second-, and third-order interaction features through a Multilayer Feature Interaction module. A Hierarchical Feature Fusion module aggregates the interaction features. To improve intermediate and high-level representations, we further introduce two plug-and-play components: a Unified Feature Interaction Module, which performs soft alignment and bidirectional feature exchange between heterogeneous token sequences, and a Self-Correcting Transformer Module, which recalibrates modality-specific and fused representations through cross-modal attention. A spatial-unit-balanced training objective first normalizes the prediction error within each region and then averages equally across regions, preventing data-rich regions from dominating the gradients. We evaluate MSTDP on four real-world benchmarks covering urban housing-price forecasting, county-level crop-yield prediction, and traffic-speed forecasting. Relative to the strongest baseline in each setting, MSTDP reduces RMSE by up to 8.6% and MAPE by up to 11.5%. Ablation studies confirm that hierarchical fusion, unified feature interaction, and self-correction provide complementary improvements. Observation-density stress tests further show that the model remains robust when entire spatial units are unavailable. These results demonstrate the effectiveness of MSTDP on the four evaluated datasets and provide empirical evidence that its interaction and correction modules can be reused across the studied housing, agriculture, and traffic forecasting tasks. Broader transferability to other multisource spatiotemporal problems remains to be established. Full article

Although deep learning-based electroencephalography (EEG) emotion recognition methods have achieved promising performance, they generally depend on complex network architectures and high computational costs, which limits their deployment in resource-constrained scenarios. Meanwhile, existing approaches mainly focus on local spatiotemporal feature extraction while insufficiently exploiting global frequency-domain dependencies and dynamic information selection during feature propagation, resulting in inadequate spectral information utilization and redundant feature representations. To address these issues, this paper proposes a Frequency Prior-Gated Lightweight Network (FPGLNet) for EEG emotion recognition. First, a Mixed Aggregation Convolution (MAC) module is introduced at the input stage to jointly encode EEG spatiotemporal characteristics through temporal and spatial convolutions. Subsequently, a Fused Fourier Convolution Mixer (FFCM) module is developed to model global frequency-domain dependencies through Fourier transform and integrate multi-scale depthwise convolutions for collaborative frequency-spatial feature representation. Furthermore, a Prior-Gated Feed Forward Network (PGFN) is proposed to dynamically generate prior-guided gating weights based on feature responses, enabling adaptive selection and recalibration of discriminative features across different brain regions and temporal segments. This mechanism enhances emotion-related representations while suppressing noise and redundant information. Owing to the lightweight architecture design, FPGLNet achieves an effective balance between feature representation capability, model complexity, and computational efficiency. Experimental results demonstrate that FPGLNet achieves classification accuracies of 96.22% and 92.69% on the public SEED-IV and SEED-V datasets, respectively. These results validate the effectiveness of the proposed frequency prior-gated strategy in improving EEG feature representation while maintaining competitive recognition performance and deployment efficiency. Full article

The number of Internet of Things (IoT) devices has been growing rapidly, driven by the demand for low-cost and low-power devices across a wide range of industrial and consumer applications. This expansion has broadened the attack surface exposed to cyber threats, making the detection of malicious activity in the IoT infrastructure a pressing research challenge. Most existing approaches are limited to binary anomaly detection, rely on a single data modality, or incur a computational overhead incompatible with realistic deployment on the edge. This paper introduces the Multi-modal Universal Informer Convolutional Neural Network (MUI-CNN) Classifier, a distributed architecture for multi-class attack detection across heterogeneous IoT sensor and network environments. It combines four components - a multi-modal universal data embedding layer that unifies numerical, categorical, and temporal features into a shared spatio-temporal representation, a CNN local pattern recognition layer, an encoder-only classifier for long-term temporal dependencies, and an optional statistical masking layer for greater sensitivity to low-amplitude signatures. The model is evaluated on the Telecommunication Operating Systems (OS) and Network Internet of Things (TON-IoT) dataset covering eight sensor types and ten attack classes. Results demonstrate F1 above 88% on all IoT sensors and up to 94.85% on the network sensor, confirming that each architectural component contributes a measurable and distinct improvement over partial baselines. Full article

The alternating current (AC) optimal power flow (AC-OPF) problem is crucial for efficient power system operation, yet it remains challenging due to its nonlinear, nonconvex structure and numerous operational constraints. This paper introduces a Physics-guided Graph Safe Reinforcement Learning (PGSRL) framework that provides accurate, scalable, and physically consistent solutions to AC-OPF. The proposed PGSRL framework first models the power system as a weighted graph, encoding voltage magnitudes, load demands, and transmission line impedances to capture existing spatiotemporal correlations among power system buses. A Graph Mamba encoder with selective State Space Modeling extracts spatiotemporal correlations, which are used to define the system states within a constrained Markov Decision Process (CMDP). To solve this CMDP, we design a physics-guided safe Twin Delayed Deep Deterministic Policy Gradient (TD3) architecture containing twin critics and deterministic actors, guided by physics-informed reward functions that promote feasibility through penalty-based soft constraints. By interacting with a simulated power system environment, the introduced PGSRL learns optimal generator dispatch strategies in continuous action spaces while honoring AC-OPF constraints. Experimental results on six test benchmarks demonstrate the superiority of PGSRL compared with traditional numerical solution and supervised data-driven baselines, rather than RL methods, in terms of estimation accuracy of AC-OPF setpoints. Full article

As Earth system science moves toward global-domain and multi-sphere coupling, traditional two-dimensional spatial information frameworks can no longer fully meet the integrated spatial computing demands of deep space, deep Earth, and deep ocean applications. Building on research in Discrete Global Grid Systems (DGGS), this paper provides a comprehensive review of Earth System Spatial Grid (ESSG) for three-dimensional Earth system modeling. It focuses on three major categories, namely shell-based, volumetric, and unstructured approaches, and examines their geometric construction, radial extension, and encoding mechanisms. The paper further summarizes representative applications of ESSG in geophysics, space sciences, digital twins, and spatial intelligence, and discusses future challenges including ellipsoid adaptation, international standardization, cloud-native computing, and four-dimensional spatiotemporal extension. Full article

Reconstructing dense and temporally continuous Landsat time series remains challenging because clear-sky observations are often sparse and irregular due to cloud and shadow contamination, limited revisit frequency, and sensor-specific data gaps such as the Landsat 7 Scan Line Corrector (SLC) failure. Although recent deep learning approaches have improved Landsat time series modeling, their performance degrades under very low observation densities, particularly when models rely primarily on temporal information. To address this limitation, this study developed a spatiotemporal Convolutional Neural Network (CNN)–Transformer framework that extends a temporal-only 1D Transformer by incorporating local spatial and spectral context from neighboring pixels. Multispectral image patches are first encoded using a ResNet-based convolutional encoder, and the resulting spatial–spectral features are combined with the central pixel sequence before temporal dependencies are modeled using a Transformer. This design allows the model to leverage neighboring observations when information at the central pixel is limited or missing. Two complementary training strategies were employed: (1) dense reference time series generated from combined Landsat–MODIS observations using Gaussian Process Regression (GPR) and (2) 30 m Harmonized Landsat–Sentinel-2 (HLS) observations. The models were evaluated using 1500 independent test time series and 319 hold-out Landsat images from three test areas in southeastern Alberta, Canada, under different seasons, land cover types, observation densities, and SLC conditions. The spatiotemporal model reduced RMSE by approximately 10% across spectral bands in the GPR-based evaluation. In the independent Landsat-image evaluation, average RMSE improvements across spectral bands were 29%, 4%, 20%, and 6% in spring, summer, fall, and winter, respectively. The spatiotemporal model also showed improved reconstruction under Landsat 7 SLC-off conditions, reducing spatial errors across heterogeneous landscapes, particularly in cropland areas. Overall, the results demonstrate that incorporating local spatial–spectral context into Transformer-based temporal modeling can improve Landsat time-series reconstruction, with the greatest benefits occurring during periods of rapid temporal change and limited clear-sky observations. Full article

Modern supply chains are highly interconnected, allowing local disruptions to spread across products, facilities, and distribution channels. This study introduces SG-Net, a spatiotemporal graph neural network that combines a relation-aware spatial encoder, a memory-augmented temporal encoder, and attention-based fusion. The model is evaluated on the SupplyGraph dataset using two daily next-step forecasting tasks—sales orders and production—and two classification tasks—anomaly detection and product-relation classification. Under the reported protocol, SG-Net achieves the strongest mean performance across the four benchmark tasks. The conclusions are restricted to daily next-step prediction and the evaluated SupplyGraph setting; no claims of long-horizon warning, empirical scalability, or demonstrated interpretability are made. Full article

Accurate wind speed forecasting (WSF) is pivotal for the secure grid integration of wind farms and the optimization of power dispatching. However, the stochastic nature of atmospheric dynamics and their complex interactions with local terrain pose significant challenges. Traditional spatiotemporal models inherently suffer from severe homogenization, irreversibly destroying fine-grained aerodynamic textures and leading to cumulative directional errors over extended horizons. To dismantle these limitations, this study proposes a novel Topography-Guided Spatio-Vertical–Temporal Transformer (TG-SVTFormer) for the multi-step forecasting of zonal and meridional wind vectors. Specifically, a dual-stage topography-guided mechanism explicitly injects static geographical variables (elevation, slope, and roughness) into both the spatial encoding and microscopic decoding phases, imposing key physical constraints on aerodynamic friction. Furthermore, the architecture employs a structured decoupled learning paradigm: a re-engineered Mix-Transformer spatial encoder deliberately eliminates deep-stage downsampling in favor of dilated convolutions—drastically expanding the global receptive field while preserving the intrinsic meteorological textures. Subsequently, the temporal evolution hub explicitly decouples the vertical dimension, utilizing upper-air atmospheric states to guide the target surface layer via simulated vertical momentum transport. The experiments showed the accuracy of TG-SVTFormer in predicting the wind speed with mean absolute error of 0.71 m/s in 3 h, 0.85 m/s in 6 h, and 1.06 m/s in 12 h. Compared to five existing baseline WSF models, the proposed TG-SVTFormer model demonstrates superior performance in wind speed forecasting and, more importantly, offers a transferable and physics-informed framework that can be adapted to other terrain-complex regions and extended to related renewable-energy and air-quality forecasting tasks. Full article

Large-scale smart grids have introduced significant challenges for false data injection attack (FDIA) detection due to the distributed nature of power system operation and the limited observability of individual regions. Existing horizontal federated learning approaches enable collaborative model training without sharing raw data; however, they rely on local information and often fail to capture the global operational characteristics needed to detect sophisticated attacks. To address this limitation, the present paper proposes VFedA, a federated vertical relational alignment framework for FDIA detection via cross-regional feature fusion. Specifically, each regional client employs a self-supervised spatiotemporal contrastive learning encoder to extract local representations from measurements. The extracted representations are uploaded to a control center and concatenated to construct global feature representations for attack detection. In order to mitigate inconsistencies across heterogeneous regional feature spaces, a relational alignment mechanism is introduced to enforce structural consistency by aligning similarity matrices derived from local representations. The proposed approach enables effective cross-regional information fusion while preserving data privacy. Extensive experiments are conducted on IEEE 30-bus and IEEE 118-bus benchmark systems. The results demonstrate that VFedA consistently outperforms conventional horizontal federated learning methods and achieves performance comparable to centralized detection models. Full article

In highly dynamic and interactive vehicular environments, accurate and efficient trajectory prediction is critical for driving safety and traffic efficiency. To overcome the limitations of existing methods, including degradation of long-range historical context, high computational cost for long sequences, attention dispersion in spatial interactions, and limited explicit modeling of future interactions, this paper proposes a Hierarchical Mamba Graph Attention Network (HM-GAT) for vehicle trajectory prediction. A Mamba-based selective state-space model (SSM) encoder with linear complexity is employed to capture nonlinear temporal dynamics and transient maneuvering behaviors. For spatial modeling, a heterogeneous topology-driven dual-layer graph attention mechanism, combined with adaptive sparse filtering and a foresighted future graph, enables precise identification of neighboring vehicles with potential collision risks. The proposed architecture integrates global intention exploration, topology bridging, and residual refinement in a coarse-to-fine manner, achieving effective spatio-temporal decoupling. Experimental results show that HM-GAT reduces the average RMSE by 33.1% relative to the best-performing baseline on NGSIM, achieves competitive prediction accuracy on Argoverse 2 without HD-map priors, and achieves lower inference latency than the compared baseline methods on both datasets. Full article

Accurate multi-step traffic forecasting requires models to respond to genuine traffic-state transitions while resisting deviations caused by missing or noisy observations. Existing deviation-aware methods characterize departures from recurrent patterns, but deviation magnitude alone does not indicate whether the resulting correction will improve a node-wise forecast. Guided by the insight that correction strength should depend on forecast gain rather than deviation magnitude, we propose Spatiotemporal Traffic Forecasting with Forecast-Gain-Guided Deviation–Invariance Routing (ST-FGDIR). ST-FGDIR constructs an anchor-conditioned stable endpoint and a deviation-corrected candidate through weight-shared current–anchor graph encoding and multi-view deviation evidence. Signed candidate-error differences provide soft forecast-gain targets for a node-wise continuous router, while selective risk extrapolation and a false-activation loss control stable-endpoint risk across perturbation environments and suppress perturbation-induced increases in residual injection. Experiments on METR-LA, PEMS-BAY, PEMS04, and PEMS08 show competitive forecasting performance relative to published baseline results. The inference gates exhibited Spearman correlations of 0.49–0.53 with realized correction gain, and the complete model yielded an average relative MAE increase of 9.38% across 12 controlled dataset–perturbation conditions. These results demonstrate that forecast-gain-guided routing coordinates stable periodic prediction with adaptive deviation correction under both clean and perturbed inputs. Full article

Electroencephalography (EEG)-based emotion recognition remains challenging due to the low signal-to-noise ratio, non-stationarity, strong inter-subject variability, and the distribution of affect-related information across multiple temporal resolutions. Although recent deep architectures have increasingly combined multi-scale local feature extraction with global dependency modeling, the adaptive integration of multi-scale representations before global encoding remains comparatively underexplored. To address this issue, we propose DSFConformer, an EEG emotion-recognition framework integrating multi-scale spatiotemporal convolution, Dynamic Sparse Feature Fusion (DSFF), and lightweight Transformer-based global modeling. The multi-scale encoder captures complementary local EEG dynamics at different temporal resolutions, while DSFF performs sample-adaptive feature reweighting and sparse suppression to reduce the influence of redundant or noise-sensitive responses. A Lite Transformer is subsequently employed to model global contextual dependencies with reduced computational cost. Experiments on the SEED dataset were conducted under complementary evaluation protocols. DSFConformer achieved an accuracy of 96.08% under chronological block-wise segment-level cross-validation. Under the stricter trial-wise protocol, it achieved 78.81% accuracy, statistically comparable to EEG-Conformer at 78.77%, with no significant paired difference. Ablation analyses further demonstrate that multi-scale convolution and DSFF provide the primary recognition-performance gains, whereas the Lite Transformer mainly improves computational efficiency while preserving comparable accuracy. Cross-session and leave-one-subject-out analyses further assess model robustness and generalization. Overall, the results indicate that adaptive sparse fusion of multi-scale EEG representations provides an effective approach for balancing local representation learning, noise-aware feature integration, and efficient global dependency modeling. Full article

Short-term typhoon wind-field prediction across multiple lead times is challenging because forecast error grows with lead time and inappropriate sample construction may leak future storm-position information. We adopt STL-Net, a lightweight spatiotemporal framework integrating convolutional encoding, long short-term memory (LSTM) temporal modeling, squeeze-and-excitation recalibration, and multi-branch fusion for 10 m wind-field prediction at 6, 12, and 24 h lead times. A fixed-t₀ protocol anchors the target patch at the initialization-time storm position, avoiding future best-track leakage. Trained in 2020–2021, validated in 2022, and tested in 2023, STL-Net achieves a root mean square error (RMSE) of 3.11, 4.06, and 5.31 m·s⁻¹ at 6, 12, and 24 h. Among deep-learning baselines, U-Net achieves the lowest errors at 6 and 12 h and convolutional neural network (CNN) the lowest 24 h RMSE and mean absolute error (MAE), while STL-Net remains competitive with 1.673 million parameters and low inference latency. Ten-seed ablation identifies multi-resolution fusion as the most consistently beneficial component. An out-of-year evaluation with 2020 as the test year reproduces the error-growth pattern, supporting cross-year robustness. Models are deterministic without uncertainty quantification; error growth is deterministic, not calibrated uncertainty. Diagnostic interpretation links the error structure to the multi-scale organization of typhoon winds, storm translation, and wind-field asymmetry. Results constitute an offline proof-of-concept, not an operational system. Full article

Precise indoor thermal environment forecasting is critical for the energy-efficient operation and model predictive control of building heating, ventilation, and air-conditioning (HVAC) systems. However, conventional predictive models often struggle to untangle the complex spatiotemporal dynamics captured by sparse sensing networks. This study proposes a Spatio-Temporal Joint-Embedding Predictive Architecture (ST-JEPA) for indoor multi-node temperature and humidity prediction. By reformulating predictive representation learning from visual domains to structured sensor-node sequences, the proposed model explicitly integrates historical multi-node observations, spatial positional encoding, and concurrent HVAC operating states. The model was evaluated at prediction horizons from 30 s to 5 min using 3568 frames recorded over approximately 29.8 h by a nine-node sensor array in a single laboratory. Experimental results indicate that ST-JEPA achieves high short-term forecasting accuracy, with temperature root mean square errors (RMSE) of 0.0490 °C at 30 s and 0.0647 °C at 1 min. In the original ablation experiment, removing HVAC operating-state inputs increased temperature RMSE by up to 68.3%. Aggregate errors remained relatively stable under the tested 30–50% node-masking conditions, although additional single-node tests revealed location-dependent sensitivity. Comparisons with simpler baselines showed no consistent superiority across targets and horizons. ST-JEPA characterizes the short-term evolution of indoor thermal fields, providing a potential forecasting basis for HVAC feedforward regulation; its effects on energy consumption and thermal comfort remain to be evaluated. Full article