Search Results (1,216)

Pulsed sounds are widespread in marine soundscapes, yet they still lack systematic characterization and information on their emitting species. This study investigates the consistency and spatio-temporal distribution of Fast Pulse Train (FPT) sounds, an abundant sound type recorded in the Venice inlets during three acoustic campaigns conducted at 40 listening points in summer 2019 and 2020. Although FPTs show variability in their temporal and spectral features, clustering analyses revealed no clear separation among extracted signals, supporting the coherence of the current classification; however, when focusing on the more characteristic sounds, three slightly acoustically different groups emerged. The role of temporal, morphological, spatial, hydrodynamic and anthropogenic variables in explaining the distribution of FPT groups was statistically evaluated using generalized linear models. Background noise in the 200–630 Hz range was the only significant predictor, exerting a negative effect on FPT occurrence, likely driven by vessel noise rather than biological chorusing. This, however, does not provide additional cues for identifying the emitting species. Although FPT characteristics remain compatible with fish vocalizations, expanding comparative studies will be essential to refine the definition of this sound type, assess potential sub-categories, identify habitat associations, and ultimately develop robust hypotheses on its biological sources. Full article

Accurate building damage assessment from satellite imagery is essential for post-disaster response and recovery. Most deep learning approaches treat this task as semantic segmentation, producing pixel-level damage maps without separating individual buildings. This formulation inherently limits reliable per-building damage quantification, which is often far more informative for operational decision-making. We present Insta-BDA, an instance-aware framework that integrates change detection with foundation-model-based instance segmentation for building-level damage assessment using only pixel-level supervision. The approach combines ChangeMamba for spatiotemporal change detection with SAM3 for zero-shot building instance extraction, reconciled through a confidence-guided fusion mechanism. The framework does not require instance-level annotations (polygons, bounding boxes, or instance masks) for training; instance-level structure is derived automatically from SAM3’s zero-shot detections, and the learnable fusion variant requires only pixel-level damage labels. We investigate two fusion strategies—a rule-based approach (Insta-BDA-RB) and a learnable variant (Insta-BDA-MLP)—using a compact multilayer perceptron on per-instance features. To improve robustness under typical satellite resolutions and annotation variability, we adopt a binary damage formulation. Experiments on xBD show that Insta-BDA reduces the aggregate building count deviation to −34.96%, compared with −47.06% for ChangeMamba, while maintaining competitive damage classification performance. The learnable fusion further improves damage F1 (0.70 vs. 0.67 for rule-based fusion). Cross-dataset evaluation on IAN-BD and IDA-BD indicates improved generalization. These results suggest that integrating foundation model segmentation with change detection offers a practical pathway toward operational, instance-level building damage assessment. Full article

With the widespread deployment of IoT devices, deploying efficient network traffic classification models on resource-constrained edge nodes is critical for real-time boundary security. However, traditional lightweight models primarily rely on macro-level structural pruning, which often sacrifices crucial feature extraction capabilities when handling complex heterogeneous traffic, leading to a severe imbalance between parameter compression and detection accuracy. To overcome this bottleneck, we propose TinyFlowNet, an ultra-lightweight multi-module fusion architecture. To prevent the parameter explosion inherent in combining CNN, LSTM, and Transformer modules, TinyFlowNet innovatively adopts an extreme operator-level reconstruction strategy. By introducing debiased computations, affine-free normalization, and a customized micro-self-attention mechanism, it comprehensively strips away underlying redundant parameters. Simultaneously, an integrated parameter-free regularization mechanism is introduced to compensate for the representational capacity lost under this extreme compression, ensuring robust spatio-temporal feature fusion. Comprehensive evaluations on the custom X-IDS-20 balanced dataset alongside the complex CICDarknet2020 and ToN_IoT public datasets demonstrate that TinyFlowNet achieves exceptional accuracies of 95.31 percent, 99.53 percent, and 97.13 percent, respectively. Furthermore, it exhibits formidable robustness against extreme class imbalances by securing a peak Matthews Correlation Coefficient of 0.9465 and an outstanding PR-AUC of 0.9834, all while strictly confining the parameter count to merely 74,600. Crucially, actual on-device hardware profiling on a commercial edge device corroborates its deployment viability, exhibiting a minimal dynamic memory footprint of 8.26 MB, an average inference latency of 0.79 ms, and a processing throughput exceeding 1200 FPS. Compared to a standard heavy Hybrid CNN-LSTM-Transformer baseline, TinyFlowNet achieves superior detection accuracy while drastically reducing the parameter footprint by over 99.3% and computational FLOPs by 95.8%. Furthermore, against mainstream lightweight benchmarks like DistilBERT and heavy baselines such as LSTM, TinyFlowNet reduces parameters by 61.4% to 94% while simultaneously achieving absolute accuracy leaps and accelerating inference speed by nearly 4× over MobileNetV2, establishing a highly efficient new paradigm for intelligent edge defense. Full article

Distributed acoustic sensing (DAS) systems capture complex, nonlinear wave propagation across fiber-optic cables. Conventional event classification architectures, constrained by static physical topologies or isolated spatial grids, fail to effectively adapt to the dynamic feature relationships associated with such events, particularly when modeling complex spatiotemporal interactions across sensor arrays. To resolve these structural limitations, we introduce the Spectro-Spatial Dynamic Graph Neural Network (S2-DyGNN), whose architecture couples a two-dimensional frequency–time convolutional front-end with a dual-matrix graph neural network (GNN). First, the convolutional module extracts spectro-temporal features, explicitly capturing localized acoustic dynamics independent of inter-sensor interference. Subsequently, the graph module constructs a dual-matrix topology, fusing a static physical distance prior with a data-driven adjacency matrix that recalculates spatial connections frame by frame from input signals. When evaluated on a highly skewed nine-class DAS field dataset, S2-DyGNN outperformed other conventional models by achieving a peak macro-averaged F1-score of 86.6% and an overall accuracy of 94.0%. The dual-matrix graph topology prevented dominant background features from washing out sparse transient events, improving the minority “openclose” class F1-score to 55.7% compared to the 48.0% ceiling of a static graph topology. These results demonstrate that explicitly coupling localized spectro-temporal representations with physically anchored spatial topologies consistently outperforms models that process these domains in isolation, providing a highly robust and scalable solution for real-world continuous monitoring systems. Full article

Gas concentration in coal mine return-air corners is affected by ventilation, mining disturbance and gas drainage conditions, and it shows strong nonstationarity, local fluctuation and dynamic multi-point correlations. To improve frequency information separation, monitoring point relationship modeling, and short-term prediction accuracy, a variational mode decomposition (VMD)-based dual-branch spatiotemporal graph method is proposed. Gas concentrations from four key monitoring points are used as inputs, and the return-air corner gas concentration is taken as the output. First, the raw series are decomposed by VMD and reconstructed into low- and high-frequency components. Then, two branches are built for different frequency components. The low-frequency branch combines adaptive graph learning, graph convolution and gated recurrent units to extract global variation features, while the high-frequency branch combines graph attention and gated recurrent units to capture local disturbance features. Finally, a feature-fusion module generates multi-step predictions, and a lightweight short-term warning strategy is developed based on the predicted values. The proposed model achieves MAE, RMSE and R² values of 0.0338, 0.0471 and 0.9499 in one-step prediction, respectively, and outperforms GRU, LSTM, GCN-GRU, GAT-GRU, VMD-GRU, Informer and STGCN under three-step and six-step conditions. Cross-dataset validation and inference time analysis indicate good adaptability and online prediction potential. Full article

An animal’s emotion is regarded as a critical indicator for assessing its welfare. At present, most canine emotion recognition models rely on static images, which leads to insufficient modeling of spatiotemporal features and low accuracy. To address this limitation, this study proposed a 3D ResESAM model based on 3D ResNet50, which is a pre-trained model on the Kinetics dataset. Specifically, the Enhanced Sequential Attention Module (ESAM) is integrated into the backbone to improve the capability of extracting and enhancing spatiotemporal features, and YOLO (You Only Look Once) was employed to detect and crop canine facial regions to reduce interference from complex backgrounds (e.g., furniture, humans, or other animals). To evaluate the effectiveness of 3D ResESAM, a canine emotion dataset named Dog-Face was constructed, which contains four emotions: happiness, sadness, anger, and calmness. Experimental results demonstrate that the proposed 3D ResESAM achieves an accuracy of 0.9239, representing a 3.25% improvement over the baseline 3D ResNet50. Furthermore, compared with other models such as C3D and R(2 + 1)D, whose accuracies are just 0.5215 and 0.8570, respectively, from different attention mechanism perspectives, the proposed model outperforms not only the baseline model without ESAM but also models incorporating other attention modules such as ECA and CBAM, with average accuracies being 0.8972 and 0.9103, respectively. In addition, cross-domain experiments are conducted on a human emotion dataset. The results show that 3D ResESAM achieves an average accuracy improvement of 2.39% over 3D ResNet50, demonstrating its significantly superior performance, effectiveness and potential in real-world applications. Full article

High-accuracy parcel-level agricultural mapping is fundamental to precision agriculture. However, in fragmented agricultural regions of the Yangtze River Delta, identifying cropping patterns at the parcel level faces two compounding challenges: asynchronous multi-source observations and mixed-pixel effects in small parcels. When historical archive records are used as training labels, inter-annual cropping pattern changes further introduce label noise that undermines model reliability. To address these challenges and the label noise issue, we propose PAST (Parcel-level Asynchronous SpatioTemporal), a parcel-level cropping pattern classification framework comprising three stages: K-Shape-based label quality control, parallel dual-branch classification, and decision-level fusion. PAST employs a dual-branch architecture: the temporal branch achieves interpolation-free cross-modal phenological fusion of Sentinel-1 and Sentinel-2 data, while the image branch extracts canopy texture features from 0.8 m high-resolution imagery to partially address mixed-pixel interference. Experiments in a typical fragmented agricultural region of the Yangtze River Delta demonstrate that PAST achieves an overall F1 score of 0.926 and a small-parcel F1 score of 0.906, outperforming mainstream time-series baselines. These results confirm that combining K-Shape label quality control at the data level with a dual-branch interference-robust architecture at the model level provides a complete integrated three-stage pipeline for fine-grained crop mapping under weakly supervised historical archive label conditions. Full article

In optical remote sensing 3D reconstruction, high-resolution satellite stereo matching is a critical task, yet it is challenged by extreme imaging geometries, texture-less and repetitive patterns, occlusions, and scene variations caused by spatio-temporal heterogeneity. To address these issues, we propose IFMA-Stereo, an innovative binocular disparity estimation method that leverages a monocular depth foundation model. Our approach constructs a multi-scale spatial information pyramid to jointly integrate the foundation model with a disparity extraction network. At the feature level, an attention interaction mechanism captures multi-dimensional contextual dependencies and transforms general scene understanding priors into long-range associative features suitable for stereo cost volume construction. At the pixel level, a cyclic iterative refinement module embeds depth information from the foundation model throughout the iteration process and performs joint optimization, enhancing the model’s adaptability in geometrically complex regions. Experiments on the US3D and GaoFen-7 datasets demonstrate that IFMA-Stereo achieves superior performance in challenging areas (texture-less regions, disparity discontinuities, repetitive patterns) and effectively mitigates prediction errors caused by spatio-temporal heterogeneity, albeit at the cost of increased inference time compared to baseline methods. Quantitatively, the method achieves an end-point error (EPE) of 1.347 and a D1 error of 7.26% on the US3D dataset, and an EPE of 1.585 and a D1 error of 13.41% on the GaoFen-7 dataset. Notably, the method also yields precise predictions for unseen urban areas, indicating strong generalization. These results confirm that IFMA-Stereo achieves state-of-the-art accuracy in remote sensing disparity estimation. Full article

Accurate detection of plant leaf diseases is essential for enhancing crop productivity and supporting global food security. In addition to disease classification, understanding how environmental and soil conditions affect model performance is important for developing robust real-world agricultural monitoring systems. Although deep learning (DL) models achieve high accuracy on benchmark datasets, their performance in real-world settings is often limited by variations in illumination, background complexity, and environmental conditions. This study proposes a smart DL framework for detecting and classifying multiple leaf diseases in tomato, potato, and pepper plants. The framework combines U²-Net-based leaf segmentation with a Convolutional Neural Network–Bidirectional Gated Recurrent Unit (CNN–Bi-GRU) architecture. MobileNetV2 is employed as the feature extraction backbone to capture spatial characteristics, while Bi-GRU layers model sequential feature dependencies, forming a spatio-temporal network whose architectural design prioritizes parameter efficiency through depthwise separable convolutions and reduced gating complexity. The model was trained and validated using the PlantVillage benchmark dataset and achieved a classification accuracy of 99.8% with a macro-averaged F1-score of 94%, outperforming several state-of-the-art architectures. To assess robustness under real-world conditions, the trained model was further tested on leaf images collected from open-field environments near Eluru, South India. The field evaluation revealed a reduction in classification accuracy to 61.97%, indicating the impact of domain shift and environmental variability. To investigate potential contributing factors, soil parameters, including pH, temperature, moisture, and NPK levels, were monitored using an IoT-based Arduino sensing system over ten consecutive days. Rather than serving as direct inputs to the disease classification model, these environmental measurements were analyzed to assess their potential influence on disease symptom expression and the observed reduction in model performance under field conditions. The results suggest that environmental conditions may influence disease symptom expression and model transferability. This study highlights the importance of integrating DL-based disease recognition with environmental monitoring for reliable field-level agricultural applications. Nevertheless, computational complexity metrics, including inference latency and memory footprint, were not evaluated in the present work and are identified as a priority for future edge deployment studies. Full article

To address the challenges of severe optical attenuation and dynamic feature extraction for moving target depth retrieval in complex underwater remote sensing environments, this paper proposes a dynamic target depth estimation method based on multi-source data fusion. Taking optical RGB imagery and neuromorphic vision (NeuroIV) data as joint inputs, the proposed method constructs a three-channel feature extraction and fusion network. By leveraging a hypergraph structure, it establishes association weights between dynamic (temporal) and static (spatial) nodes to capture spatiotemporal correlations. To efficiently process the high-dimensional multi-modal data, the traditional dot-product attention is replaced with element-wise multiplication, significantly reducing computational complexity. Furthermore, a lightweight deformable attention pyramid (DAP) and diffusion model is introduced to refine depth image edges, effectively suppressing discontinuities and abruptness in the estimation results. Compared to single-modality optical imagery, the fused multi-modal data yields a superior signal-to-noise ratio and foreground contrast, achieving an improvement of over 20% in the MAE index. These results validate the effectiveness and superiority of the proposed multi-modal fusion strategy for dynamic target observation and depth retrieval in aquatic environments. Full article

Forecasting fine particulate matter with a diameter of 2.5 μm (PM_2.5) is critically important due to its adverse effects on human health and environmental sustainability. Although ground-based monitoring stations provide accurate measurements, their limited spatial coverage restricts large-scale PM_2.5 assessment, especially in complex urban regions. Consequently, aerosol optical depth (AOD) derived from satellite imagery, combined with advanced deep learning (DL) techniques, has emerged as an effective alternative by offering wide spatial coverage and rich spatio-temporal information. This paper proposed an optimization-driven fuzzy transformer–deep belief network (ODFT-DBN) for accurate PM_2.5 air pollution prediction. The proposed framework integrates a fuzzy inference module to model uncertainty and nonlinear environmental relationships, a transformer encoder to capture long-range spatio-temporal dependencies, and a DBN to extract hierarchical features and improve prediction robustness. In addition, a novel multi-objective gray wolf optimizer (NMOGWO) is employed to jointly optimize the model hyper-parameters and fuzzy membership functions. The proposed approach is implemented for the city of Tehran, Iran, using meteorological variables, topographical features, ground-based PM_2.5 measurements, and satellite-derived AOD data. The ODFT-DBN model is compared with several benchmark methods, including bidirectional encoder representations from transformers (BERT), transformer, long short-term memory (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN), DBN, and extreme gradient boosting (XGBoost). Experimental results demonstrate that the proposed framework achieves superior predictive performance, attaining an R² value of 0.94 and root mean square error (RMSE) of 0.8 μg/m³. Scatter plot analyses indicate a strong agreement between predicted and observed PM_2.5 values, while the proposed model exhibits low variance, stable convergence behavior, and acceptable computational time. Overall, the results confirm the effectiveness, robustness, and practical applicability of the proposed ODFT-DBN framework for spatio-temporal PM_2.5 forecasting. Full article

This study examines the evolving patterns and zoning characteristics of gas migration and storage zones during coal seam mining, taking the 215 fully mechanized longwall face at Huangling No. 2 Coal Mine as the engineering background. By integrating theoretical analysis, physical similarity simulation experiments, and field measurements, the research systematically explores the zonal linkage evolution mechanism of mining-induced depressurization gas migration and storage zones, together with the associated depressurization gas extraction technology. A flow regime determination equation, driven by the fracture expansion coefficient and permeability, is established on the basis of the fluid Reynolds number criterion. According to differences in gas flow states and medium morphology, the mining-induced fracture field is divided into five distinct zones: a high-permeability zone dominated by turbulent transport, a medium-to-high permeability zone with transitional flow as the secondary dominant region, a low-permeability zone featuring linear laminar flow with micro-permeability, an extremely low-permeability zone characterized by linear laminar flow in a locked state, and a zone of abrupt permeability change associated with gas enrichment. The dynamic evolution of depressurization gas migration and storage zones and their regional linkage mechanisms are clarified. On the basis of these findings, a dynamic targeted layout strategy for high-level boreholes is proposed that is consistent with the spatiotemporal evolution of the overburden permeability field. Field engineering practice shows that the optimized high-level borehole layout maintains the overall gas extraction rate at the drilling site stably above 70%, with a peak value of 93.7%, thereby ensuring safe and efficient mining of the working face. Full article

Existing seizure detection methods cannot fully exploit the spatiotemporal features of multimodal signals. They also fail to capture deep associations among cross-modal features. This limits their ability to learn unified representations of spatiotemporal dependencies. This work proposes CMEpiNet (Complex-valued Multimodal Epilepsy detection Network model) to address this issue. CMEpiNet first uses complex-valued convolutions for feature extraction. It explicitly models phase synchronization, phase shifts, and cross-frequency coupling. Thus, EEG, ECG, and EMG features are represented in the complex-valued domain. During feature fusion, CMEpiNet uses a two-level semantic alignment-based fusion method. It applies cross-modal consistency constraints in a shared alignment space. It also performs distribution-level alignment in an epilepsy-related semantic latent space. These operations ensure the consistency of multimodal features in the global semantic structure. Finally, CMEpiNet uses a spatial attention-guided 3D convolutional classifier. The classifier jointly models the temporal, feature, and modality dimensions. Experimental results on the SeizeIT2 dataset show that CMEpiNet improves seizure detection sensitivity, reduces the false alarm rate, and maintains stable performance under perturbations. Full article

Accurate break size assessment is critical for the safety response of small modular reactors (SMRs) during loss-of-coolant accidents (LOCAs). Traditional methods struggle with the rapid transient features, strong spatiotemporal coupling, and complex uncertainty characteristics of SMR-LOCA, leading to low accuracy and poor stability. To address these issues, this study proposes an entropy-driven intelligent diagnosis approach based on a CNN-LSTM-Attention hybrid model. The framework adopts information entropy for data uncertainty quantification, adaptive weighting, and loss constraint, so as to realize high-precision break size assessment. A time-series dataset covering break sizes from 0.05 to 10 cm² was constructed using the PCTRAN/SMART platform. The CNN module extracts spatial coupling features of multi-sensor parameters, the LSTM module captures long-term temporal dependencies, and the attention mechanism dynamically weights key information to enhance feature representation under high uncertainty. Experimental results show that the model achieves a mean absolute error (MAE) of 0.096311, reducing errors by over 64.4% compared with baseline models; more than 90% of prediction errors are within ±5%, and the correlation coefficient reaches 0.994902. Based on the well-validated PCTRAN/SMART simulation platform, the proposed entropy-informed spatiotemporal learning framework provides a technical solution for intelligent LOCA diagnosis, uncertainty quantification, and safety assessment of SMRs. Full article

The increasing sophistication of cyber attacks has created a growing demand for effective Network Intrusion Detection Systems (NIDSs). Although deep learning has improved NIDS performance, existing models often lack adaptive mechanisms for spatiotemporal feature fusion and struggle with complex traffic distributions characterized by severe intra-class heterogeneity and inter-class overlap. Meanwhile, current interpretability methods mainly rely on feature importance analysis and provide limited insight into the model’s decision process. To address these challenges, we propose HARMONI, a two-stage hybrid learning framework that enhances both detection accuracy and model interpretability. In the first stage, a dual-branch Convolutional Neural Network and Gated Recurrent Unit (CNN-GRU) architecture extracts spatiotemporal features, which are dynamically fused through a lightweight adaptive gating network. The representation learning process is jointly optimized using a Dynamic Class-Center Loss to enforce intra-class compactness and inter-class separability in the latent space. In the second stage, the learned deep representations are concatenated with raw traffic features and fed into an ensemble classifier. This residual-style design mitigates information loss during deep encoding while leveraging the non-linear modeling capability of ensemble learning. We further develop a multi-level interpretability framework based on SHapley Additive exPlanations (SHAP) that analyzes global feature importance, individual feature contributions, and feature interactions to provide quantitative insights into the model’s decision mechanisms. Experiments on four benchmark datasets show that HARMONI consistently outperforms state-of-the-art baselines, achieving 80.19% and 78.24% accuracy on NSL-KDD and UNSW-NB15 respectively, surpassing representative deep learning and ensemble methods. Full article