Search Results (490)

In this work, a lightweight framework enabled by the modified YOLOv11s-seg model for tea shoot detection and plucking point localization is proposed. Detecting tea shoots and localizing plucking points with higher accuracy generally require larger model size and more model parameters, making it difficult to balance accuracy and lightweighting. To overcome this limitation, a modified lightweight YOLOv11s-seg model is developed. First, the multi-scale edge information enhancement is introduced into the conventional YOLOv11s-seg to extract edge feature better and improve the detection accuracy of tea shoots. Meanwhile, context anchor attention is utilized to modify the cross stage partial spatial attention module in a backbone network to improve the detection capability for small objects. Moreover, the detail calibration reconstruction feature pyramid network is proposed. It utilizes spatial and contextual semantic information to reconstruct and calibrate features in key regions, enhancing the capability for object fusion and recognition at various scales. Furthermore, with the modified model performing instance segmentation to acquire the contour of each tea shoot, the coordinates of the three lowest pixel points in the contour are captured to localize the plucking point based on the average coordinates. In addition, the layer-adaptive magnitude-based pruning (LAMP) method is used to lighten the model. The experimental results show that the LAMP-pruned modified YOLOv11s-seg model with a speedup ratio of 1.5 achieves a mAP@0.5 of 86.5% for tea shoot detection, exhibiting a 4.7 percentage point improvement over the conventional YOLOv11s-seg model. Moreover, it exhibits an accuracy of 81.9% for plucking point localization on the validation and test subsets with 232 images in total, and its number of parameters, model size and floating point operations (FLOPs) separately achieve reductions of 67.3%, 66.2%, and 24.9% over the conventional model as well. Therefore, the proposed LAMP-pruned modified model shows good balance between lightweighting and detection accuracy. Finally, the modified LAMP-pruned YOLOv11s-seg model is deployed on a Jetson Orin NX edge module and measured in a tea plantation, with the measured results exhibiting a detection speed of 34.1 FPS and verifying its availability in practical applications. Full article

Detecting objects in drone-captured aerial imagery is particularly formidable due to challenges such as the prevalence of numerous small targets and their dense spatial distribution. To bridge this gap, this paper introduces YOLO-UTD (YOLO-UAV Traffic Detection), a dedicated small object detector tailored for drone traffic surveillance. Built upon the YOLOv8 framework, the proposed model incorporates three principal enhancements. First, a specialized small-object detection head replaces the original large-object head to increase the sensitivity to fine-grained features. Second, we introduce a shallow-augmented feature pyramid network (SFPN) into the neck module. The SFPN enriches the semantic content of high-resolution shallow features via dense multiscale interactions and CARAFE upsampling, boosting performance on small targets. Finally, a C2fA layer is integrated into the deep backbone stages to adaptively fuse spatial details and semantic context through a dual-path architecture and a cross-attention mechanism, thereby dynamically refining features critical for small objects. Extensive experiments on the VisDrone2019 dataset validate that YOLO-UTD achieves a 3.6% higher mean average precision (mAP) than YOLOv8 while preserving a low parameter footprint, with a particularly significant gain of 5.3% in vehicle detection accuracy. These findings confirm the model’s efficacy and strong potential for application in smart city drone surveillance. Full article

Spatially explicit water-quality information is critical for precision management in pond aquaculture but point sampling alone cannot capture pond-to-pond heterogeneity in multi-unit farms. This single-date, single-farm study re-evaluated the potential of UAV hyperspectral imagery for water-quality screening in inland aquaculture ponds in Shanghai, China, using site-matched extraction from a 138-band orthomosaic (450–998 nm, Cubert S185) acquired during a single UAV survey on 24 August 2023 and matched with 23 GPS-registered sampling sites. Eight water-quality parameters were analyzed: chemical oxygen demand (COD), total phosphorus (TP), total nitrogen (TN), ammonium (${\mathrm{NH}}_4^{+}$ ), nitrite (${\mathrm{NO}}_2^{-}$), nephelometric turbidity unit (NTU), chlorophyll-a (Chla), and total suspended solids (TSS). Raw single-band correlations were modest ($\left|\mathrm{r}\right|$= 0.236–0.417), but two-band difference spectral indices (DSI), normalized spectral indices (NSI), and ratio spectral indices (RSI) substantially improved sensitivity, with $\left|\mathrm{r}\right|$ reaching 0.558–0.928. Quadratic inversion models were calibrated on the full dataset and assessed using three validation layers: two-fold cross-validation, nested leave-one-pond-out (LOPO) validation with within-fold predictor reselection, and extraction-window sensitivity tests. Bootstrap 95% confidence intervals for calibration (Cal) R² characterize small-sample uncertainty (n = 23). Three parameters satisfied all three defensibility criteria (Cal R² > 0.5, CV R² > 0.2, and LOPO R² > 0.2): ${\mathrm{NH}}_4^{+}$ (Cal R² = 0.836 [0.61, 0.94]; LOPO R² = 0.420), COD (0.607 [0.34, 0.82]; 0.328), and NTU (0.862 [0.77, 0.96]; 0.204). TP, TN, ${\mathrm{NO}}_2^{-}$, TSS, and Chla showed overfit behavior under nested holdout and were demoted to exploratory products. A TreeSHAP analysis confirmed that band-to-band contrast carried more explanatory power than raw reflectance magnitude. Extraction-sensitivity tests further demonstrated that positional uncertainty (±2-pixel offset: ΔCV $R^2$= 0.23–0.41) exceeded averaging-window sensitivity (3 × 3→10 × 10: ΔCV $R^2$ ≤ 0.11), identifying geolocation control as the dominant robustness constraint. This single-date, single-farm reanalysis suggests that UAV hyperspectral imagery may support exploratory pond-scale screening of ${\mathrm{NH}}_4^{+}$, COD, and NTU. However, robust quantitative inversion and broader transferability remain unverified and will require denser sampling, improved geolocation control, pond-edge masking, multi-site observations, and multi-temporal calibration. Full article

Wire-arc directed energy deposition (arc-DED) of maraging 250 (M250) steel is of growing interest for aerospace, tooling, and defense applications, yet systematic process characterization data remain limited. This study presents a mixed quantitative–qualitative factorial comparison of three Fronius synergic transfer modes, GMAW-CMT-Mix, GMAW-CMT-Universal, and GMAW-Pulsed-Arc, for single-bead M250 deposition across wire feed speeds of 4.45 to 8.26 m/min and travel speeds of 0.3 to 1.5 m/min. Bead geometry and process behavior are characterized using non-contact optical profilometry and destructive methods (i.e., metallographic sectioning, optical microscopy, and Vickers microhardness). The material feed rate ratio, R_wt, is introduced as a unifying process descriptor; heat input and cross-sectional area scale linearly with R_wt, while travel speed primarily governs bead height and wire feed speed primarily governs bead width. At the highest travel speed tested, GMAW-CMT-Mix and GMAW-Pulsed-Arc exhibit bead humping, rendering those conditions unsuitable, while GMAW-CMT-Universal maintains stable deposition with consistent dilution and the lowest heat input at equivalent R_wt. GMAW-CMT-Mix yielded the highest dilution and hardness. Linear regression of process responses against R_wt gives $R^2$ exceeding 0.83 for both height and width across all modes. These results establish a characterization baseline supporting future multi-layer studies. Full article

Early depression screening is important for timely intervention, and electroencephalography (EEG) offers an objective and potentially portable sensing modality for computer-aided assessment. However, recognition from fixed three-channel frontal EEG remains difficult because of limited spatial information and incomplete modeling of temporal–frequency characteristics and temporal dependencies. This study proposes DEP-TFDualNet for acquisition-constrained frontal resting-state EEG. The framework integrates multi-scale convolution, dual-domain channel attention, temporal modeling derived from the independent recurrent neural network (IndRNN) architecture, and decision-stage fusion of deep representations with low-order statistical descriptors through a Kolmogorov–Arnold Network (KAN)-based nonlinear projection layer. Experiments were conducted on the publicly available three-channel frontal EEG subset of the MODMA dataset. After additional quality control, 48 subjects were retained (22 patients with major depressive disorder, 26 healthy controls). Under subject-wise stratified five-fold cross-validation, DEP-TFDualNet achieved 85.42% accuracy, 85.26% macro-F1, 81.82% sensitivity, 88.46% specificity, an AUC of 0.82, and a Brier score of 0.121. It achieved the best threshold-based subject-level performance and the lowest Brier score among the evaluated models. These results provide preliminary evidence that simplified frontal EEG sensing may support depression recognition in acquisition-constrained settings, although larger and external validation is still required. Full article

Global agriculture is seriously threatened by plant diseases, which result in large losses in both productivity and quality. Timely and accurate disease detection is essential for effective crop management and food security. This work presents an improved MobileNetV2 architecture with Modified Lite Attention (MLA) Model for detecting plant leaf disease. Our methodology incorporates pre-processing, feature extraction through attention model, convolution layers, and classifying into diseased or healthy categories. Further, multiclassification of diseases is performed on a dataset comprising 4432 samples including whitefly, leaf spot, leaf curl, yellowish and healthy leaves. The proposed attention model is compared with existing attention models like CBAM (Convolutional Block Attention Model), SE (Squeeze and Excitation), ECA (Efficient Channel Attention) and SDMnet (Spatially Dilated Multi-Scale Network) to validate our hybrid MLA feature extraction technique. Customizing the categorization with fully connected layers and utilisation of a pre-trained MobileNetV2 model allow the system to achieve excellent results. Findings show encouraging accuracy, surpassing 97% compared to existing techniques for multiclass dataset classification. The integration of MobileNetV2 with custom dense layers enables robust detection even with limited datasets, making it ideal for use in mobile or low-resource agricultural environments. Further, the proposed method is tested on the PlantVillage dataset consisting of 10,836 samples using K-Fold cross-validation for K = 5 and K = 4 to obtain an average accuracy of 98.4% and 98.69%, respectively. Full article

Wearable-sensor human activity recognition (HAR) plays a key role in health monitoring, elderly care, and human–computer interaction. Deep learning dominates the field, but two limitations remain. CNNs with fixed kernels cannot capture cross-scale temporal events such as gait cycles and postural transitions in a single layer, and softmax attention on small sensor datasets is often diluted by common-mode background responses across the sequence. We propose DMSCNet, an end-to-end framework with two modules. The Dilated Multi-Scale Branch Block (DMSB) combines a shared bottleneck, parallel dilated convolutions, a pooling bypass, and SE-based channel recalibration to widen the temporal receptive field under a controlled parameter budget. The Contrastive Temporal Attention (CTA) module adopts a dual-path differential design, in which the two paths learn overlapping but non-identical attention patterns and their subtraction suppresses shared low-level responses while preserving the discriminative positions each path locks onto, encoded with opposite signs. DMSB and CTA are cascaded into a DMSC Block and stacked residually. On UCI-HAR, USC-HAD, and RealWorld, DMSCNet reaches F1-scores of 97.65%, 91.80%, and 99.05%, outperforming nine baselines. Ablations confirm that SE acts along the channel axis and CTA along the temporal axis, and visualization reveals a dynamic–static dichotomy together with a signed bipolar encoding pattern produced by the dual-path subtraction. Full article

Cross-laminated timber (CLT) is increasingly used in multi-story timber construction, but its combustible nature requires reliable fire protection, particularly in layered wall assemblies with concealed cavities. This study compares two medium-scale cross-laminated timber (CLT) sandwich wall assemblies exposed to radiant heat flux of 20 kW/m² for 90 min: an uncoated reference assembly and an assembly with PROMADUR^® intumescent coating applied to the CLT surfaces. Both specimens consisted of a 90 mm three-ply CLT panel encapsulated with 12.5 mm gypsum-fiber boards fixed to a wooden stud frame forming a 40 mm installation cavity. Fire-test observations were supplemented by simultaneous thermal analysis (STA), i.e., thermogravimetry (TG)/differential thermogravimetry (DTG)/differential scanning calorimetry (DSC), of uncoated and coated CLT specimens under oxidative conditions. During the applied medium-scale radiant exposure, the unexposed-face temperatures of both assemblies remained below the insulation temperature-rise limits defined in STN EN 1363-1; however, these limits were used only as a comparative benchmark and the test does not represent a formal fire-resistance classification. The coated assembly showed improved thermal protection during the early and intermediate stages of exposure, delaying a critical thermal event near the wooden stud by approximately 35 min. However, flaming combustion of the stud occurred at about 75 min and led to degradation of the intumescent char within the cavity. In contrast, the uncoated assembly reached higher early CLT surface temperatures but showed no flaming combustion during the test. STA results supported the fire-test interpretation: the coated specimen showed a 37% reduction in peak DTG rate, a higher residual mass at the end of the test, and substantially greater mass loss in the 150–280 °C range, consistent with intumescent activation and volatile release. The results indicate that, under the tested medium-scale exposure, the intumescent coating improved early and intermediate thermal protection of the CLT surface, but did not prevent late-stage cavity flaming involving the wooden stud. Therefore, the behavior of intumescent-coated CLT in partially enclosed cavities with combustible framing should be validated under replicated, standardized and larger-scale fire exposure. Full article

Cardiovascular diseases impose a substantial global health burden and often require timely detection, creating strong demand for real-time electrocardiogram (ECG) monitoring on resource-constrained devices. Although portable single-lead wearable ECG devices are valuable for daily monitoring, their diagnostic performance is limited by spatial information loss and hardware constraints. Moreover, conventional lightweight models lack interpretable analysis beyond coarse classification. This study proposes an edge–cloud collaborative ECG-assisted analysis method combining lightweight ECG model distillation with large language models. At the algorithmic level, a cross-lead distillation framework transfers knowledge from a 12-lead InceptionTime–Transformer teacher to an ultra-lightweight single-lead student via a hybrid loss integrating hard-label, temperature-scaled soft-label, and auxiliary multi-label objectives. At the system level, a three-layer architecture integrates edge-side real-time screening with cloud-side report generation through a LoRA-fine-tuned Qwen3-8B model. Experiments on PTB-XL show that, under 123.7× parameter compression and 12-to-1 lead reduction, the student retains 92.8% of the teacher’s Macro-F1 and 94.7% of its AUC-ROC. After 8-bit integer (INT8) quantization, the TFLite file is 20.8 KB; QEMU-based Cortex-M4 simulation shows approximately 63.0 KB SRAM usage and 11.6 ms latency, suggesting potential on-device deployment under simulated conditions. Validation on physical hardware—including power consumption, BLE latency, and motion artifacts—remains necessary. Full article

To address the issues of insufficient single-modality robustness and limited multi-scale object detection accuracy in remote sensing image detection (RSID) in complex environments, this paper proposes a multimodal RSID network named VI-MSFFN. The model adopts a symmetric parallel dual-branch architecture to achieve independent extraction and collaborative modeling of visible and infrared modal features. A cross-modal multi-scale sparse cross-attention fusion module is proposed and applied to the P4 and P5 feature layers, and a high-low-level feature collaborative cross-modal fusion strategy was constructed to achieve efficient and robust cross-modal feature fusion while enhancing multi-scale object modeling capability and suppressing feature redundancy and noise. Additionally, a progressive feature interaction and fusion architecture was designed to combine spatial and frequency domain information to strengthen deep object representation. The experimental results on the VEDAI and Drone Vehicle datasets demonstrate that VI-MSFFN achieves state-of-the-art (SOTA) performance in detection accuracy, robustness, and generalization ability. The proposed method effectively solves the detection challenges of RSID and has significant application value in the field of multi-modal RSID. Full article

Accurate viral protein function annotation is essential for genomic surveillance, yet conventional retrieval-augmented generation (RAG) pipelines often fragment biological evidence into fixed-length text chunks, disrupting relationships among ORFs, annotations, structural domains, sequence motifs, residue mappings, and model-derived attention evidence. We propose ViroBioTree, a tree-structured biological evidence retrieval framework for downstream viral protein evidence review rather than a new primary annotation classifier. Built as an evidence organization layer on ViralMultiNet-derived ORF-level predictions and annotations, ViroBioTree converts sequence, annotation, structure, and attention evidence into typed biological nodes and traceable edges, then performs deterministic multi-channel recall, evidence-aware reranking, balanced TopK selection, rule-based verification, and node-cited report generation. In a demo benchmark, ViroBioTree achieved its strongest deterministic proxy performance on structure-explanation tasks, with Precision@K = 1.0, Recall@K = 1.0, and diversity = 0.52; these values reflect expected node-type and tag agreement rather than independent biological correctness. A bounded full-scale SARS-CoV-2 index contained 39,800 ORF rows, 80,000 attention records, 199,418 nodes, and 495,886 edges. In a stratified full20k diagnostic evaluation, ViroBioTree showed task-dependent advantages over LlamaIndex vector retrieval for conflict detection, evidence retrieval, and structure explanation, while LlamaIndex remained competitive or stronger for annotation-rich function annotation. A cross-family Influenza A Virus (IAV) diagnostic audit showed that the schema can represent IAV evidence namespaces while explicitly exposing missing formal ORF inputs, missing attention evidence, and unavailable residue/PDB assertions. Supplementary robustness, external sanity-check, diversity-risk, expert-evaluation, domain-tool positioning, and cross-family audit analyses supported traceability, report quality, and conservative evidence handling, but also showed that stable Precision@K under query perturbation does not necessarily imply stable retrieved evidence sets. ViroBioTree operates offline and deterministically, but does not address raw-read assembly, base calling, primary ORF prediction, or wet-lab validation. Its results should be interpreted as proxy and expert-reviewed evidence for traceable viral protein evidence retrieval and report generation rather than as direct validation of biological function annotation. Full article

With the rapid proliferation in communication devices and the expansion of applications, future sixth-generation (6G) networks are expected to enable a truly connected world. They will allow large-scale use cases, such as the Internet of Things (IoT) and unmanned aerial vehicles (UAVs), providing significantly faster and more innovative services ubiquitously. However, challenges remain, particularly in security. The growing number of devices and increased connectivity may lead to a larger attack surface. Many emerging technologies are actively addressing these security and privacy concerns, ensuring that we can benefit from the advantages of 6G networks and applications without falling victim to malicious attacks. In this paper, we conduct a comprehensive literature review of emerging technologies for secure communication in 6G networks, including artificial intelligence (AI) and machine learning (ML), blockchain technology, quantum-safe communication, and physical-layer security. First, we discuss the architecture of 6G networks from a security perspective. Second, we review existing surveys on 6G security issues and provide a quantitative analysis to identify research gaps, including technology-driven silos and domain fragmentation. Third, we develop a hierarchical taxonomy of security challenges and attacks in 6G networks, covering physical-layer attacks, network-level threats, device vulnerabilities, data privacy concerns, and emerging application-specific risks. We then examine the roles of key enabling technologies and present a mapping between security threats and corresponding technological solutions, along with a unified evaluation framework to facilitate cross-technology comparison. Furthermore, we propose an integrated multi-technology security framework and discuss practical deployment challenges by bridging the gap between simulation-based studies and real-world implementations. Finally, we outline concrete future research directions for advancing secure 6G communication systems. Full article

Automatic building extraction plays an important role in various remote sensing applications, such as seismic disaster investigation, seismic hazard risk assessment, urban planning, and photogrammetry. Despite the substantial progress, state-of-the-art building extraction methods are still limited by two issues: (i) existing approaches leverage convolutional layers or non-local self-attention to encode the position-aware dependencies, while they cannot flexibly adapt to the complex background contexts and varied structure patterns of buildings; and (ii) the local details cannot be well preserved by existing hierarchical decoders due to the imperfect feature aggregation, yielding unsatisfactory segmentation outputs in the local adjacent region. To address these issues, we propose Relative Position Aggregated Transformer (RPAFormer), which is capable of modeling the relative position dependencies of buildings and producing accurate local details using a dual attention transformer network. Specifically, we propose a Relative Position-aware Self-attention (RPSA) framework to learn the token dependencies within the local window. A transformer decoder network consisting of multiple Cross Masked Attention (CMA) blocks is also introduced to fuse the multi-scale features. Extensive experiments demonstrate the superior performance of the proposed method and its great promise for real-world engineering deployment. Full article

Accurate volatility forecasting and timely risk early warning are foundational requirements of financial risk management: Value-at-Risk estimates, portfolio risk limits, derivative hedging ratios, and stress-test scenario calibrations all depend on forward-looking volatility signals that remain reliable when market conditions depart from average. This paper develops a public-data causal multiscale wavelet spillover learning (CMWSL) framework that jointly addresses stock-index volatility forecasting and high-volatility early warning under strict walk-forward evaluation. CMWSL integrates three components: a heterogeneous autoregressive (HAR) persistence block as the dominant linear baseline, causal stationary wavelet transform (SWT) summaries that encode within-index multiscale market dynamics, and a cross-index spillover layer that tests whether medium- and long-scale wavelet energy from peer indices carries incremental risk-relevant information. The empirical analysis covers the S&P 500, Nasdaq-100, and Dow Jones Industrial Average over a 2513-step out-of-sample evaluation period from 2016 to 2025, with forecast horizons $h\in \{1,5,10\}$ and OHLC-based volatility targets. All preprocessing, wavelet decomposition, calibration rules, and warning thresholds are re-estimated inside each rolling training window to eliminate look-ahead bias. HAR remains the strongest average model in the main Rogers–Satchell specification, confirming that daily index volatility risk is highly persistence-driven. The multiscale extension delivers statistically significant improvements at longer horizons, in richer public macro-financial information environments, and under the Parkinson target. Clark–West tests detect significant spillover gains in five of nine index–horizon cells (CW $=4.83$, $p<0.001$ for S&P 500 at $h=10$). Critically, tail-conditioned and rolling-window diagnostics show that multiscale and cross-index gains concentrate in upper-volatility regimes and synchronized stress episodes—precisely the conditions in which risk management decisions are most consequential. For market-risk early warning, a logistic classifier built on the same causal feature pipeline delivers the most stable precision–recall performance across all settings, providing an interpretable and operationally auditable alert mechanism suitable for practical risk monitoring. Full article

With the widespread adoption of Wi-Fi 7 in campus networks, high-density access and large-scale research data transmission challenge traditional congestion control algorithms. TCP-bottleneck bandwidth and round-trip propagation time (BBR) lacks deep link awareness and service semantic breadth, leading to misinterpreting non-congestive packet loss and inter-flow unfairness in complex wireless scenarios. To address this, this paper proposes semantic-aware and cross-layer collaborative optimized BBR (SAC-BBR), a semantic-aware cross-layer optimization mechanism for high-density heterogeneous campus networks. It leverages an Extended Berkeley Packet Filter (eBPF) to capture physical link characteristics in real time within the Linux kernel, accurately distinguishing random loss from congestion loss. It then constructs a lightweight semantic identification engine to classify traffic and establish a service satisfaction utility model. Finally, a deep reinforcement learning-based dynamic gain regulator maps cross-layer states and service priorities to the action space, enabling millisecond-level intelligent tuning of pacing_gain and cwnd_gain. Experimental results show that SAC-BBR improves throughput by over 22% compared to BBRv3 and reduces average round-trip time (RTT) by 17% while suppressing RTT jitter by over 60% in high-density scenarios. Furthermore, it enhances the Jain fairness index to 0.93 under multi-protocol competition, ensuring high-performance and equitable transmission. Full article