Search Results (467)

Spaceborne microwave imagers are vital for monitoring global precipitation due to their wide swath and high sensitivity. This study proposes a deep learning approach that integrates a U-Net with a multi-task learning (MTL) framework. The model was separately trained over land and ocean using GPM Microwave Imager (GMI) brightness temperatures, with collocated precipitation rates and types from the Dual-frequency Precipitation Radar (DPR) as labels. This combines the accuracy of radars with the coverage of imagers to produce high-precision, wide-swath precipitation estimates. In the MTL setup, near-surface precipitation rate retrieval is the main task, and precipitation type classification is the auxiliary task. A composite loss (weighted MSE and quantile regression) is used for the main task, and weighted cross-entropy for the auxiliary task. Residual blocks and an attention mechanism are incorporated to improve physical representation and generalization, thereby significantly enhancing the model’s capability to retrieve heavy precipitation. The model was trained on 2015–2024 GPM data and evaluated on an independent six-month 2025 GMI dataset. Compared to a standard U-Net, the MTL model achieved significant gains: Pearson Correlation Coefficient (PCC) increased by 9.7% (ocean) and 13.7% (land), and Critical Success Index (CSI) by 10.7% (ocean) and 10.8% (land). The method was also applied to the FY-3G Microwave Radiation Imager (MWRI-RM). In case studies, it outperformed the official product, achieving average increases of 20.1% in PCC and 15.7% in CSI, respectively. Validation against FY-3G Precipitation Measurement Radar (June–August 2024) yielded over ocean PCC = 0.757, RMSE = 1.588 mm h⁻¹, MAE = 0.355 mm h⁻¹; over land PCC = 0.691, RMSE = 2.007 mm h⁻¹, MAE = 0.692 mm h⁻¹. The study demonstrates that the MTL-enhanced U-Net significantly improves the accuracy of spaceborne microwave imager rainfall retrieval and shows robust practical applicability. Full article

The scarcity and incompleteness of well-log data pose a critical challenge to deep learning-based reservoir prediction. To address this small-sample problem and improve prediction quality, we propose a novel semi-supervised asymmetric deep co-training framework integrated with a shape context descriptor. This method leverages abundant unlabeled seismic data as well as complementary information on related physical properties. Specifically, we introduce a shape context descriptor to encode seismic waveform morphology and spatial context, thereby improving the lateral continuity and interpretability of predictions while mitigating issues inherent in the sequence-to-point paradigm, wherein three-dimensional seismic data are used as input and a single target point is predicted. To overcome data limitations, a sliding-window resampling strategy is employed to expand the training samples. For co-training, we design an asymmetric dual-task architecture wherein one model performs porosity regression while the other conducts reservoir type classification, thereby enabling synergistic learning. The proposed framework is validated using real three-dimensional seismic data from the Yinggehai Basin in the South China Sea through ablation experiments. The results demonstrate superior performance in prediction accuracy, spatial consistency, and training stability compared to baseline methods. Full article

Task allocation for heterogeneous unmanned aerial vehicle (UAV) swarms requires complex spatiotemporal coordination. While traditional algorithms struggle to interpret abstract semantic intents, general large language models (LLMs) often suffer from physical hallucinations and superficial tactical reasoning. To address these limitations, we propose a generative task allocation paradigm augmented by a heterogeneous toolchain, shifting the approach from rigid numerical optimization toward tool-grounded semantic planning. To implement this and overcome domain data scarcity, we design a decoupled dual-model architecture. This architecture is optimized through an execution-manifold-anchored orthogonal evolution training method. By utilizing simulated self-play within a stable execution environment, this approach prevents gradient conflicts and autonomously generates abundant training data. Furthermore, to resolve the credit assignment problem in long-horizon scenarios, we develop a Recursive Causal Probe (RCP) algorithm. By tracing failures backward through the simulation, RCP synthesizes counterfactual preference data, effectively translating tactical mistakes into precise corrections for the planning model. Extensive simulations demonstrate that our method achieves an 82.34% mission success rate in complex scenarios, requiring significantly fewer interactive corrections than general LLMs, fully verifying its physical feasibility and practical robustness. Full article

In pedestrian re-identification (ReID) tasks, existing models face dual challenges: first, surveillance cameras capture images at long distances with low resolution and blurriness; second, image data suffers from insufficient samples, limited poses, and cross-domain adaptation issues. To address these issues, we propose a wavelet-convolution-based person re-identification framework assisted by a Stable Diffusion-based identity-preserving image generation module used only at inference time. This approach employs a dual-channel wavelet convolutional neural network for multi-scale feature extraction of pedestrian images, combined with cross-attention and gating mechanisms for dynamic data fusion. Additionally, we incorporate a pre-trained Pose2ID-based auxiliary generation branch that synthesizes identity-preserving pedestrian views with diverse poses under human keypoint guidance. These generated views are used only at inference time, where their WSNet features are fused with the feature of the original image to provide pose-complementary representation enhancement. Experiments on the Market-1501 and MSMT17 benchmark datasets demonstrate that our method achieves an mAP of 92.1% and a Rank-1 accuracy of 96.5% on Market-1501, and an mAP of 60.1% and a Rank-1 accuracy of 81.2% on MSMT17, with a WSNet backbone of 2.66 M parameters. Compared with the baseline models, the proposed method improves mAP by 5.1 and 7.6 percentage points on Market-1501 and MSMT17, respectively. Full article

Coral reefs are among the most biodiverse marine ecosystems, playing irreplaceable roles in maintaining marine ecological balance and coastal services. Under dual pressures of global climate change and human activities, coral bleaching and degradation have become increasingly frequent, creating an urgent need for large-scale, long-term, and highly automated monitoring technologies. In recent years, advances in underwater imaging and deep learning have made visual recognition a core approach for coral classification and health assessment. However, most studies only focus on isolated model accuracy optimization, lacking systematic full-chain analysis integrating datasets, model evolution, cross-domain generalization, engineering constraints, and ecological adaptation, which severely hinders large-scale cross-regional and long-term application. This paper systematically reviews coral visual recognition technologies. It summarizes underwater image acquisition, public dataset characteristics, and annotation system evolution, then compares traditional feature engineering and deep learning in key tasks, highlighting their differences in feature representation and generalization. Four core challenges are identified: class imbalance, poor underwater image quality, weak cross-device/region generalization, and mismatched algorithm metrics with ecological needs. Finally, feasible solutions based on self-supervised pre-training, domain adaptation, and multimodal fusion are discussed to enhance model robustness and ecological interpretability, providing methodological support for intelligent coral reef monitoring systems. Full article

Background: Stroke-related impairments in balance and gait are among the most common and disabling sequelae, significantly limiting functional independence and increasing fall risk. This study investigated the effects of short-term dynamic balance training on balance and gait in post-stroke hemiplegic patients. Methods: In this randomized controlled pilot trial, 16 post-stroke hemiplegic patients (intervention group, n = 8; control group, n = 8; mean age ≈ 58 years; predominantly male) were assigned to either a control group receiving conventional rehabilitation or an intervention group receiving additional daily dynamic balance training using the Prokin-252 system (30 min/day, 5 days/week, 3 weeks). Primary outcome measures included balance performance (Berg Balance Scale, mini-BESTest, single-leg stance), center-of-pressure (COP) parameters, gait performance (Timed Up and Go Test), and surface electromyography (sEMG) activity. Results: Following the intervention, both groups demonstrated improvements; however, the intervention group showed significantly greater gains in balance and gait outcomes. Specifically, Berg Balance Scale scores improved significantly (p = 0.012), as did mini-BESTest scores (p = 0.004). Eyes-closed single-leg stance time increased significantly on both sides (p < 0.05). COP analysis revealed reductions in sway area and trajectory length under challenging conditions. sEMG analysis indicated increased activation of the affected-side gluteus medius. In terms of gait performance, the intervention group demonstrated greater improvements in Timed Up and Go Test performance (p = 0.002), dual-task walking, and gait phase symmetry. Conclusions: Supplementing conventional rehabilitation with dynamic balance training effectively enhances balance and gait function in post-stroke patients, potentially through improved neuromuscular control. The integration of sensor-based COP analysis and sEMG provides additional mechanistic insight into rehabilitation outcomes. Full article

Background/Objectives: Institutionalized aging is associated with severe physical deconditioning, a high risk of falls, and a pervasive fear of falling. Physical exercise mitigates these factors, but the comparative efficacy of different training methodologies in this specific population remains unclear. The objective of this study was to compare the impact of a multicomponent exercise program versus a dual-task (cognitive-motor) training program on reducing fall risk, decreasing the fear of falling, and improving physical performance in institutionalized older adults. Methods: A randomized, parallel group controlled trial involving 21 older adults residing in a nursing home (Mean age = 83.67 ± 6.17 years). Participants were allocated to either a Multicomponent Group (n = 11) or a Dual-Task Group (n = 10) for a 12-week intervention (2 sessions/week). Fall risk, fear of falling, and global physical performance were assessed at baseline and post-intervention. Results: No significant improvements were observed in fall risk assessment execution time for either group. The Multicomponent Group showed a significant reduction in the fear of falling (−29.1%; 95% CI [−17.27, −1.27], p = 0.025) and a clinically significant improvement in physical performance (+40.9%; 95% CI [1.11, 3.43], p < 0.001), supported by large time effects (FES-I: F(1, 19) = 4.52, η²p = 0.192; SPPB: F(1, 19) = 13.68, η²p = 0.419). The Dual-Task Group achieved no significant changes in these dimensions. Furthermore, a marginally significant time-by-group interaction was observed for physical performance, favoring the multicomponent approach (F(1, 19) = 3.83, p = 0.065, η²p = 0.168 [large effect]). Conclusions: Multicomponent training proved superior in improving physical performance and reducing the fear of falling. In a frail, institutionalized population, the attentional cost demanded by dual-task training appears to limit the physical and psychological benefits of exercise. Full article

Background: Complications developing in individuals with Type 2 Diabetes Mellitus (T2DM) lead to functional impairments and losses in postural balance; however, changes in cognitive functions are also observed and are often overlooked. Dual-task exercises allow simultaneous engagement of balance and cognitive functions. Therefore, this study aimed to investigate the effects of dual-task exercise training on cognitive functions, balance, and functional status in individuals with T2DM. Methods: In this study, 40 individuals diagnosed with T2DM were randomly assigned to three groups: the dual-task exercise group (DTEG, n = 13), the single-task exercise group (STEG, n = 13), and the control group (CG, n = 14). Over eight weeks, balance exercises were administered to the STEG, while simultaneous balance and cognitive exercises were applied to the DTEG, twice weekly under the supervision of a physiotherapist. Participants in the control group received no intervention. Dual-task performance, cognitive functions, balance, and functional status were assessed at baseline and at the end of eight weeks. Dual-task performance was defined as the primary outcome. Results: After the intervention, for the primary outcome, dual-task performance (TUG single-task condition and TUG dual-task condition), both exercise groups showed greater improvements than controls. Both exercise groups also demonstrated significant improvements in balance, functional status, and cognitive outcomes compared to the control group. In the between-group comparisons, both exercise groups showed significant improvements in several cognitive functions compared with the control group (p < 0.05). In addition, the MoCA total score was significantly higher in the DTEG compared with the other groups. Conclusions: Both dual-task and single-task exercises improve cognitive function, balance, and functional status in individuals with T2DM. Full article

Underwater sonar object detection is challenging because targets are often small, boundaries are blurred, background clutter is strong, and labeled sonar data are limited. To address these issues, we propose T2C-DETR, a detector built on RT-DETR with three task-oriented improvements: (i) a Transformer–Convolution dual-channel backbone (TCDCNet) for complementary global-context and local-detail modeling, (ii) a Noise Filtering Module (NFM) inserted before neck fusion to suppress noise-dominated activations, and (iii) a stage-wise transfer-learning strategy tailored to small sonar datasets. We evaluate the method under three pre-training sources (COCO 2017, DOTA, and an infrared dataset) and then fine-tune on a self-built sonar dataset. Experimental results show that T2C-DETR achieves AP50 of 97.8%, 98.2%, and 98.5% at 72–73 FPS, consistently outperforming the RT-DETR baseline, YOLOv5-Imp, and MLFFNet in the accuracy–speed trade-off. These results indicate that combining global–local representation learning with targeted noise suppression is effective for practical real-time sonar detection. Full article

Background and purpose: Balance and gait problems pose a significant burden in Parkinson’s disease (PD), and they are often poorly treated with levodopa. We intended to summarize evidence of mid- and long-term impact of various physiotherapeutic interventions (≥3 months post-intervention) on dynamic balance, gait, and general motor function in patients with PD. Method: A systematic search was conducted across the PubMed, Cochrane Library, and Scopus databases to identify controlled clinical trials on sustained effects of various exercise interventions in PD on the outcomes of interest (lasting ≥ 3 months after completion of the exercise program). We conducted meta-analyses on commonly used clinical measures of dynamic balance and gait ability, as well as on UPDRS-III scores using the Comprehensive Meta-Analysis Software (CMA). Results: A total of 26 studies were included in meta-analyses, with a total of 1261 participants in the experimental and 989 participants in the control groups. Positive cumulative effects at the post-exercise follow-up (3 to 23 months) were shown in favor of the intervention group regarding balance (SMD = 0.512, 95% CI [0.240, 0.785], p < 0.001, I² = 87%), gait (SMD = 0.614, 95% CI [0.301, 0.926], p < 0.001, I² = 75%), and general motor function (SMD = 0.922, 95% CI [0.559, 1.285], p < 0.001, I² = 87%). Heterogeneity among studies was high for all three outcomes, apparently reflecting diversity with regard to patient characteristics, type, and duration of intervention, as well as the method of outcome assessment. The certainty of evidence was consequently judged as ‘’low’’ to ‘’moderate,’’ according to the GRADE system. Subgroup analyses revealed that balance can sustainably improve mostly through multimodal rather than targeted balance-oriented exercise but also through dual-task exercise, tai chi, and Pilates. Gait showed improvement at follow-up mainly through multimodal exercise, aerobic exercise, dual-task exercise, and Pilates, with benefits confined to early- and mid-stage disease. Sustained UPDRS-III improvement could be achieved through multimodal exercise, which showed a large overall effect but also through aerobic, resistance, and dual-task training, tai chi and qigong. Conclusions: Exercise interventions can improve balance and gait, as well as preserve the overall motor function in patients with PD, also in the mid- and long-term post-intervention periods. Full article

Background: Occult lymph node metastasis (LNM) occurs in 30–80% of patients with clinically node-negative papillary thyroid carcinoma (cN0-PTC), partly owing to the limited sensitivity of current preoperative nodal assessment, and may contribute to postoperative recurrence. Conventional sentinel lymph node (SLN) biopsy, typically performed with a single tracer, has limited reliability for detecting occult metastatic nodes, which can result in either overtreatment or undertreatment with lymph node dissection. We aimed to develop a highly accurate multimodal prediction framework to accurately identify second-echelon lymph node metastasis (SeLNM) and non-sentinel lymph node metastasis (NsLNM). Methods: We prospectively enrolled 301 patients with cN0-PTC between April and October 2024, of whom 131 met the inclusion criteria. Intraoperatively, a dual-tracer technique combining carbon nanoparticles and indocyanine green was applied, and near-infrared imaging was used to record the entire SLN visualization process in real time. For each case, a 3 min video clip (150 frames) was captured. Two senior surgeons delineated regions of interest to generate 19,650 mask images. A total of 2048 spatial features and 20 temporal features were extracted, combined with 32 clinical variables, including demographics, ultrasound characteristics, and gene mutation status. Nine deep learning models were developed and evaluated using 10-fold cross-validation. Model performance was quantified using receiver operating characteristic curves, decision curve analysis curves, calibration curves, precision–recall curves, learning curves, and 12 metrics. Statistical comparisons were performed using the DeLong test, and models were further evaluated using a probability-based ranking approach. Shapley Additive Explanations (SHAP) analysis was applied to interpret key predictive features. The primary outcomes were SeLNM and NsLNM, defined based on postoperative histopathology. Results: The Long Short-Term Memory (LSTM) + Transformer model showed the best performance for both prediction tasks, with stable AUCs across training and testing (SeLNM: 0.980/0.982; NsLNM: 0.986/0.983). In the testing set, the model reached the same accuracy for both outcomes (94.7%) and showed strong sensitivity/specificity for SeLNM (94.7%/94.6%) and NsLNM (96.4%/91.5%). SHAP analysis indicated that time-series fluorescence flow features were the most influential predictors, followed by spatial structural features and SLN status. Conclusions: Dual-tracer SLN mapping with deep learning demonstrated encouraging intraoperative prediction of lymph node metastasis with interpretable features in this single-center cohort. Independent multicenter validation and prospective outcome studies are needed before considering clinical adoption. Full article

This study addresses the challenge of the difficult identification of organ structures in the standard sections of fetal abdominal ultrasounds. A deep learning-based multi-task model named SEG-FAUSP was developed to segment the core anatomical structures of seven key fetal abdominal ultrasound sections. We collected fetal abdominal ultrasound images from pregnant women in the mid-pregnancy period (18–24 weeks) using various mainstream ultrasound devices, and professional physicians annotated key anatomical structures (e.g., umbilical veins, gastric bubbles, spine) in the images. Based on an improved deep learning framework, the model accurately segments and locates the target organ structures through a parallel dual-branch semantic segmentation network, which avoids the over-reliance on large-scale pre-trained data in traditional methods. Experimental results show that the model achieves excellent performance in anatomical structure segmentation, with the intersection over union of the bladder and gastric bubble both reaching above 0.84; its segmentation accuracy for complex structures such as the inferior vena cava is also significantly superior to the baseline model. As an end-to-end model, it simplifies the clinical interpretation process of fetal abdominal ultrasound, reduces the risk of missed diagnoses caused by unclear organ identification, provides an efficient auxiliary tool for prenatal screening in grassroots medical institutions, and is of great significance for improving the quality of newborns. Full article

Metro train timetable rescheduling (TTR) is a critical task for ensuring the reliability of urban rail transit systems. However, with the increasing density of railway networks and the growing number of operational trains, TTR has evolved into a typical high-dimensional and large-scale optimization problem. Traditional mathematical programming and heuristic approaches often struggle with the “curse of dimensionality” and fail to provide real-time responses under stochastic disturbances. To address these challenges, this paper proposes a novel framework based on Multi-Agent Deep Reinforcement Learning (MADRL). Specifically, we model the TTR problem as a decentralized cooperative process and utilize the Multi-Agent Advantage Actor-Critic (MAA2C) algorithm to optimize train schedules dynamically. The proposed framework is implemented within the Flatland simulation environment, which allows for the representation of complex arbitrary topologies. We design a composite reward function that minimizes total delay deviation while maximizing passenger satisfaction, subject to constraints such as headway, operating time, and train capacity. Furthermore, to enhance the robustness of the model against high-dimensional state uncertainties, random disturbances following a negative exponential distribution are introduced during training. Experimental results across various scenarios—ranging from simple dual-track to complex random networks—demonstrate that the MAA2C-based approach significantly outperforms traditional baselines. It not only achieves faster convergence in small-scale scenarios but also demonstrates superior computational efficiency and scalability in large-scale environments, effectively minimizing passenger waiting times. This study validates the potential of MADRL in solving high-dimensional traffic control problems for intelligent transportation systems. Full article

Extracting road networks from Synthetic Aperture Radar (SAR) data represents a core challenge in remote sensing scene analysis, particularly for applications in traffic monitoring and emergency management. The task is complicated by several inherent limitations: speckle noise degrades image quality, geometric distortions arise from the side-looking acquisition geometry, and roads often exhibit weak radiometric separation from surrounding terrain. Traditional processing pipelines and recent single-branch deep learning frameworks have shown insufficient performance when global contextual reasoning and fine-scale spatial detail must both be addressed. This work presents DINOv3-PEFT, a parameter-efficient dual-encoder network designed specifically for SAR road segmentation. The architecture employs two complementary processing streams tailored to SAR characteristics: one stream utilizes adapter-based fine-tuning applied to pre-trained DINOv3 weights (kept frozen), which captures long-distance spatial relationships crucial for maintaining network connectivity despite speckle corruption. The second stream, based on convolutional operations, focuses on extracting localized geometric features that preserve the narrow, elongated structure and sharp boundaries typical of road infrastructure. Feature fusion occurs through the Topological-Geometric Feature Integration (TGFI) Module, which synthesizes multi-scale representations hierarchically. This mechanism proves effective at bridging fragmented road segments and recovering geometric accuracy in scenarios with heavy shadow casting or signal interference. Performance evaluation on the GF-3 satellite dataset across four spatial resolutions (1 m, 3 m, 5 m, and 10 m) demonstrates the proposed method achieves an 82.61% F1-score, a 76.51% IoU, and a 98.08% overall accuracy, all averaged across the four resolutions. When benchmarked against six state-of-the-art methods, DINOv3-PEFT demonstrates substantial improvements in road class segmentation quality and topological connectivity preservation, supporting its robustness for operational SAR road mapping tasks. Full article

Indoor pedestrian navigation tasks, as a key part of smart cities and navigation services, face dual challenges of accuracy and cost under complex building environments. Currently, neural inertial navigation is at the vanguard of current research in indoor pedestrian navigation, and existing related studies have achieved positive results. However, the exploration of deep learning solutions is still not sufficient, mainly reflected in the lack of explorations of model training configurations. Based on testing results under different deep learning schemes, this paper proposes EDIN, an enhanced deep inertial navigation approach. This method benefits from a proprietary neural network based on ResNeXt with Convolutional Block Attention Module (CBAM) to predict the relationship between inertial data and motion trajectory. Compared to existing projects, this paper also makes improvements in the model training process, thereby improving the predictive effect of the trained model. Specifically, this paper innovatively uses Logcosh as the loss function and combines data rotation and additional noise as data augment methods. To assess EDIN’s performance, extensive tests were conducted using three publicly available datasets: RoNIN, OXIOD, and RIDI. The results clearly indicate EDIN’s superior performance relative to other neural inertial navigation systems. Notably, localization accuracy improved significantly, with an average enhancement of 16.06% compared to the RoNIN-ResNet method. Full article