Search Results (2,017)

Semantic segmentation of land use and land cover (LULC) in arid regions remains challenging due to severe class imbalance, fragmented spatial distributions, and high spectral similarity among different land cover types. These characteristics often lead to an information bottleneck in deep segmentation networks and hinder the extraction of discriminative semantic representations. To address these issues, we propose SDS-Former, a lightweight semantic segmentation network specifically designed for remote sensing imagery in arid environments. SDS-Former incorporates an SSM-inspired Lightweight Semantic Enhancement (LSE) module to strengthen contextual modeling and alleviate the loss of discriminative information in deep features. To tackle scale variations, a Dynamic Selective Feature Fusion (DSFF) module is employed in the decoder to adaptively weight and fuse high-level semantics with low-level spatial details. Furthermore, a Feature Refinement Head (FRH) is introduced to enhance boundary localization and improve the recognition of small-scale and sparsely distributed land cover objects. Extensive ablation and comparative experiments demonstrate that SDS-Former consistently outperforms representative semantic segmentation methods across multiple evaluation metrics. On the Tarim Basin dataset, the proposed network achieves a mean Intersection over Union (mIoU) of 82.51% and an F1 score of 86.47%, indicating its superior effectiveness and robustness. Qualitative results further verify that SDS-Former exhibits clear advantages in distinguishing spectrally similar land cover types and preserving the spatial continuity of ground objects in complex arid-region scenes. Full article

Automated detection of defects in welds are inevitable in the assurance of structural integrity, but this faces serious challenges due to the microscopic characteristics of the discontinuities, low visual contrast and infrequent occurrence of defect samples. Conventional deep learning methods, while accurate, often lack interpretability and exhibit low recall for rare defects. This paper proposes a novel hybrid system combining a Generative Adversarial Network (GAN), a Convolutional Neural Network (CNN), and Extreme Gradient Boosting (XGBoost 2.0.0) to enhance weld defect classification performance and transparency. Firstly, a Deep Convolutional GAN (DCGAN) creates synthetic images of the minority classes; thus, the problem of class imbalance is resolved. Then, a pretrained ResNet50V2 CNN is used to extract features of the deep layers from the original images as well as from the generated ones. After that, these features are fed into an XGBoost classifier, which uses tree-based learning to optimize classification results and make the process more understandable to the user. Furthermore, interpretation is also facilitated by Grad-CAM rendering of the CNN regions of interest and SHAP analysis to measure the involvement of the features in XGBoost. Experiments using the available LoHi-WELD datasets show that the overall accuracy is significantly improved, the per-class recall of the rare defects is also enhanced, and the robustness is also improved. The proposed hybrid method not only achieves better results but also generates visual/explainable output, which is very valuable when the system is implemented in industrial welding inspection systems. This paper serves as a liaison between the latest AI technology and the practical interpretability requirements of the mechanical and welding engineering fields. Full article

Accurate soil type classification is fundamental to geotechnical engineering, yet traditional laboratory methods are often time consuming and labor intensive. This study investigates the potential of a Transformer-based deep learning framework for the automated classification of complex soil compositions. An image database for geotechnical analysis is constructed using six distinct geotechnical samples comprising gravel, sand, silt, and clay systematically blended into 80 ternary mixtures. To address the inherent class imbalances in the multi-component dataset, the Synthetic Minority Oversampling Technique (SMOTE) is employed, ensuring robust representation across all categories. The proposed framework utilizes a Vision Transformer (ViT) architecture, leveraging its self-attention mechanism to capture both intricate textural patterns and long-range structural dependencies within the soil matrices. Experimental results demonstrate that the SMOTE–ViT pipeline achieved an overall accuracy of 95.83%, with high precision and recall across diverse ternary compositions. This interdisciplinary approach provides a scalable and high-precision alternative for soil characterization, offering significant potential for real-time decision-making in geotechnical investigation workflows. Full article

Durian (Durio zibethinus Murray) is Thailand’s most economically significant fruit export, yet foliar diseases pose a major threat to productivity and crop quality. Early-stage symptoms of several durian leaf diseases are visually similar, making reliable diagnosis difficult for farmers and even trained agronomists. This study aims to develop and evaluate an automated deep learning-based system for durian leaf disease classification under realistic field conditions. A dataset of 6119 leaf images representing six classes—Leaf_Healthy, Leaf_Colletotrichum, Leaf_Algal, Leaf_Phomopsis, Leaf_Blight, and Leaf_Rhizoctonia—was compiled from public datasets and field-collected samples. Six convolutional neural network (CNN) architectures—ConvNeXt, ResNet, DenseNet201, InceptionV3, EfficientNet-B3, and MobileNetV3—were benchmarked using a unified transfer-learning training protocol. Class imbalance was addressed using weighted cross-entropy loss, and performance was evaluated on a stratified held-out test set using accuracy, precision, recall, and F1-score metrics. The results show that ConvNeXt achieved the highest performance with 98.00% accuracy and a weighted F1-score of 0.98, followed by ResNet (96.82%) and DenseNet201 (96.09%), while efficiency-oriented models plateaued near 91%. Confusion matrix analysis revealed consistent misclassification among visually similar disease categories—Leaf_Algal, Leaf_Blight, and Leaf_Phomopsis—indicating biological similarity in lesion appearance rather than model limitations. The best-performing model was deployed as a publicly accessible web application using Gradio, enabling real-time disease diagnosis with an average inference time of approximately 0.54 s per image. Unlike prior studies, this work combines large-scale architecture benchmarking, class imbalance mitigation, and real-world deployment within a single unified framework. These findings demonstrate that modern CNN architectures can provide highly accurate and scalable disease detection tools, supporting precision agriculture by enabling early diagnosis, reducing inappropriate pesticide use, and improving decision-making for durian farmers. Full article

In intensive care units, large-scale clinical time-series data are continuously accumulated through electronic medical records and bedside monitoring systems. However, direct utilization of such data for clinical decision-making remains challenging due to irregular sampling, pervasive missingness, unstructured diagnostic information, and incomplete ICD labeling. Automated ICD coding constitutes an extreme multi-class classification problem with thousands of long-tailed categories, while intervention prediction tasks, such as mechanical ventilation management, involve rare transition events and severe class imbalance. To address these challenges, we propose CAGE, a hierarchical Clinical Decision Support System framework that integrates diagnosis, time-series signals, and intervention prediction. The framework first infers admission-level diagnostic context using a partial-label Automated ICD Coding module that combines DCNv2 with an Adaptive CLPL loss, producing probability-weighted diagnostic embeddings. These embeddings are subsequently fused with ICU time-series tensors and processed by a multi-branch Temporal Convolutional Network equipped with an ICD-conditioned gating mechanism to predict future ventilation state transitions. The experimental results demonstrate that DCNv2 achieves consistent superiority across all hit@k and probability concentration metrics for ICD coding. For intervention prediction, the proposed method substantially outperforms existing baselines, achieving a Macro-AUC of 98.2, Macro-AUPRC of 77.4, and F1-score of 79.4. These findings indicate that reinjecting diagnostic context as a conditioning variable, together with imbalance-aware loss design, effectively enhances rare-event detection and improves the practical applicability of clinical decision support systems. Full article

Ramadan fasting substantially alters meal timing, sleep patterns, and daily activity, thereby increasing the risk of hypoglycaemia in adults with type 1 diabetes (T1D). Although continuous glucose monitoring (CGM) systems provide real-time alerts, these are largely reactive or limited to short prediction horizons, offering insufficient warning under fasting-related behavioural and circadian disruption. This study aims to evaluate whether behaviour-aware, temporally enriched recurrent deep learning models, leveraging multimodal CGM and wearable-derived signals, can forecast hypoglycaemia one hour ahead during Ramadan and the post-fasting period. In an observational, free-living cohort study conducted in Qatar, 33 adults with T1D were monitored using CGM and a wrist-worn wearable during Ramadan 2023 and the subsequent month. Multimodal data were aggregated into hourly features and organised into rolling 36 h sequences. In addition to physiological signals, explicit temporal and circadian proxy features were engineered, including cyclic time encodings, day–night indicators, and Ramadan-specific behavioural windows (e.g., pre-iftar, iftar, post-iftar, and fasting phases). Recurrent models, including LSTM and BiLSTM architectures, were trained using patient-wise, leak-free splits, with focal loss applied to address class imbalance. Model performance was evaluated on a held-out, naturally imbalanced test set using ROC AUC, precision–recall AUC, recall, and probability calibration, alongside cross-phase evaluation between Ramadan and post-fasting periods. Following quality control, 1164 participant-days were retained, with hypoglycaemia accounting for approximately 4% of hourly observations. Temporal feature enrichment and the use of a 36 h lookback window improved both discrimination and calibration, with performance stabilizing beyond this horizon. On the imbalanced test set, the best-performing multimodal model achieved an ROC AUC of 0.867 and a precision–recall AUC of 0.341, identifying 77% of next-hour hypoglycaemic events at a sensitivity-focused operating point (precision = 0.14). The selected BiLSTM model demonstrated good probability calibration (Brier score ≈ 0.03). Models trained using wearable-derived inputs alone achieved comparable discrimination and, in some configurations, higher precision–recall AUC than CGM-only baselines. Notably, models trained on the original imbalanced data outperformed resampled variants, suggesting that temporal and behavioural features provided sufficient discriminatory signal without requiring aggressive class balancing. Cross-phase evaluation indicated robust generalisation, particularly for the BiLSTM model. Overall, behaviour-aware, temporally enriched multimodal models can provide calibrated, hour-ahead hypoglycaemia risk estimates during Ramadan fasting in adults with T1D, enabling proactive intervention beyond reactive CGM alerts. Explicit modelling of circadian and behavioural dynamics enhances predictive performance under real-world class imbalance. Furthermore, integrating wearable-derived behavioural and physiological signals adds predictive value beyond CGM alone, supporting robustness across varying levels of contextual data availability. External validation and prospective clinical evaluation are required prior to deployment. Full article

Hyperspectral unmixing aims to estimate endmember spectra and their corresponding abundance fractions at the subpixel scale, which is a critical preprocessing step for quantitative analysis of hyperspectral remote sensing imagery. While deep learning-based methods have achieved remarkable progress, three fundamental challenges remain: (i) reliance on a single shared spatial prior that cannot decouple the heterogeneous spatial patterns of different land covers; (ii) the lack of synergy in jointly optimizing endmember extraction and abundance estimation; (iii) the poor robustness of unsupervised training to complex mixtures, noise, and class imbalance. To address these issues, we propose a novel unsupervised unmixing framework that integrates adaptive orthogonal multi-faceted graph representation with curriculum learning. Specifically, we design an Adaptive Orthogonal Multi-Faceted Graph Generator (AOMFG) to learn a set of independent orthogonal graph structures, achieving spatially informed decoupling of land cover patterns. Then, a dual-branch collaborative optimization network is constructed: a Graph Convolutional Network (GCN) branch that incorporates the learned spatial topological priors for abundance estimation, and a 1D Convolutional Neural Network (1DCNN) branch that employs a query-attention mechanism to adaptively aggregate pure spectral features for endmember extraction. Finally, we introduce a three-stage curriculum learning strategy that progressively fine-tunes the model, which significantly enhances its performance. Extensive experiments on three widely used real-world benchmark datasets demonstrate that our proposed framework consistently outperforms state-of-the-art methods in both endmember extraction and abundance estimation accuracy. Comprehensive ablation studies, parameter sensitivity analysis, and noise robustness tests further validate the effectiveness of each core component. Full article

Customer churn refers to the termination of a customer’s business relationship with a bank, representing a direct loss of future revenue. Product downsell manifests as a reduction in the number of financial products held or a downgrade in service tier, often signaling early customer disengagement. Accurately identifying customers at risk of these two behaviors has become a cornerstone of profitable growth in the competitive retail banking industry as downsell frequently serves as a precursor to total churn. However, the existing research typically treats these highly correlated behaviors as independent prediction tasks, overlooking their intrinsic link and failing to address the critical challenges of class imbalance and regulatory demands for model interpretability. To tackle these problems, we propose an enhanced multi-task learning network (EMTL-Net), a deep learning framework specifically designed to capture the nuanced interplay between churn and downsell behaviors. EMTL-Net introduces an explicit feature interaction module to enhance the modeling of high-order feature relationships and utilizes a shared representation layer to extract universal customer risk patterns, enabling the joint prediction of churn and downsell. Furthermore, we employ Focal Loss as the training objective to dynamically adjust sample weights, effectively mitigating the class imbalance problem. Critically, to meet financial compliance requirements, we implement a SHAP-based interpretation mechanism that is compatible with multi-task outputs, providing preliminary insights into feature importance. Formal validation of interpretability claims remains an important direction for future research. The experimental results on a publicly available pedagogical bank customer benchmark dataset demonstrate that EMTL-Net achieves excellent performance on both tasks. For churn prediction, the model achieves an AUC of 0.8259, an accuracy of 0.8361, and an F1-score of 0.6235, significantly outperforming the existing baseline models. For downsell prediction (noting that the downsell label is rule-derived from the number of products held), the model achieves an AUC of 0.8932, an accuracy of 0.8571, and an F1-score of 0.7504. Ablation studies confirm the critical contributions of the explicit feature interaction module, Focal Loss, and the residual structure to model performance. Crucially, the interpretability analysis corroborates business intuition by identifying customer age, account balance, and product holdings as dominant churn drivers—a consistency that reinforces the model’s credibility and practical utility in high-stakes financial environments. Full article

Accurately estimating the severity of crash injuries among older drivers is paramount for enhancing traffic safety, a task challenged by class imbalance and label noise. Traditional predictive paradigms often struggle to identify rare severe cases, as they tend to prioritize global accuracy, thereby compromising sensitivity to high-risk outcomes. To overcome these limitations, this study develops a Log-Loss Cleaned and Probability-Calibrated Cascade (L-CSC) framework by strategically integrating existing advanced algorithmic components for robust and reliable severity prediction. Initially, a Log-Loss-based noise filtering mechanism is implemented to purge outliers and ambiguous samples from the training data, thereby enabling higher-quality representation learning. Subsequently, a two-stage cascade architecture is designed to decouple the classification task. Stage I employs a Preliminary Screening Model, optimized via Bayesian optimization for F2-score, to specifically maximize the recall for severe and fatal cases. In Stage II, a Stacking ensemble classifier is deployed to achieve a fine-grained classification of injury levels among the cases identified in the initial screening. Finally, Isotonic Regression is employed to calibrate the output probabilities from both stages, ensuring that the resulting risk estimations are statistically sound and reliable. Empirical evaluations demonstrate that the L-CSC framework effectively balances overall performance with critical risk detection, achieving a robust Macro-F1 of 0.7296. Specifically, compared to the best-performing baseline, the recall and F1-score for the critical severe and fatal category showed relative improvements of over 82% and 62%, respectively. Ablation analyses further substantiate the vital contributions of both the data cleaning and calibration modules. This research demonstrates that the cascaded framework effectively mitigates the biases inherent in imbalanced datasets, providing a robust algorithmic foundation to potentially support future traffic safety interventions. Full article

Background/Objectives: The manual analysis of sperm morphology, crucial for male infertility diagnosis, is subjective and time-consuming. Automated methods using deep learning, offer a promising alternative; however, standard deep models are prone to overfitting when applied to small, heavily unbalanced clinical datasets, limiting their generalization capability. This study proposes a knowledge distillation approach that functions as a strong regularizer, improving the robustness of automated sperm morphology analysis. Methods: We utilize soft distillation to transfer knowledge from a set of high-capacity teacher models to a smaller student model (SwinV2-base). The teacher architectures include SwinV2-large, EfficientNetV2-m, and ConvNeXtV2-large. To maximize performance, we investigated two distillation strategies: a single-teacher approach, where the student learns from one specific architecture, and a multi-teacher approach, where the student learns from an averaged response of multiple teachers. The models were trained on the imbalanced Hi-LabSpermMorpho dataset, which comprises 18 different sperm morphology categories derived from three differently stained (BesLab, Histoplus, GBL) sample sets. We adopted a cross-dataset training approach in which the teacher models were fine-tuned using the combination of two stained datasets, and the student model was trained on the third, distinct stained dataset. The global loss function combined cross-entropy loss with Kullback–Leibler divergence, employing the teacher’s soft probabilities to prevent the student from over-confidence. Results: The experimental results demonstrate that the student model trained in a multi-teacher setup with augmentation and soft distillation attains higher accuracies (70.94% on BesLab, 73.61% on Histoplus, 71.63% on GBL) than the baseline models. Conclusions: This approach mitigates challenges associated with data scarcity and heavily unbalanced sperm morphology datasets, providing consistent improvements and offering a highly generalizable solution for clinical diagnostics. Full article

Background: Accurate mortality prediction remains a major challenge in public health due to the complex interactions among demographic, socioeconomic, behavioral, and medical factors. This problem is particularly relevant for identifying high-risk groups and improving preventive healthcare strategies. While existing studies demonstrate strong predictive performance, they mainly rely on clinically structured data and focus on model performance. Challenges such as misclassification and atypical cases remain less explored. Methods: Using the Integrated Public Use Microdata Series National Health Interview Survey (IPUMS-NHIS) 2010 and 2015 datasets (193,765 records, 104 features), this study investigates mortality prediction through comparative Machine Learning. Data preprocessing included feature engineering, categorical encoding, and removal of missing entries. Class imbalance was addressed using SMOTE and SMOTE-ENN resampling, followed by hyperparameter tuning. Three models—Logistic Regression, Random Forest, and XGBoost—were trained to classify mortality, with recall prioritized to ensure accurate identification of deceased cases. Results: Results showed that XGBoost achieved the best performance (Recall = 69%, F1 = 0.39, AUC = 0.92), outperforming other models in balancing sensitivity and specificity. Feature importance and permutation analyses highlighted age, employment status, self-reported health, and lifestyle indicators as key predictors. Misclassification analysis combined with Isolation Forest revealed atypical profiles not captured by standard models. Conclusions: The findings underscore XGBoost’s effectiveness and demonstrate the value of integrating anomaly detection with classification to improve mortality prediction and inform public health planning. Full article

As the cornerstone of contemporary electronics, the quality of electronic substrates—including Printed Circuit Boards (PCBs) and Ceramic Package Substrates (CPSs)—is intrinsic to product reliability. However, automated inspection is currently impeded by two persistent obstacles: the drastic multi-scale variation in defects and the acute class imbalance within defect datasets. Conventional deep learning approaches often fail to reconcile these challenges simultaneously, leading to suboptimal recognition of rare defect categories. To bridge this gap, we propose Multi-scale Partial Vision Transformer—Simple, Parameter-free Attention Module (MulPViT-SimAM), a robust framework designed for class-imbalanced electronic substrate defect detection. Our method features a novel multi-scale backbone (MulPViT) that synergizes partial convolutions with hierarchical attention mechanisms, facilitating the efficient extraction of both fine-grained local textures and global contextual dependencies. Additionally, we embed the Simple, Parameter-free Attention Module (SimAM) into the feature fusion stage to adaptively highlight defect-specific features while dampening background noise. To further mitigate data imbalance, we utilize the Equalized Focal Loss (EFL) function, which employs a category-specific modulating factor to dynamically equilibrate the learning focus across different classes. Comprehensive benchmarking reveals state-of-the-art performance, achieving mAP@0.5 scores of 95.7% on the standard PKU-MARKET-PCB dataset and 54.2% on the highly challenging CPS2D-AD dataset. Significantly, our approach effectively mitigates class imbalance, narrowing the performance deviation of rare categories to just 4.3% on the PKU-Market-PCB dataset and 1.4% on the CPS2D-AD dataset, compared to 11.8% and 7.5% in baseline models. These findings position MulPViT-SimAM as a viable and efficient solution for industrial quality control. Full article

To address the challenges of scarce unmanned aerial vehicle (UAV) track samples, severe class imbalance, and high motion similarity between UAVs and birds in low-altitude radar recognition, this paper proposes a trajectory classification method integrating a TCN-Transformer model with a physics-constrained TimeGAN (PC-TimeGAN) data augmentation framework. Specifically, the PC-TimeGAN generates high-quality, kinematically compliant UAV trajectories to alleviate data scarcity and class imbalance. A multi-scale TCN-Transformer is then constructed to comprehensively extract features, utilizing multi-kernel dilated convolutions for local temporal correlations and self-attention mechanisms for global temporal dependencies, thereby improving the discrimination between UAV and bird trajectories with similar motion patterns. Furthermore, a joint loss function combining Focal Loss and Triplet Loss is employed to optimize the decision boundaries and feature space, enhancing model robustness and generalization. Experiments on a measured dataset demonstrate that, under the 15-dimensional input setting, the proposed method achieves a UAV recall of 80.00%, an FAR of 3.15%, a precision of 64.00%, and an F1-score of 0.7111. Compared to baseline methods (e.g., SVM, LSTM, GRU, Transformer, and 1D-CNN), the proposed approach significantly improves UAV recall under limited trajectory information while keeping the false-alarm rate of misclassifying birds as UAVs low. Ultimately, this method markedly enhances the comprehensive performance of rapid track-level target classification for low-altitude surveillance radars. Full article

This study presents a representation-centric evaluation of audio foundation models for fine-grained musical instrument analysis, focusing on cymbal classification. A confound-aware comparison of CLAP and MERT embeddings is conducted to examine how each latent space supports recoverability of acoustically and semantically relevant information. To support this analysis, the study introduces a representation-centric, confound-aware multi-stage evaluation framework that separates exploratory geometry, leakage-safe probing, and supporting unsupervised clustering evidence. The methodology is applied to a challenging cymbal dataset characterized by hierarchical labels, class imbalance, and subtle acoustic variation. Results reveal a target-dependent profile of representational strengths rather than a single overall winner. CLAP exhibits stronger variance concentration and more label-consistent local neighborhood organization, and it outperforms MERT on fine-grained, strike-related targets. MERT, however, retains a small but consistent advantage on higher-level cymbal-type classification. Unsupervised analyses show that these advantages reflect local neighborhood structure, not strong global cluster formation, and confound diagnostics indicate that size-related information remains largely type-mediated. Overall, the findings underscore the importance of structured, multi-stage evaluation for disentangling embedding geometry, recoverability, and confound effects while demonstrating the complementary strengths of AFMs in complex audio classification settings. Full article

Background/Objectives: Postoperative pulmonary complications (PPCs) following abdominal surgery are associated with prolonged hospitalization, delayed recovery, and increased mortality. Because nursing surveillance is essential for early detection and timely intervention, this study aimed to develop a holistic nursing surveillance decision support system integrating PPC risk prediction with structured nursing action recommendations. Methods: In this retrospective cohort study, electronic medical record (EMR) data from approximately 6900 adult patients who underwent abdominal surgery at a single institution between January 2015 and September 2023 were analyzed. The study protocol was approved by the Institutional Review Board, and the requirement for informed consent was waived because of the retrospective study design. PPC risk was predicted using a tabular multilayer perceptron (MLP) encoder with SHapley Additive exPlanations (SHAP)-based feature weighting and a random forest classification head optimized via Optuna. Class imbalance was addressed using weighted sampling, class weighting in BCE(Binary Cross Entropy) With Logits Loss, and decision-threshold optimization. For clinical decision support, a large language model generated structured nursing surveillance recommendations in an action–evidence–rationale JSON format and was aligned through supervised fine-tuning (SFT) using human-evaluated cases. Results: The prediction model achieved an AUROC of 0.810, with an accuracy of 0.811, precision of 0.547, and recall of 0.545. In expert evaluation, the SFT-aligned model improved recommendation quality, reducing incorrect nursing actions from 19.3% to 8.0%. Conclusions: The proposed system demonstrates the feasibility of an end-to-end nursing surveillance decision support framework linking PPC risk prediction with structured clinical recommendations. The findings suggest its potential to support more accurate risk prediction and more actionable nursing surveillance for patients undergoing abdominal surgery. Full article