Search Results (12,849)

Mapping alteration minerals from spaceborne hyperspectral imagery in drylands is hindered by mixed pixels, spectrally similar mafic minerals, and false positives from vegetation, soil and iron staining. We present a knowledge-driven workflow combining physical spectral evidence, geological specificity constraints and field-trained random forest statistics. A bedrock and signal-quality mask restricts retrieval to exposed rock; continuum-removed absorption depth, adaptive coherence estimation and endmember orthogonalization provide sub-pixel evidence, while an epidote slope gate and a VNIR copper-carbonate > iron-oxide rule reduce the iron-stain–copper-carbonate overlap from 52.6% to 9.8%. A one-vs-rest random forest trained on 248 field points rescales the evidence, and a union budget confines fifteen channels to ~9% of the bedrock-exposed area. Because the forest is fitted on all field points, the map-level recall of 49.5% is a consistency measure; a leave-one-spatial-band-out cross-validation gives 58.9% (95% CI 48.9–68.3%), against 47.4% for the physical evidence alone, and among the 95 altered points the mineral-specific evidence ranks the recorded mineral above the others with AUC 0.71–0.89. The field labels are macroscopic hand-lens observations rather than laboratory-confirmed mineralogy, and the forest scores are not calibrated probabilities (Brier 0.39, ECE 0.45). In East Junggar the phyllic–argillic assemblage follows the Halasu intermediate–felsic bodies and serpentine–talc along the Kalatongke mafic–ultramafic trend, which separates the two mineral systems; the product is intended for regional screening. Full article

Federated learning (FL) trains shared models without centralizing raw data, but provides limited post-incident accountability or evidence preservation. We present FL-Forensics, a Flower extension for malicious-client and compromised-aggregator settings. Client-signed commitments are witnessed before server acceptance and later anchored, in a simulated four-validator quorum, for independent verification. We define a Forensic Quality Index (FQI) that combines evidence quality, attribution quality, and detection $F_1$. In experiments with 20 clients, Dirichlet non-IID data, and 15% malicious clients, $F_1$ on MNIST, Fashion-MNIST, and CIFAR-100 is 0.950, 0.873, and 0.660, with corresponding FQI values of 0.964, 0.929, and 0.815 (the MNIST and Fashion-MNIST values are means over two runs; the CIFAR-100 values are from one run). Detection weakens as the federation grows to 60 clients and against adaptive or colluding attackers, and Trimmed Mean, Krum, or FLTrust matches or exceeds it in some settings, so the contribution is the evidence layer rather than superior detection. Across six repeated tampering runs of 132 trials each, all eight manipulation types were detected and recovered; in the reported CIFAR-100 run, responsible-party attribution succeeded in 87/132 trials across five of the eight types. A counterfactual ablation on captured evidence illustrates the role of the independent witness; it is not an end-to-end ablation. Under secure aggregation, commitments remain verifiable but in-round screening is largely unavailable; post-incident content attribution requires auditor escrow, which exposes individual updates to an authorized auditor. Secure aggregation is simulated, and end-to-end overhead against a matched baseline is not reported. Full article

Virtual learning environments log time-stamped learner–platform events, from which a soft sensor can estimate course outcomes early. A leakage-audited protocol synchronizes all channels at 20%, 40%, or 60% landmarks, excludes already-withdrawn registrations, and uses student-disjoint splits. We tested an Enhanced Heterogeneous Graph Transformer (EHGT) soft sensor that fuses four student–student relations with temporal encoding and adaptive relation attention. On the Open University Learning Analytics Dataset, we compared EHGT with ten baselines in three modules and retested exploratory patterns with new seeds. Across all seven modules at the 40% landmark and under two training budgets, EHGT had a higher equally weighted mean macro-F1 than a graph convolutional network (GCN) and a heterogeneous graph attention network, but its heterogeneous graph transformer backbone had a higher mean. EHGT had a lower mean than logistic regression and an adapted temporal graph network in every module and never outperformed a multilayer perceptron when retested. Its first layer has no skip connection for a student’s own features, so its relations replace them. In a relational GCN that keeps these features, the relations raised mean macro-F1 only under the shorter budget. Multi-source graph fusion must therefore be tested against strong tabular and event-stream baselines under adequate training. Full article

Detecting intrusions in database systems is challenging due to the use of legitimate credentials and the scarcity of labeled attack data. This paper presents a specification-based intrusion detection system (IDS) for role-based access control (RBAC) database environments that automatically derives behavioral policies from normal transaction traces and enforces them at runtime. The proposed framework employs a specification check followed by a novel-bigram gate, i.e., a deterministic specification layer that checks role-specific read and write permissions and read-before-write dependencies. An optional novel-bigram gate then flags adjacent operation pairs that never occur in normal training for that role. Isolation Forest was evaluated as a further residual scorer and is not part of the recommended detector. On a synthetic RBAC holdout, the specification layer alone achieves 98.4% accuracy, 100% precision, and 86.0% recall, with zero false positives. The novel-bigram gate raises accuracy to 98.8% and recall to 89.6% while keeping precision at 100%. Isolation Forest detects no further attack and adds eight false positives. These results show that the learned rules and the bigram gate carry the detection and that the tested one-class model does not justify its cost on this corpus. Each detection is accompanied by a human-readable reason, which is useful for audit. The study does not claim efficacy against adaptive adversaries or production DBMS workloads. Full article

Background/Objectives: Deep networks trained to recognise oral squamous cell carcinoma (OSCC) in histopathology perform well on held-out images from the same collections, but a recent study found that such a network lost most of its discriminative ability on lymph node tissue. We examined whether encoder fine-tuning or label-free adaptation changes this, and how many labelled target patches are needed to recover performance. Methods: The source domain comprised 23,290 images of oral and oropharyngeal mucosa from three public collections. The target domain was PatchCamelyon, a public dataset of sentinel lymph node patches from breast cancer patients; its metastases are breast carcinoma, so the transfer crosses both tissue type and cancer histology. We used 40,000 target patches from 216 slides, split into slide-disjoint adaptation and evaluation pools. We compared nine encoders in ten configurations, including three pathology foundation models (Phikon, Phikon-v2 and the recent Midnight model). We evaluated direct transfer with frozen and fine-tuned encoders, label-free adaptation, and few-shot adaptation, in which we compute class prototypes from k labelled target patches using a frozen encoder. All results are at patch level. Results: Direct transfer failed for every configuration (external AUROC 0.35–0.64), including fine-tuned networks (mean over three seeds 0.44–0.52) and fine-tuned Phikon (0.54–0.58). Label-free methods did not close the gap; a zero-shot vision–language model reached 0.77. With a frozen Phikon encoder, prototypes from 10 and 50 labelled patches per class reached 0.869 ± 0.056 and 0.930 ± 0.015, against 0.82–0.86 at 50 patches for encoders without histology pre-training and 0.977 for a classifier trained on 20,000 labelled patches. Logistic regression on the same patches performed similarly at one and two patches per class and better from five upward (0.955 at 50). Midnight reached 0.922 with prototypes at 50 patches per class, similar to Phikon. The results were similar when support and evaluation patches came from different scanning centres. Conclusions: For the methods and datasets evaluated, no classifier trained on oral tissue transferred to this cross-tissue, cross-cancer target. Our analyses suggest that the change in what the classes contain mattered more than stain or scale, but they do not separate the causal contributions of tissue, cancer histology and acquisition. Frozen pathology foundation models made label-efficient classification on the target domain inexpensive; this replaces the oral-trained classifier and does not use oral images. The findings concern patch-level classification of breast cancer metastases and do not establish clinical performance for OSCC lymph node metastases. Full article

Tool wear directly affects the accuracy and reliability of machined workpieces; accurately predicting wear trends is crucial for developing reasonable tool-change strategies and avoiding sudden tool failure. Traditional parametric Inverse Gaussian (IG) processes can effectively characterize the non-negativity and stochastic fluctuations of incremental tool wear; however, the mean functions in such models often adopt an analytically defined power-law form that is arbitrarily preset, making it difficult to adapt to the nonlinear evolution patterns of tools at different stages. Consequently, their generalization capabilities are limited in cross-tool prediction scenarios. To address this, this paper proposes an IG-LSTM tool wear prediction method designed for cross-tool forecasting. First, Variational Mode Decomposition (VMD) is employed to extract the degradation trend of wear, and a Long Short-Term Memory (LSTM) neural network is trained based on this. Next, the trend predicted by the LSTM is used as the cumulative degradation function of the IG process to construct a degradation model for wear increment; simultaneously, the model parameters of the IG process are estimated via maximum likelihood estimation based on the training data. Finally, two sets of experiments—single-tool prediction and cross-tool prediction—were designed using the PHM2010 dataset. The experimental results show that, in both scenarios, the proposed method significantly outperforms traditional parametric inverse Gaussian models and purely data-driven LSTM models, achieving an organic integration of data-driven adaptive trend extraction and physics-driven stochastic degradation modeling. Full article

Construction safety inspections require identifying violated rules and localizing the associated workers or hazardous areas. Although multimodal large language models can describe hazards, supervision based solely on structured answers does not explicitly constrain the correspondence between rule records and visual targets. This study proposes RTR-Align, a Rule Token-to-Region Alignment method for construction safety violation recognition and localization. The method links the specific occurrence position of the Rule Token in each violation record to the corresponding target region, imposing separate spatial constraints on different instances associated with the same rule. It jointly learns rule assessment and target localization during LoRA adaptation of Qwen3-VL-8B-Instruct. The study uses the ConstructionSite10k dataset, which contains 10,013 construction site images with violation question-answering and localization annotations, and produces instance-level annotations for training and evaluation by separating grouped instances, adjusting bounding-box boundaries, and adding annotations for missed targets. In the main run, RTR-Align achieved an F1 score of 62.72% and a total intersection over union (Total IoU) of 39.40% on the 3004 test images, exceeding the base model by 48.11 and 32.55 percentage points, respectively. RTR-Align produced fewer false positives than the base model, but its recall was 1.19 percentage points lower. Analyses across rules and scenarios, along with parameter and qualitative analyses, revealed performance differences and remaining challenges. These results support joint learning of violation recognition and rule-associated localization, enabling the model to identify relevant targets with fewer false positives and provide spatial evidence for manual verification. Full article

Accurate road extraction from high-resolution satellite imagery is important for transportation planning, infrastructure monitoring, and geospatial updating, but blurred boundaries and fragmented segments still limit practical large-area applications. This study proposes EORENet, an edge-optimized road extraction network for 2 m GF-1 imagery and large-area road product generation. EORENet decouples road-surface segmentation and road-edge delineation into task-specific branches, introduces deformable convolution for adaptive spatial aggregation, and uses a dual-task cross-feature fusion module to exchange semantic and boundary cues. A GF-1 road sample set, the Cangling Road Dataset, was constructed and evaluated together with the Ottawa and DeepGlobe Road datasets. Compared with eight representative methods, EORENet achieved the best overall IoU, F1 score, and edge-buffered F1 score; on the Cangling dataset, it reached 55.11% IoU, 71.06% F1, and 56.21% edge-buffered F1. The trained model was then applied to a spatially independent large-area GF-1 inference collection. Scene-stratified product-level assessment with 1200 independently annotated 512 × 512 patches showed systematic scene-dependent differences, with F1 scores ranging from 62.80% in mountainous forest areas to 74.57% in high-density urban cores and a balanced scene-wise mean F1 of 69.48%. City-scale extraction over four metropolitan regions further visually illustrated the workflow’s potential for large-area road-surface product generation. Full article

Remote sensing semantic segmentation often produces spatially concentrated errors around object boundaries, small targets, shadowed areas, and ambiguous land-cover transitions. This study investigates whether learned spatial uncertainty can guide additional supervision of such difficult pixels. We propose Spatially Aware Uncertainty Feedback for Hard Region Learning (SUF-HRL), which refines a maximum-softmax-probability prior through a residual uncertainty branch attached to a SegFormer backbone. Local alignment and boundary concentration objectives encourage a task-adaptive spatial difficulty map; after a warm-up stage, its highest-scoring pixels receive additional cross-entropy supervision. Experiments use the Potsdam and Vaihingen benchmarks of the International Society for Photogrammetry and Remote Sensing and the LoveDA land-cover dataset. Under the controlled SegFormer-B2 protocol, mean intersection over union (mIoU) increases by 0.69, 1.14, and 1.01 percentage points, respectively. Averaging boundary-region mIoU over 3-, 5-, and 7-pixel bands gives gains of approximately 1.44, 1.32, and 1.32 percentage points. Component ablations support contributions from both spatial objectives. Learned uncertainty improves the reported spatial uncertainty criteria on Vaihingen, although it does not achieve the best precision–recall or calibration score. On Potsdam, training time per iteration increases by approximately 68.5% while the reported parameter totals remain nearly unchanged. The additional inference latency is small in the tested configuration. These results support learned spatial uncertainty as a useful training signal for hard-region segmentation. Full article

This study examines whether an explicit separation of navigation and avoidance yields useful control structure and performance gains for small unmanned surface vehicles in simulation. The policy-splitting proximal policy optimization actor (PSN-PPO) combines a navigation mean, a bounded avoidance residual, and a geometric risk gate with a learnable exponent, while retaining the PPO update and a separate critic. Five independent models per method are trained in each of five scenarios and evaluated on common nominal and perturbed cases. Crossing arrival is 100% for PSN-PPO versus 80% for PPO, but arrival without soft-domain intrusion is $25.6\%$ versus $37.6\%$; static-obstacle arrival is $50.4\%$ versus $80\%$. A 48-run equal-interaction study in static and crossing tasks finds no consistent advantage of full gating over always-on correction. Navigation-only control also achieves full crossing arrival, and the learned gate exponent remains close to its initial value. Simulated braking, turning, and paired threshold tests characterize the numerical maneuvering and risk scales. The contribution is an explicitly constrained actor decomposition and an empirical assessment of its limits; the results do not establish an additional performance benefit from the complete adaptive gate in these benchmark families. Full article

Geosmin (GSM) and 2-methylisoborneol (2-MIB) are dominant taste-and-odor (T&O) substances in drinking water and have extremely low human perceptual thresholds (0.004 and 0.01 µg/L, respectively). While nontoxic, their continued presence generates consumer rejection, damages utility credibility, and results in enormous economic losses. The management of GSM and 2-MIB throughout the drinking water supply chain, from source water to tap, is investigated here through an integrative review of representative T&O occurrences across multiple continents, GSM and 2-MIB-producing organisms, environmental influencing factors, and the inadequate removal achieved by strictly conventional treatment processes (coagulation/filtration), which improves only to 36–64% (GSM) and 38–59% (2-MIB) with the addition of low-dose pre-ozonation (0–1 mg/L). An analysis of pre-outbreak detection methods (remote sensing, qPCR, ML, and phycocyanin-based fluorescence detectors) and novel sensors (bio-electronic nose, electronic tongue, and odorant binding protein (OBP2a)-based biosensors) was conducted. Notably, advanced biosensors such as the bio-electronic nose can detect both GSM and 2-MIB below their respective human odor thresholds (<10 ng/L) without the need for sample pre-treatment. In contrast, OBP2a-based platforms show affinity for 2-MIB (but not GSM) and require signal amplification or preconcentration to reach trace-level detection, while the electronic tongue operates near these threshold levels (~20 ng/L). This review presents the first integrated comparison of conventional, biological, integrated, and advanced oxidation treatments in terms of removal efficiency, operating cost, energy consumption, toxicity of byproducts, and technology readiness levels, and concludes that no single current technology can simultaneously achieve >95% removal, maintain operating costs below US$ 0.50/m³, prevent the formation of toxic byproducts, and ensure long-term operational stability. Consequently, the implementation of multi-barrier treatment systems is required. Particular attention is given to scalable and economical treatment alternatives for small- and medium-sized plants in developing countries, such as biological aerated filtration coupled with ultrafiltration and permanganate/bisulfite processes. Future research priorities include life cycle assessment of treatment trains, harmonized ecotoxicological assessment of transformation products, real-time sensor-integrated adaptive treatment, and climate-resilient treatment strategies for intensified eutrophication. Full article

Echolocation enables individuals with impaired or no vision to recognize spatial features by analyzing sound reflections. However, acquiring this skill requires extensive training, and environmental factors influence its reliability. Moreover, users must continuously emit sounds while interpreting auditory feedback. In previous work, we introduced the LIDAR-Assisted Spatial Sensing (LASS) system, which detects the surrounding environment and translates it into stereo sound with varying pitches. Pitch encodes distance and stereo panning encodes direction, enhancing spatial perception. A key limitation of LASS is its passive scanning methodology: it restricts user control and requires a fixed 13-s full scan. In addition, the previous LASS evaluation was conducted with seated participants performing stationary localization judgments; it did not test whether the system could support closed-loop, mobile navigation. This paper presents an active LASS (ALASS) system that enables users to select regions of interest via a tactile handheld controller. This mechanism reduces regional scan time and gives users greater control, enabling adaptive scanning strategies similar to biological echolocation. A proof-of-concept study (N = 2) provides preliminary validation that ALASS supports closed-loop navigation in structured environments, where the user continuously senses, decides, and acts, steering based on audio feedback, with the resulting motion changing what the system subsequently senses. Full article

Agricultural machinery has extensive blind zones, and reliable proximity alerts need distance information that no single sensor provides across the working range. We fused a depth camera and a 16-channel LiDAR by distance-adaptive inverse-variance weighting, with weights derived from nominal, specification-based error models that cross at 3.87 m. YOLO11m, trained on five classes, detected objects, and the pipeline ran at 26 frames per second on an edge computing platform. Distances to Person and Tractor targets were evaluated against real-time kinematic GNSS in 2521 samples from front and rear tractor approaches at nominal speeds of 3, 5, and 7 km/h. Fusion reduced the root mean square error (RMSE) to 0.211 m, 30.9% below that of the depth camera alone, with the largest reduction (35.8%) at 7–12 m and no change within 7 m. The LiDAR alone was more accurate (0.167 m) but valid in only 82.7% of frames. With both sensors valid, the fused RMSE (0.141 m) was lower than that of either sensor. For detected targets, alert levels agreed with the reference in 99.0% of frames, but a Person recall of 0.72 limits overall reliability. Target geometry and LiDAR resolution strongly affect the benefit of fusion, which may guide sensor selection for other agricultural machines. Full article

Deformation of a phased array antenna surface can degrade its electrical performance. Conventional modal reconstruction depends on the accuracy of the structural model, while purely data-driven methods require extensive experimental data. We propose a reconstruction method that combines a modal prior, an input-dependent low-rank correction network, and Gaussian process data migration. Finite element modal analysis informs the baseline strain-to-displacement mapping. The correction network adapts this mapping to the input strain. Gaussian processes trained on experimental displacement residuals generate pseudo-experimental labels and confidence weights for fine-tuning. In the panel case study, the proposed method achieved a mean absolute error of 0.1596 mm and a maximum absolute error of 0.5963 mm. The corresponding errors for the conventional modal method were 1.5922 mm and 5.7291 mm. The proposed method retains an explicit strain-to-displacement mapping and uses a small experimental dataset. Full article

Forecasting heterogeneous vehicle–environment temperatures requires models that accommodate differences in amplitude, persistence and meteorological response while keeping component-model proliferation manageable. This study presents a channel-adaptive component forecasting framework and directed reuse of frozen dominant-component predictors. RIME selects complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) parameters using decomposition quality. A common component predictor combines static feature weighting, dilated causal convolution, historical-window bidirectional long short-term memory encoding, temporal attention and an autoregressive highway (DConvFTA-BiLSTM-AR). A field campaign in Xuzhou, China, provides 14,217 consecutive one-minute records from 64 vehicle and environmental temperature channels. Fourteen channels spanning contrasting materials and thermal regimes were benchmarked using a chronological 70/20/10 split, train-fitted scaling and decomposition restricted to the train–validation prefix. Under observed future meteorological inputs, macro MAE was 0.264 °C and macro MSE was 0.430 °C². The MAE was 13.7% lower than that of CEEMDAN-LSTM, the strongest neural baseline by macro MAE, in the evaluated interval. Directed reuse tests identify compatible source–recipient pairs while retaining recipient-specific predictors for non-dominant components and the residual. For 13 exposed-surface channels, a fixed-emissivity Planck transformation gives mean self-emitted 8–14 μm radiant-exitance MAE and channel RMSE of 0.593 and 1.280 W/m². These results establish component forecasting and partial predictor reuse as useful tools for the evaluated heterogeneous thermal array and provide temperature inputs for downstream infrared analysis. Full article