Rapid and reliable identification of multivariate geochemical anomalies is critical for delineating prospective mineralized zones and reducing uncertainty in mineral exploration targeting. Extended isolation forest (EIF) is a powerful unsupervised ensemble learning algorithm that efficiently isolates anomalies from high-dimensional geochemical datasets using randomly oriented hyperplane partitions. Previous studies have demonstrated the effectiveness of EIF in multivariate geochemical anomaly detection and mineral potential modeling. However, its performance can be significantly affected by stochastic variability arising from random partitioning and random subsampling during isolation tree construction, which may result in unstable anomaly patterns and inconsistent exploration targets in complex geological environments. To mitigate this limitation, we developed a robust unsupervised framework for the identification of multivariate geochemical anomalies associated with gold mineralization in the Southwestern Yilgarn Craton, Australia. The proposed framework integrates robust factor analysis (RFA), a Jaccard-based stability index and EIF to enhance the reliability and reproducibility of anomaly detection. RFA was first applied to compositional soil geochemical data to identify the most significant pathfinder elements associated with gold mineralization, which were subsequently used as input variables for the EIF model. The model was then optimized using a Jaccard-based stability criterion to ensure consistent anomaly detection across repeated independent runs. Model performance was assessed using area under the receiver operating characteristic curve (AUC). The obtained AUC value of 0.82 indicates strong predictive capacity, confirming that the generated anomaly map effectively delineates mineralization-related geochemical patterns and provides a reliable proxy for mineral prospectivity mapping. Overall, the proposed framework offers a robust and reproducible unsupervised approach for multivariate geochemical anomaly detection with strong applicability in both greenfield and brownfield mineral exploration settings.
LiDAR-based 3D object detection is essential for autonomous driving. This paper investigates the distinctive capability of FMCW LiDAR for short-term predictive perception. Given a single-frame FMCW point cloud with per-point radial velocity measurements, we study predictive object detection (POD), a detection-centric task that estimates future 3D bounding boxes from the current observation alone. Unlike conventional motion forecasting methods that typically rely on historical frames, trajectories, or tracking results, POD uses the instantaneous velocity cue of the current scan and therefore avoids additional temporal buffering. To this end, we design a velocity-aware framework that compensates radial velocity, generates virtual future points along the radial direction, and forms a virtual two-frame point cloud from a single scan. The virtual current-future points are encoded by sparse 4D voxel encoders and then separated into bird’s-eye-view (BEV) features for current-frame detection and future-frame prediction. We instantiate this framework with two representative 4D encoders, namely 4D SparseConv VoxelNet and 4D Voxel Transformer, and evaluate their accuracy-efficiency trade-offs. Experiments on data from this modality show that the proposed framework can perform short-horizon future-box detection while maintaining practical response time. Additional baseline, complexity, and failure-case analyses further clarify the benefits and limitations of single-frame POD.
Detecting particles from electromagnetic calorimeter (EMC) readouts is a crucial step for physics analysis and discovery. In this study, we propose Anti-Neutron Transformer (ANT), a representation learning approach that formats each energy hit as a token unit and converts EMC readouts to hit-token sequences. The key to building such a representation is the position cloze procedure (PCP), a self-supervised pre-training strategy in which the model removes and then recovers the positional information of selected high-energy hits. With 11 million unlabeled particle collision events, ANT learns the implicit spatial-structure prior of hits, thereby enabling representation of the sparse, irregular and non-local hit patterns. Experiments validate that ANT equipped with linear detection layers reduces the position prediction error from 17.31∘ to 9.17∘, and realizes anti-neutron momentum regression using deep learning for the first time, achieving a relative error of 20.78%.
The rapid evolution of autonomous driving has placed increasingly rigorous demands on 3D object detection within point cloud environments. High-quality datasets with efficient, accurate annotations are essential for improving detection performance. To address the challenges of high manual labeling costs and low efficiency, this paper proposes a model named O3DAA (Offline 3D Point Cloud Automatic Annotation). The proposed method integrates sparse and offline detectors to exploit long-term point cloud sequences, generating high-quality object bounding boxes and trajectories. Moreover, the branches of shape refinement and trajectory refinement are designed to extract semantic features of object shapes and motion patterns, thereby significantly enhancing the accuracy of 3D bounding box predictions. Experiments on the Waymo Open Dataset demonstrate that O3DAA surpasses most existing methods in tracking performance and long-range object detection, showcasing its superior performance and application potential in autonomous driving point cloud scenarios. Consequently, O3DAA offers an efficient and practical framework for large-scale point cloud annotation and the development of high-performance detection models.
Promptable segmentation foundation models can return visually plausible masks even when they fail, especially under cross-site shift. Existing reliability estimators often require labelled calibration data, model modification, or pixel-level uncertainty maps. We propose Confidence-Balanced Structural Prompt-Response Consistency (SPRC-CB), a training-free and label-free score that audits a predicted mask by measuring how its structure changes under small, approximately semantics-preserving prompt perturbations, including scale-relative box jitter and a boundary-near negative point. Component, hole, boundary, area-response and negative-prompt instability cues are fused with the model’s own confidence through a balanced rank rule. Across four endoscopic polyp benchmarks, SPRC-CB raises pooled AUPRC over SAM confidence by 0.12–0.19 across three failure definitions (paired bootstrap, all p ≤ 0.005) while retaining comparable or higher AUROC. Rejecting the highest-risk 10 % of masks recovers 73 % of failures versus 51 % for confidence alone. The advantage also persists under noisy boxes, detector-derived prompts and a supporting MedSAM check, as detailed in Supplementary Material S1.
Water conservancy and hydropower projects enhance regional climate resilience and watershed water security yet inevitably trigger large-scale involuntary reservoir resettlement. As representative involuntarily displaced populations, reservoir resettlees' social identity directly impacts local social stability and regional sustainable socioeconomic development. Based on a ten-year longitudinal qualitative investigation including in-depth interviews and participant observation in Village Y of Wuxikou Reservoir, Jiangxi Province, this paper divides the entire resettlement process into three stages: relocation, stabilization and development. From the dual perspectives of host community identity and out-groups identity, this study explores the dynamic evolutionary rules of resettlees' social identity. The results indicate that resettlees sequentially develop alienated identity, superficial adaptive identity and segregated identity across different phases. On this basis, the core concept of differential identity is proposed, which features a dual structure of vertical temporal differentiation and horizontal spatial differentiation. This paper expands the applicable scope and explanatory power of the “differential mode of association” and social identity theory in involuntary resettlement contexts. Grounding on the differential identity framework, this paper puts forward targeted integrated governance solutions to break intergroup segregation, facilitate cross-group integration and build resilient resettlement communities consistent with Sustainable Development Goals (SDGs) 11 and 13.
The discrimination of ore deposit types is primarily based on geological, geochemical, and isotopic characteristics. Conventionally, these types are identified using specific element diagrams. However, traditional geochemical methods often fail to determine scheelite deposit types of the complex Xuefengshan Sb-Au-W metallogenic belt in China, where mineralization resulted from the superposition of multiphase geological events. Machine learning (ML) methods, have been increasingly applied to identify deposit genesis by establishing relationships between deposit characteristics and genetic types using extensive datasets. However, inaccurate data labels, the limitations of single models, and poor model interpretability lead to decreased accuracy. This study proposes a ML framework based on interpretable ensemble learning. We collects geochemical element data from typical orogenic and magmatic-hydrothermal scheelite deposits globally. Deep clustering is used to filter data and overcome the subjectivity of original data labels. An ensemble learning model is used to construct a classifier to improve the model's robustness and generalization ability. An interpretable model is introduced to analyze the contribution of individual feature elements, revealing the metallogenic genesis. This method demonstrates high accuracy on the test set. According to this method, the scheelite deposit type of the Xuefengshan metallogenic belt is primarily magmatic-hydrothermal in origin, with orogenic superposition. This helps resolve a long-standing controversy in the region and establishes a repeatable and interpretable new paradigm for ML-based discrimination of ore deposit genetic types.
Though volcanogenic massive sulfide (VMS) deposits are major global sources of indium (In), the physicochemical mechanisms and key factors controlling its significant enrichment remain poorly understood. To address the issue, this study investigates the Tiemurt VMS Pb-Zn-Cu deposit, utilizing detailed petrography, in-situ LA-ICP-MS analysis, and thermodynamic modeling to reveal the In enrichment mechanisms in VMS deposits. Petrographic observations identified two distinct generations of sphalerite corresponding to different mineralization stages. The early-stage sphalerite (Sp1) is euhedral-subhedral, associated with pyrite, and displays darker colors (red to brown), whereas the late-stage sphalerite (Sp2) is anhedral, intimately intergrown with chalcopyrite, and shows lighter colors (mainly yellow). The trace element results demonstrate that Sp1 has a significantly higher In content (average 317 ppm) than Sp2 (average 220 ppm). Additionally, In concentrations positively correlate with Fe contents. Because Fe is the primary chromophore that darkens sphalerite, this strong coupled enrichment mechanism allows macroscopic sphalerite color (red > brown > yellow) to serve as a reliable indicator for In concentration. Crystallization temperatures calculated using the GGIMFis thermometer range from 344 to 382 °C for Sp1 and 312 to 355 °C for Sp2, indicating a cooling trend during fluid evolution. Thermodynamic modeling data showed that in the early-stage hydrothermal fluids (≥360 °C), Zn2+ preferentially complexes with Cl−, leaving InCl2+ or In3+ as unstable species, and In efficiently precipitates into Sp1 under the environment of log fO2 = −32 to −26 and pH = 6–8. As the fluids cool at ~340 °C, weakened Zn2+ competition allows In3+ to form stable InCl3, and In precipitates into Sp2 under the conditions of log fO2 = −42 to −32 and pH = 5.5–11. We therefore conclude that the key factor controlling the difference in In content between Sp1 and Sp2 is the precipitation mechanism rather than migration capacity. This may be different from the In enrichment mechanism associated with magmatic hydrothermal systems, where In is mainly present as InCl3 complexes with strong migration capacity. These new findings enable us to understand how the physicochemical conditions of fluids control the enrichment of In in VMS deposits, and also highlight that the color of sphalerite can be used to target potential In resources in PbZn deposits.
Few-shot segmentation aims to segment unlabeled images from unseen classes given only a few labeled images. Although significant advances have been achieved in few-shot segmentation, many approaches still exist the problem of the generalization ability of the model on unseen classes due to inaccurate correlation and location between support and query images. In this work, We propose the Task-Aware Meta-Learner Network (TAML), a meta-learning framework that dynamically captures task-specific support-query correlations through adaptive feature learning. Specifically, our network incorporates two novel designs: (1) a dynamic affinity module that can adaptively establish multi-scale correlations between support and query images based on the input task, which makes the propagation of the target information more flexible and effective; (2) a dynamic attention module that responds to different inputs in a highly class-specific manner to localize the target object accurately. Extensive experiments on PASCAL-5i and COCO-20i demonstrate the effectiveness of our proposed approach.
Gait recognition has important potential in industrial access-control applications. However, its application in industrial access-control scenarios remains challenging due to complex illumination, uniform workwear, safety helmets, turnstile-induced occlusion, and view variations caused by bidirectional passage. Existing public gait datasets cannot sufficiently characterize these real-world industrial conditions. To bridge this gap, we construct an industrial access-control gait dataset consisting of 150 workers, where gait sequences are collected from two overhead passage views: a forward overhead view and a backward overhead view. Each subject is recorded with five video sequences under each view, reflecting industrial passage characteristics. Furthermore, we propose two lightweight modules, namely Denoising Progressive Compression (DPC) and Segment-wise Quality Aggregation (SQA). DPC performs local denoising and progressive compression for 31-channel skeleton heatmaps, while SQA performs response-driven segment-wise temporal aggregation for occluded gait sequences. Experiments conducted on DeepGaitV2, SkeletonGait, and SkeletonGait++ demonstrate that the proposed dataset reveals the recognition challenges of industrial access-control gait. The results show that the proposed improvements enhance recognition robustness under occlusion and cross-view conditions.
Transformer-based methods have shown strong potential for image deraining through flexible contextual modeling. However, rain streaks vary substantially in scale and orientation, while multi-scale restoration remains sensitive to intra-scale representation and inter-scale feature transfer. To address these issues, we propose MinT, a multi-scale implicit neural Transformer for image deraining. Each branch follows a Transformer-style residual organization in which directional context aggregation serves as a rain-oriented token mixer. Shared-encoder INR bridges establish continuous correspondence between adjacent resolutions, and adaptive feature fusion selectively combines their features. Experiments on five synthetic and real-world benchmarks demonstrate strong and consistent restoration performance. Compared with the closest multi-scale INR baseline, MinT reduces FLOPs and trainable parameters by 13.9% and 14.0%, respectively. The source code is available at .
