TerraMosaic Daily Digest: Feb 15, 2026
Daily Summary
From 150 selected papers (950 new papers analyzed), the clearest signal is methodological convergence: hazard studies are no longer separating process understanding, sensing, and decision analytics. Instead, they are coupling them into operational chains that run from observation to warning.
The evidence is concrete. High-impact studies refine physically interpretable triggers and failure pathways (e.g., clustered landslide rainfall thresholds, anti-dip rock-slope seismic failure, deep-earthquake displacement fields), while parallel work hardens deployment under real constraints: sparse labels, non-stationary dynamics, correlated infrastructure failures, and real-time response needs. In short, this corpus marks a shift from hazard description toward testable, deployable hazard intelligence.
Key Trends
- Mechanism and observation remain the center of gravity: 108/150 papers fall in Concepts & Mechanisms or Detection & Monitoring, indicating that model performance is increasingly judged by whether it recovers physical failure behavior, not by accuracy alone.
- Precursor science is becoming quantifiable and actionable: Multiple studies move from empirical heuristics to probabilistic trigger design (rainfall-threshold matrices, b-value anomaly tracking, uncertainty-aware risk prediction), improving direct use in warning operations.
- Cascading risk is treated as a systems problem: Urban flood and seismic studies now model interdependent infrastructure, correlated damage, and recovery trajectories, linking hazard intensity to network-level service loss and intervention priorities.
- Field-constrained sensing pipelines are maturing: UAV LiDAR, seismic-geodetic data, and multi-source Earth-observation workflows are used to recover event-scale dynamics under incomplete observations, then feed forecasting and assessment models in near real time.
- Advanced AI is being filtered through deployment constraints: Graph, diffusion, and foundation-style models are increasingly designed for non-stationarity, data scarcity, and computational efficiency, signaling a transition from prototype algorithms to operational geohazard tooling.
Selected Papers
This digest features 150 selected papers from 950 new papers analyzed. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Critical rainfall thresholds for landslides based on extreme rainfall–induced clustered landslides and characteristic rainfall parameter analysis: a case study in Western Qinling Mountains, China
Core Problem: Traditional rainfall threshold models for landslides suffer from dependence on historical stochastic events and subjectivity in parameter selection, making them less effective for forecasting clustered landslides triggered by extreme precipitation.
Key Innovation: Proposes a novel critical rainfall threshold determination method based on landslide density-probability analysis, identifying characteristic rainfall parameters (3-day antecedent effective rainfall and 5-h current triggering rainfall), developing nonlinear models, and establishing four-tiered probabilistic alert thresholds and a comprehensive meteorological alert matrix.
2. Seismic Failure Mechanisms of Anti-dip Rock Slopes with Weak Interlayers: Insights from Large-scale Shaking Table Tests and Discrete Element Modeling
Core Problem: Poor understanding of seismic failure mechanisms in high-steep, anti-dip rock slopes with weak interlayers, common in southwestern China, hindering disaster risk management.
Key Innovation: Combined large-scale shaking table tests and discrete element modeling to reveal a progressive five-phase seismic failure evolution (microcracking to overall instability). Identified significant seismic amplification zones and, counter-intuitively, found that weak interlayers decisively enhanced seismic stability by dissipating energy and attenuating waves, reducing deformation range and magnitude.
3. Deep Earthquakes Can Generate Larger Co‐Seismic Displacements Than Shallow Events
Core Problem: The extent to which deep earthquakes contribute to Earth's surface deformation compared to shallow earthquakes, particularly the counter-intuitive observation that deep events might cause larger co-seismic displacements over broad regions.
Key Innovation: The finding, based on comparative GNSS analysis and modeling, that deep earthquakes can generate larger co-seismic displacements than shallow events across broad regions, primarily due to the 3D displacement pattern, highlighting deep earthquakes as substantial sources of surface deformation.
4. Detecting Spatiotemporal b-Value Anomalies with a Progressive Deep Learning Architecture
Core Problem: Identifying systematic seismicity patterns, specifically spatiotemporal anomalies in b-values, that precede large earthquakes, which is a central challenge in statistical seismology.
Key Innovation: Proposes a methodological framework using a hybrid deep-learning architecture (spatial convolutional encoder + temporal convolutional network) to detect spatiotemporal b-value anomalies. It constructs continuous 2+1D b-value fields and employs a progressive meta-epoch training scheme mirroring operational deployment to mitigate nonstationarity issues.
5. A Lightweight LLM Framework for Disaster Humanitarian Information Classification
Core Problem: Timely classification of humanitarian information from social media is critical for disaster response, but deploying large language models (LLMs) for this task is challenging in resource-constrained emergency settings.
Key Innovation: Develops a lightweight, cost-effective framework for disaster tweet classification using parameter-efficient fine-tuning (LoRA, QLoRA) on Llama 3.1 8B, demonstrating high accuracy with minimal parameter training and memory cost, and showing that RAG can degrade performance due to label noise.
6. Fault volume digital twin to reproduce the full slip spectrum, scaling and statistical laws
Core Problem: Existing mechanisms for fault zones explain only some observed slip dynamics, scaling, and statistical laws, lacking a unified framework to reproduce the full spectrum of seismic and geodetic observations.
Key Innovation: Demonstrates that incorporating an off-fault damage zone (with power-law distributed cracks) in a 2D shear fault zone model can reproduce a natural continuum from slow to fast ruptures, including Omori law, Gutenberg-Richter scaling, and the emergence of tremors, VLFEs, LFEs, SSEs, and earthquakes.
7. Accelerating tropical cyclone wave height estimation via machine learning and deep latent surrogates
Core Problem: Classical PDE-based wave models for estimating Significant Wave Height (Hs) during tropical cyclones are computationally expensive, and existing ML models often underestimate extreme Hs values, hindering real-time coastal hazard assessment.
Key Innovation: Development of machine learning and deep learning surrogates (specifically a PCA-TCN-LSTM model) to accelerate tropical cyclone wave height estimation, incorporating a data pre-processing pipeline (oversampling, weighted loss functions, PCA) to explicitly target and improve the estimation of extreme Hs values, achieving significant runtime reduction with high accuracy.
8. Terrain controls on snow accumulation and avalanche characteristics: a case study of the February 6, 2024, Bukongla Mountain Avalanche, Southern Himalayas
Core Problem: Understanding the geomorphic controls and dynamic mechanisms of extreme snow avalanches, particularly how terrain influences their destructive power and runout distance, as exemplified by the Bukongla Mountain Avalanche.
Key Innovation: Identifies a direct causal relationship between a two-level planation surface geomorphology (upper plateau accumulation, steep acceleration track) and the formation of a decoupled, far-reaching airblast, which extended the destructive path by over 25% and was the main cause of distal damage.
9. Rapid hazard prediction and assessment of post-fire debris flows using UAV lidar: Eaton Fire, California
Core Problem: Limited robust prediction systems for the magnitude and timing of post-wildfire debris flows due to sparse historical data, hindering effective hazard mitigation and emergency response.
Key Innovation: Deploys UAV lidar to map topographic change before and after rainfall events, quantifying debris flow volumes, measuring dry ravel volumes, and demonstrating a data-driven approach to predict hazards by comparing pre- and post-rainstorm surveys to evaluate debris flow sources.
10. Dynamic Response of Coal Under Different Lateral Pressure Coefficients: Insights into Rockburst Mechanisms in Coal Mines
Core Problem: Inadequate understanding of rockburst mechanisms in deep mining, especially regarding dynamic effects and the synergistic influence of loading rate and confinement, leading to underestimation of risk by traditional quasi-static criteria.
Key Innovation: Conducted triaxial SHPB tests on coal, revealing that dynamic strength and elastic strain energy increase linearly with loading rate, and confinement enhances energy storage. Identified a synergistic effect where rapid energy input under strong confinement leads to explosive failure. Proposed a modified rockburst hazard assessment framework incorporating triaxial confinement and rate dependence.
11. Modeling and analysis of urban critical infrastructure dynamic cascading failures in urban floods based on Simulink
Core Problem: Urban extreme flood disasters trigger complex dynamic cascading failures across critical infrastructure systems (CISs), making their prediction and analysis challenging for effective emergency management.
Key Innovation: Developing a MATLAB Simulink model for urban infrastructure dynamic cascading failures, using Guangzhou as a case study, to analyze system performance, facility anomaly causes, failure chain evolution, and spatiotemporal impacts of a 500-year flood event.
12. An enhanced system reliability framework for regional seismic risk assessment considering inter-structural correlations
Core Problem: Regional seismic risk assessment (RSRA) is crucial, but adequately capturing correlations in ground motion intensity measures (IMs) and structural damage measures (DMs) across buildings remains a major challenge.
Key Innovation: Presents an enhanced system reliability-based RSRA framework, reformulating the problem as a dependent k-out-of-n system reliability model and developing a numerical scheme integrating Gaussian copula sampling with the probability density evolution method (PDEM) for correlated cases.
13. Permafrost degradation: A critical driver of aboveground carbon sink loss in China's boreal forests
Core Problem: The specific influence of permafrost degradation on the carbon budget of boreal ecosystems, particularly the mechanisms driving aboveground biomass (AGB) changes, remains inadequately understood, and current Earth system models often misrepresent these processes.
Key Innovation: Reconstructed annual AGB stocks using multi-source satellite data (1988-2023) to demonstrate that permafrost degradation, transitioning from energy-limited to water-limited growth, is a critical constraint on boreal forest carbon accumulation, highlighting model biases and the need to incorporate thaw-related hydrological processes.
14. Optimization of cellular automata-based routing using discrete Boltzmann model for urban pedestrian evacuation during flooding
Core Problem: Urban flooding is a severe disaster threatening lives and property, necessitating effective route planning strategies for pedestrian evacuation to safe areas.
Key Innovation: Proposes a rainfall model using the multi-speed discrete Boltzmann model (DBM) in conjunction with a cellular automaton-based emergency evacuation and risk avoidance model to optimize urban pedestrian evacuation routes during flooding, demonstrating advantages over traditional methods.
15. Seismic displacement demand quantification of self-centering structures with metallic dampers considering collapse scenario under far-field ground motions
Core Problem: Existing design methods for Self-Centering Metallic-Energy-Dissipating Systems (SC-MEDS) inadequately address the P-Δ effect, which can lead to stiffness degradation and collapse under severe earthquakes, or ignore the unique trilinear hysteresis of SC-MEDSs, resulting in significant errors in nonlinear displacement evaluation.
Key Innovation: Analytically characterized the dynamic response of SC-MEDS with six independent parameters and quantitatively evaluated the P-Δ effect on nonlinear displacement ratio (CR) spectra. Proposed an accurate two-step deep learning-based prediction method: an SVM-based integrity classification model to estimate collapse and an optimized neural network for fast CR-R-T relation prediction.
16. DuoCast: Duo-Probabilistic Diffusion for Precipitation Nowcasting
Core Problem: Existing deep learning approaches for short-term precipitation forecasting (nowcasting) struggle to balance global structural consistency with local detail preservation, especially under complex meteorological conditions, which is critical for disaster response.
Key Innovation: DuoCast, a dual-diffusion framework that decomposes precipitation forecasting into low- and high-frequency components modeled in orthogonal latent subspaces, theoretically proven to reduce prediction error and achieving superior accuracy in both spatial detail and temporal evolution on benchmark radar datasets.
17. PeakWeather: MeteoSwiss Weather Station Measurements for Spatiotemporal Deep Learning
Core Problem: There is a need for high-quality, comprehensive spatiotemporal weather datasets to advance machine learning-based weather forecasting and related applications, especially in complex topographies.
Key Innovation: Introduces PeakWeather, a high-quality dataset of 8+ years of 10-minute surface weather observations from 302 MeteoSwiss stations across Switzerland's complex topography, complemented with topographical indices and NWP ensemble forecasts, serving as a real-world benchmark for spatiotemporal ML tasks relevant to weather forecasting and hazard prediction.
18. Dynamic numerical analysis of liquefiable silty fine sand reinforced by gravel pile
Core Problem: Silty fine sand foundations are prone to liquefaction under seismic loading, leading to significant displacement and pore pressure buildup in engineering structures like canals.
Key Innovation: Demonstrates through FLAC 3D numerical analysis that gravel piles effectively mitigate seismic-induced pore water pressure buildup and reduce plastic displacement, enhancing liquefaction resistance by creating drainage pathways.
19. A New Apparatus to Measure the Coefficient of Friction and Changes in Fracture Permeability
Core Problem: The need for a more comprehensive understanding of conditions controlling subsurface fault slip and seismic events, specifically the ability to accurately measure frictional properties and changes in fracture permeability simultaneously during fluid injection.
Key Innovation: Developed a new triaxial apparatus capable of independently controlling confining stress, axial shear stress, and pore pressure. Uniquely measures dynamic friction coefficient, stick-slip events, and changes in fracture permeability during and after frictional slipping, providing additional insight into subsurface changes due to fault slip and the effects of pore pressure and fluid flow.
20. Quantifying Predictive Resilience: A Fault-Tolerant Forecasting Framework with Operational Boundaries for Natural Hazards
Core Problem: Natural hazards severely disrupt transportation systems by inducing nonstationary traffic dynamics and concurrent sensor failures, creating a critical gap in obtaining reliable, real-time forecasts during ongoing emergencies.
Key Innovation: Pioneers a framework to enhance predictive resilience, featuring an Early-to-Late Prediction paradigm and a novel Graph-based Hybrid MoE-enhanced PDFormer (GHMoE-PDFormer) architecture for fault-tolerant forecasting, quantifying predictive resilience and defining operational boundaries.
21. Numerical investigation of TBM shield jamming mechanisms and mitigation measures using a GPU-Parallelized 3D FDEM method
Core Problem: Tunnel Boring Machines (TBMs) are prone to jamming in adverse ground conditions (e.g., loose fractured zones, high-stress areas), leading to significant economic losses and construction delays, and the underlying jamming mechanisms are not fully understood.
Key Innovation: Analyzed a TBM jamming incident using a GPU-parallelized 3D Finite Discrete Element Method (FDEM) integrated with field data to reveal shield jamming mechanisms, and proposed a grouting reinforcement algorithm and evaluated combined reinforcement measures to effectively reduce shield-rock interaction forces and mitigate jamming risks.
22. An entropy-like computer vision method for post-earthquake damage assessment of nonstructural components within indoor scenes
Core Problem: Quantifying the collective damage patterns of diverse nonstructural components (NSCs) in indoor environments after earthquakes, which is challenging for traditional assessment methods.
Key Innovation: Proposes an entropy-like computer vision method that integrates Structural Similarity Index (SSIM) for scene-level disorder detection and a multi-object segmentation-tracking module for identifying severe NSC damage (sliding/overturning). The method was validated using shaking table tests and real earthquake surveillance videos.
23. Long Unrest (2007–2025) at Laguna del Maule: Linking Strain Field and Seismicity From GNSS and Seismic Data
Core Problem: Understanding the interplay between crustal deformation, magma dynamics, and seismicity during prolonged volcanic unrest, specifically at the Laguna del Maule volcanic field which has been uplifting at exceptional rates.
Key Innovation: Integrated GNSS and local seismic observations (2013-2024) to model the reservoir strain field, relocate earthquakes, and determine focal mechanisms. The study identified a shallow spheroidal reservoir undergoing significant overpressure and volume change, linked seismicity clusters to dilatational strain, and distinguished two phases of unrest, demonstrating how magma input, crustal deformation, and faulting interact during long-term volcanic deformation.
24. Atmospheric and Deposition Responses of 10Be to Volcanic Eruptions Inferred From the Aerosol‐Climate Model ECHAM6.3‐HAM2.3‐SALSA2.0: 10Be
Core Problem: Interpretation of 10Be records for solar reconstructions is complicated by volcanic influences, requiring better understanding of atmospheric and deposition responses to eruptions.
Key Innovation: Used the ECHAM6.3-HAM2.3-SALSA2.0:10Be model to assess the impacts of major volcanic eruptions (Agung, El Chichón, Pinatubo) on 10Be, showing enhanced stratospheric sedimentation, tropospheric increases coinciding with stratosphere-troposphere exchange, and nonlinear increases with higher SO2 injections.
25. MAUNet-Light: A Concise MAUNet Architecture for Bias Correction and Downscaling of Precipitation Estimates
Core Problem: Satellite-derived data products and climate model simulations of precipitation often exhibit systematic biases and coarse resolution compared to in-situ measurements, hindering their use in operational systems.
Key Innovation: Develops MAUNet-Light, a compact and lightweight neural network architecture that performs both bias correction and spatial downscaling of precipitation estimates with reduced computational requirements and comparable accuracy to state-of-the-art models, leveraging knowledge transfer.
26. Selection of CMIP6 Models for Regional Precipitation Projection and Climate Change Assessment in the Jhelum and Chenab River Basins
Core Problem: Accurate water resource management and climate change impact studies require reliable regional precipitation projections, but different General Circulation Models (GCMs) from CMIP6 produce contrasting results, necessitating a robust selection method without relying on in-situ data.
Key Innovation: Develops an envelope-based method, incorporating machine learning, for selecting CMIP6 GCMs for regional hydroclimate change impact studies, enabling selection without in-situ reference data. Identifies NorESM2 LM and FGOALS g3 as suitable models for the Jhelum and Chenab River Basins and highlights highly vulnerable regions under climate change scenarios.
27. PromptDepthAnything++: Accurate 4K Metric Depth Estimation via Pattern-Agnostic Prompting
Core Problem: Achieving accurate, high-resolution metric depth estimation, especially at 4K, is challenging, and effectively leveraging prompts in depth foundation models for this task is an unexplored paradigm.
Key Innovation: Prompt Depth Anything++, a novel prompting paradigm for metric depth estimation that uses low-cost LiDAR as a prompt to guide depth foundation models for accurate 4K metric depth output, incorporating a concise prompt fusion design, a scalable data pipeline, and a new pattern-agnostic prompting mechanism.
28. Surrogate-based dimension reduction for reliability analysis and LRFD calibration: Breakwater foundations in depth-varying soils
Core Problem: Traditional reliability analysis and LRFD calibration for breakwater foundations with spatially variable and depth-varying soils suffer from high dimensionality and computational cost, making it difficult to accurately assess failure probability under seismic loading.
Key Innovation: Proposes an efficient framework using a correlation-based sensitivity filter for dimension reduction and adaptive ensemble artificial neural network metamodels for probabilistic simulations, achieving significant computational savings and improved accuracy in assessing failure probability and calibrating resistance factors for breakwater foundations under seismic loading.
29. The July 14, 2025 Mw 5.3 earthquake in the NE Alboran Sea (Spain): insights into the causative source from seismic relocation and moment tensor analysis
Core Problem: Understanding the causative source and mechanisms of a significant Mw 5.3 earthquake in a historically active but previously unmapped offshore area, and improving the accuracy of earthquake locations in the region.
Key Innovation: Computes the seismic moment tensor and relocates seismic sequences using 3D Earth models and probabilistic/double-difference approaches, providing a better-constrained earthquake catalog and insights into the strike-slip faulting mechanism and deeper hypocenters of the 2025 event.
30. Assessment of social vulnerability in terms of natural hazards in Türkiye based on entropy method
Core Problem: A lack of recent, objective, and comprehensive assessments of social vulnerability to disasters at the provincial level in Türkiye, hindering effective disaster risk reduction planning.
Key Innovation: Assesses social vulnerability at the provincial level in Türkiye using an objective entropy method, determining 9 sub-indicators and 34 variables across demographic, economic, and adaptive capacity dimensions, providing a weighted index and regional vulnerability insights for decision-makers.
31. Synergistic biopolymer-fiber reinforcement for mechanical optimization of EICP- solidified sand
Core Problem: EICP-solidified soil suffers from brittle failure and a lack of nucleation sites, limiting its mechanical performance and applicability in geotechnical engineering.
Key Innovation: Proposed a synergistic biopolymer (xanthan gum) and fiber reinforcement method for EICP-solidified sand. Demonstrated that xanthan gum enhances urease immobilization and provides nucleation sites, while fibers offer additional immobilization sites and a bridging effect, leading to a 187.51% increase in UCS and excellent water stability.
32. Volcanic arsenic sources and seasonal hydrogeochemistry in the El Salado Basin, Jalisco, Mexico
Core Problem: Understanding the hydrogeochemical processes that regulate groundwater and surface water quality, particularly the sources and mobility of arsenic and boron, in volcanic-geothermal regions like the El Salado Basin is crucial for risk assessment.
Key Innovation: Applied Piper diagrams, principal component analysis, clustering, and spatial interpolation to seasonal hydrogeochemical data. Identified that sodium-bicarbonate facies and elevated As-B concentrations are controlled by geothermal water-rock interaction. Revealed how seasonal hydrology (dilution, redox, evapoconcentration) modulates trace-element mobility, clarifying the geogenic origin of arsenic and providing a methodological framework for interpreting trace-element mobilization in volcanic aquifers.
33. Stiffness and cyclic responses of kaolin clay at a wide range of strains: the role of structure
Core Problem: A fundamental understanding of how soil structure influences the stiffness parameters and cyclic responses of kaolin clay across a wide range of strains, crucial for predicting soil behavior under dynamic loading.
Key Innovation: Investigated the stiffness and cyclic responses of kaolin clay with controlled structures through resonant column and cyclic triaxial tests, demonstrating the significant role of soil structure in stiffness properties at small to intermediate strains and its influence on pore water pressure generation and deformation under cyclic loading.
34. Experimental and Numerical Study on the Triaxial Failure Mechanisms of Shotcrete–Rock Combined Body: Effects of Confining Pressure and Interface Inclination
Core Problem: Insufficient understanding of the mechanical response and failure mechanisms of shotcrete-rock combined bodies (SRCBs) with inclined interfaces under triaxial loading, especially in deep-buried tunnels where complex orientations are common.
Key Innovation: Conducted triaxial compression tests on SRCBs with varying interface inclinations and confining pressures, categorizing three distinct failure modes. Showed that peak strength, elastic modulus, and Poisson's ratio vary significantly with inclination and confining pressure. Developed and verified discrete element models, providing guidance for design and safety assessment of deep-buried tunnels.
35. A hybrid unsupervised-to-supervised machine learning framework for fracture segmentation in natural gas hydrate-bearing sediments
Core Problem: Reliable segmentation and quantitative analysis of low-contrast and discontinuous fracture networks in hydrate-bearing sediments (HBS) from CT images for geomechanical stability assessment and hazard prediction.
Key Innovation: Developed a hybrid unsupervised-to-supervised machine learning framework (SCD-UNet) for robust fracture segmentation in HBS, achieving high accuracy (Dice score 0.9533) and efficiency, providing quantifiable inputs for geomechanical stability assessment and hazard prediction.
36. A 3D coupled multibody dynamics (MBD)-FDEM framework with GPGPU parallelization for simulating TBM-ground interaction and orientation behavior
Core Problem: Accurately simulating complex TBM-ground interaction, including rock fracturing and fragmentation, and efficiently capturing TBM component motions and mechanical couplings, is challenging but crucial for TBM orientation prediction and guidance in complex geologies.
Key Innovation: Proposed a 3D coupled Multibody Dynamics (MBD)-Finite-Discrete Element Method (FDEM) framework with GPGPU parallelization to simulate TBM-ground interaction and orientation response, effectively capturing fracturing behaviors and TBM component motions, demonstrating its potential for TBM orientation prediction in complex geologies.
37. Assessing and optimizing high-resolution global river streamflow estimates with triple collocation analysis
Core Problem: Accurate global river discharge data are essential for water resource management, flood forecasting, and drought mitigation, but global river gauge networks are sparse and hydrological models require systematic evaluation and improvement.
Key Innovation: Assesses and optimizes high-resolution global river streamflow estimates from three hydrological models using observed data and two data fusion methods (Triple Collocation and simple averaging), demonstrating that Triple Collocation offers the highest overall performance and high-resolution models are more accurate.
38. A new method for anti-floating of underground structures: experimental study on active–passive combined anti-floating
Core Problem: Designing effective anti-floating measures for underground structures in complex hydrogeological conditions, where buoyancy from basal water pressure poses a significant stability challenge.
Key Innovation: Developed an active–passive combined anti-floating test system that integrates active pressure-limiting drainage with passive dead weight. Experimental and numerical (COMSOL Multiphysics) results demonstrated that active pressure-limiting drainage effectively reduces basal water pressure, and smaller pressure-limiting heads increase the active anti-floating contribution, providing a reference for design.
39. A coupled train-track-subgrade dynamic framework integrating the continuous surface cap model for ballastless track damage analysis
Core Problem: Accurately predicting damage in ballastless tracks under dynamic train loads, considering concrete's nonlinear behavior and the bidirectional coupling between concrete damage and dynamic responses, especially under conditions like differential subgrade settlement.
Key Innovation: A novel dynamic damage analysis framework integrating the calibrated continuous surface cap model (CSCM) into a three-dimensional train-track-subgrade coupled dynamic model (TTSDM). This framework provides an accurate and efficient tool for dynamic damage simulation, offering valuable insights for safety assessment and durability design in high-speed railway engineering, particularly in analyzing the effects of differential subgrade settlement.
40. Matching of SAR and optical images based on transformation to shared modality
Core Problem: The significant differences in physical principles underlying SAR and optical image acquisition make their precise co-registration (matching) difficult, hindering multi-modal remote sensing applications.
Key Innovation: Proposes a novel approach for SAR and optical image matching by transforming both image types into a shared, non-degenerate modality that preserves significant features. This enables the use of pre-trained general image matching models (like RoMa) and demonstrates superior matching quality and versatility.
41. CBEN -- A Multimodal Machine Learning Dataset for Cloud Robust Remote Sensing Image Understanding
Core Problem: Clouds in optical satellite imagery hinder remote sensing, especially for time-sensitive applications like natural disasters, and existing machine learning datasets often exclude cloudy images, leading to methods that fail under real-world cloudy conditions.
Key Innovation: Introduction of CloudyBigEarthNet (CBEN), a multimodal dataset of paired optical and radar images with cloud occlusion, and demonstration that adapting state-of-the-art methods to cloudy optical data during training significantly improves performance on cloudy test cases, enabling more robust remote sensing for natural disaster applications.
42. Physics-Informed Laplace Neural Operator for Solving Partial Differential Equations
Core Problem: Purely data-driven neural operators for parametric partial differential equations (PDEs) require extensive training data and generalize poorly in small-data regimes or under unseen input functions.
Key Innovation: Proposes the Physics-Informed Laplace Neural Operator (PILNO), which embeds governing physics (PDE, boundary/initial conditions) into training, leverages virtual inputs for physics-only supervision, and uses temporal-causality weighting to improve accuracy, reduce variability, and achieve stronger out-of-distribution generalization in small-data settings.
43. GSM-GS: Geometry-Constrained Single and Multi-view Gaussian Splatting for Surface Reconstruction
Core Problem: 3D Gaussian Splatting, despite its speed and rendering quality, suffers from unstructured and irregular Gaussian point clouds, leading to challenges in reconstruction accuracy and loss of high-frequency details in complex surface microstructures.
Key Innovation: Proposes GSM-GS, a synergistic optimization framework integrating single-view adaptive sub-region weighting constraints (using image gradient features and depth discrepancy) and multi-view spatial structure refinement (geometry-guided cross-view point cloud association with dynamic weight sampling) to enhance geometric detail characterization and reinforce multi-view consistency for improved surface reconstruction.
44. Universal Transformation of One-Class Classifiers for Unsupervised Anomaly Detection
Core Problem: One-class classifier-based anomaly detectors are susceptible to training label noise because they assume the training data consists solely of nominal values.
Key Innovation: Presenting a dataset folding method that transforms an arbitrary one-class classifier-based anomaly detector into a fully unsupervised method by utilizing multiple independently trained instances to filter the training dataset for anomalies, achieving state-of-the-art performance.
45. Learning functional components of PDEs from data using neural networks
Core Problem: Partial differential equations (PDEs) often contain unknown functions that are difficult or impossible to measure directly, hampering the ability to derive predictions from the model.
Key Innovation: Embeds neural networks into PDEs to approximate unknown functions from steady-state data, demonstrating recovery of interaction kernels and external potentials in nonlocal aggregation-diffusion equations, and showing how various factors affect recovery success.
46. Barron-Wiener-Laguerre models
Core Problem: Classical Wiener-Laguerre models provide only deterministic point estimates for causal operator learning, lacking uncertainty quantification, despite their structural efficiency and interpretability.
Key Innovation: A probabilistic extension of Wiener-Laguerre models that combines Laguerre-parameterized causal dynamics with probabilistic Barron-type nonlinear approximators, enabling Bayesian inference on the nonlinear map and providing posterior predictive uncertainty for time-series modeling and nonlinear systems identification.
47. Random Forests as Statistical Procedures: Design, Variance, and Dependence
Core Problem: Random forests are typically described algorithmically, lacking a finite-sample, design-based statistical formulation that clarifies how their underlying randomized construction impacts predictive variability and dependence.
Key Innovation: A finite-sample, design-based formulation of random forests that provides an exact variance identity, separating finite-aggregation variability from structural dependence, and decomposes single-tree dispersion and inter-tree covariance to clarify how design mechanisms (resampling, feature randomization, split selection) govern forest behavior.
48. Spatio-Temporal driven Attention Graph Neural Network with Block Adjacency matrix (STAG-NN-BA) for Remote Land-use Change Detection
Core Problem: Standard deep neural networks are limited to Euclidean domains, restricting their application in remote sensing for land-use monitoring, which involves non-Euclidean and spherical data. Existing methods struggle to effectively capture irregular spatial relationships and spatio-temporal dynamics in satellite imagery.
Key Innovation: Proposes novel Graph Neural Network architectures, SAG-NN and STAG-NN-BA, for spatial and spatio-temporal land-use change detection using satellite imagery. These methods leverage superpixels as graph nodes and introduce block adjacency matrices to combine unconnected region adjacency graphs, enabling effective information propagation and outperforming graph and non-graph baselines.
49. Beyond All-to-All: Causal-Aligned Transformer with Dynamic Structure Learning for Multivariate Time Series Forecasting
Core Problem: Most multivariate time series forecasting methods use an undifferentiated 'all-to-all' paradigm, which struggles to identify variable-specific causal influences and often entangles causally relevant information with spurious correlations.
Key Innovation: Proposes an 'all-to-one' forecasting paradigm and the Causal Decomposition Transformer (CDT), which constructs a Structural Causal Model, partitions historical sequences based on inferred causal structure, integrates a dynamic causal adapter for structure learning, and applies projection-based output constraints to mitigate collider bias, significantly improving multivariate time series forecasting.
50. Self-Supervised Temporal Super-Resolution of Energy Data using Generative Adversarial Transformer
Core Problem: Bridging the temporal granularity gap in energy system models requires resampling time series, but conventional upsampling methods lose information, and advanced models face challenges like reliance on supervised learning, which is paradoxical when high-resolution data is intrinsically absent.
Key Innovation: Introduces a new self-supervised method utilizing Generative Adversarial Transformers (GATs) for temporal super-resolution of time series data, which can be trained without ground-truth high-resolution data, significantly reducing RMSE in upsampling tasks and improving accuracy in application scenarios.
51. Context-Specific Causal Graph Discovery with Unobserved Contexts: Non-Stationarity, Regimes and Spatio-Temporal Patterns
Core Problem: Real-world problems, especially in climate applications, require causal reasoning on spatially gridded time series data, but existing methods struggle with non-stationarity, imperfectly recoverable structure, and the complexity of encoding system-states.
Key Innovation: Developed a modular framework that modifies constraint-based causal discovery approaches on the level of independence testing, allowing for context-specific causal graph discovery in the presence of unobserved contexts, non-stationarity, and spatio-temporal patterns.
52. Language-in-the-Loop Culvert Inspection on the Erie Canal
Core Problem: Human inspection of culverts is challenging due to age, geometry, poor illumination, weather, and lack of easy access, making frequent inspections for safe operation difficult.
Key Innovation: VISION, an end-to-end, language-in-the-loop autonomy system, couples a web-scale vision-language model (VLM) with constrained viewpoint planning for autonomous culvert inspection, using brief prompts for open-vocabulary ROI proposals, stereo depth for scale, and a planner for targeted close-ups, achieving high agreement with subject-matter experts without domain-specific fine-tuning.
53. Learning to Feel the Future: DreamTacVLA for Contact-Rich Manipulation
Core Problem: Vision-Language-Action (VLA) models struggle with contact-rich manipulation tasks due to their blindness to physical contact, limiting their effectiveness in open-ended physical environments like disaster response.
Key Innovation: DreamTacVLA, a framework that grounds VLA models in contact physics by learning to feel the future, uses a hierarchical perception scheme and a tactile world model to predict future tactile signals, enabling robust, touch-aware robotic agents for contact-rich manipulation tasks, including disaster response, and outperforming state-of-the-art VLA baselines.
54. A real-time hybrid testing strategy for offshore wind turbines under wind–wave–earthquake loads using a frequency-decoupled wind load simulator
Core Problem: Rigorous and accurate testing of offshore wind turbines under combined wind, wave, and earthquake loads is challenging due to Froude-Reynolds scaling conflicts and the need for real-time simulation of multi-hazard interactions.
Key Innovation: A novel Real-Time Hybrid Testing (RTHT) strategy featuring a frequency-decoupled Wind Load Simulator (WLS) with a physical-frequency decoupling design (Fluctuating WLS for high-frequency, Average WLS for static loads) and a dual-loop control framework, demonstrating efficacy in replicating structural dynamic responses under multi-hazard conditions.
55. Local resilience in the rural Northern Sweden: challenges and opportunities in disaster management
Core Problem: Rural areas face unique challenges in effective disaster management due to vast geography, demographic decline, and limited resources, leading to disparities between formal demands and local capacities.
Key Innovation: Identifies key challenges (formal demands, geographical constraints, resource limitations) and innovative local strategies (dual roles, informal working methods, collaborative work, skilled personnel, organizational spirit) employed by rural municipalities in Northern Sweden to enhance disaster management and resilience.
56. Particle size and grain fabric effect on mechanical behavior of crystalline rocks using grain-based DEM
Core Problem: Conventional grain-based models (GBM) cannot alter mineral grain orientation and shape, limiting their ability to fully represent natural rock, and the effects of particle size and fabric on crystalline rock mechanical behavior need further investigation.
Key Innovation: Introduced an improved GBM (IGBM) allowing alteration of mineral grain orientation and shape. Demonstrated that uniaxial compressive strength (UCS) is lowest when maximum shear stress aligns with grain boundary orientation and increases with decreasing particle size. Advised maintaining a grain-to-particle size ratio of at least 8 for accurate DEM simulations in rock cracking.
57. Mechanism of Microstructural Damage and Macromechanical Deterioration of Deep Sandstone Under Dry–Wet Erosion
Core Problem: Lack of comprehensive understanding of the multi-scale mechanisms (mineral, microstructure, pore structure) by which cyclic dry-wet erosion and environmental humidity deteriorate deep roadway sandstone, impacting its macroscopic mechanical properties and stability.
Key Innovation: Integrated MTS-815, SEM, NMR, and XRD to show that dry-wet cycles and humidity weaken intergranular bonding, cause feldspar decomposition, loosen pore structures, and lead to a "damage–linkage–invasion" effect. Quantified mineral depletion (ηm) and demonstrated its correlation with sandstone deterioration, providing insights for deep underground roadway safety and stability assessment.
58. A Probabilistic Variational Inference Method for Anomaly Detection in Unmanned Aerial Vehicle Flight Data
Core Problem: Accurate anomaly detection in complex, multidimensional UAV flight data is challenging, and current static threshold-based methods fail to adapt to dynamic patterns or provide end-to-end detection.
Key Innovation: A probabilistic variational inference model for UAV flight data anomaly detection, utilizing a conditional encoder and a Transformer-based VAE to extract contextual information and generate expected distributions, providing anomaly probabilities and confidence intervals for improved decision-making.
59. Epistemic and aleatoric uncertainty in optical vegetation trait retrieval: Concepts, Methods, and Outlook
Core Problem: The uncertainty of vegetation trait retrievals from remote sensing is often under-reported or ambiguously defined, hindering reliable assessments of ecosystem health, crop productivity, and climate impacts, and making data assimilation and risk-aware decision-making challenging.
Key Innovation: A comprehensive scoping review that clarifies and operationalizes the distinction between aleatoric and epistemic uncertainty in vegetation trait retrieval, synthesizing methodologies for their quantification, disentanglement, and integration into retrieval frameworks, and advocating for routine release of transparently calibrated uncertainty layers for improved data interpretation and decision-making.
60. Density–Inertia Coupling Drives Solute Trapping at Vertical Fracture Intersections
Core Problem: The coupled impact of density-driven flow and fluid inertia on solute transport and trapping in fracture networks remains underexplored, despite their individual effects being recognized.
Key Innovation: Integration of pore-to-network-scale dye visualization experiments and numerical simulations to reveal that density-induced convection and inertia-driven vortices cause localized solute retention at fracture intersections, with maximum trapping occurring under specific flow imbalance conditions.
61. On the Beam Characteristics of X‐Ray Bursts Observed in Rocket‐Triggered Lightning
Core Problem: Limited understanding of the X-ray emission characteristics and beam patterns during lightning, specifically from leader propagation.
Key Innovation: Detected X-ray bursts during rocket-triggered lightning, innovatively integrating optical imaging with 3D lightning channel reconstruction (DAS) to provide geometric evidence of beam-like radiation along the leader path and the first quantitative estimation of the X-ray photon beam half-angle width (40°-46°).
62. Projected Late 21st Century Warming Unprecedented in Northwest China in a Holocene Context
Core Problem: A scarcity of regional paleorecord syntheses limits understanding of natural long-term climate variability in Northwest China and hinders the contextualization of contemporary warming.
Key Innovation: Presented paleorecord syntheses for summer and annual temperatures during the Holocene in Northwest China, projecting that late 21st-century warming will exceed early-Holocene peak warmth (>1.0°C warmer than 20th century mean) even under a low emission scenario.
63. Diverging Spring Warming and Growing Season Shifts Across Eurasia and North America Under Future Climate
Core Problem: Large model uncertainty limits climate risk assessment for springtime warming over Northern Mid-High-latitude Land, primarily due to model divergence in surface-albedo feedback linked to historical snowmelt sensitivity.
Key Innovation: Developed a novel emergent constraint targeting snowmelt sensitivity, which halved the spread of projected warming and revealed a pronounced geographical asymmetry, leading to refined projections that substantially alter ecological outcomes for growing season shifts.
64. High-dimensional Level Set Estimation with Trust Regions and Double Acquisition Functions
Core Problem: Accurately classifying whether an unknown function's value exceeds a specified threshold in high-dimensional spaces using active learning with limited initial data.
Key Innovation: Proposes TRLSE, an algorithm for high-dimensional LSE that uses trust regions and dual acquisition functions (global and local) to identify and refine regions near the threshold boundary, demonstrating superior sample efficiency.
65. Safe Reinforcement Learning via Recovery-based Shielding with Gaussian Process Dynamics Models
Core Problem: Reinforcement learning (RL) often lacks provable safety guarantees for safety-critical applications, especially with unknown and non-linear continuous dynamical systems.
Key Innovation: Introduces a novel recovery-based shielding framework for safe RL, integrating a backup policy with Gaussian Process (GP) based uncertainty quantification to predict and recover from potential safety constraint violations, enabling unrestricted exploration and sample-efficient learning without compromising safety.
66. LiDAR-Anchored Collaborative Distillation for Robust 2D Representations
Core Problem: Pre-trained 2D image encoders lack robustness in noisy and adverse weather conditions, which is crucial for reliable visual perception in real-world scenarios.
Key Innovation: A novel self-supervised approach, Collaborative Distillation, that leverages 3D LiDAR as self-supervision to improve the robustness of 2D image encoders to noisy and adverse weather conditions, while retaining original capabilities and enhancing 3D awareness.
67. Self-Supervised JEPA-based World Models for LiDAR Occupancy Completion and Forecasting
Core Problem: Autonomous driving requires robust world models for long-term planning, but learning such models scalably demands self-supervised approaches, and existing methods may not fully leverage unlabeled data for spatiotemporal evolution from LiDAR.
Key Innovation: AD-LiST-JEPA, a self-supervised world model for autonomous driving that predicts future spatiotemporal evolution from LiDAR data using a Joint-Embedding Predictive Architecture (JEPA) framework, demonstrating improved performance on LiDAR-based occupancy completion and forecasting (OCF).
68. Power Interpretable Causal ODE Networks: A Unified Model for Explainable Anomaly Detection and Root Cause Analysis in Power Systems
Core Problem: Existing machine learning models for time series anomaly detection in cyber-physical systems like power grids often operate as black boxes, providing only binary outputs without explanations (e.g., anomaly type, origin, or shape characterization).
Key Innovation: Proposes Power Interpretable Causality Ordinary Differential Equation (PICODE) Networks, a unified, causality-informed architecture that jointly performs anomaly detection along with explanations, including root cause localization, anomaly type classification, and anomaly shape characterization, achieving competitive detection performance with improved interpretability and reduced reliance on labeled data.
69. Coden: Efficient Temporal Graph Neural Networks for Continuous Prediction
Core Problem: Existing Temporal Graph Neural Networks (TGNNs) are primarily designed for one-time predictions, leading to significant computational overhead or prediction quality issues when adapted for continuous prediction scenarios on large dynamic graphs.
Key Innovation: Coden, a Temporal Graph Neural Network model specifically designed for efficient and effective continuous predictions on dynamic graphs, which innovatively overcomes key complexity bottlenecks while preserving comparable predictive accuracy, substantiated by theoretical analyses.
70. Channel-Aware Probing for Multi-Channel Imaging
Core Problem: Training and evaluating vision encoders on Multi-Channel Imaging (MCI) data is challenging due to varying channel configurations, limiting reuse of pre-trained encoders and leading to weak probing performance when directly transferring strategies from other domains.
Key Innovation: Channel-Aware Probing (CAP), which exploits inter-channel diversity in MCI datasets by controlling feature flow at both encoder (Independent Feature Encoding) and probe (Decoupled Pooling) levels. CAP consistently improves probing performance, matches fine-tuning from scratch, and significantly reduces the gap to full fine-tuning.
71. Adaptive Structured Pruning of Convolutional Neural Networks for Time Series Classification
Core Problem: Deep learning models for Time Series Classification (TSC) have high computational and memory requirements, limiting deployment on resource-constrained devices, and existing structured pruning methods rely on manually tuned hyperparameters.
Key Innovation: Proposes Dynamic Structured Pruning (DSP), a fully automatic, structured pruning framework for convolution-based TSC models. DSP introduces an instance-wise sparsity loss during training and a global activation analysis to identify and prune redundant filters without predefined ratios, achieving significant compression while maintaining accuracy.
72. Closing the Loop: A Control-Theoretic Framework for Provably Stable Time Series Forecasting with LLMs
Core Problem: Existing LLM-based time series forecasting approaches use a naive open-loop autoregressive generation strategy, leading to inevitable error accumulation (exposure bias) and significant trajectory drift over long horizons.
Key Innovation: Reformulates autoregressive forecasting using control theory, proposing F-LLM (Feedback-driven LLM), a novel closed-loop framework that actively stabilizes the trajectory via a learnable residual estimator (Observer) and a feedback controller, providing theoretical guarantees for uniformly bounded error.
73. Bootstrapping MLLM for Weakly-Supervised Class-Agnostic Object Counting
Core Problem: Fully-supervised object counting methods require costly point-level annotations, while existing weakly-supervised methods are often limited to counting a single category, hindering broad applicability.
Key Innovation: Proposes WS-COC, the first MLLM-driven weakly-supervised framework for class-agnostic object counting, incorporating a divide-and-discern dialogue tuning strategy, a compare-and-rank count optimization strategy, and a global-and-local counting enhancement strategy to achieve performance comparable to or surpassing many fully-supervised methods with reduced annotation costs.
74. EPRBench: A High-Quality Benchmark Dataset for Event Stream Based Visual Place Recognition
Core Problem: There is a scarcity of dedicated high-quality benchmark datasets for event stream-based Visual Place Recognition (VPR), hindering the development of robust VPR solutions for challenging conditions (low illumination, overexposure, high-speed motion) and limiting the integration of LLMs.
Key Innovation: Introduces EPRBench, a high-quality benchmark dataset (10K event sequences, 65K frames) for event stream-based VPR, collected under diverse real-world conditions, including LLM-generated and human-refined scene descriptions. It also proposes a novel multi-modal fusion paradigm for VPR leveraging LLMs for interpretable and accurate place recognition.
75. Unleashing MLLMs on the Edge: A Unified Framework for Cross-Modal ReID via Adaptive SVD Distillation
Core Problem: Practical cloud-edge deployment of Cross-Modal Re-identification (CM-ReID) using Multi-Modal Large Language Models (MLLMs) is challenging due to the need for a fragmented ecosystem of specialized cloud models and the lack of effective knowledge distillation strategies for edge devices.
Key Innovation: Proposes MLLMEmbed-ReID, a unified cloud-edge framework that adapts a foundational MLLM into a state-of-the-art cloud model for unified cross-modal embedding (RGB, infrared, sketch, text) using instruction-based prompting and LoRA-SFT. For edge deployment, it introduces a novel adaptive SVD distillation strategy (Principal Component Mapping loss and Feature Relation loss) to create a lightweight, high-performing edge model.
76. Drift-Aware Variational Autoencoder-based Anomaly Detection with Two-level Ensembling
Core Problem: Identifying anomalies in vast amounts of unlabeled, streaming data is challenging, especially in nonstationary environments where concept drift degrades model performance.
Key Innovation: Introduces VAE++ESDD, a novel method combining incremental learning and two-level ensembling (VAEs for anomaly prediction and statistical drift detectors) to effectively detect anomalies in nonstationary streaming data with concept drift.
77. DynaGuide: A Generalizable Dynamic Guidance Framework for Unsupervised Semantic Segmentation
Core Problem: Existing unsupervised image segmentation methods struggle to balance global semantic structure with fine-grained boundary accuracy, especially in domains with scarce labeled data.
Key Innovation: Introduces DynaGuide, an adaptive framework combining global pseudo-labels from zero-shot models with local boundary refinement using a lightweight CNN, employing a dynamic multi-component loss for high-precision unsupervised semantic segmentation.
78. Learning Image-based Tree Crown Segmentation from Enhanced Lidar-based Pseudo-labels
Core Problem: Automatically segmenting individual tree crowns from aerial imagery is challenging due to factors like texture and overlaps, and manual annotation is costly for training deep learning models.
Key Innovation: Develops a method to train deep learning models for image-based tree crown segmentation using enhanced pseudo-labels derived from aerial laser scanning (ALS) data, specifically by enhancing them with a zero-shot instance segmentation model (SAM 2), achieving superior performance without manual annotation.
79. EXCODER: EXplainable Classification Of DiscretE time series Representations
Core Problem: Deep learning models for time series classification lack explainability, and existing XAI techniques are often hindered by the high dimensionality and noise in raw time series data.
Key Innovation: Investigating whether discrete latent representations (from VQ-VAE/DVAE) enhance explainability in time series classification by reducing redundancy, and proposing Similar Subsequence Accuracy (SSA) as a novel metric to quantitatively assess XAI-identified salient subsequences.
80. LongStream: Long-Sequence Streaming Autoregressive Visual Geometry
Core Problem: Existing autoregressive models for long-sequence streaming 3D reconstruction fail due to attention decay, scale drift, and extrapolation errors when processing thousands of frames.
Key Innovation: Introduces LongStream, a gauge-decoupled streaming visual geometry model that predicts keyframe-relative poses, uses orthogonal scale learning to disentangle geometry from scale, and employs cache-consistent training with periodic cache refresh to achieve stable, metric-scale reconstruction over kilometer-scale sequences.
81. LatentAM: Real-Time, Large-Scale Latent Gaussian Attention Mapping via Online Dictionary Learning
Core Problem: Building scalable latent feature maps from streaming RGB-D observations for open-vocabulary robotic perception is challenging, often requiring distillation of high-dimensional VLM embeddings using model-specific decoders.
Key Innovation: LatentAM, an online 3D Gaussian Splatting (3DGS) mapping framework that uses an online dictionary learning approach to build scalable latent feature maps from streaming RGB-D observations, enabling model-agnostic and pretraining-free integration with different VLMs at test time, achieving near-real-time speed and high feature reconstruction fidelity.
82. Annealing in variational inference mitigates mode collapse: A theoretical study on Gaussian mixtures
Core Problem: Mode collapse in variational inference, where models fail to capture all modes of a multimodal target distribution, specifically analyzed in the context of Gaussian mixtures.
Key Innovation: A mathematical analysis showing that an appropriately chosen annealing scheme can robustly prevent mode collapse in variational inference, providing a sharp formula for the probability of mode collapse and demonstrating its qualitative extension to neural network-based models.
83. TFTF: Training-Free Targeted Flow for Conditional Sampling
Core Problem: The challenge of conditional sampling in flow matching models, particularly the weight degeneracy issue of naive importance sampling in high-dimensional settings.
Key Innovation: A training-free conditional sampling method (TFTF) for flow matching models that uses importance sampling with a sequential Monte Carlo-based resampling technique and a stochastic flow with adjustable noise, providing asymptotic accuracy and outperforming existing methods on various conditional sampling tasks.
84. LTSM-Bundle: A Toolbox and Benchmark on Large Language Models for Time Series Forecasting
Core Problem: Training Large Time Series Models (LTSMs) for Time Series Forecasting (TSF) on heterogeneous data is challenging due to diverse frequencies, dimensions, and patterns, and various design choices aimed at enhancing LTSM capabilities are typically studied and evaluated in isolation.
Key Innovation: Introduces LTSM-Bundle, a comprehensive toolbox and benchmark that modularizes and benchmarks LTSMs across multiple dimensions, including pre-processing, model configurations, and dataset configurations. Demonstrates that combining the most effective design choices achieves superior zero-shot and few-shot performances compared to state-of-the-art LTSMs and traditional TSF methods.
85. Generating Physical Dynamics under Priors
Core Problem: Generating physically feasible dynamics in a data-driven context is challenging, as existing methodologies often overlook the integration of physical priors, leading to violations of basic physical laws and suboptimal performance.
Key Innovation: A novel framework that seamlessly incorporates physical priors (distributional and physical feasibility, including conservation laws and PDE constraints) into diffusion-based generative models, enabling the efficient generation of high-quality, physically realistic dynamics for various physical phenomena.
86. FISHER: A Foundation Model for Multi-Modal Industrial Signal Comprehensive Representation
Core Problem: Effectively analyzing heterogeneous multi-modal industrial signals (M5 problem) and detecting abnormal states is challenging, with previous works focusing on sub-problems and failing to utilize synergies or scaling laws.
Key Innovation: Proposes FISHER, a Foundation model for multi-modal Industrial Signal compreHEnsive Representation, which models M5 signals in a unified manner by considering STFT sub-bands as modeling units and using a teacher-student SSL framework, demonstrating versatile and outstanding capabilities in health management tasks.
87. Instruction-based Time Series Editing
Core Problem: Existing diffusion-based time series editors rely on rigid, predefined attribute vectors, limiting flexibility in condition format and control over editing strength.
Key Innovation: Introduces Instruction-based Time Series Editing and InstructTime, the first instruction-based time series editor that allows users to specify edits using natural language, learns a structured multi-modal representation space for controllable editing strength, and uses multi-resolution encoders for local and global edits, achieving high-quality and generalizable edits.
88. Bridging Generalization Gap of Heterogeneous Federated Clients Using Generative Models
Core Problem: Data heterogeneity among participants in Federated Learning (FL) often compromises performance and generalization, especially in scenarios with heterogeneous model architectures, as traditional methods are unsuitable.
Key Innovation: Proposes a model-heterogeneous FL framework where clients share feature distribution statistics to train variational transposed convolutional neural networks, which generate synthetic data to fine-tune local models, significantly improving generalization ability while reducing communication costs and memory consumption.
89. ATLAS: Adaptive Topology-based Learning at Scale for Homophilic and Heterophilic Graphs
Core Problem: Graph neural networks (GNNs) struggle with heterophilic graphs where connected nodes do not imply similarity, and iterative message passing limits scalability due to neighborhood expansion overhead.
Key Innovation: Introduced ATLAS (Adaptive Topology-based Learning at Scale), a propagation-free framework that encodes graph structure through multi-resolution community features, employing modularity-guided adaptive search to identify informative community scales for improved accuracy and scalability on both homophilic and heterophilic graphs.
90. MDAFNet: Multiscale Differential Edge and Adaptive Frequency Guided Network for Infrared Small Target Detection
Core Problem: Infrared small target detection (IRSTD) methods suffer from degradation of target edge pixels with increasing network layers and struggle to differentiate frequency components, leading to background interference and false detections from noise.
Key Innovation: MDAFNet, which integrates a Multi-Scale Differential Edge (MSDE) module to compensate for edge information loss and a Dual-Domain Adaptive Feature Enhancement (DAFE) module to adaptively enhance high-frequency targets and suppress high-frequency noise.
91. The Conditions of Physical Embodiment Enable Generalization and Care
Core Problem: Artificial agents struggle to generalize across distribution shifts and lack intrinsic motivation to preserve the well-being of others, particularly in open-ended physical environments like disaster response.
Key Innovation: The paper argues that generalization and care in AI agents arise from conditions of physical embodiment (being-in-the-world and being-towards-death), necessitating a homeostatic drive. It outlines a reinforcement-learning framework where homeostatic mortal agents continually learning in open-ended environments may offer efficient robustness and trustworthy alignment for tasks like disaster response.
92. Underwater side-scan sonar object detection with multi-dimensional feature enhancement and adaptive attention routing network
Core Problem: Underwater side-scan sonar (SSS) object detection struggles to balance accuracy and efficiency due to inherent imagery artifacts like diffused edges, low signal-to-noise ratio, multi-scale variations, and complex backgrounds, limiting perception for resource-constrained platforms.
Key Innovation: Proposes MEAR-EffiNet, a multi-scale edge-aware adaptive refinement network with a multi-scale edge refiner, cross-stage macro-micro routing attention, and a novel neck architecture, achieving state-of-the-art accuracy on challenging SSS datasets.
93. Machine learning-assisted modeling of nonlinear dispersive waves in ocean engineering
Core Problem: Accurately modeling nonlinear dispersive wave phenomena, such as long wave propagation, tidal bores, and internal waves, which are complex and crucial in ocean engineering.
Key Innovation: Introduction of a hybrid method combining Riccati sub-equation neural networks to derive both exact and approximate solutions for the (1+1)-dimensional modified Benjamin-Bona-Mahony equation, enabling the capture of nonlinear wave phenomena like solitary and periodic waves with high accuracy.
94. Experimental investigation and constitutive model on tensile properties of different steels considering seawater corrosion
Core Problem: Seawater corrosion significantly challenges the long-term safety and reliability of steel structures in marine engineering, necessitating a better understanding of material property evolution and selection.
Key Innovation: This study investigates the evolution laws of material properties (carbon steel, low alloy structural steel, stainless steel, weathering steel) during initial seawater corrosion through tests. It defines and compares mechanical reduction and pitting corrosion influence coefficients, and proposes and verifies two simplified constitutive models for corroded materials, providing a basis for reliability research of marine structures.
95. ForestScan: a unique multiscale dataset of tropical forest structure across 3 continents including terrestrial, UAV and airborne LiDAR and in-situ forest census data
Core Problem: There is a need for comprehensive, multiscale datasets of tropical forest structure (including various LiDAR types and in-situ data) to calibrate and validate Earth Observation (EO) estimates of forest biomass and provide broader insights into forest structure, especially for international initiatives like GEO-TREES.
Key Innovation: Presents the ForestScan project, a unique open-access multiscale dataset of tropical forest 3D structural measurements from three continents, integrating terrestrial, UAV, and airborne LiDAR with in-situ tree census data, along with detailed collection protocols and recommendations, serving as a critical resource for EO calibration/validation and broader forest science.
96. Experimental study to evaluate the effect of specimen size and bar diameter on the test results using SHPB
Core Problem: The influence of specimen size and bar diameter on the dynamic mechanical properties and axial inertia effects of rock-like materials in Split Hopkinson Pressure Bar (SHPB) tests is not fully understood.
Key Innovation: Conducted dynamic compression experiments on sandstone with varying specimen sizes and bar diameters, proposing a dimensionless stress difference concept. Found that the oscillation zone duration is primarily constrained by bar diameter and impact intensity, providing theoretical references for SHPB test design and optimizing impact-resistant rock engineering structures.
97. Changes in tensile failure characteristics of granite under combined microwave-liquid nitrogen treatment
Core Problem: Deep underground engineering requires efficient hard rock breaking technology, necessitating a better understanding of granite's structural damage and tensile failure properties under combined microwave and liquid nitrogen (LN2) thermal shock.
Key Innovation: Experimentally investigated granite's tensile failure under microwave-LN2 treatment, showing significant reductions in tensile strength and splitting modulus with increased microwave time and thermal shock cycles. Established a thermal-mechanical model to simulate cracking, providing new insights for increasing speed and reducing cost in deep hard formation drilling.
98. Probabilistic Prediction of Rock Mass Deformation Modulus Using Vine Copula-SVR Fusion Model
Core Problem: The need for a robust, data-driven approach to probabilistically predict and quantify uncertainty in rock mass deformation modulus (Em), which is essential for reliable rock engineering design and risk-informed decision-making.
Key Innovation: Proposed a Vine Copula-SVR fusion framework to quantify Em uncertainty using six input indicators. The Vine Copula captured joint dependency, and SVR characterized nonlinear mapping. Generated 10,000 synthetic samples for probabilistic prediction, showing a smooth Gaussian distribution for Em. Identified Depth and UCS as dominant influential indicators, enhancing risk-informed decision-making.
99. Nested adaptive Kriging-based reliability analysis and computational resource allocation for complex systems
Core Problem: Efficient reliability analysis of complex engineering systems faces significant challenges due to multiple subproblems, multidisciplinary coupling, high-dimensionality, and the computational burden of time-consuming simulations.
Key Innovation: Proposing a nested adaptive Kriging-based method for complex system reliability analysis, integrating system decomposition with adaptive surrogates, featuring a nested reliability-oriented acquisition function and a cost-effectiveness strategy for computational resource allocation.
100. Hazard Network Taxonomy: A Systemic Approach to Risk Analysis in Complex Sociotechnical Systems
Core Problem: Traditional point-based risk assessments in complex socio-technical systems often overlook the systemic role of hazards, their position in the network, and their potential to initiate or amplify cascades.
Key Innovation: Presents a novel Hazard Network Taxonomy (HNT) using directed, weighted graphs to represent hazards and their cause-and-effect interactions, developing quantitative indicators (Initiation Score, Amplification Score, Cascade Vulnerability Score) for systemic risk analysis.
101. RoofX-Net: A tailored approach to accurate multi-type rooftop segmentation in remote sensing images using edge and scale awareness
Core Problem: Existing image segmentation models struggle with accurate multi-type rooftop segmentation in high-resolution remote sensing images, particularly due to ambiguous edges and significant variations in rooftop scales, hindering effective solar PV capacity planning.
Key Innovation: Introduction of RoofX-Net, a novel decoder within an encoder–decoder framework, incorporating an Edge Extraction Module (using hand-crafted kernels) and a Scale Awareness Module (generating geometric awareness at different scales) to significantly improve precision in rooftop segmentation, achieving superior performance on dedicated datasets.
102. Minimizing sun glint impacts on Sentinel-2 <em>Sargassum</em> mapping
Core Problem: Sentinel-2 MultiSpectral Instrument (MSI) imagery used for Sargassum mapping is frequently impacted by strong sun glint, leading to data gaps or inconsistencies in estimated Sargassum density and reducing the reliability of monitoring efforts.
Key Innovation: Development of an empirical correction scheme that uses overlapping Sentinel-2A and Sentinel-2B pixels (where only one is glint-affected) to force glint-impacted Sargassum estimates to agree with minimal-glint estimates, based on surface roughness, substantially reducing glint-induced overestimation and improving quantification accuracy.
103. Mapping drawdown-zone bathymetry using SWOT observations: implications for global monitoring of lake inundation and storage changes
Core Problem: Existing approaches for high-precision mapping of lake drawdown topography are limited by insufficient spatial coverage, high costs, or inability to capture dynamic terrain exposure caused by water level fluctuations, and an effective method for integrating multi-temporal SWOT observations is lacking.
Key Innovation: Developed and validated the Spatial Iterative Filtering and Weighted Average Fusion (SIF-WAF) method, which leverages multi-temporal SWOT observations to generate drawdown topography with high completeness and accuracy (MAE < 1 m), enabling robust global monitoring of lake inundation and storage changes.
104. Detecting Water‐Surface Superelevation in Meandering Rivers Using Surface Water and Ocean Topography (SWOT) Satellite Data
Core Problem: Difficulty in obtaining accurate field measurements of curvature-induced water-surface superelevation in meandering rivers, which is crucial for understanding complex fluid motion patterns.
Key Innovation: Demonstrating that SWOT satellite data can reliably detect distinct water-surface superelevation in large river bends, providing a basis for global assessments of hydrodynamics in large meandering rivers.
105. Insertion Network for Image Sequence Correspondence
Core Problem: Precisely matching and localizing specific 2D slices within a 3D volume or determining anatomical coverage from 2D slices, which is difficult due to the independent treatment of slices in existing methods like body part regression.
Key Innovation: Proposes an 'insertion network' that learns to insert a slice from one sequence into another by encoding contextual representations and using a slice-to-slice attention mechanism, significantly reducing slice localization errors (from 8.4 mm to 5.4 mm) in medical imaging.
106. Vehicle behaviour estimation for abnormal event detection using distributed fiber optic sensing
Core Problem: Distributed fiber-optic sensing (DFOS) systems are effective for wide-area traffic monitoring but struggle to detect single-lane abnormalities (e.g., lane changes) that precede congestions.
Key Innovation: Presents a method to detect single-lane abnormalities by tracking individual vehicle paths and detecting lane changes along a road section using DFOS, estimating vehicle position and fitting paths with clustering, and monitoring changes in spectral centroid of vehicle vibrations, achieving 80% accuracy for lane change detection.
107. Efficient Personalized Federated PCA with Manifold Optimization for IoT Anomaly Detection
Core Problem: Current federated principal component analysis (PCA) methods for anomaly detection in distributed IoT environments lack the integration of personalization and robustness, which are critical for effective security.
Key Innovation: FedEP, an efficient personalized federated PCA method for anomaly detection in IoT networks, which achieves personalization through $\ell_1$-norm for element-wise sparsity and robustness via $\ell_{2,1}$-norm for row-wise sparsity, solved by a manifold optimization algorithm with theoretical convergence guarantees.
108. Multi-Task Learning with Additive U-Net for Image Denoising and Classification
Core Problem: Improving training stability and controlled information flow in U-Net architectures for single-task image denoising and multi-task learning (denoising-centric classification) while maintaining competitive reconstruction performance.
Key Innovation: Additive U-Net (AddUNet), which replaces concatenative skips with gated additive fusion, constraining shortcut capacity and stabilizing joint optimization, leading to competitive reconstruction, improved training stability, and implicit task decoupling through task-aware redistribution of learned skip weights in multi-task learning.
109. VimRAG: Navigating Massive Visual Context in Retrieval-Augmented Generation via Multimodal Memory Graph
Core Problem: Traditional Retrieval-augmented Generation (RAG) methods struggle with long-context tasks, especially those involving information-sparse yet token-heavy visual data in iterative reasoning scenarios, hindering effective multimodal information retrieval and understanding for agentic systems.
Key Innovation: Introduces VimRAG, a framework for multimodal Retrieval-augmented Reasoning across text, images, and videos. It models the reasoning process as a dynamic directed acyclic graph, uses a Graph-Modulated Visual Memory Encoding mechanism to dynamically allocate high-resolution tokens, and employs a Graph-Guided Policy Optimization strategy for fine-grained credit assignment, achieving state-of-the-art performance on multimodal RAG benchmarks.
110. Towards reconstructing experimental sparse-view X-ray CT data with diffusion models
Core Problem: It is unclear whether training data mismatch (domain shift) or forward model mismatch complicate the successful application of diffusion models to experimental sparse-view X-ray CT data, as most studies use synthetic data.
Key Innovation: Demonstrates that domain shift plays a nuanced role (diverse priors outperform well-matched but narrow ones) and forward model mismatch causes artifacts but can be mitigated. Emphasizes the need for validation against real-world benchmarks for performance gains to translate from synthetic to experimental data.
111. PixelRush: Ultra-Fast, Training-Free High-Resolution Image Generation via One-step Diffusion
Core Problem: Pre-trained diffusion models are limited by native training resolution, and existing training-free approaches for high-resolution image generation incur substantial computational overhead, taking minutes to produce a single 4K image.
Key Innovation: Presents PixelRush, a tuning-free framework for practical high-resolution text-to-image generation that enables efficient patch-based denoising within a low-step regime, using a seamless blending strategy and noise injection to mitigate artifacts and over-smoothing, achieving significant speedup (10x-35x) while maintaining visual fidelity.
112. X-VORTEX: Spatio-Temporal Contrastive Learning for Wake Vortex Trajectory Forecasting
Core Problem: Tracking and forecasting wake vortex trajectories from sparse LiDAR measurements is difficult due to fading signatures, atmospheric turbulence, and expensive point-wise annotation, with existing methods overlooking temporal structure.
Key Innovation: X-VORTEX, a spatio-temporal contrastive learning framework grounded in Augmentation Overlap Theory, which learns physics-aware representations from unlabeled LiDAR point cloud sequences to achieve superior vortex center localization and accurate trajectory forecasting with minimal labeled data.
113. Robustness of Object Detection of Autonomous Vehicles in Adverse Weather Conditions
Core Problem: Object detection models in autonomous vehicles lack robustness and reliability under adverse weather and lighting conditions, posing safety challenges for widespread adoption.
Key Innovation: Proposes a method using data augmentation to simulate progressive intensity levels of adverse conditions, enabling the evaluation and comparison of object detection model robustness (e.g., YOLOv5s, Faster R-CNN) by measuring average first failure coefficients (AFFC).
114. Detecting Object Tracking Failure via Sequential Hypothesis Testing
Core Problem: Real-time online object tracking systems lack formal safety assurances to reliably convey when tracking is failing, relying instead on heuristic measures of model confidence.
Key Innovation: Proposes interpreting object tracking as a sequential hypothesis test (e-process) to quickly identify tracking failures while provably containing false alerts, offering a computationally lightweight, model-agnostic, and statistically grounded mechanism for safety assurances.
115. Towards Universal Video MLLMs with Attribute-Structured and Quality-Verified Instructions
Core Problem: Existing video-instruction data for universal video understanding models is often incomplete, lacks fine-grained organization, and has unreliable annotations, limiting model performance.
Key Innovation: Introduces ASID-1M (a collection of one million structured, fine-grained audiovisual instruction annotations), ASID-Verify (a scalable data curation pipeline for automatic verification and refinement), and ASID-Captioner (a video understanding model trained on ASID-1M), demonstrating improved fine-grained caption quality and instruction following.
116. Multimodal Classification via Total Correlation Maximization
Core Problem: Existing multimodal learning methods often overfit certain modalities, leading to performance degradation, and lack an information-theoretic understanding of modality competition.
Key Innovation: Proposes TCMax, a hyperparameter-free loss function that maximizes total correlation between multimodal features and labels, alleviating modality competition and capturing inter-modal interactions, outperforming state-of-the-art methods.
117. Prior-Guided Symbolic Regression: Towards Scientific Consistency in Equation Discovery
Core Problem: Existing Symbolic Regression (SR) methods often produce equations that fit observations but are inconsistent with fundamental scientific principles (Pseudo-Equation Trap) due to a lack of explicit scientific constraints.
Key Innovation: Proposes PG-SR, a prior-guided SR framework that incorporates a prior constraint checker and a Prior Annealing Constrained Evaluation (PACE) mechanism to steer discovery towards scientifically consistent regions, reducing Rademacher complexity and outperforming baselines.
118. TCRL: Temporal-Coupled Adversarial Training for Robust Constrained Reinforcement Learning in Worst-Case Scenarios
Core Problem: Existing robust Constrained Reinforcement Learning (CRL) approaches primarily focus on single-step or temporally independent adversarial perturbations, failing to explicitly model robustness against more realistic temporally coupled perturbations in safety-critical domains.
Key Innovation: Proposes TCRL, a novel temporal-coupled adversarial training framework that introduces a worst-case-perceived cost constraint function to estimate safety costs under temporally coupled perturbations and establishes a dual-constraint defense mechanism on the reward, outperforming existing methods in robustness.
119. Efficient Plug-and-Play method for Dynamic Imaging Via Kalman Smoothing
Core Problem: Traditional Kalman smoothing (KS) methods lack expressivity due to not incorporating spatial prior information, and existing PnP-ADMM methods for dynamic imaging can be computationally inefficient for large numbers of timesteps.
Key Innovation: Proposal of PnP-KS-ADMM, an efficient Plug-and-Play algorithm that combines Kalman smoothing with deep learning denoisers, improving computational efficiency for 2D+t imaging problems compared to standard PnP-ADMM.
120. SIEFormer: Spectral-Interpretable and -Enhanced Transformer for Generalized Category Discovery
Core Problem: Enhancing feature adaptability and interpretability in Vision Transformers (ViT) for challenging Generalized Category Discovery (GCD) tasks, by better leveraging the attention mechanism.
Key Innovation: Introduction of SIEFormer (Spectral-Interpretable and -Enhanced Transformer), a novel ViT approach that uses spectral analysis to reinterpret and enhance the attention mechanism, composed of an implicit branch (graph Laplacians, Band-adaptive Filter) and an explicit branch (Maneuverable Filtering Layer using Fourier transform) for joint optimization, achieving state-of-the-art performance in image recognition.
121. Unified Multi-Domain Graph Pre-training for Homogeneous and Heterogeneous Graphs via Domain-Specific Expert Encoding
Core Problem: Existing graph pre-training methods are typically designed for either homogeneous or heterogeneous graphs, hindering unified modeling across diverse graph types and making them susceptible to distribution shifts between upstream and downstream tasks.
Key Innovation: Proposing GPH², a unified multi-domain graph pre-training method that uses a Unified Multi-View Graph Construction to encode both graph types and domain-specific expert encoding to capture domain knowledge and mitigate cross-domain discrepancies.
122. Which Algorithms Can Graph Neural Networks Learn?
Core Problem: Existing work on neural algorithmic reasoning with Graph Neural Networks (GNNs) is largely empirical or focuses solely on expressivity, lacking formal guarantees on when and how such architectures generalize beyond finite training sets.
Key Innovation: Proposing a general theoretical framework that characterizes sufficient conditions under which Message-Passing GNNs (MPNNs) can learn an algorithm from small instances and provably approximate its behavior on inputs of arbitrary size, along with impossibility results and more expressive MPNN-like architectures.
123. A Gradient Boosted Mixed-Model Machine Learning Framework for Vessel Speed in the U.S. Arctic
Core Problem: Understanding how environmental and operational conditions influence vessel speed in the Arctic is crucial for characterizing navigational conditions, but treating zero and positive speeds as a single continuous process can obscure important patterns.
Key Innovation: A two-stage machine learning framework using gradient boosted decision trees with random effects to model both the probability of vessel speed greater than zero and the speed conditional on being positive, integrating AIS data with various environmental and operational covariates to empirically characterize Arctic vessel operating regimes.
124. Grandes Modelos de Linguagem Multimodais (MLLMs): Da Teoria \`a Pr\'atica
Core Problem: The need for a comprehensive understanding of Multimodal Large Language Models (MLLMs), encompassing their theoretical fundamentals, practical application techniques, and current challenges and trends.
Key Innovation: A chapter presenting the main fundamentals of MLLMs, emblematic models, practical techniques for preprocessing, prompt engineering, and building multimodal pipelines with LangChain and LangGraph, along with a discussion of challenges and promising trends.
125. Variational Green's Functions for Volumetric PDEs
Core Problem: Green's functions, essential for tasks like shape analysis and physical simulation, remain computationally prohibitive to evaluate on arbitrary geometric discretizations, especially for resolving sharp singularities.
Key Innovation: Variational Green's Function (VGF), a method that learns a smooth, differentiable representation of Green's functions for linear self-adjoint PDE operators by decomposing them into an analytic free-space component and a learned corrector, enabling fast, differentiable evaluation and natural imposition of boundary conditions.
126. Represent Micro-Doppler Signature in Orders
Core Problem: The minimal distinctiveness of micro-Doppler signatures between similar indoor human activities and the large scale of time-frequency spectrograms, which creates challenges for model training and inference efficiency in through-the-wall radar (TWR).
Key Innovation: The Chebyshev-time map, a method that characterizes micro-Doppler signatures using polynomial orders by mapping time-frequency spectrum slices into a robust Chebyshev-time coefficient space, effectively compressing data scale while preserving morphological details for distinguishing human activities.
127. MaskInversion: Localized Embeddings via Optimization of Explainability Maps
Core Problem: Vision-language foundation models like CLIP excel at global vision-language alignment but show limitations in creating precise, context-aware representations for specific image regions.
Key Innovation: Proposes MaskInversion, a method that generates context-aware embeddings for a query image region specified by a mask at test time. It refines an embedding token by minimizing the discrepancy between its explainability map (derived from a frozen foundation model) and the query mask, enabling localized representations for tasks like open-vocabulary class retrieval and referring expression comprehension.
128. Post-hoc Probabilistic Vision-Language Models
Core Problem: Vision-language models (VLMs) deterministically map inputs, failing to capture uncertainties over concepts that arise from domain shifts, which is crucial for safety-critical applications.
Key Innovation: A post-hoc uncertainty estimation method for VLMs that does not require additional training, leveraging a Bayesian posterior approximation over the last layers to analytically quantify uncertainties over cosine similarities, leading to improved and well-calibrated predictive uncertainties.
129. Easy-Poly: An Easy Polyhedral Framework For 3D Multi-Object Tracking
Core Problem: Recent 3D multi-object tracking (3D MOT) methods often suffer from high false positives, missed detections, and identity switches, especially in crowded and small-object scenarios, hindering robust perception for autonomous driving.
Key Innovation: Easy-Poly, a filter-based 3D MOT framework with four key innovations: a novel Camera-LiDAR fusion detection method (CNMSMM), dynamic track-oriented (DTO) data association, dynamic motion modeling (DMM) using a confidence-weighted Kalman filter, and an extended life-cycle management system, achieving state-of-the-art performance in complex driving environments.
130. Leveraging Noisy Manual Labels as Useful Information: An Information Fusion Approach for Enhanced Variable Selection in Penalized Logistic Regression
Core Problem: In large-scale supervised learning, robust variable selection in penalized logistic regression is critical, but label noise from manual annotation is typically dismissed, and efficient, reproducible distributed learning for such scenarios is challenging.
Key Innovation: The paper theoretically demonstrates that noisy manual labels, intrinsically linked to classification difficulty, can enhance variable selection in PLR by refining non-zero coefficient estimation. It proposes a novel partition-insensitive parallel ADMM algorithm for efficient, globally convergent information fusion in large-scale distributed settings.
131. CNN and ViT Efficiency Study on Tiny ImageNet and DermaMNIST Datasets
Core Problem: There is a need to systematically evaluate the trade-offs between convolutional and transformer-based architectures for image classification, particularly concerning inference latency and model complexity, to identify efficient options for resource-constrained environments.
Key Innovation: The study demonstrates that appropriately fine-tuned Vision Transformers can match or exceed ResNet-18 performance, achieve faster inference, and operate with fewer parameters on both general and medical image classification benchmarks, highlighting their viability for deployment in resource-constrained settings.
132. Diffusion-Based Scenario Tree Generation for Multivariate Time Series Prediction and Multistage Stochastic Optimization
Core Problem: Efficient decision-making in uncertain systems requires estimating the full distribution of future scenarios, and existing methods for scenario tree generation may not provide superior representations of uncertainty for control tasks.
Key Innovation: Proposes Diffusion Scenario Tree (DST), a general framework that uses diffusion-based probabilistic forecasting models to recursively sample future trajectories and organize them into a non-anticipative tree via clustering, improving stochastic optimization performance in uncertain systems.
133. Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models
Core Problem: Inherent temporal heterogeneity (varying sampling densities, periodic structures) poses substantial challenges for zero-shot generalization in Time Series Foundation Models (TSFMs), leading to reliance on massive parameterization.
Key Innovation: Proposes Kairos, an adaptive and parameter-efficient TSFM that decouples temporal heterogeneity from model capacity using a dynamic patching tokenizer, a mixture-of-size encoding, and multi-granularity positional embedding based on dynamic rotary encodings, achieving superior zero-shot performance with fewer parameters.
134. VDW-GNNs: Vector diffusion wavelets for geometric graph neural networks
Core Problem: Developing effective geometric graph neural networks for analyzing data with complex geometric structures, particularly in tangent bundles, and ensuring desirable theoretical properties.
Key Innovation: Introduces Vector Diffusion Wavelets (VDWs) and VDW-GNNs, demonstrating their effectiveness on various data types (including point clouds and wind-field measurements) and proving their desirable frame theoretic and symmetry properties.
135. Active Learning with Task-Driven Representations for Messy Pools
Core Problem: Active learning in messy, uncurated data pools is hindered by the reliance of state-of-the-art methods on fixed, unsupervised representations, which often fail to capture task-relevant information.
Key Innovation: Proposes using task-driven representations that are periodically updated during the active learning process, demonstrating significant empirical performance improvement over unsupervised or pretrained representations for messy data pools.
136. Imitation Learning for Combinatorial Optimisation under Uncertainty
Core Problem: Approximating policies for large-scale combinatorial optimization problems formulated as sequential decision problems (SDPs) using imitation learning (IL) is challenging, and existing studies lack a unifying framework to characterize expert constructions, their computational properties, and impact on learning performance.
Key Innovation: Introduced a systematic taxonomy of experts for IL in combinatorial optimization under uncertainty, classified by treatment of uncertainty, level of optimality, and interaction mode, and proposed a generalized Dataset Aggregation (DAgger) algorithm supporting multiple expert queries and flexible interaction strategies.
137. R3DPA: Leveraging 3D Representation Alignment and RGB Pretrained Priors for LiDAR Scene Generation
Core Problem: Scarcity of 3D data for robotic tasks and limitations in leveraging image-pretrained priors and self-supervised 3D representations for LiDAR scene generation.
Key Innovation: R3DPA, a LiDAR scene generation method that aligns intermediate features with self-supervised 3D features, transfers knowledge from large-scale image-pretrained generative models, and enables point cloud control at inference.
138. SimpleMatch: A Simple and Strong Baseline for Semantic Correspondence
Core Problem: Existing semantic correspondence methods require high-resolution inputs, leading to computational overhead, and suffer from irreversible fusion of adjacent keypoint features due to deep downsampling.
Key Innovation: SimpleMatch, a framework that uses a lightweight upsample decoder to recover spatial detail at low resolutions and a multi-scale supervised loss, combined with sparse matching and window-based localization to optimize memory and achieve strong performance at low resolutions.
139. Bayesian Ego-graph Inference for Networked Multi-Agent Reinforcement Learning
Core Problem: Networked Multi-Agent Reinforcement Learning (MARL) methods struggle with adaptability to dynamic or heterogeneous environments due to assumptions of static communication neighborhoods or impractical reliance on centralized frameworks.
Key Innovation: BayesG, a decentralized actor-framework for Networked-MARL that learns sparse, context-aware interaction structures via Bayesian variational inference, enabling agents to jointly learn interaction topology and decision-making strategies for improved scalability and performance in dynamic environments.
140. Zero-Shot Enhancement With Cross-Modal Applicability for Low-Light Vis-$\mu$OCT Images
Core Problem: Low-light conditions in visible-light micro-OCT (vis-$\\mu$OCT) lead to decreased image brightness, limited penetration, and restricted application.
Key Innovation: Proposed Dif-NIR, a novel zero-shot framework for enhancing low-light OCT images using a neural implicit representation network with custom loss functions, demonstrating strong generalization and revealing deep-layer information.
141. Active Diffusion Matching: Score-Based Iterative Alignment of Cross-Modal Retinal Images
Core Problem: Accurately aligning Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs) is challenging due to substantial differences in viewing range and amorphous retinal appearance, with existing methods lacking accuracy.
Key Innovation: Proposed Active Diffusion Matching (ADM), a novel cross-modal alignment method integrating two interdependent score-based diffusion models to jointly estimate global transformations and local deformations, achieving state-of-the-art accuracy.
142. Generation of angular-normalized, cloud-filled, 0.01°-downscaled land surface temperature from 2018 to 2023 based on official FY-4A dataset
Core Problem: Official Land Surface Temperature (LST) datasets (e.g., FY-4A) are limited by unaddressed thermal radiation directionality effects, spatial discontinuities due to clouds, and coarser resolution compared to polar-orbiting satellites, hindering their utility for various environmental research.
Key Innovation: Proposes a novel framework to generate an angular-normalized, cloud-filled, and 0.01°-downscaled LST (ANCFDS-LST) product from FY-4A data, using a time-evolving kernel driven model for angular normalization, a CatBoost model for clear-sky prediction and cloud radiation force correction for cloudy-sky LST, and an improved hybrid downscaling algorithm, resulting in enhanced angular consistency, spatial continuity, and finer resolution.
143. The LULUCF Data Hub: translating global land use emissions estimates into the national GHG inventory framework
Core Problem: Discrepancies between global estimates of anthropogenic land use CO2 fluxes and national GHG inventories due to conceptual differences in definitions, hindering comparability and consensus on emissions estimates.
Key Innovation: Development of the 'LULUCF Data Hub,' an online platform that compiles and 'translates' various country-level CO2 emissions/removals datasets from LULUCF (GCB, GFW, NGHGI) to enhance comparability and foster consensus on anthropogenic land-use CO2 flux estimates.
144. Effect of Water-Deck Length on Blast-Induced Crack Propagation in Axial Decoupled Charges
Core Problem: Optimizing blast-induced crack propagation and fragmentation efficiency in axial decoupled charges, specifically investigating the effect of water-deck length, to improve precision-controlled blasting in deep mining and tunneling.
Key Innovation: Developed an in situ visualization method to study dynamic fracturing, showing that water-decks significantly enhance crack velocity and fragmentation efficiency, with dual-end water-deck achieving peak initial crack velocity and damage index. Numerical simulations confirmed water-deck induced stress concentration and improved energy transfer, increasing internal energy and blasting damage effect.
145. A Modular Digital Twin Architecture for Offshore Safety and Decision-Making: A North Sea Platform Case Study
Core Problem: Adoption of digital twin technology in offshore operations is constrained by inconsistent data, fragmented systems, and usability gaps, hindering improvements in safety and efficiency.
Key Innovation: A modular digital twin architecture that unifies data acquisition, edge analytics, twin processing, and visualization, standardizing data, enforcing checks, and using predictive models for traceable, explainable recommendations to improve offshore safety and decision-making.
146. Explainable urban renewal prediction at building-scale using hierarchical graph neural networks
Core Problem: Existing urban renewal prediction methods overlook complex spatial dependencies and lack explainability, hindering evidence-based urban planning.
Key Innovation: Proposes URHGN, an explainable hierarchical graph network, to model building and community-level spatial interactions for urban renewal prediction, achieving high accuracy and quantifying driving factors.
147. MMP-Mapper: Multi-modal priors enhancing vectorized HD road map construction from aerial imagery
Core Problem: Existing methods for HD road map extraction from aerial imagery process single image tiles, failing to leverage crucial spatial context and multi-modal semantic priors, leading to inaccuracies and discontinuities.
Key Innovation: Proposes MMP-Mapper, a novel framework using Contextual Image Fusion and Map-Guided Fusion modules to integrate spatial continuity from neighbor tiles and semantic/geometric priors from SD maps, significantly enhancing HD road map construction accuracy and generalization.
148. Grain size and Rare Earth Element analyses reveal the influence of aeolian-fluvial interactions on sand transport along the southern margin of the Taklimakan Desert
Core Problem: The role of fluvial processes and sediment provenance in aeolian-fluvial interactions and sand transport along desert margins, particularly the southern Taklimakan Desert, remains insufficiently understood.
Key Innovation: Integrated grain size, Rare Earth Element geochemistry, and remote sensing to clarify the dual role of rivers and the spatial heterogeneity of sediment sources, depositional structures, and transport mechanisms, deepening understanding of aeolian-fluvial dynamics and dune evolution.
149. Enhancing Bayesian model averaging ensemble model performance via feature partitioning and set pair analysis for freshwater-saltwater interface prediction
Core Problem: Traditional Bayesian Model Averaging (BMA) for freshwater-saltwater interface prediction suffers from static, empirically-derived or subjective prior weights, limiting accuracy under shifting hydrological conditions.
Key Innovation: Proposes a novel framework integrating Feature Partitioning (FP) and Set Pair Analysis (SPA) to generate dynamic, data-driven prior weights for BMA, significantly improving the reliability and transferability of freshwater-saltwater interface elevation (FSIE) prediction.
150. ASR-PINN: Adaptive step-size runge–kutta physics-informed neural network for multi-component reactive solute transport
Core Problem: Standard Physics-Informed Neural Networks (PINNs) struggle with complex, multi-component, two-dimensional reactive solute transport in heterogeneous porous media due to optimization bias, imbalanced loss terms, strong nonlinear reactions, and multiscale behavior, limiting their applicability to simplified scenarios.
Key Innovation: Developed ASR-PINN, an Adaptive Step-size Runge–Kutta Physics-Informed Neural Network, which integrates a high-order implicit Runge–Kutta scheme with two adaptive time-stepping strategies and a lightweight recursive mechanism. This significantly improves predictive accuracy and computational efficiency (65–95% error reduction, 2–8x speedup) for complex reactive transport problems.