TerraMosaic Daily Digest: Feb 2, 2026
Daily Summary
This digest synthesizes 247 selected papers and focuses on landslide process mechanics and slope evolution, flood generation, routing, and hydroclimatic forcing, seismic source-to-ground response pathways. Top-ranked studies examine earthquake-triggered slope response and liquefaction, mass-movement initiation and runout dynamics, and flood generation and hydroclimatic forcing.
Across the full set, evidence converges on mechanism-constrained analysis with operational relevance, especially for high-resolution remote-sensing monitoring workflows and infrastructure-focused hazard performance. The strongest contributions pair interpretable process evidence with monitoring or forecasting workflows that support warning design and risk prioritization.
Key Trends
- Landslide studies increasingly resolve process chains: Contributions connect triggering conditions, slope deformation, and mobility outcomes, improving the basis for warning thresholds and scenario testing.
- Flood analyses are becoming event-specific and process-based: Papers emphasize precipitation structure, antecedent wetness, and catchment controls rather than static hazard descriptors.
- Seismic hazard research links source behavior to ground response: Recurring topics connect rupture or loading conditions with geotechnical performance and consequence assessment.
- Monitoring workflows rely on integrated remote-sensing products: Multi-source satellite and airborne observations are used for deformation retrieval, change detection, and rapid post-event mapping.
- Infrastructure-facing outputs are increasingly decision-ready: Asset performance is evaluated with uncertainty-aware frameworks to support mitigation and maintenance prioritization.
Selected Papers
This digest features 247 selected papers from 1283 deduplicated papers analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Self-Arresting and Runaway Earthquakes:Nucleation, Propagation, Gutenberg-Richter law and Dragon-King Events
Core Problem: A unified physical framework is needed to explain the coexistence of self-arresting and runaway earthquake ruptures, and to link rupture physics to observed seismicity statistics like the Gutenberg-Richter law.
Key Innovation: A dissipation-based framework for earthquake rupture that explicitly separates the onset of unstable slip from self-sustained propagation, identifying distinct nucleation and propagation radii, explaining self-arresting and runaway events, deriving the Gutenberg-Richter law, and interpreting runaway ruptures as dragon-king events, providing a unified physical basis for earthquake dynamics.
2. SPIDER: Scalable Probabilistic Inference for Differential Earthquake Relocation
Core Problem: Traditional Bayesian inference methods for earthquake relocation do not scale well to large seismicity catalogs and high-dimensional parameter spaces, and neglect residual correlation between observations, biasing uncertainty estimates.
Key Innovation: Introduces SPIDER, a scalable Bayesian inference framework for double-difference hypocenter relocation that uses a physics-informed neural network Eikonal solver and Stochastic Gradient Langevin Dynamics to generate posterior samples for entire catalogs, designed to whiten residual correlation and provide robust uncertainty quantification.
3. A numerical study of the transition process from rock-ice avalanche to debris flow
Core Problem: The increasing frequency and impact of transitions from rock-ice avalanches to debris flows necessitate a better understanding of the underlying physical processes.
Key Innovation: Applied a three-phase depth-averaged model, incorporating particle size, water cohesion, and pore water saturation, solved with a second-order numerical method, to effectively capture and explain the transition process from rock-ice avalanches to debris flows, demonstrating how fluid viscosity and particle size influence the transformation rate.
4. Assessing residential buildings vulnerability on a regional scale to rainfall–earthquake-induced landslides on Mt. Umyeon, Korea
Core Problem: Limited research has been conducted on assessing regional vulnerability to landslides induced by the simultaneous occurrence of rainfall and seismic activity, despite the increasing threat to people and infrastructure.
Key Innovation: Introduces an innovative framework for evaluating regional vulnerability to rainfall-earthquake-induced landslides, combining machine learning and physical models for hazard analysis, Flow-R for runout propagation, and integrating physical vulnerability with quantitative approaches to compute a vulnerability index.
5. Integrating TRIGRS and RAMMS for the spatiotemporal prediction of rainfall induced landslides and landslide trajectory: a case study
Core Problem: Most existing studies on landslide susceptibility and debris-flow hazards address either initiation or propagation in isolation, lacking a holistic assessment of the entire hazard sequence.
Key Innovation: Integrates the TRIGRS model for landslide initiation and the RAMMS tool for debris-flow propagation, allowing for a holistic assessment under realistic rainfall scenarios, and uniquely incorporates field-measured saturated and unsaturated soil properties for quantitative validation of both initiation and runout predictions.
6. Hydro-gravitational dominance and differential landslide deformation unveiled by SBAS-InSAR-enhanced susceptibility mapping in Wushan, Three Gorges
Core Problem: Conventional landslide susceptibility maps in the Three Gorges region primarily rely on static inventories and slope-gradient analysis, potentially overlooking other crucial contributing factors and active deformation.
Key Innovation: Uses seven machine-learning models with eleven conditioning factors, finding elevation and monsoon rainfall as dominant factors, and incorporates five years of Sentinel-1A SBAS-InSAR data to reveal active deformation, resulting in susceptibility upgrades for 43% of pixels and reclassifying planned villages as very-high-hazard areas, demonstrating hydro-gravitational loading as a primary triggering mechanism.
7. Impact of seepage on the breaching of non-cohesive landslide dams with different grain size distributions
Core Problem: Most existing studies on landslide dam breaching focus solely on overtopping-induced mechanisms, neglecting the potential influence of pre-breaching seepage, which may alter dam erodibility, stability, and material composition.
Key Innovation: Through flume experiments, investigates breaching mechanisms of cohesionless landslide dams under coupled seepage-overtopping conditions, demonstrating that seepage significantly reduces downstream slope stability, increases material erodibility, shortens breaching duration, amplifies peak discharge, and exacerbates channel erosion, especially for smaller median particle sizes.
8. Seismic landslide susceptibility assessment based on the SCM-ANFIS model: A case study of the Wenchuan earthquake area
Core Problem: Current predictive models for earthquake-triggered landslide susceptibility often face limitations in accuracy (typically 80-90%), hindering effective emergency response and post-disaster reconstruction.
Key Innovation: Introduces a new hybrid machine learning framework, SCM-ANFIS, for seismic landslide susceptibility assessment, achieving a high prediction accuracy of 98.5% and correctly identifying 97.89% of historical landslides in the Wenchuan earthquake region, demonstrating superior performance and scalability.
9. Integrating engineering and non-engineering mitigation for secondary hazards at Mount Marapi: Sabo dams and irrigation intakes for debris flow control
Core Problem: Volcanic eruptions at Mount Marapi deposit material into rivers, reducing flow capacity and leading to deadly flash floods and debris flows, damaging infrastructure and irrigation, necessitating comprehensive management of post-eruption floods.
Key Innovation: Proposes a comprehensive approach integrating engineering (sabo dams, adapted irrigation intakes) and non-structural mitigation for secondary hazards like debris flows and flash floods resulting from volcanic eruptions, aiming to minimize sediment transport, enhance agricultural systems, and improve regional resilience.
10. Noise-resistant automatic seismic framework for monitoring rockslide slope
Core Problem: Denser local seismic networks for monitoring prone slopes near human settlements suffer from significant noise interference, making it difficult to reliably detect small-scale slope failures.
Key Innovation: Developed a noise-resistant automatic algorithm (detection, noise elimination, classification) for local seismic networks, effectively filtering diverse noise and purifying datasets before machine learning classification, achieving high recall rates for rockslides and earthquakes.
11. Structural fragility curves for brick-concrete buildings subjected to debris flow loading in northern China using momentum flux approach
Core Problem: Lack of reliable vulnerability assessment frameworks for brick-concrete buildings subjected to increasing debris flows in northern China.
Key Innovation: Develops momentum flux-based structural fragility curves for brick-concrete buildings, empirically estimating the model parameter β from field survey data to enhance predictive accuracy for debris flow disaster prevention and mitigation.
12. A Joint Probability Model for Multi-Hazard Intensity in Earthquake-Induced Rockfall Scenarios
Core Problem: Lack of quantitative load models for earthquake-induced rockfall multi-hazard scenarios limits the ability to evaluate structural reliability under such conditions.
Key Innovation: Proposes a methodology to establish a joint probability model for multi-hazard intensity in earthquake-induced rockfall scenarios, using numerical simulations and Copula theory to accurately capture the energy characteristics and provide a reference for risk assessment.
13. Deformation and early warning of reservoir landslide subjected to water level fluctuation and rainfall
Core Problem: Understanding the mechanisms of reservoir landslide evolution under water level fluctuations and rainfall, and developing an effective early warning model that can accurately capture step-like evolutionary characteristics.
Key Innovation: Designed a physical model experiment to investigate the complete evolutionary process and failure mechanisms of reservoir landslides, identifying hydrodynamic actions, fine particle loss, and bank reconstruction as main instability factors. Proposed a novel statistical early warning model that effectively addresses the issue of conventional methods failing to capture step-like landslide evolution.
14. Stacking ensemble learning-driven risk assessment framework of rockfall in karst terrains: A case study in Guilin, China
Core Problem: Insufficient research on rockfall hazard risk assessment in karst regions, which experience frequent rockfalls due to complex topography, geology, and human activities.
Key Innovation: Developed a stacking ensemble learning model (integrating RF, XGBoost, SVM, ANN) for high-precision rockfall susceptibility prediction, outperforming individual and other ensemble models. Constructed a comprehensive "hazard-vulnerability" assessment using AHP and enhanced model interpretability with SHAP value analysis.
15. Mechanisms and scenarios of the unprecedent flooding event in South Brazil 2024
Core Problem: The unprecedented May 2024 flooding in South Brazil, particularly impacting complex river–estuary–lagoon systems, requires understanding its mechanisms and assessing the effectiveness of potential hydraulic flood control interventions.
Key Innovation: The first detailed hydrodynamic assessment of the May 2024 South Brazil flood, using a validated model (with SWOT satellite observations) to investigate governing mechanisms (e.g., river peak synchronization) and assess scenarios for hydraulic flood control interventions, revealing their varying and often limited influence.
16. Landslide mapping from polarimetric SAR images using deep learning and morphological model
Core Problem: The need for rapid and accurate landslide mapping after triggering events, especially in cloud-prone mountainous regions where optical imagery is limited.
Key Innovation: Developed a deep learning framework with a morphological optimization model for precise landslide detection from single post-event polarimetric SAR images, demonstrating high accuracy (95.2% with ALOS-2 quad-pol) and applicability for operational disaster assessment.
17. Integrating Fourier Neural Operators with Diffusion Models to improve Spectral Representation of Synthetic Earthquake Ground Motion Response
Core Problem: Generating realistic synthetic earthquake ground motion for nuclear reactor building design is challenging due to complex earthquake physics, epistemic uncertainties, and prohibitive computational costs for model calibration, leading to frequency biases in current numerical tools.
Key Innovation: An AI physics-based approach combining a neural operator (approximating elastodynamics Green's operator) with a denoising diffusion probabilistic model to correct ground motion time series, significantly enhancing the realism of synthetic seismograms by mitigating mid-frequency spectral falloff and improving Goodness-Of-Fit scores.
18. A VAE Approach to Sample Multivariate Extremes
Core Problem: Generating accurate extreme samples from observational data is crucial for estimating risks associated with future extreme events (e.g., natural disasters), but standard generative machine learning approaches do not apply without adaptation, and existing extreme value theory (EVT) methods have limitations.
Key Innovation: A variational autoencoder (VAE) approach for sampling multivariate heavy-tailed distributions, which improves the learning of the dependency structure between extremes compared to standard VAE and competing EVT-based generative approaches, demonstrating potential for flood risk assessment on real data from the Danube river network.
19. Learning and extrapolating scale-invariant processes
Core Problem: It is challenging to effectively regress scale-free processes (e.g., earthquakes, avalanches) using machine learning, particularly for predicting large, rare events that require extrapolation capabilities, given the peculiar nature of scale invariance as a symmetry.
Key Innovation: Performs experiments on various architectures (U-net, Riesz network, wavelet-decomposition GNN, Fourier embedding, Fourier-Mellin Neural Operator) on paradigmatic self-similar problems, identifying main issues related to spectral biases and coarse-grained representations, and discussing how to alleviate them with relevant inductive biases for extrapolating rare events.
20. The potential triggering causes of dam collapses: insights into the rock mass weaknesses and given hydraulic or geophysical processes with projection to October 8, 2023, uphill lake dam collapse at Mbankolo, Cameroon
Core Problem: Understanding the potential triggering causes of dam collapses, specifically focusing on rock mass weaknesses and hydraulic/geophysical processes, exemplified by the Mbankolo dam collapse.
Key Innovation: Probes rock mass weaknesses using GSI and Hoek-Brown criterion, identifying that weathering of discontinuity surfaces can reduce the factor of safety, making the rock mass predisposed to failure, where low-magnitude seismic tremors or heavy rain become sufficient accelerating factors.
21. Assessing hurricane-induced flood risks in urban areas using Bayesian networks and GIS applications
Core Problem: Traditional indicator-based flood risk assessment approaches are limited in handling new data and uncertainty in dynamic urban environments, especially in the context of climate change and increasing impervious surfaces.
Key Innovation: Developed a Bayesian network (BN) model based on an established indicator-based framework, allowing for dynamic incorporation of new evidence and explicit probabilistic quantification of uncertainties, demonstrating 83.71% agreement with the indicator-based approach and identifying key impactful factors.
22. Influence of input motion uncertainty in developing slope-specific seismic fragility curves based on nonlinear finite element simulations
Core Problem: Input motion selection for nonlinear dynamic finite element (FE) analyses in seismic fragility assessment for geotechnical structures (slopes) has solely relied on the incremental dynamic analysis (IDA) method, potentially leading to uncertainties.
Key Innovation: Investigates uncertainties in seismic fragility curves of slopes arising from different input motion selections (CA, IDA, MSA) for nonlinear dynamic FE analysis, finding that an MSA-based method with logarithmically spaced PGA levels (Set 4) is more computationally efficient and yields curves stochastically closer to the CA-based method, offering a practical alternative to improve precision.
23. Mechanism of Progressive Failure in Surrounding Rock of Large-Diameter Subsea Shield Tunnels Crossing Fault Fracture Zones Under Seepage Effect
Core Problem: Large-diameter shield tunnels crossing water-rich fault fracture zones are highly vulnerable to progressive failure and water/mud inrush due to coupled effects of cutterhead disturbance, excavation unloading, and high-water pressure.
Key Innovation: Developed a large-scale fluid–solid coupling physical model to clarify the underlying failure mechanisms, revealing the formation and evolution of a neo-pressure arch, stage-dependent variations in seepage pressure and stress, and the role of particle migration and grain size redistribution in accelerating failure.
24. Graphical Assessment of Strain Burst Hazard in Hard-Rock Deep Underground Excavations to Improve Support Design
Core Problem: Despite existing methods, there is a necessity for further work to enhance the prediction and prevention of strain bursts and to effectively decide necessary support modifications in deep underground excavations.
Key Innovation: Proposed a new strain burst graph based on time-explicit numerical modeling, considering energy indicators and the IMASS failure criterion, to assess both the intensity and susceptibility of strain burst occurrence, providing a straightforward approach for hazard estimation and support optimization.
25. Role of Coal Properties and Surrounding Rock Stiffness in Coal Burst Triggering Under Superimposed Dynamic Loads During Deep Mining
Core Problem: Understanding the coupled material–structural influence on coal bursts in deep coal roadways, specifically the roles of coal bursting proneness, surrounding rock stiffness, and dynamic load disturbances in triggering these events.
Key Innovation: Systematically examined the roles of coal bursting proneness, surrounding rock stiffness, and dynamic load disturbances, proposing a material–structural coupling mechanism for roadway failure and identifying critical instability evaluation parameters under dynamic conditions, providing a theoretical basis for coal burst prevention.
26. An integrated approach to evaluating and prioritizing socio-physical flooding mitigation planning to enhance resilience in a community
Core Problem: Existing models for flood mitigation planning either separate physical and societal impacts or integrate social vulnerability superficially, failing to provide a detailed, integrated evaluation of mitigation strategies at the neighborhood level.
Key Innovation: Develops an integrated mathematical model for formulating and prioritizing flood mitigation strategies in the preplanning stage, considering both physical performance degradation and loss of residents' societal capabilities. It uses a fine-grained physical co-simulation and evaluates both physical and societal efficacy.
27. Dynamic evolutionary pathway analysis of urban rail transit flood risks and intelligent decision support based on knowledge graphs
Core Problem: Urban rail transit systems are highly vulnerable to rainstorm-induced flood disasters, and effectively extracting and utilizing accumulated domain-specific knowledge from historical events for risk identification and emergency management is challenging.
Key Innovation: Proposes a method integrating Knowledge Graph and Natural Language Processing (NLP) technologies to analyze risk evolution mechanisms and establish a knowledge-driven decision support model for urban rail transit flood risks, enhancing knowledge utilization and identifying potential risk paths.
28. Quantifying the Value of Seismic Structural Health Monitoring for post-earthquake recovery of electric power system in terms of resilience enhancement
Core Problem: Post-earthquake recovery of electric power networks is hindered by prolonged and imprecise manual inspections for damage diagnosis, and the system-level value of Seismic Structural Health Monitoring (SSHM) for enhancing resilience remains underexplored.
Key Innovation: Develops an integrated probabilistic simulation framework to quantify the system-level value of SSHM in enhancing EPN resilience by expediting recovery through improved damage awareness, demonstrating significant reduction in Lack of Resilience (LoR).
29. Mechanical degradation and microcrack evolution of weathered basalt under freeze–thaw cycles
Core Problem: Insufficient understanding of the degradation behavior of basalt with different weathering degrees under freeze-thaw (F-T) conditions, which impacts slope stability in cold regions.
Key Innovation: Investigated the mechanical degradation and microcrack evolution of weathered basalt under freeze-thaw cycles, demonstrating that moderately weathered basalt degrades more severely and developing an empirical model for strength degradation, providing practical support for rock slope engineering.
30. Diachronic Stereo Matching for Multi-Date Satellite Imagery
Core Problem: Existing stereoscopic 3D reconstruction pipelines fail for multi-date satellite imagery due to strong seasonal, illumination, and shadow changes that violate standard assumptions.
Key Innovation: The first Diachronic Stereo Matching method for satellite imagery, enabling reliable 3D reconstruction from temporally distant pairs by fine-tuning a deep stereo network with monocular depth priors on a curated dataset of diverse diachronic image pairs.
31. Under-Canopy Terrain Reconstruction in Dense Forests Using RGB Imaging and Neural 3D Reconstruction
Core Problem: Mapping terrain and understory beneath dense forest canopies is challenging, typically requiring expensive specialized sensors like airborne LiDAR or AOS, limiting accessibility and cost-effectiveness.
Key Innovation: Introduces a novel approach for canopy-free, photorealistic ground view reconstruction using conventional RGB images and Neural Radiance Fields (NeRF). It includes specific image capture considerations, a low light loss, and two complementary approaches to remove occluding canopy elements, demonstrating value for search and rescue and forest inventory tasks as a cost-effective alternative.
32. To See Far, Look Close: Evolutionary Forecasting for Long-term Time Series
Core Problem: The prevailing Direct Forecasting (DF) paradigm in Long-term Time Series Forecasting (LTSF) is computationally prohibitive for varying horizons and suffers from an optimization pathology where conflicting gradients from distant futures cripple the learning of local dynamics.
Key Innovation: Uncovers that models trained on short horizons, when coupled with the proposed Evolutionary Forecasting (EF) paradigm, significantly outperform those trained directly on long horizons, mitigating a fundamental optimization pathology and achieving robust asymptotic stability.
33. 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification
Core Problem: Efficiently modeling large 2D contexts in vision tasks (like Giga-Pixel WSIs and remote sensing) is challenging due to the quadratic complexity of Transformers and the computational inefficiency of existing 2D SSMs.
Key Innovation: 2DMamba, a novel 2D selective SSM framework that incorporates 2D spatial structure into Mamba with a highly optimized hardware-aware operator, achieving both spatial continuity and computational efficiency, improving performance on WSI classification, survival analysis, and natural imaging tasks, including semantic segmentation.
34. QuiZSF: A Retrieval-Augmented Framework for Zero-Shot Time Series Forecasting
Core Problem: Zero-shot time series forecasting is challenging due to the rapid emergence of new domains and scarce labeled history data, and existing methods lack dynamic external knowledge incorporation or are not extended beyond text.
Key Innovation: Presents QuiZSF, a retrieval-augmented forecasting framework that integrates search and forecasting for time series data by retrieving structurally similar sequences from a hierarchical tree-structured database (ChronoRAG Base), capturing multi-grained dependencies, and adapting retrieved knowledge to TSPMs, consistently outperforming baselines in zero-shot settings.
35. PYVALE: A Fast, Scalable, Open-Source 2D Digital Image Correlation (DIC) Engine Capable of Handling Gigapixel Images
Core Problem: Existing Digital Image Correlation (DIC) packages often have limitations such as operating-system restrictions, lack of support for computing clusters, and poor scalability to gigapixel-scale images common in Scanning Electron Microscopy DIC (SEM-DIC).
Key Innovation: Pyvale, a fast, scalable, open-source 2D DIC engine with a multithreaded, reliability-guided algorithm, capable of correlating gigapixel-scale image pairs in under 5 minutes on high-specification desktop workstations with metrological performance comparable to existing commercial and open-source codes, positioning it as a platform for large-scale deformation measurements.
36. Physics-Informed Neural Networks and Neural Operators for Parametric PDEs
Core Problem: Traditional numerical methods for solving parametric Partial Differential Equations (PDEs) require re-solving for each parameter, making parameter space exploration prohibitively expensive in science and engineering.
Key Innovation: Critically analyzes and unifies Physics-Informed Neural Networks (PINNs) and neural operators, demonstrating that neural operators can achieve computational speedups of 10^3 to 10^5 times faster than traditional solvers for multi-query scenarios while maintaining comparable accuracy, offering a comprehensive framework for parametric PDE solving.
37. Coastal erosion in Eastern Bali: impacts of the Indonesian throughflow
Core Problem: The contribution of large-scale oceanic currents, particularly the Indonesian Throughflow (ITF), to coastal erosion in Eastern Bali is poorly understood, despite its significant threat to ecosystems and communities.
Key Innovation: Assessed long-term shoreline changes using Landsat imagery and DSAS, integrated with reanalysis datasets, providing the first evidence of ITF-driven shoreline retreat in Bali and highlighting its strong link to a westward ITF branch, exacerbated by coastal infrastructure and steep bathymetry.
38. Multiple-hazard fragility assessment of bridges subjected to successive earthquake and flood events
Core Problem: Existing methodologies for bridge fragility assessment focus on single or sequential hazards without adequately considering cumulative damage from successive events.
Key Innovation: Introduces a novel framework for Damage State-Dependent Multiple-Hazard Fragility (DSMF) curves, incorporating the bridge's damage state after the first event into the fragility analysis for subsequent hazards, revealing significant differences in vulnerability based on hazard sequence and type.
39. To move or not to move? An investigation of the phenomenon of relocation of settlements in the lower Yellow River floodplain area
Core Problem: Settlements in the lower Yellow River floodplain face escalating safety and developmental risks due to flood-channel positioning and inadequate infrastructure, necessitating an understanding of relocation patterns and drivers.
Key Innovation: Examines 25 years of relocation patterns for 639 resettled settlements using spatial dynamics, gravity shift, autocorrelation, and geographic detector models, identifying policy dominance, community willingness, and the interaction between policy and economic development as key drivers, and revealing the formation of a linear resettlement belt beyond the floodplain.
40. Assessing the Temporal Qualities of Indirect Impacts of Flooding on Traffic
Core Problem: The temporal nature and indirect/cascading impacts of flooding on traffic, particularly with the addition of rainfall, are not fully understood or realistically modeled in flood impact assessments.
Key Innovation: Assessed the temporal qualities of indirect flood impacts on traffic using coupled HEC-RAS and SUMO models, proposing a penalty model to realistically simulate indirect impacts, and identifying significant increases in time delay and distances traveled under various flood and rainfall scenarios.
41. Beyond waterlogging: Evaluating the impact of extreme rainfall on the road network
Core Problem: Existing research on extreme rainfall impacts on transport networks primarily examines waterlogging, underexploring the contributions of reduced visibility and traffic-signal power outages at the network scale.
Key Innovation: Gauged the impacts of waterlogging, reduced visibility, and traffic-signal power outages on road network connectivity and efficiency during extreme rainfall, demonstrating that focusing solely on waterlogging significantly overestimates road capacities and underestimates network efficiency impacts.
42. Fusion of vibration velocity and excavation displacements for inversion of mechanical parameters of heterogeneous hillslope
Core Problem: Traditional displacement back analysis for hillslope mechanical parameters is costly and inefficient, often resulting in overly smooth estimations of elastic modulus when only surface displacement data is used.
Key Innovation: Proposed a method fusing monitored excavation displacements at the hillslope surface and vibration velocities after impact-loading to invert the spatial distribution of elastic modulus. Demonstrated that this fusion significantly improves the resolution of the estimated elastic modulus field and leads to accurate slope stability predictions, comparable to using borehole data.
43. Image Processing based Infrared Precursor Recognition for NPR Cable Anchored Rock with Inclination Angles
Core Problem: Ensuring safe operation of deep underground engineering by recognizing rock fracture precursors, especially in anchored rock specimens with varying inclination angles.
Key Innovation: Proposed a new index, Enhanced Infrared Matrix of Damage (EIMD), and a correlative index, Damage Infrared Energy Response (DIER), derived from noise-suppressed IR temperature matrices. Developed an integrated isolation forest model for anomaly detection, achieving early warning within 13–51 seconds (89%–99% peak stress), and demonstrated the stabilizing role of NPR cables in delaying infrared precursors.
44. Numerical analysis of roof layer presplitting in underground coal mines: An empty-hole controlled blasting technology
Core Problem: Conventional blasting in underground coal mines often results in disordered crack propagation, failing to guarantee penetrating presplitting surfaces needed to prevent rockbursts and strong seismic events.
Key Innovation: Proposed an innovative satellite empty-hole controlled blasting (SECB) technique, which uses empty-holes to reflect blast waves and promote longer, directional cracks. Numerical simulations showed SECB increased maximum crack length by 27.61% in the controlled direction, generated higher and more slowly decaying tensile stress, and identified optimal blasting parameters for effective crack control.
45. A hybrid physics–AI approach using universal differential equations with state-dependent neural networks for learnable, regionalizable, spatially distributed hydrological modeling
Core Problem: Traditional conceptual hydrological models face structural uncertainties and lack scale-relevant theories, while high-resolution models are crucial for predicting extreme events, necessitating spatially distributed hybrid modeling.
Key Innovation: A hybrid physics–AI framework embedding neural networks into a spatialized, regionalizable, and fully differentiable process-based hydrological model via universal differential equations (UDEs), which refines internal water fluxes and achieves more accurate streamflow simulations for flood modeling.
46. Is Training Necessary for Anomaly Detection?
Core Problem: Existing multi-class unsupervised anomaly detection (MUAD) methods rely on training reconstruction models, facing a fidelity-stability dilemma and requiring task-specific training.
Key Innovation: Retrieval-based Anomaly Detection (RAD), a training-free approach that stores anomaly-free features in memory and detects anomalies through multi-level retrieval, achieving state-of-the-art performance and overturning the assumption that MUAD requires task-specific training.
47. HierLoc: Hyperbolic Entity Embeddings for Hierarchical Visual Geolocation
Core Problem: Visual geolocalization remains challenging due to global scale, visual ambiguity, and the inherently hierarchical structure of geography, with existing paradigms being storage-intensive or ignoring geographic continuity.
Key Innovation: Introduces an entity-centric formulation of geolocation that replaces image-to-image retrieval with a compact hierarchy of geographic entities embedded in Hyperbolic space, reducing mean geodesic error by 19.5% and improving fine-grained accuracy by 43%.
48. FlowCalib: LiDAR-to-Vehicle Miscalibration Detection using Scene Flows
Core Problem: Angular misalignments of LiDAR sensors (LiDAR-to-vehicle miscalibration) can lead to safety-critical issues, but current methods primarily focus on correcting sensor-to-sensor errors without considering individual sensor miscalibration.
Key Innovation: Introduces FlowCalib, the first framework to detect LiDAR-to-vehicle miscalibration using motion cues from the scene flow of static objects, leveraging systematic bias in the flow field and integrating a neural scene flow prior with a dual-branch detection network.
49. Forward and Inverse Mantle Convection with Neural Operators
Core Problem: Thermal state reconstructions of the mantle are computationally expensive, and traditional methods struggle with observational noise in inverse problems, limiting understanding of mantle convection and its relation to seismic tomography and plate tectonics.
Key Innovation: Transforms numerical solvers into neural operators (Fourier Neural Operators) to significantly accelerate forward and inverse mantle convection modeling, demonstrating their ability to learn complex mappings and proposing a joint inversion technique that overcomes observational noise limitations for large-scale thermal state inversion problems.
50. Particle-Guided Diffusion Models for Partial Differential Equations
Core Problem: Existing generative methods for solving Partial Differential Equations (PDEs) often produce solution fields with higher numerical error and may not ensure generated samples remain physically admissible.
Key Innovation: Introduces a guided stochastic sampling method that augments sampling from diffusion models with physics-based guidance derived from PDE residuals and observational constraints, embedded within a Sequential Monte Carlo framework, yielding a scalable generative PDE solver with lower numerical error.
51. Decoupled Diffusion Sampling for Inverse Problems on Function Spaces
Core Problem: Existing plug-and-play diffusion posterior samplers for inverse PDE problems implicitly represent physics through joint coefficient-solution modeling, requiring substantial paired supervision and leading to over-smoothing.
Key Innovation: Proposes DDIS (Decoupled Diffusion Inverse Solver), a data-efficient, physics-aware generative framework in function space for inverse PDE problems. It uses a decoupled design where an unconditional diffusion learns the coefficient prior and a neural operator explicitly models the forward PDE for guidance, enabling superior data efficiency and avoiding guidance attenuation.
52. AI Decodes Historical Chinese Archives to Reveal Lost Climate History
Core Problem: Converting qualitative descriptions of climate events from historical archives into quantitative, high-resolution climate records has remained a fundamental challenge.
Key Innovation: Introduces a generative AI framework that infers quantitative climate patterns associated with documented historical events, producing a sub-annual precipitation reconstruction for southeastern China (1368-1911 AD), quantifying iconic extremes and mapping El Niño influence, with broader implications for climate and social sciences.
53. Comparing and Contrasting DLWP Backbones on Navier-Stokes and Atmospheric Dynamics
Core Problem: It is unclear which Deep Learning Weather Prediction (DLWP) architectures are most suitable for weather forecasting and future model development due to differences in training protocols, forecast horizons, and data choices.
Key Innovation: A detailed empirical analysis comparing prominent DLWP models and their backbones on synthetic Navier-Stokes and real-world global weather dynamics, identifying FNO for synthetic data, ConvLSTM/SwinTransformer for short-to-mid-range forecasts, and GraphCast/Spherical FNO for long-range stability and physical soundness.
54. Online Navigation Refinement: Achieving Lane-Level Guidance by Associating Standard-Definition and Online Perception Maps
Core Problem: Lane-level navigation relies on expansive global HD maps that cannot adapt to dynamic road conditions, while online perception maps lack global topology, making it hard to refine SD-map-based routes into accurate lane-level guidance.
Key Innovation: Introduces Online Navigation Refinement (ONR) and contributes the first ONR benchmark (OMA dataset) and MAT, a transformer with path-aware and spatial attention, to associate SD maps with OP maps, handling many-to-one lane-to-road mappings and severe misalignments, enabling low-cost, up-to-date lane-level navigation.
55. Citizen Science in Action: Building Trust in Deforestation Data Through Continuous Reliability Enhancements
Core Problem: Ensuring the reliability of data produced by citizen science projects for applications like deforestation detection, which depend on precise annotations and can be affected by image quality and campaign design.
Key Innovation: Demonstrated that continuous reliability enhancements in citizen science campaigns, such as using higher resolution Sentinel-2 imagery and a "transfer segmentation" approach, significantly improve volunteer classification accuracy for deforestation detection (from 57.77% to 93.33%), highlighting the value of iterative design and collaborative efforts.
56. Access to the Fagradalsfjall 2021–2023 volcanic eruptions: a case-study of pedestrian routes mapping with least-cost-path analyses
Core Problem: Identifying optimal and safe pedestrian routes for public access during dynamic volcanic eruptions, accommodating evolving lava flows and hazard assessments.
Key Innovation: Illustrates the potential of least-cost-path analyses (LCPAs) to quickly identify optimal routes and estimate hiking durations over dynamic volcanic terrain, accounting for lava flow evolution, hazard assessments, and terrain ruggedness, thereby enhancing safety and accessibility for visitors in crisis management.
57. Post-seismic turbidity response in the Goksu Stream after the 2023 Kahramanmaras earthquakes
Core Problem: Understanding the impacts of major seismic activity on water quality, specifically turbidity levels in critical drinking water sources like the Goksu Stream, is crucial for water security.
Key Innovation: Investigates post-seismic turbidity variations using satellite imagery (PlanetScope), orthophotos, and in situ measurements, demonstrating that turbidity levels significantly spiked after the Kahramanmaras earthquakes and that both seismic shaking and rainfall contribute to these spikes, highlighting the need for continuous monitoring.
58. Coupled hydromechanical elastoplastic framework to assess stress state and stability of unsaturated soil
Core Problem: The conventional local factor of safety (LFS) method for slope stability assessment in unsaturated soil leads to considerable errors at low matric suction conditions due to an inaccurate assumption about Mohr's circle for stress.
Key Innovation: Proposes a coupled hydromechanical elastoplastic framework to evaluate the stress state and stability of unsaturated soil, demonstrating that at small matric suction, the radius of the Mohr's circle for stress increases significantly (amplification effect) as it shifts leftward, improving accuracy compared to conventional LFS methods.
59. A Stratified Double-Difference Dynamic Tomography of 3D Velocity Field for Deep-Buried Large Underground Cavern
Core Problem: Time-lapse seismic exploration is impractical and conventional passive seismic inversion methods struggle to converge for obtaining dynamic 3D velocity models in deep-buried large underground caverns due to ill-posed sensor networks and unknown microseismic event parameters.
Key Innovation: Proposed a new stratified double-difference (SDD) tomography method for 3D dynamic velocity field inversion based on microseismic events, which solves the challenge of unknown spatial-temporal parameters and mitigates adverse effects from sensor/source distribution.
60. Dynamic Behavior and Energy Evolution of Coal with Varying Filling Materials Under Impact Loading
Core Problem: Evaluating the dynamic stability of coal and rock formations, particularly in roadways and backfilled boreholes, under impact loads, and optimizing filling materials to enhance stability.
Key Innovation: Demonstrated that filling materials significantly enhance dynamic strength, modify crack propagation, and inhibit cracking in coal, with polyurethane grouting being more effective than cement grout in reducing stress concentrations and improving impact resistance.
61. A maximum entropy moment quadrature method for quasi-real-time flight duration reliability assessment of UAVs
Core Problem: There is a need for efficient, quasi-real-time assessment of UAV flight duration reliability under multi-source uncertainties for critical applications like emergency response.
Key Innovation: Proposes an efficient reliability assessment framework for fixed-wing UAV flight duration, integrating a physics model with moment quadrature and maximum entropy methods to reconstruct the probability density function, enabling fast and accurate near real-time mission risk estimation with minimal computational demand, applicable to flood surveys and wildfire monitoring.
62. Physics-constrained digital twin framework for deformation analysis and safety assessment of high earth-rock dams
Core Problem: Conventional surrogate-assisted optimization for dam deformation analysis suffers from non-unique solutions, limited accuracy, and numerical divergence due to neglecting intrinsic physical parameter correlations.
Key Innovation: Proposes a physics-constrained digital twin framework utilizing a β-variational autoencoder to extract parameter correlations as prior knowledge, integrated with multi-objective optimization, to achieve high-fidelity, stable, and physically consistent deformation analysis for dam safety.
63. Spatial heterogeneity of soil functions in the gully vicinity within a watershed: coupling effects of landscape characteristics and ridge practices
Core Problem: Lack of systematic exploration into the coupling effects of landscape characteristics and field management on the spatial heterogeneity of soil functions, particularly in the vicinity of eroded gullies at the watershed scale.
Key Innovation: A composite sampling and modeling approach (mixed-effects and structural equation models) to evaluate five critical soil functions across watershed positions, orientations, and distances to gullies, revealing that distance from gullies is the dominant factor shaping soil function patterns and highlighting the significant role of eroded gullies and ridging practices.
64. Physics-based digital twin system for artificial ground freezing: implementation in Bangkok tunnel rehabilitation
Core Problem: Urban underground infrastructure projects, especially tunnel rehabilitation in complex geological and environmental constraints, demand high-precision and adaptive technologies beyond traditional numerical simulations.
Key Innovation: Developed and implemented the first physics-based digital twin (DT) system for artificial ground freezing (AGF), synergizing finite element modeling with real-time data assimilation to achieve accurate full-field predictions (MAE <0.6 °C) and optimize freezing parameters in complex urban geotechnical environments.
65. A Global Drought Dataset from Clustering-Based Event Identification with Integrated Population, and GDP Exposure and Socioeconomic Impacts
Core Problem: Precisely detecting drought events and linking physical drought indicators to socioeconomic consequences (population affected, economic losses) remains a key challenge.
Key Innovation: Introduced a robust two-step framework integrating clustering-based drought detection with impact analysis. Identified coherent drought events using SPI/SPEI at various timescales and linked them to high-resolution population and GDP exposure, as well as disaster impact data.
66. GeoDS (v.1.0): a simple Geographical DownScaling model for long-term precipitation data over complex terrains
Core Problem: Global climate models have insufficient spatial resolution for long-term precipitation data over complex terrains, and existing downscaling techniques have limitations for multi-millennia simulations.
Key Innovation: GeoDS (v.1.0), a simple topography-based model that downscales precipitation fields in complex areas for paleoclimate studies, using a topographic exposure index based on large-scale winds and terrain configuration, demonstrating robustness and computational efficiency.
67. Revealing the causes of groundwater level dynamics in seasonally frozen soil zones using interpretable deep learning models
Core Problem: Accurately characterizing and understanding the underlying causes of groundwater level dynamics in seasonally frozen soil regions is crucial for water resource management and ecosystem protection.
Key Innovation: An integrated framework of "simulation–classification–interpretation" using an LSTM model for groundwater level simulation and the Expected Gradients (EG) method to quantitatively identify dominant factors and mechanisms of different groundwater level variation types in seasonally frozen soil zones, revealing the regulatory role of frozen-thaw processes.
68. Multitask Learning for Earth Observation Data Classification with Hybrid Quantum Network
Core Problem: The increasing computational demands for analyzing large Earth observation (EO) data with complex deep learning models present a bottleneck, motivating the exploration of quantum computing for EO data classification.
Key Innovation: Presents a hybrid model incorporating multitask learning for efficient data encoding and a location weight module with quantum convolution operations to extract valid features for Earth observation data classification, demonstrating the potential of quantum machine learning in this domain.
69. Knowledge-Informed Kernel State Reconstruction for Interpretable Dynamical System Discovery
Core Problem: Existing methods for recovering governing equations from data often fail under noisy/partial observations or rely on black-box latent dynamics, hindering interpretable scientific discovery.
Key Innovation: Introduces MAAT (Model Aware Approximation of Trajectories), a framework for symbolic discovery built on knowledge-informed Kernel State Reconstruction. It formulates state reconstruction in a reproducing kernel Hilbert space, directly incorporating structural and semantic priors (e.g., conservation laws) to yield smooth, physically consistent state estimates with analytic time derivatives, improving downstream symbolic regression.
70. PromptMAD: Cross-Modal Prompting for Multi-Class Visual Anomaly Localization
Core Problem: Visual anomaly detection in multi-class settings poses significant challenges due to the diversity of object categories, scarcity of anomalous examples, and presence of camouflaged defects.
Key Innovation: PromptMAD, a cross-modal prompting framework for unsupervised multi-class visual anomaly detection and localization that integrates semantic guidance via vision-language alignment using CLIP-encoded text prompts, enriching visual reconstruction and achieving state-of-the-art pixel-level performance.
71. Benchmarking Long Roll-outs of Auto-regressive Neural Operators for the Compressible Navier-Stokes Equations with Conserved Quantity Correction
Core Problem: Deep learning solutions for numerical approximation of PDE solutions struggle to perform well over long prediction durations due to the accumulation of auto-regressive error and the inability of models to conserve physical quantities.
Key Innovation: Presents 'conserved quantity correction', a model-agnostic technique for incorporating physical conservation criteria within deep learning models. This demonstrates consistent improvement in the long-term stability of auto-regressive neural operator models, regardless of architecture.
72. Cross-Domain Few-Shot Learning for Hyperspectral Image Classification Based on Mixup Foundation Model
Core Problem: Cross-domain few-shot learning (CDFSL) for hyperspectral image (HSI) classification often relies on unrealistic data augmentation, involves many parameters prone to overfitting, and has not explored the strength of foundation models for generalization.
Key Innovation: MIFOMO (MIxup FOundation MOdel), built upon a remote sensing (RS) foundation model for generalizable features. It introduces coalescent projection (CP) for quick adaptation and mixup domain adaptation (MDM) to address extreme domain discrepancy, along with label smoothing, achieving significant improvements.
73. FOTBCD: A Large-Scale Building Change Detection Benchmark from French Orthophotos and Topographic Data
Core Problem: Existing building change detection benchmarks are geographically constrained to single cities or limited regions, which limits large-scale benchmarking and evaluation under geographic domain shift.
Key Innovation: FOTBCD, a large-scale building change detection dataset derived from authoritative French orthophotos and topographic data, spanning 28 departments across mainland France. It provides FOTBCD-Binary (28,000 image pairs with pixel-wise masks) and FOTBCD-Instances, demonstrating improved cross-domain generalization.
74. Full-Graph vs. Mini-Batch Training: Comprehensive Analysis from a Batch Size and Fan-Out Size Perspective
Core Problem: Choosing between full-graph and mini-batch Graph Neural Network (GNN) training approaches is challenging due to distinct system design demands and an insufficient understanding of the impact of batch and fan-out sizes on performance and efficiency.
Key Innovation: This paper systematically compares full-graph vs. mini-batch GNN training through empirical and theoretical analyses, providing a novel generalization analysis using Wasserstein distance. It uncovers non-isotropic effects of batch and fan-out sizes on GNN convergence and generalization, offering practical guidance for hyperparameter tuning and showing that full-graph training is not always superior.
75. OOVDet: Low-Density Prior Learning for Zero-Shot Out-of-Vocabulary Object Detection
Core Problem: Zero-shot out-of-vocabulary detection (ZS-OOVD) methods tend to overfit in-vocabulary (IV) classes, causing out-of-vocabulary (OOV) or undefined classes to be misclassified with high confidence.
Key Innovation: OOVDet, a novel zero-shot OOV detector, addresses this by synthesizing region-level OOV prompts from low-likelihood regions of class-conditional Gaussian distributions and mining pseudo-OOV image samples using a Dirichlet-based gradient attribution mechanism. This constructs an OOV decision boundary through a low-density prior constraint, significantly improving OOV detection performance.
76. DAVIS: OOD Detection via Dominant Activations and Variance for Increased Separation
Core Problem: Most post-hoc Out-of-Distribution (OOD) detection methods operate on penultimate feature representations derived from global average pooling (GAP), which discards valuable distributional statistics from activation maps.
Key Innovation: DAVIS, a simple and broadly applicable post-hoc technique, enriches feature vectors by incorporating channel-wise variance and dominant (maximum) activations, directly addressing information loss from GAP. It sets a new benchmark across diverse architectures, achieving significant reductions in false positive rate (e.g., 48.26% on CIFAR-10 using ResNet-18).
77. Discovering Scaling Exponents with Physics-Informed M\"untz-Sz\'asz Networks
Core Problem: Standard neural networks fail to explicitly capture and identify governing scaling exponents in physical systems exhibiting power-law scaling near singularities, interfaces, and critical points.
Key Innovation: Physics-informed M"untz-Sz'asz Networks (MSN-PINN), a power-law basis network that treats scaling exponents as trainable parameters, enabling the explicit discovery and unique recovery of these exponents with high accuracy, and producing learned parameters with direct physical meaning.
78. Synthetic Abundance Maps for Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images
Core Problem: Hyperspectral single image super-resolution (HS-SISR) methods are mostly supervised, requiring ground truth data that is often unavailable in practice.
Key Innovation: A novel unsupervised training framework for HS-SISR based on synthetic abundance data, where a neural network is trained using synthetic abundance maps (generated from a dead leaves model) to perform abundance super-resolution, then used to reconstruct a super-resolved hyperspectral image.
79. FarmMind: Reasoning-Query-Driven Dynamic Segmentation for Farmland Remote Sensing Images
Core Problem: Farmland remote sensing image segmentation methods are limited by a static paradigm, relying on single input patches and lacking reasoning capability for ambiguous scenes, unlike human experts who query auxiliary data.
Key Innovation: FarmMind, a reasoning-query-driven dynamic segmentation framework for FRSIs, which introduces a mechanism to dynamically and on-demand query external auxiliary images (e.g., higher-resolution, multi-temporal) to compensate for insufficient information and resolve segmentation ambiguities, mimicking human expert reasoning.
80. How Much of a Model Do We Need? Redundancy and Slimmability in Remote Sensing Foundation Models
Core Problem: The direct transfer of scaling assumptions from computer vision (CV) to remote sensing (RS) foundation models (FMs) has not been adequately examined, leading to a hypothesis that RS FMs enter an overparameterized regime at smaller scales, inducing redundant representations.
Key Innovation: Uses post-hoc slimming as a diagnostic tool to measure representational redundancy across six state-of-the-art RS FMs, empirically supporting the hypothesis that RS FMs are highly redundant compared to CV FMs. It also demonstrates that learned slimmable training can improve models and provides mechanistic explanations for high redundancy, establishing slimmability as a practical deployment strategy and diagnostic tool.
81. When Anomalies Depend on Context: Learning Conditional Compatibility for Anomaly Detection
Core Problem: Anomaly detection often assumes abnormality is intrinsic, but in many real-world settings, it's context-dependent (contextual anomaly detection), which is underexplored in the visual domain.
Key Innovation: Introduces CAAD-3K, a benchmark for contextual anomalies, and proposes a conditional compatibility learning framework leveraging vision-language representations to model subject-context relationships under limited supervision. This method substantially outperforms existing approaches on CAAD-3K and achieves state-of-the-art on MVTec-AD and VisA, demonstrating the value of modeling context dependence.
82. Multi-Cue Anomaly Detection and Localization under Data Contamination
Core Problem: Visual anomaly detection in industrial settings faces two major limitations: methods are often trained on purely normal data (assuming no contamination, which is rare in practice) and assume no access to labeled anomaly samples, leading to poor performance in distinguishing anomalies.
Key Innovation: Proposes a robust anomaly detection framework that integrates limited anomaly supervision into adaptive deviation learning. It introduces a composite anomaly score combining deviation, entropy-based uncertainty, and segmentation-based scores for accurate detection and gradient-based localization. It also incorporates a small set of labeled anomalies while mitigating contaminated samples through adaptive instance weighting, outperforming state-of-the-art baselines under various contamination levels.
83. Adaptive Benign Overfitting (ABO): Overparameterized RLS for Online Learning in Non-stationary Time-series
Core Problem: Extending the recursive least-squares (RLS) framework to the overparameterized regime for online learning in non-stationary time-series while ensuring numerical stability and adaptability.
Key Innovation: Introduces Adaptive Benign Overfitting (ABO), a QR-based exponentially weighted RLS algorithm that combines random Fourier feature mappings with forgetting-factor regularization, enabling online adaptation under non-stationary conditions, maintaining bounded residuals, and achieving high accuracy and speed in forecasting.
84. Learning Reward Functions for Cooperative Resilience in Multi-Agent Systems
Core Problem: Multi-agent systems in dynamic and uncertain environments need to pursue individual goals while safeguarding collective functionality, with cooperative resilience being a critical yet underexplored property in Multi-Agent Reinforcement Learning.
Key Innovation: Introduces a novel framework that learns reward functions from ranked trajectories, guided by a cooperative resilience metric, for multi-agent systems. It demonstrates that a hybrid reward strategy significantly improves robustness under disruptions without degrading task performance and reduces catastrophic outcomes.
85. Nested Slice Sampling: Vectorized Nested Sampling for GPU-Accelerated Inference
Core Problem: Scalable inference for model comparison and calibrated uncertainty quantification is challenging for complex, multimodal targets, and traditional Nested Sampling methods are difficult to implement efficiently on accelerators due to their sequential nature and hard constraints.
Key Innovation: Nested Slice Sampling (NSS), a GPU-friendly, vectorized formulation of Nested Sampling that uses Hit-and-Run Slice Sampling for constrained updates, enabling efficient parallel execution, improving high-dimensional behavior, and demonstrating robustness on challenging multimodal Bayesian inference problems.
86. On The Relationship Between Continual Learning and Long-Tailed Recognition
Core Problem: Real-world datasets, including those in geohazards, often exhibit long-tailed distributions where "Tail" classes are severely underrepresented, leading to biased learning and poor generalization for these critical, often rare, events.
Key Innovation: A theoretical framework revealing a connection between Long-Tailed Recognition (LTR) and Continual Learning (CL), leading to CLTR, a principled approach that uses standard CL methods to sequentially learn Head and Tail classes without forgetting, mitigating gradient saturation and improving Tail learning while maintaining Head performance on imbalanced datasets.
87. TorchCP: A Python Library for Conformal Prediction
Core Problem: While Conformal Prediction (CP) offers guaranteed coverage for prediction intervals, existing CP libraries often lack the model support and scalability required for integrating state-of-the-art CP algorithms with large-scale deep learning models used in various applications.
Key Innovation: TorchCP, a PyTorch-native library that integrates state-of-the-art Conformal Prediction algorithms into deep learning techniques (DNNs, GNNs, LLMs), offering CP-specific training, online prediction, GPU-accelerated batch processing (up to 90% inference time reduction), and full scalability for enhanced uncertainty quantification.
88. A Library for Learning Neural Operators
Core Problem: A lack of a high-quality, tested, open-source Python library for operator learning that combines cutting-edge models with ease of use and customizability, hindering the development and deployment of neural operators.
Key Innovation: NeuralOperator, an open-source Python library for operator learning, providing tools for training, deploying, and developing neural operator models that generalize neural networks to function spaces and handle various discretizations, satisfying discretization convergence properties.
89. SuperPoint-SLAM3: Augmenting ORB-SLAM3 with Deep Features, Adaptive NMS, and Learning-Based Loop Closure
Core Problem: The widely adopted ORB-SLAM3 falters in extreme viewpoint, scale, and illumination variations due to its reliance on hand-crafted ORB keypoints.
Key Innovation: Introduces SuperPoint-SLAM3, a drop-in upgrade that replaces ORB with the self-supervised SuperPoint detector-descriptor, enforces spatially uniform keypoints via adaptive non-maximal suppression, and integrates a lightweight NetVLAD place-recognition head for learning-based loop closure, significantly improving accuracy while preserving real-time operation.
90. YOLO26: Key Architectural Enhancements and Performance Benchmarking for Real-Time Object Detection
Core Problem: The continuous need for more efficient, accurate, and deployable real-time object detection models, especially for edge and low-power devices.
Key Innovation: YOLO26, a new YOLO variant with architectural enhancements (e.g., DFL removal, NMS-free inference, ProgLoss, STAL, MuSGD) and multi-task capabilities, demonstrating superior efficiency and accuracy on edge devices compared to previous YOLO versions and transformer-based detectors.
91. Joint Learning of Depth, Pose, and Local Radiance Field for Large Scale Monocular 3D Reconstruction
Core Problem: Photorealistic 3D reconstruction from monocular video collapses in large-scale scenes when depth, pose, and radiance are solved in isolation, leading to scale-ambiguous depth, long-horizon pose drift, and inability to model large content with a single global NeRF.
Key Innovation: A joint learning framework that couples a Vision-Transformer (ViT) depth network with metric-scale supervision, a multi-scale feature bundle-adjustment (BA) layer for pose refinement, and an incremental local-radiance-field hierarchy (hash-grid NeRFs) for scene representation, achieving metric-scale, drift-free 3D reconstruction and high-fidelity novel-view synthesis from a single uncalibrated RGB camera in city-block-scale scenes.
92. Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations
Core Problem: Most existing time series generation (TSG) methods assume regular sampling and fixed output resolutions, which are often violated in practice where observations are irregular and sparse, while downstream applications require continuous and high-resolution time series.
Key Innovation: Diff-MN, a continuous TSG framework that enhances Neural Controlled Differential Equation (NCDE) with a Mixture-of-Experts (MoE) dynamics function and a decoupled architectural design, employing a diffusion model to parameterize the NCDE temporal dynamics parameters, consistently outperforming baselines on irregular-to-regular and irregular-to-continuous TSG tasks.
93. CaLiV: LiDAR-to-Vehicle Calibration of Arbitrary Sensor Setups
Core Problem: Existing LiDAR calibration methods for autonomous systems often require overlapping fields of view, external sensing devices, or feature-rich environments, and most do not support Sensor-to-Vehicle calibration, limiting their applicability for arbitrary multi-LiDAR setups.
Key Innovation: Introduction of CaLiV, a novel target-based technique for extrinsic Sensor-to-Sensor and Sensor-to-Vehicle calibration of multi-LiDAR systems that works for non-overlapping fields of view without external sensing, using motion-induced overlaps, Kalman filtering, and GMM-based registration to accurately solve for both translational and rotational sensor extrinsics.
94. M-SGWR: Multiscale Similarity and Geographically Weighted Regression
Core Problem: Traditional local regression models (GWR, MGWR) quantify spatial relationships solely through geographic proximity, which may be insufficient to capture how locations are interconnected, as different phenomena exhibit distinct spatial patterns.
Key Innovation: Proposes M-SGWR, a new multiscale local regression framework that characterizes spatial interaction across two dimensions: geographic proximity and attribute (variable) similarity, combining separate geographic and attribute-based weight matrices with an optimized parameter (alpha) for each predictor, consistently outperforming existing GWR variants.
95. Beyond Water: Mapping Sediment Bars to Enhance Satellite Monitoring of River Dynamics
Core Problem: Enhancing satellite monitoring of river dynamics by accurately mapping not just water, but also sediment bars, to track sustained river changes.
Key Innovation: An automated approach using Sentinel-2 imagery to classify river water and sediment bars, enabling the separation and tracking of active channel lateral mobility.
96. Insights on the Working Principles of a CNN for Forest Height Regression From Single-Pass InSAR Data
Core Problem: Understanding the "black-box" nature of AI models, specifically CNNs, used for Earth Observation applications like forest height estimation from InSAR data, to build trust and improve reliability.
Key Innovation: Proposed a multifaceted XAI approach for a CNN-based model estimating forest height from TanDEM-X InSAR data, revealing that the model implicitly compensates for SAR acquisition geometry distortions and that mean phase center height, local variability, interferometric coherence, and backscatter maps are key informative predictors.
97. A Deep Learning-Based Model for Forest Canopy Height Mapping Using Multisource Remote Sensing Data
Core Problem: Accurately assessing forest carbon storage requires precise forest canopy height, but existing methods struggle with complex nonlinear relationships and adaptive feature extraction from multisource remote sensing data.
Key Innovation: Proposed a deep learning-based neural network-guided interpolation (NNGI) model that fuses GEDI LiDAR data with multisource remote sensing features, using a dual-network architecture to dynamically learn interpolation weights based on environmental similarity, achieving significantly higher accuracy in forest canopy height mapping across complex environments.
98. Controllable Reference-Guided Diffusion With Local–Global Fusion for Real-World Remote Sensing Super Resolution
Core Problem: Existing reference-based super-resolution (RefSR) methods for remote sensing images struggle with cross-sensor resolution gaps and land-cover changes, leading to undergeneration or overreliance on the reference image.
Key Innovation: Proposed CRefDiff, a controllable reference-guided diffusion model for real-world remote sensing image SR, which uses a powerful generative prior, a dual-branch local-global fusion mechanism, and an accelerated inference strategy, achieving state-of-the-art performance and improving downstream tasks like scene classification and semantic segmentation.
99. SFCFNet: A Spatial–Frequency Cross-Attention Fusion Network for Hyperspectral Image Classification
Core Problem: Most existing HSI classification methods primarily focus on spectral-spatial feature fusion, neglecting the complementary information contained in frequency-domain features, which limits classification performance.
Key Innovation: Proposed SFCFNet, a spatial–frequency cross-attention fusion network, that jointly models spectral, spatial, and frequency-domain features using multiscale convolutions, triple-branch representation (global, local, multiscale frequency), and dual-domain cross-attention fusion, achieving higher overall accuracy on multiple HSI datasets.
100. Sensitive Band Selection Method for Hyperspectral Images of Pine Wilt Disease Infected Trees Based on Adaptive Subspace Partitioning
Core Problem: Early detection of Pine Wilt Disease (PWD) infected trees is challenging due to subtle physiological alterations not visible in conventional imagery, and existing methods struggle to adaptively identify pathology-sensitive spectral bands while reducing data redundancy.
Key Innovation: Developed PwdBSNet, a novel hyperspectral band selection model for PWD-infected trees, which uses an adaptive subspace partitioning (ASP) mechanism to identify pathology-sensitive bands and a spectral reconstruction network (CNN–ViT with multiattention) to enhance spectral feature learning, significantly outperforming mainstream methods in early PWD detection.
101. Investigation of the land and atmospheric processes associated with the heavy rainfall events over the complex terrains of the Northwestern Himalayan States of India
Core Problem: Numerical Weather Prediction (NWP) models fail to capture Heavy Rainfall Events (HREs) with adequate lead-time over complex Himalayan terrains, hindering disaster preparedness.
Key Innovation: Utilizes WRF model with high-resolution land data assimilation (LDAS) to better represent soil moisture and evapotranspiration, leading to elevated near-surface relative humidity, increased moisture residence time, and stronger convective updrafts, thereby enhancing the accuracy of rainfall prediction.
102. Assessment of precipitation extremes and risk factors in the Himalayan foothills: a machine learning approach for hydro-meteorological hazard analysis
Core Problem: Understanding precipitation patterns and the influence of physiographic characteristics on precipitation variability in topographically contrasting districts of the Himalayan foothills is crucial for hydro-meteorological hazard analysis.
Key Innovation: Utilizes IMDAA reanalysis data and Random Forest (RF) classification to examine precipitation patterns, identifying elevation as the dominant predictor (68% variance) and revealing a non-linear relationship with peak precipitation occurring between 1500-2500m, advancing understanding of precipitation classification in complex terrain.
103. Spatial and temporal distribution characteristics of low-visibility phenomena in Xinjiang based on instrument measurements
Core Problem: Understanding the spatiotemporal distribution characteristics of low-visibility phenomena, including dust-related events, in Xinjiang is crucial for environmental and transportation management, but instrument measurements offer improved fine-scale identification compared to manual observations.
Key Innovation: Systematically analyzed differences between manual and instrument observations for six low-visibility phenomena over 20 years in Xinjiang, revealing that instrument measurements significantly improved fine-scale identification, and demonstrating that terrain factors strongly influence spatial distribution, with dust-related phenomena radiating from the Taklimakan Desert.
104. Minimum Amount of Stress Magnitude Data Records For Reliable Geomechanical Modeling
Core Problem: High cost and limited representativity of in-situ stress measurements lead to sparse datasets, introducing uncertainty in geomechanical-numerical models used for subsurface projects.
Key Innovation: A numerical framework to determine the minimum number of stress magnitude data records needed for reliable stress modeling, constrained by formation stiffness variability, and to objectively identify outliers linked to local stiffness anomalies.
105. Using a Coupled FLAC3D–PFC Model to Investigate the Impact of Inclined Weak Beddings on Mechanical and Flow Behavior in Oil Sands
Core Problem: Inclined heterolithic strata (IHS) may hinder steam chamber development and negatively impact bitumen production in oil sands recovery, requiring a better understanding of their impact on deformation, failure, and permeability evolutions.
Key Innovation: Developed a coupled FLAC3D–PFC model, integrated with a CFD solver, to investigate the flow-geomechanical behavior of IHS, proposing a novel permeability model that leverages DEM to analyze permeability variations in conjunction with crack propagation and failure modes.
106. Dependence of the Temperature on the Fatigue Properties of Sandstones Under Graded Cyclic Loading and Unloading of the Intermediate Principal Stress
Core Problem: The impact of temperature on the fatigue characteristics of sandstone under graded cyclic loading and unloading of intermediate principal stress is not well understood, despite its importance for deep underground engineering.
Key Innovation: Revealed that increasing temperature slightly enhances sandstone fatigue life and significantly improves deformability (irreversible strain and elastic modulus), prolonging fatigue damage evolution, with implications for underground construction and mining.
107. A People-Centric Framework for Worst-case Disruption Analysis of Interdependent Infrastructure Systems
Core Problem: Existing worst-case disruption analyses for interdependent critical infrastructure systems overlook the logical interdependency created by people who depend on multiple services, leading to underestimation of affected populations.
Key Innovation: Proposes a people-centric worst-case disruption modeling framework to identify failure scenarios leading to the largest impacts on people under localized and non-localized disruptions, capturing logical interdependencies and showing significant underestimation by traditional methods.
108. Textural features of sand fraction grains as a source of information on the transfer of alluvial sediments in step-pool channels in small forested mountain catchments
Core Problem: Understanding the interplay between hillslope and fluvial processes and interpreting sediment supply, transport, and deposition in fluvial systems, particularly whether sand grain texture can indicate sediment processing and transfer in step-pool mountain streams.
Key Innovation: Confirmation that using textural features of sand fraction grains (Q, QW, S) is an effective method for analyzing sediment transfer in small mountain catchments, revealing that grain features reflect origin and evolve downstream, and precisely identifying the transition zone between hillslope and fluvial systems.
109. Fluvial-aeolian interactions in northeastern South America: Implications for provenance and paleoenvironmental interpretations
Core Problem: The need to reconstruct environmental changes and understand fluvial-aeolian interactions in northeastern South America from the Last Glacial Maximum through the Holocene, to interpret provenance and paleoenvironmental conditions, and address knowledge gaps in geomorphological and paleoclimatic dynamics.
Key Innovation: An integrated morphostratigraphic approach combining sedimentological, geochemical, and OSL analyses of various deposits, revealing sedimentation pulses synchronous with global climate events, alternation between erosional and aggradational phases, and strong regulation of the hydro-sedimentary regime by climatic teleconnections.
110. Coupled thermo-mechanical simulation of lining cracking evolution and sealing system mechanical response in CAES lined rock caverns using finite-discrete element method
Core Problem: Clarifying the lining's cracking pattern and achieving coordinated performance with the sealing layer in lined rock caverns (LRCs) for compressed air energy storage (CAES), especially under coupled thermo-mechanical conditions.
Key Innovation: Proposed a coupled thermo-mechanical numerical framework based on the finite-discrete element method to effectively predict random cracking and crack evolution, and investigated the effects of thermal effects, surrounding rock stiffness, and reinforcement parameters on the lining-sealing system.
111. An experimental insight into water-driven fracture of granite under coupled stress–temperature conditions
Core Problem: Rock fracture toughness testing under coupled stress-temperature conditions remains unaddressed by ISRM methods, limiting understanding of water-injection-driven rock fractures in deep horizons.
Key Innovation: Performed hydraulic fracturing experiments on granite under coupled stress-temperature conditions mimicking deep burial, revealing that heated water drastically reduces fracture toughness via thermochemical reactions, and that fracture toughness peaks near 3000m due to competition between geostress strengthening and water-induced degradation.
112. Deep learning for the probabilistic prediction of semi-continuous hydrological variables – An application to streamflow prediction across CONUS
Core Problem: Accurately predicting semi-continuous hydrological variables like streamflow, especially low flows, while respecting the physical lower bound of zero, and providing probabilistic forecasts rather than just deterministic ones.
Key Innovation: Introducing two end-to-end probabilistic deep learning models (rectified Gaussian and hurdle models) that couple LSTMs with probabilistic output layers to predict streamflow parameters, with the hurdle model outperforming benchmarks in low flow accuracy and overall probabilistic performance, offering a new state-of-the-art for daily streamflow prediction.
113. Crack mitigation and moisture retention in sand-clay mixtures under drought: Effects of biopolymer and fiber reinforcement
Core Problem: Mitigating soil cracking and enhancing moisture retention in sand-clay mixtures under drought conditions, which are exacerbated by global warming and lead to land degradation.
Key Innovation: Demonstrated that biopolymer (BMG) fills pores and enhances moisture retention, reducing crack intensity, while palm fibers (PFs) alleviate brittleness by forming a stress-dissipating network. The combined use of BMG and PFs extended evaporation duration and balanced brittleness/heterogeneity, offering an eco-friendly approach to improve drought resilience and soil stabilization.
114. Stabilization of carbonate sand with filamentous fungi: Optimization and mechanical evaluation
Core Problem: Carbonate sand, used as fill material, is vulnerable to erosion and instability due to its high porosity, angular grains, and brittle structure, especially in tropical/subtropical environments.
Key Innovation: Optimized mycelium growth conditions for filamentous fungi to stabilize carbonate sand. Demonstrated that fungal reinforcement significantly suppressed excess pore water pressure, enhanced peak strength (up to 79%), and initial stiffness (up to 369%). Microscopic analysis showed fungal hyphae form a flexible, fiber-like network, reducing brittleness and offering a mechanistically distinct bio-based reinforcement pathway.
115. Next-generation Metop ASCAT Surface Soil Moisture datasets
Core Problem: The operational near real-time (NRT) ASCAT surface soil moisture (SSM) product lagged behind historical offline data records in algorithmic improvements, and a higher-resolution product was needed.
Key Innovation: Released next-generation ASCAT SSM datasets, applying the latest retrieval algorithm and consistent grid to both NRT and historical data, creating a unified product. Introduced a new high-resolution 6.25 km sampling product, improved dry/wet backscatter reference estimation using a moving-window approach, and added a subsurface scattering flag.
116. Global satellite gravity data products for prompt detection of short-term Mass Change (MC)
Core Problem: A long-standing gap exists in investigating sub-monthly surface mass change from satellite gravimetry, limiting the ability to promptly detect and characterize short-term hydrological extremes.
Key Innovation: Presented the globally available Line-of-sight Gravity Differences (LGD) dataset, derived directly from GRACE-FO Level-1B observations, providing instantaneous, in situ gravity change at satellite altitude. Demonstrated its potential for monitoring sub-monthly terrestrial water storage variability and characterizing hydrological extremes.
117. An original approach combining biogeochemical signatures and a mixing model to discriminate spatial runoff-generating sources in a peri-urban catchment
Core Problem: While hydrograph separation can vertically decompose flow, its application to spatially decompose flow and identify contributions linked to land use, geology, and contaminant sources remains limited.
Key Innovation: An original approach combining a Bayesian mixing model with a wide range of biogeochemical tracers (classical and innovative) to discriminate eight spatial runoff-generating sources in a peri-urban catchment, improving a perceptual hydrological model and demonstrating potential for validating distributed hydrological models.
118. The general formulation for mean annual runoff components estimation and their change attribution
Core Problem: A general framework to quantify and attribute mean annual runoff components (surface flow, baseflow, total runoff) is lacking, hindering understanding of precipitation partition and runoff generation.
Key Innovation: A general formulation (MPS model) derived from observational data and theoretical derivation, which characterizes mean annual runoff components as a function of available water with one parameter, demonstrating high accuracy across 662 catchments and elucidating responses of surface flow and baseflow to available water and environmental factors.
119. Joint calibration of multi-scale hydrological data sets using probabilistic water balance data fusion: methodology and application to the irrigated Hindon River Basin, India
Core Problem: Hydrological datasets, despite their potential for water resource management, are subject to uncertainties, and a robust method for their joint calibration and bias correction is needed.
Key Innovation: A monthly probabilistic water balance data fusion approach that automatically bias-corrects and noise-filters multi-scale hydrological data by linking them through the water balance, generating hydrologically consistent estimates of precipitation, evaporation, storage, irrigation canal water imports, and river discharge, and downscaling them to pixel-scale.
120. Do Open-Vocabulary Detectors Transfer to Aerial Imagery? A Comparative Evaluation
Core Problem: Open-vocabulary object detection models, despite strong performance on natural images, exhibit severe domain transfer failure when applied to aerial imagery, primarily due to semantic confusion and brittleness to imaging conditions.
Key Innovation: Presents the first systematic benchmark evaluating state-of-the-art OVD models on aerial datasets under zero-shot conditions, revealing critical limitations and the need for domain-adaptive approaches in aerial OVD.
121. SHED Light on Segmentation for Dense Prediction
Core Problem: Existing dense prediction methods treat it as an independent pixel-wise prediction, often resulting in structural inconsistencies in real-world scenes despite strong inherent geometric structure.
Key Innovation: Introduces SHED, a novel encoder-decoder architecture that explicitly enforces geometric prior by incorporating segmentation into dense prediction through bidirectional hierarchical reasoning. It improves depth boundary sharpness, segment coherence, cross-domain generalization, and 3D reconstruction quality.
122. Hybrid Cross-Device Localization via Neural Metric Learning and Feature Fusion
Core Problem: Achieving robust and accurate cross-device localization in diverse environments, particularly for challenges requiring high recall and accuracy, remains a complex task.
Key Innovation: A hybrid cross-device localization pipeline integrating a shared retrieval encoder, a classical geometric branch (feature fusion and PnP), and a neural feed-forward branch (MapAnything). It includes a neural-guided candidate pruning strategy and depth-conditioned localization, leading to significant improvements in recall and accuracy.
123. DELNet: Continuous All-in-One Weather Removal via Dynamic Expert Library
Core Problem: All-in-one weather image restoration methods are valuable but depend on pre-collected data and require retraining for unseen degradations, leading to high cost and limiting practical deployment.
Key Innovation: DELNet, a continual learning framework for weather image restoration. It integrates a judging valve to measure task similarity and a dynamic expert library that stores experts trained on different degradations, enabling continuous optimization without retraining existing models and achieving superior performance and efficiency.
124. TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstruction
Core Problem: Streaming recurrent models for 3D reconstruction suffer from catastrophic memory forgetting over long sequences, hindering long-term stability.
Key Innovation: TTSA3R, a training-free framework, leverages temporal state evolution and spatial observation quality for adaptive state updates. It uses a Temporal Adaptive Update Module and a Spatial Contextual Update Module, significantly improving long-term reconstruction stability with only a 15% error increase on extended sequences.
125. ExpAlign: Expectation-Guided Vision-Language Alignment for Open-Vocabulary Grounding
Core Problem: Open-vocabulary grounding requires accurate vision-language alignment under weak supervision, but existing methods either lack fine-grained expressiveness or rely on explicit supervision/heavy cross-attention.
Key Innovation: ExpAlign, an expectation-guided vision-language alignment framework, introduces an Expectation Alignment Head for attention-based soft MIL pooling over token-region similarities, enabling implicit token and instance selection. It also uses an energy-based multi-scale consistency regularization, achieving 36.2 AP$_r$ on LVIS minival, outperforming state-of-the-art methods.
126. Fire on Motion: Optimizing Video Pass-bands for Efficient Spiking Action Recognition
Core Problem: Spiking Neural Networks (SNNs) underperform on dynamic video tasks compared to ANNs because standard spiking dynamics act as a temporal low pass, emphasizing static content and attenuating motion-bearing bands.
Key Innovation: The Pass-Bands Optimizer (PBO), a plug-and-play module, optimizes the temporal pass-band toward task-relevant motion bands by suppressing static components. PBO introduces only two learnable parameters and a lightweight consistency constraint, yielding over ten percentage points improvement on UCF101 and consistent gains on other complex video tasks.
127. Do Transformers Have the Ability for Periodicity Generalization?
Core Problem: Large language models (LLMs) based on Transformers exhibit substantial limitations in out-of-distribution (OOD) generalization, particularly regarding periodicity, compared with humans.
Key Innovation: A unified interpretation of periodicity from abstract algebra and reasoning is introduced, and a controllable generative benchmark (Coper) for composite periodicity is constructed. Experiments reveal that Transformers' periodicity generalization is limited, as they can memorize periodic data but struggle to generalize to unseen composite periodicity, highlighting a gap in OOD generalization.
128. Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification
Core Problem: Multimodal deep learning (MDL) is hindered by incomplete multimodal data, leading to a "discarding-imputation dilemma" where discarding modalities loses information and recovering them introduces noise.
Key Innovation: Proposes DyMo, an inference-time dynamic modality selection framework that adaptively identifies and integrates reliable recovered modalities. It uses a novel selection algorithm maximizing task-relevant information, guided by a principled reward function, and a flexible network architecture, outperforming state-of-the-art methods across various missing-data scenarios.
129. Calibrated Multivariate Distributional Regression with Pre-Rank Regularization
Core Problem: Achieving multivariate calibration in probabilistic prediction remains challenging, and while pre-rank functions assess specific aspects of multivariate calibration, their use has been limited to post-hoc evaluation.
Key Innovation: Proposes a regularization-based calibration method that enforces multivariate calibration during training of multivariate distributional regression models using pre-rank functions. It introduces a novel PCA-based pre-rank that projects predictions onto principal directions, substantially improving multivariate pre-rank calibration without compromising accuracy and revealing dependence-structure misspecifications.
130. Uncertainty-Aware Extrapolation in Bayesian Oblique Trees
Core Problem: Decision trees struggle in regression tasks requiring reliable extrapolation and well-calibrated uncertainty, as piecewise-constant leaf predictions are bounded by training targets and can be overconfident under distribution shift.
Key Innovation: Proposes a single-tree Bayesian model that extends VSPYCT by equipping each leaf with a GP predictor. Bayesian oblique splits provide uncertainty-aware partitioning, while GP leaves model local functional behavior and enable principled extrapolation. It includes an efficient inference scheme and a gating mechanism for GP-based extrapolation, showing improvements in predictive performance and substantial gains in extrapolation scenarios.
131. Environment-Conditioned Tail Reweighting for Total Variation Invariant Risk Minimization
Core Problem: Out-of-distribution (OOD) generalization is challenging when models simultaneously encounter correlation shifts across environments and diversity shifts driven by rare or hard samples, as existing invariant risk minimization (IRM) methods primarily address environment-level spurious correlations but overlook sample-level heterogeneity.
Key Innovation: Proposes Environment-Conditioned Tail Reweighting for Total Variation Invariant Risk Minimization (ECTR), a unified framework that augments TV-based invariant learning with environment-conditioned tail reweighting to jointly address both types of distribution shift. It also extends to scenarios without explicit environment annotations by inferring latent environments, demonstrating consistent improvements in worst-environment and average OOD performance across various benchmarks.
132. SplineFlow: Flow Matching for Dynamical Systems with B-Spline Interpolants
Core Problem: Current flow matching methods are not well-suited for modeling dynamical systems, as they construct conditional paths using linear interpolants that may not capture underlying state evolution, especially from irregular sampled observations.
Key Innovation: Introduces SplineFlow, a theoretically grounded flow matching algorithm that jointly models conditional paths across observations via B-spline interpolation, exploiting B-spline smoothness and stability to learn complex underlying dynamics.
133. MeshGraphNet-Transformer: Scalable Mesh-based Learned Simulation for Solid Mechanics
Core Problem: Standard MeshGraphNets suffer from inefficient long-range information propagation on large, high-resolution meshes, limiting their scalability for industrial-scale solid mechanics simulations.
Key Innovation: Presents MeshGraphNet-Transformer (MGN-T), combining Transformers' global modeling with MeshGraphNets' geometric inductive bias via a physics-attention Transformer, enabling efficient and accurate learning on high-resolution meshes for complex solid mechanics problems like impact dynamics.
134. A Generalized Analytical Heat Transfer Model for Enhanced Geothermal Systems: Capturing Fracture Interactions and Correcting Classical Optimistic Predictions
Core Problem: Classical analytical heat transfer models for enhanced geothermal systems rely on simplified assumptions and systematically overestimate thermal performance, leading to unrealistic engineering decisions.
Key Innovation: Presents a generalized analytical model based on Green's functions that explicitly captures thermal interactions between fractures while preserving analytical tractability. The model corrects optimistic bias and provides more reliable predictions of production temperature and energy recovery, validated against numerical simulations.
135. Amortized Simulation-Based Inference in Generalized Bayes via Neural Posterior Estimation
Core Problem: Existing Generalized Bayesian Inference (GBI) methods typically rely on costly MCMC or SDE-based samplers and must be re-run for each new dataset and each temperature value, lacking amortization.
Key Innovation: Introduces the first fully amortized variational approximation to the tempered posterior family in GBI by training a single (x,β)-conditioned neural posterior estimator. This enables sampling in a single forward pass without simulator calls or inference-time MCMC, achieving competitive posterior approximations over a wide range of temperatures.
136. OneFlowSBI: One Model, Many Queries for Simulation-Based Inference
Core Problem: Traditional simulation-based inference often requires task-specific models or retraining for different inference tasks (e.g., posterior sampling, likelihood estimation, conditional distributions).
Key Innovation: OneFlowSBI, a unified framework that learns a single flow-matching generative model over the joint distribution of parameters and observations, supporting multiple inference tasks without task-specific retraining by leveraging a query-aware masking distribution during training.
137. Graph Attention Network for Node Regression on Random Geometric Graphs with Erd\H{o}s--R\'enyi contamination
Core Problem: Despite the widespread use and perceived robustness of Graph Attention Networks (GATs), rigorous statistical guarantees demonstrating their provable advantage over non-attention Graph Neural Networks (GNNs) for node regression, especially under simultaneous covariate and edge corruption, are scarce.
Key Innovation: A carefully designed, task-specific GAT that constructs denoised proxy features for node regression on random geometric graphs with Erd H{o}s--R H{e}nyi contamination, provably achieving lower error asymptotically compared to OLS and vanilla GCNs, and demonstrating the effectiveness of the attention mechanism in noisy graph settings.
138. Symmetrical Flow Matching: Unified Image Generation, Segmentation, and Classification with Score-Based Generative Models
Core Problem: Existing generative models and flow matching frameworks do not efficiently unify image generation, segmentation, and classification while preserving semantic information and ensuring bi-directional consistency.
Key Innovation: Introduces Symmetrical Flow Matching (SymmFlow), a new formulation that unifies semantic segmentation, classification, and image generation within a single model using a symmetric learning objective and a new training objective to explicitly retain semantic information, enabling efficient one-step segmentation and classification.
139. From Label Error Detection to Correction: A Modular Framework and Benchmark for Object Detection Datasets
Core Problem: Label errors compromise the quality of object detection datasets, affecting training and evaluations, and systematic, scalable correction methods are lacking.
Key Innovation: Introduces Rechecked, a semi-automated framework for label error correction that builds on existing detection methods and uses crowd-sourced microtasks for review, applied to the KITTI dataset to detect and correct errors, and provides a benchmark to motivate further research.
140. GMOR: A Lightweight Robust Point Cloud Registration Framework via Geometric Maximum Overlapping
Core Problem: Existing point cloud registration methods for high outlier ratios are either computationally expensive (graph-based) or suffer from local optima (multi-stage BnB search).
Key Innovation: Proposes GMOR, a geometric maximum overlapping registration framework via rotation-only BnB search, which decomposes rigid transformation and solves for optimal rotation and translation efficiently, achieving superior accuracy and efficiency.
141. VideoNSA: Native Sparse Attention Scales Video Understanding
Core Problem: Video understanding in multimodal language models is limited by context length, causing models to miss key transition frames and struggle with coherence across long time scales.
Key Innovation: VideoNSA, which adapts Native Sparse Attention (NSA) to video-language models using a hardware-aware hybrid attention approach, achieving improved performance on long-video understanding, temporal reasoning, and spatial benchmarks, and reliably scaling to 128K tokens.
142. Omni-View: Unlocking How Generation Facilitates Understanding in Unified 3D Model based on Multiview images
Core Problem: Integrating 3D scene understanding and generation from multiview images to achieve synergistic interaction and improve holistic understanding of 3D scenes.
Key Innovation: Omni-View, a unified multimodal understanding and generation model for 3D scenes based on multiview images. It jointly models scene understanding, novel view synthesis, and geometry estimation, leveraging spatiotemporal modeling and explicit geometric constraints. It achieves state-of-the-art 3D understanding while delivering strong performance in novel view synthesis and 3D scene generation.
143. Hyperspectral Image Data Reduction for Endmember Extraction
Core Problem: The high computational cost of self-dictionary methods for endmember extraction from large-scale hyperspectral images limits their applicability, despite their high accuracy.
Key Innovation: A data reduction technique that removes pixels not containing endmembers, preserving those close to endmembers, and integrating it with a self-dictionary method based on a linear programming formulation, substantially reducing computational time without sacrificing endmember extraction accuracy.
144. From Tokens to Photons: Test-Time Physical Prompting for Vision-Language Models
Core Problem: Extending Vision-Language Models (VLMs) from web images to sensor-mediated physical environments requires improving their robustness to real-world capture variations, which is challenging for conventional test-time adaptation.
Key Innovation: MVP (Multi-View Physical-prompt for Test-Time Adaptation), a forward-only framework that treats the camera exposure triangle (ISO, shutter speed, aperture) as 'physical prompts' to acquire and select multiple physical views, significantly improving VLM robustness in sensor-mediated environments without gradients or model modifications.
145. State Estimation Using Sparse DEIM and Recurrent Neural Networks
Core Problem: The Sparse Discrete Empirical Interpolation Method (S-DEIM) for state estimation from sparse observations requires knowledge of governing equations and suffers from convergence issues in its data assimilation step for inferring the optimal kernel vector.
Key Innovation: Introduction of an equation-free S-DEIM framework that utilizes recurrent neural networks (RNNs) to estimate the optimal kernel vector from sparse observational time series, demonstrating improved state estimations for complex spatiotemporal systems without requiring governing equations.
146. Deep Ensembles for Epistemic Uncertainty: A Frequentist Perspective
Core Problem: Decomposing prediction uncertainty into aleatoric and epistemic components is critical, but approximating the parameter posterior for epistemic uncertainty is computationally challenging, and a theoretical understanding of deep ensembles' empirical success from a frequentist perspective is limited.
Key Innovation: Connects deep ensembles to a bootstrap-based estimator for epistemic uncertainty (proven asymptotically correct) by decomposing it into data variability and training stochasticity, showing that deep ensembles capture the majority of epistemic uncertainty through the training stochasticity component.
147. CAOS: Conformal Aggregation of One-Shot Predictors
Core Problem: One-shot prediction, while enabling rapid adaptation of pretrained foundation models, lacks principled uncertainty quantification, and standard split conformal methods are inefficient in this setting due to data splitting and reliance on a single predictor.
Key Innovation: Proposes Conformal Aggregation of One-Shot Predictors (CAOS), a conformal framework that adaptively aggregates multiple one-shot predictors and uses a leave-one-out calibration scheme to fully exploit scarce labeled data, achieving valid marginal coverage and substantially smaller prediction sets than baselines.
148. Research on bow slamming loads and whipping response's characteristics in hydroelastic experiment by FBG method
Core Problem: Investigating the characteristics of slamming loads and whipping response on ship bows under severe sea states, which severely threaten ship safety.
Key Innovation: Utilizing the Fiber Bragg Grating (FBG) method in hydroelastic model experiments to measure and analyze ship whipping response, confirming its reliability and quantifying the impact of slamming loads on structural stress.
149. Coordinate-Aware Multiscale Feature Fusion Network for Spacecraft Component Recognition
Core Problem: Existing deep learning methods for spacecraft component recognition struggle with detecting small-scale components due to significant size disparity and lack of multiscale representation and spatial sensitivity.
Key Innovation: Proposed CAMFFN, a coordinate-aware multiscale feature fusion network, which uses a multibranch feature enhancement module and a multiscale coordinate-aware pyramid aggregation module to enhance multiscale representation and spatial sensitivity, outperforming state-of-the-art methods in recognizing fine-grained and small-scale components.
150. Dual-Perception Detector for Ship Detection in SAR Images
Core Problem: Ship detection in SAR images is challenging due to complex scattering from the background and difficulty in precisely distinguishing ship contours.
Key Innovation: Introduced a dual-branch detection framework with scattering characteristic perception (using a conditional diffusion model) and convex contour perception (two-stage coarse-to-fine pipeline), combined with a cross-token integration module, to construct discriminative features and achieve superior performance in oriented ship detection in SAR images.
151. A Hybrid Machine Learning Framework for Water Quality Index Prediction Using Feature-Based Neural Network Initialization
Core Problem: Existing water quality index prediction models suffer from unstable performance and reduced interpretability due to arbitrary weight initialization and limited use of ensemble learning.
Key Innovation: Introduced a hybrid machine learning framework combining SHAP-based neural network initialization with gradient boosting (XGBoost), achieving superior accuracy (86.9%) and interpretability in water quality index prediction by leveraging feature significance scores and iterative refinement.
152. Enhancing Estimation Performance of Winter Wheat Chlorophyll Fluorescence Parameter Fv/Fm by Fusing Spectral, Textural, and Physiological Features
Core Problem: Accurate remote sensing estimation of winter wheat chlorophyll fluorescence parameter Fv/Fm (a critical indicator of photosynthetic health) remains challenging, especially in integrating vegetation indices, texture features, and chlorophyll content.
Key Innovation: Demonstrated that fusing spectral (VIs from AAV multispectral), textural (TFs from AAV RGB), and physiological (SPAD) features with Gaussian process regression significantly enhances Fv/Fm estimation performance (R2 from 0.70 to 0.85), providing a scalable tool for precision agriculture.
153. Automated Extraction of 3-D Windows From MVS Point Clouds by Comprehensive Fusion of Multitype Features
Core Problem: Existing methods for 3D window extraction from street imagery or laser scanning data suffer from compromised accuracy and completeness due to reliance on limited feature types and challenges from shadows and geometric decorations.
Key Innovation: Proposed an automatic method for 3D window extraction from MVS point clouds that comprehensively fuses multitype features (color, geometry, HSV, depth distances, global arrangement) within an adaptive divide-and-combine pipeline, achieving high precision (92.7%) and integrity (89.81%) for buildings with varied appearances.
154. AMFC-DEIM: Improved DEIM With Adaptive Matching and Focal Convolution for Remote Sensing Small Object Detection
Core Problem: Small object detection in remote sensing imagery faces challenges from complex backgrounds and insufficient feature representation, leading to detection degradation.
Key Innovation: Proposed AMFC-DEIM, a novel architecture for remote sensing small object detection, which introduces an adaptive one-to-one matching mechanism, a focal convolution module for fine-grained features, and an enhanced normalized Wasserstein distance, achieving remarkable performance on benchmark datasets with low parameter count.
155. Influence of water on the physical properties of marl, calcareous marl, clayey marl, and limestone: a comparative study
Core Problem: Characterization and prediction of physical parameters of marls are difficult due to their variable properties, and the influence of water on these properties needs further consideration.
Key Innovation: Physically and texturally characterizes Eocene marly lithologies and limestones, comparing their properties (density, water absorption, UCS, etc.) in air-dry and water-saturated states, outlining new equations for the link between physical properties and micro-fabrics, and demonstrating the significant decrease in UCS of calcareous marl when water-saturated.
156. Mean stress effects and energy-based modeling of fatigue behavior in artificially cemented rock-like materials under cyclic loading
Core Problem: Soft rocks in construction environments are susceptible to fatigue failure under cyclic loading, and the explicit consideration of mean stress effects in fatigue life evaluation is necessary but often overlooked.
Key Innovation: Investigates the influence of cyclic mean stress on the fatigue behavior of artificially cemented sandstone specimens under uniaxial cyclic compression, establishing stress-life (S–N) curves and developing an empirical fatigue life prediction model based on cumulative dissipated energy at failure, explicitly incorporating mean stress effects.
157. Enhancing the Robustness of Cyber-Physical Power Systems against Cross-domain Cascading Failures: Cyber-Physical Dynamic Reconfiguration
Core Problem: Traditional defense mechanisms for Cyber-Physical Power Systems (CPPS) are insufficient to address complex cross-domain cascading failures, where failures propagate and amplify between cyber and physical domains.
Key Innovation: Proposes a cyber-physical joint dynamic reconstruction method based on Software-Defined Networking (SDN) to block cross-domain cascading failures, utilizing a Hidden Markov Model for communication network routing and optimal power flow for physical layer recovery, significantly improving CPPS robustness.
158. Temporal Causal Graph-based Attention Gated Recurrent Unit for Interpretable Fault Diagnosis in Nuclear Power Plants
Core Problem: Intelligent fault diagnosis in Nuclear Power Plants (NPPs) is challenging due to complex architectures and dynamic time-series data, and conventional data-driven models lack interpretability.
Key Innovation: Proposes a novel temporal causal graph-based gated recurrent unit with cell-level graph attention (GraphAttGRU-Cell) network for interpretable and efficient system-level fault diagnosis in NPPs, integrating domain expertise for causal graph learning and outperforming baselines in accuracy and interpretability.
159. A geometric Cross-Propagation-Calibration method for SAR constellation based on the graph theory
Core Problem: Ensuring high geometric accuracy of SAR constellation products requires long-term calibration and monitoring of system electronic delays, which is challenging without any calibrators.
Key Innovation: Proposes a geometric cross-propagation-calibration method for SAR constellations based on graph theory, which calibrates slant ranges without calibrators. The method constructs a graph from SAR images, estimates cumulative calibration error along optimal paths, and achieves geometric accuracy of less than 5m for two microsatellites.
160. Sedimentary and environmental changes of terminal lake in the arid region of Mongolia during the last two millennia
Core Problem: The need to understand past environmental and climate changes in arid regions like the Valley of Gobi Lake, Mongolia, to interpret their influence on human history and predict future trends, particularly regarding sedimentary and environmental shifts over the last two millennia.
Key Innovation: Analysis of two sediment cores from Boontsagaan Lake using grain size, radiocarbon, and OSL dating, revealing dramatic lake level lowering, subsequent recovery, and intensified aeolian input, thereby capturing key environmental transitions and regional climate changes over two millennia.
161. Mid-Holocene wet optimum and Early-Late Holocene arid phases shaped steppe-forest vegetation and human societies of Uzbekistan: Multi-proxy evidences from Lake Fazilman
Core Problem: The limited understanding of Holocene climate dynamics, land surface processes, and human-environment interactions in drylands of the Uzbekistan piedmont due to a lack of well-dated, high-resolution records.
Key Innovation: The first Holocene-scale, multi-proxy paleoenvironmental reconstructions from Lake Fazilman, combining various sediment, pollen, and biomarker analyses, to track hydro-climatic variability, vegetation dynamics, and land-use patterns, revealing aridification trends, centennial-scale climate events, and early human impact.
162. Unearthing historical pedology: An analysis of soil science concepts in 1200 years of Persian poetry
Core Problem: The lack of investigation into the historical semantic understanding of soil, from a pedological perspective, within cultural repositories like Persian poetry.
Key Innovation: An analysis of 1200 years of Persian poetry to investigate the semantic understanding of soil, revealing ten distinct thematic categories including a significant focus on soil erosion, and demonstrating soil's dual status as both a tangible element and a profound symbolic concept.
163. Simulating the recent drought-induced mortality of European beech (Fagus sylvatica L.) and Norway spruce (Picea abies L.) in German forests
Core Problem: Understanding and simulating drought-related tree mortality in German forests, which is a critical driver of forest dynamics, and improving model performance by incorporating factors like soil properties and bark beetle damage.
Key Innovation: Used a process-based model (ForClim v4.2, incorporating a predisposing-inciting framework and bark beetle module) to simulate drought-induced mortality of European beech and Norway spruce across German forests. Demonstrated the framework's ability to capture general mortality patterns, highlighted the critical role of soil water holding capacity and heterogeneity in modulating drought responses, and showed that the bark beetle submodel improved simulations for Norway spruce.
164. Early Pleistocene ecosystem turnover in South Siberia linked to abrupt regional cooling
Core Problem: Quantifying Earth system feedbacks that amplify greenhouse gas forcing, especially on land where long paleoclimate records are scarce, and understanding the timescale of ecosystem changes in the Arctic and subarctic.
Key Innovation: Reconstruction of warm-season temperatures and vegetation at Lake Baikal over 8.6 million years, documenting an abrupt transition approximately 2.7 million years ago to severe cold temperatures, replacement of forests by steppe–tundra, and probable permafrost extension, supporting nonlinear regional climate response and hypothesizing roles for vegetation albedo and permafrost carbon storage in amplifying glacial–interglacial cycles.
165. Dynamic monitoring of grassland land cover types of Inner Mongolia 1990–2023 and testing the causal relationship between meteorological data and grassland area
Core Problem: The continuous decline of global grassland ecosystems and the need for dynamic monitoring of grassland land cover types and understanding the causal relationship with meteorological data.
Key Innovation: Compared classification performance of Random Forest and other methods for dynamic monitoring of grassland land cover in Inner Mongolia from 1990-2023 and investigated the causal relationship between meteorological data and grassland area.
166. Conformal Prediction for Generative Models via Adaptive Cluster-Based Density Estimation
Core Problem: Conditional generative models lack calibrated uncertainty estimates, undermining trust in individual outputs for high-stakes applications, and existing conformal prediction methods may be sensitive to outliers or have high structural complexity.
Key Innovation: Proposes CP4Gen, a systematic conformal prediction approach tailored to conditional generative models that leverages adaptive clustering-based density estimation to construct prediction sets that are less sensitive to outliers, more interpretable, and of lower structural complexity, demonstrating superior performance in prediction set volume and simplicity, including for climate emulation tasks.
167. Score-based Integrated Gradient for Root Cause Explanations of Outliers
Core Problem: Traditional approaches for identifying the root causes of outliers struggle under uncertainty and high-dimensional dependencies in causal inference and anomaly detection.
Key Innovation: SIREN, a novel and scalable method that attributes root causes of outliers by estimating score functions of data likelihood via integrated gradients, enabling tractable and uncertainty-aware attribution in complex causal models.
168. CoDCL: Counterfactual Data Augmentation Contrastive Learning for Continuous-Time Dynamic Network Link Prediction
Core Problem: The rapid growth and continuous structural evolution of dynamic networks make effective predictions challenging, requiring models to be robust to emerging structural changes.
Key Innovation: CoDCL, a plug-and-play dynamic network learning framework that combines counterfactual data augmentation with contrastive learning, and a comprehensive strategy to generate high-quality counterfactual data, significantly improving state-of-the-art baseline models in dynamic network link prediction.
169. Temporal Graph Pattern Machine
Core Problem: Prevailing methods for temporal graph learning are task-centric and rely on restrictive assumptions (e.g., short-term dependency, static neighborhood), hindering the discovery of transferable temporal evolution mechanisms.
Key Innovation: Temporal Graph Pattern Machine (TGPM), a foundation framework that learns generalized evolving patterns by conceptualizing interactions as patches from temporally-biased random walks, processed by a Transformer-based backbone, and enhanced with self-supervised pre-training tasks, achieving state-of-the-art performance in link prediction and cross-domain transferability.
170. DRL-Enabled Trajectory Planing for UAV-Assisted VLC: Optimal Altitude and Reward Design
Core Problem: Optimizing three-dimensional trajectory planning in UAV-assisted Visible Light Communication (VLC) systems to minimize UAV flight distance and maximize data collection efficiency is a challenging mixed-integer non-convex optimization problem.
Key Innovation: Derives a closed-form optimal flight altitude and optimizes the UAV horizontal trajectory by integrating a novel pheromone-driven reward mechanism with the twin delayed deep deterministic policy gradient algorithm, significantly reducing flight distance and convergence steps.
171. Can 3D point cloud data improve automated body condition score prediction in dairy cattle?
Core Problem: Conventional visual body condition scoring (BCS) in dairy cattle is subjective and labor-intensive, and while 3D point cloud data offers richer geometric characteristics, its direct head-to-head comparison with depth image-based approaches for BCS prediction remains limited.
Key Innovation: A comprehensive comparison of top-view depth image and point cloud data for BCS prediction under four settings. Results indicate that depth image-based models consistently achieve higher accuracy or comparable performance, and point cloud-based predictions are more sensitive to noise and model architecture.
172. EndoCaver: Handling Fog, Blur and Glare in Endoscopic Images via Joint Deblurring-Segmentation
Core Problem: Endoscopic image analysis is compromised by real-world conditions such as lens fogging, motion blur, and specular highlights, which severely affect automated polyp detection.
Key Innovation: EndoCaver, a lightweight transformer with a unidirectional-guided dual-decoder architecture, enabling joint multi-task image deblurring and segmentation. It integrates a Global Attention Module, a Deblurring-Segmentation Aligner, and a cosine-based scheduler, achieving superior performance on degraded data with significantly reduced parameters.
173. Learning to Defer in Non-Stationary Time Series via Switching State-Space Models
Core Problem: Learning to Defer (L2D) in non-stationary time series is challenging due to partial feedback, time-varying expert availability, and the need for effective cross-expert information transfer.
Key Innovation: Models signed expert residuals using L2D-SLDS, a factorized switching linear-Gaussian state-space model with context-dependent regime transitions, a shared global factor, and per-expert idiosyncratic states. Proposes an IDS-inspired routing rule that trades off predicted cost against information gained.
174. FedDis: A Causal Disentanglement Framework for Federated Traffic Prediction
Core Problem: Federated learning for traffic prediction is challenged by the non-identically and independently distributed (non-IID) nature of decentralized traffic data, which often entangles globally shared patterns with client-specific local dynamics.
Key Innovation: FedDis, a novel causal disentanglement framework with a dual-branch design: a Personalized Bank to capture client-specific factors and a Global Pattern Bank to distill common knowledge. A mutual information minimization objective enforces informational orthogonality, achieving state-of-the-art performance.
175. Heterogeneous Graph Alignment for Joint Reasoning and Interpretability
Core Problem: Effectively integrating information across collections of heterogeneous graphs with differing topologies, scales, and semantics, often in the absence of shared node identities, remains a significant challenge in multi-graph learning.
Key Innovation: The Multi-Graph Meta-Transformer (MGMT), a unified, scalable, and interpretable framework. It applies Graph Transformer encoders to each graph, selects task-relevant supernodes via attention, and builds a meta-graph connecting functionally aligned supernodes, enabling joint reasoning and providing built-in interpretability.
176. GaussianOcc3D: A Gaussian-Based Adaptive Multi-modal 3D Occupancy Prediction
Core Problem: 3D semantic occupancy prediction in autonomous driving faces challenges with single-modality trade-offs, modality heterogeneity, spatial misalignment, and computationally heavy/lossy representations.
Key Innovation: GaussianOcc3D, a multi-modal framework using a memory-efficient, continuous 3D Gaussian representation to bridge camera and LiDAR data, incorporating specialized modules for feature aggregation, smoothing, fusion, and global context for robust 3D occupancy prediction.
177. Cascaded Flow Matching for Heterogeneous Tabular Data with Mixed-Type Features
Core Problem: Generating mixed-type features (combining discrete states with continuous distributions) in tabular data using generative models remains challenging.
Key Innovation: A cascaded flow matching approach for heterogeneous tabular data, which first generates a low-resolution representation and then leverages this in a high-resolution flow matching model via a guided conditional probability path and data-dependent coupling, enabling more faithful generation of mixed-type features.
178. Unconditional flow-based time series generation with equivariance-regularised latent spaces
Core Problem: While flow-based models are successful for time-series generation, designing latent representations with desirable equivariance properties for time-series generative modelling remains underexplored.
Key Innovation: Proposes a latent flow-matching framework that explicitly encourages equivariance through a simple regularization of a pre-trained autoencoder, introducing an equivariance loss. This improves generation quality and preserves computational advantages, outperforming diffusion-based baselines and achieving faster sampling.
179. OptiMAG: Structure-Semantic Alignment via Unbalanced Optimal Transport
Core Problem: In Multimodal Attributed Graphs (MAGs), there's a discrepancy between the implicit semantic structure of modality embeddings and the explicit graph structure, leading to aggregation of dissimilar features and modality-specific noise during message passing.
Key Innovation: Proposes OptiMAG, an Unbalanced Optimal Transport-based regularization framework that uses Fused Gromov-Wasserstein distance to explicitly guide cross-modal structural consistency within local neighborhoods, mitigating structural-semantic conflicts. It also includes a KL divergence penalty for adaptive inconsistency handling, consistently outperforming baselines across various graph-centric and multimodal-centric tasks.
180. Development of Domain-Invariant Visual Enhancement and Restoration (DIVER) Approach for Underwater Images
Core Problem: Underwater images suffer severe degradation (attenuation, scattering, non-uniform illumination) that varies across water types and depths, and existing state-of-the-art methods perform poorly in deep, unevenly illuminated, or artificially lit conditions.
Key Innovation: Proposes DIVER, an unsupervised Domain-Invariant Visual Enhancement and Restoration framework that integrates empirical correction with physics-guided modeling. It uses IlluminateNet, Spectral Equalization Filter, Adaptive Optical Correction Module, and Hydro-OpticNet with a composite loss function, consistently achieving best or near-best performance across diverse datasets and improving robotic perception.
181. Synthetic Time Series Generation via Complex Networks
Core Problem: Access to high-quality time series data is often limited due to privacy, acquisition costs, and labeling challenges, hindering the development of robust machine learning models.
Key Innovation: Presents a framework for generating synthetic time series by leveraging complex networks mappings, specifically investigating whether time series transformed into Quantile Graphs (QG) and then reconstructed can preserve statistical and structural properties. This methodology offers a competitive and interpretable alternative to GANs for synthetic time series generation.
182. DINO-SAE: DINO Spherical Autoencoder for High-Fidelity Image Reconstruction and Generation
Core Problem: Existing approaches using pretrained Vision Foundation Models (VFMs) like DINO for generative autoencoders suffer from limited reconstruction fidelity due to the loss of high-frequency details.
Key Innovation: Presents DINO Spherical Autoencoder (DINO-SAE), which bridges semantic representation and pixel-level reconstruction. It introduces a Hierarchical Convolutional Patch Embedding module and a Cosine Similarity Alignment objective to enhance detail preservation and semantic consistency. Leveraging the hyperspherical nature of SSL-based VFM representations, it employs Riemannian Flow Matching to train a Diffusion Transformer (DiT) on this manifold, achieving state-of-the-art reconstruction quality and efficient convergence.
183. FlexLoRA: Entropy-Guided Flexible Low-Rank Adaptation
Core Problem: Low-Rank Adaptation (LoRA) in parameter-efficient fine-tuning (PEFT) has a fixed-rank design, and existing dynamic rank allocation methods rely on heuristic, element-level metrics, lacking mechanisms to expand capacity in layers needing more adaptation.
Key Innovation: Proposes FlexLoRA, an entropy-guided flexible low-rank adaptation framework that evaluates matrix importance via spectral energy entropy, supports rank pruning and expansion under a global budget, and employs zero-impact initialization for stability. It consistently outperforms state-of-the-art baselines across benchmarks.
184. Scalable Topology-Preserving Graph Coarsening with Graph Collapse
Core Problem: Graph coarsening methods often preserve spectral or spatial characteristics, but topology-preserving methods, which help maintain GNN predictive performance, suffer from exponential time complexity.
Key Innovation: Proposes Scalable Topology-Preserving Graph Coarsening (STPGC) by introducing graph strong collapse and graph edge collapse concepts. STPGC comprises three new algorithms (GStrongCollapse, GEdgeCollapse, NeighborhoodConing) that eliminate dominated nodes/edges while rigorously preserving topological features. It proves GNN receptive field preservation and develops approximate algorithms, demonstrating efficiency and effectiveness for node classification with GNNs.
185. Learning to Execute Graph Algorithms Exactly with Graph Neural Networks
Core Problem: Understanding what Graph Neural Networks (GNNs) can learn, especially their ability to execute algorithms exactly under bounded-degree and finite-precision constraints, remains a central theoretical challenge.
Key Innovation: Proves exact learnability results for graph algorithms by training an ensemble of MLPs to execute local instructions and using them as the update function within a GNN, demonstrating error-free execution with high probability for algorithms like BFS, DFS, and Bellman-Ford.
186. How well do generative models solve inverse problems? A benchmark study
Core Problem: A need to compare the effectiveness of state-of-the-art generative learning models (conditional GANs, INNs, Conditional Flow Matching) against traditional Bayesian inverse approaches for solving inverse problems.
Key Innovation: Benchmarks three generative learning models against a traditional Bayesian inverse approach for gas turbine combustor design, proposing several evaluation metrics and identifying Conditional Flow Matching as consistently outperforming competing approaches.
187. Training-Free Test-Time Adaptation with Brownian Distance Covariance in Vision-Language Models
Core Problem: Vision-language models suffer performance degradation under domain shift, and existing test-time adaptation methods are computationally intensive, rely on back-propagation, or focus on single modalities.
Key Innovation: Proposes TaTa (Training-free Test-Time Adaptation with Brownian Distance Covariance), a method that leverages Brownian Distance Covariance to dynamically adapt VLMs to new domains without training or back-propagation, combined with attribute-enhanced prompting and dynamic clustering for improved efficiency and stability.
188. VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation
Core Problem: Recent video diffusion models (VDMs) struggle to maintain 3D structural consistency, often resulting in object deformation or spatial drift, because standard denoising objectives lack explicit incentives for geometric coherence.
Key Innovation: Introduces VideoGPA (Video Geometric Preference Alignment), a data-efficient self-supervised framework that leverages a geometry foundation model to automatically derive dense preference signals, guiding VDMs via Direct Preference Optimization to enhance temporal stability, physical plausibility, and motion coherence.
189. Deep Lightweight Unrolled Network for High Dynamic Range Modulo Imaging
Core Problem: High-dynamic range (HDR) modulo imaging recovery is a non-convex and ill-posed problem, and recent recovery networks suffer in high-noise scenarios.
Key Innovation: Formulates the HDR reconstruction task as an optimization problem incorporating a deep prior, unrolling it into a lightweight convolutional denoiser network. It also introduces a Scaling Equivariance term for self-supervised fine-tuning, achieving superior performance and quality compared to state-of-the-art algorithms.
190. Simulation-based Bayesian inference with ameliorative learned summary statistics -- Part I
Core Problem: The exact likelihood function for simulation-based Bayesian inference is often difficult to obtain in a closed form or is computationally intractable, especially for complex simulation models and large datasets.
Key Innovation: Introduces a simulation-based inference framework using learned summary statistics as an empirical-likelihood, leveraging a transformation technique based on the Cressie-Read discrepancy criterion under moment restrictions to preserve statistical power, handle weakly dependent data, and be suitable for distributed computing.
191. High-Definition 5MP Stereo Vision Sensing for Robotics
Core Problem: High-resolution (5MP+) stereo vision systems for robotics require commensurately higher calibration accuracy and faster processing than conventional methods can provide to realize their full potential in generating dense and accurate 3D point clouds.
Key Innovation: Develops a novel, advanced frame-to-frame calibration and stereo matching methodology for 5MP camera imagery, achieving both high accuracy and speed, and introduces a new approach to evaluate real-time performance, demonstrating that high-quality point clouds from high-pixel-count cameras require high-accuracy calibration.
192. Generative and Nonparametric Approaches for Conditional Distribution Estimation: Methods, Perspectives, and Comparative Evaluations
Core Problem: The inference of conditional distributions is fundamental for prediction and uncertainty quantification, with a wide range of methodologies, but a systematic comparison and understanding of their distinct advantages and limitations is needed.
Key Innovation: Reviews and systematically compares representative generative and nonparametric approaches for conditional distribution estimation (e.g., single-index, basis-expansion, generative simulation-based methods) using a unified evaluation framework, providing insights into their performance, flexibility, and computational costs.
193. Neural Backward Filtering Forward Guiding
Core Problem: Inference in non-linear continuous stochastic processes on trees is challenging, especially with sparse observations and complex topologies, as exact smoothing is intractable and particle-based methods degrade in high dimensions.
Key Innovation: Neural Backward Filtering Forward Guiding (NBFFG), a unified framework that constructs a variational posterior by leveraging an auxiliary linear-Gaussian process to guide the generative path, and learns a neural residual to capture non-linear discrepancies, enabling efficient and accurate inference in complex stochastic processes.
194. Posterior Label Smoothing for Node Classification
Core Problem: The potential of label smoothing for node classification in graph-structured data, spanning homophilic to heterophilic graphs, remains largely unexplored.
Key Innovation: Introduces posterior label smoothing, a novel method for transductive node classification that derives soft labels from a posterior distribution conditioned on neighborhood labels, demonstrating consistent improvements in classification accuracy across various graph datasets.
195. FC-KAN: Function Combinations in Kolmogorov-Arnold Networks
Core Problem: Existing Kolmogorov-Arnold Networks (KANs) could benefit from more flexible and powerful function combinations to improve performance and leverage low-dimensional data effectively.
Key Innovation: Introduces FC-KAN, a KAN that leverages combinations of popular mathematical functions (e.g., B-splines, wavelets, RBFs) through element-wise operations, outperforming other KANs and MLPs on MNIST and Fashion-MNIST datasets.
196. The Narrow Gate: Localized Image-Text Communication in Native Multimodal Models
Core Problem: Understanding how vision-language models (VLMs) process visual information and transfer it to the textual domain, especially comparing native and non-native multimodal VLMs, is crucial for improving their capabilities.
Key Innovation: Reveals that native multimodal VLMs rely on a single 'narrow gate' post-image token for visual information transfer, unlike non-native VLMs' distributed communication, and shows that ablating this token significantly degrades image-understanding performance, enabling fine-grained control.
197. Decentralized Domain Generalization with Style Sharing: Formal Model and Convergence Analysis
Core Problem: Existing work on federated learning (FL) and domain generalization (DG) lacks formal mathematical analysis of DG objectives and limits DG research in FL to star-topology architectures, failing to address distribution shifts in decentralized settings.
Key Innovation: StyleDDG, a decentralized DG algorithm that allows peer-to-peer networks to achieve DG by sharing style information, providing the first systematic analysis of style-based DG training in decentralized networks, and demonstrating significant accuracy improvements across target domains with minimal communication overhead.
198. Spatially-Adaptive Gradient Re-parameterization for 3D Large Kernel Optimization
Core Problem: Naively increasing kernel size in large kernel convolutions for 3D volumetric analysis leads to optimization instability, and existing methods are complex.
Key Innovation: Introduces Rep3D, a framework that employs a lightweight modulation network to generate receptive-biased scaling masks, adaptively re-weighting kernel updates within a plain encoder architecture, improving 3D segmentation performance and robust local-to-global convergence.
199. Learning Hierarchical Sparse Transform Coding for 3DGS Compression
Core Problem: Current 3DGS (3D Gaussian Splatting) compression methods largely forego neural analysis-synthesis transforms, overburdening entropy coding and reducing rate-distortion performance.
Key Innovation: Proposes a training-time transform coding (TTC) method with a hierarchical design (channel-wise KLT and sparsity-aware neural transform) that optimizes the transform jointly with 3DGS representation and entropy model, delivering strong R-D performance and fast decoding.
200. Influence Functions for Edge Edits in Non-Convex Graph Neural Networks
Core Problem: Existing graph influence prediction methods for GNNs rely on strict convexity assumptions, only consider edge deletions, and fail to capture changes in message propagation caused by modifications.
Key Innovation: Proposes a proximal Bregman response function specifically tailored for GNNs, relaxing convexity, explicitly accounting for message propagation effects, and extending influence prediction to both edge deletions and insertions in a principled way.
201. DREAMS: Preserving both Local and Global Structure in Dimensionality Reduction
Core Problem: Existing dimensionality reduction methods typically preserve either local or global data structure well, but none can effectively represent both aspects simultaneously.
Key Innovation: Presents DREAMS (Dimensionality Reduction Enhanced Across Multiple Scales), a method that combines local structure preservation of t-SNE with global structure preservation of PCA via a simple regularization term, generating a spectrum of embeddings that efficiently balance both local and global structure preservation.
202. Filtering with Confidence: When Data Augmentation Meets Conformal Prediction
Core Problem: While synthetic data augmentation is effective, controlling the bias introduced by poor-quality generations is critical for its effectiveness.
Key Innovation: Proposes 'conformal data augmentation,' a principled data filtering framework that leverages conformal prediction to generate diverse synthetic data while provably controlling the risk of poor-quality generations, leading to consistent performance improvements across various tasks.
203. Fidel-TS: A High-Fidelity Multimodal Benchmark for Time Series Forecasting
Core Problem: The evaluation of time series forecasting models is hindered by a critical lack of high-quality benchmarks, suffering from data contamination, temporal/description leakage, and biases.
Key Innovation: Introduces Fidel-TS, a new large-scale, high-fidelity multimodal benchmark for time series forecasting, built from live APIs with data sourcing integrity and leak-free design, revealing flaws in previous benchmarks and providing new insights into forecasting models.
204. TAP: Two-Stage Adaptive Personalization of Multi-Task and Multi-Modal Foundation Models in Federated Learning
Core Problem: Personalized fine-tuning of foundation models in federated learning is challenging, especially with heterogeneity across clients in data, tasks, and modalities, and a lack of understanding of how to address this.
Key Innovation: Proposes TAP, a two-stage adaptive personalization framework for federated learning that leverages mismatched model architectures for selective replacement and post-FL knowledge distillation, demonstrating effectiveness across diverse datasets and tasks while providing convergence analysis.
205. FedLLM-Align: Feature Extraction From Heterogeneous Clients
Core Problem: Federated learning faces a major obstacle in handling heterogeneous tabular data across clients, where schema mismatches and incompatible feature spaces prevent straightforward aggregation.
Key Innovation: Proposes FedLLM-Align, a federated learning framework that leverages pretrained transformer-based language models to extract semantically aligned embeddings from serialized tabular data, enabling federated training with heterogeneous clients, outperforming baselines and reducing communication overhead.
206. FrameOracle: Learning What to See and How Much to See in Videos
Core Problem: Existing frame sampling strategies for video understanding in Vision-Language Models (VLMs) are inefficient and fail to adapt to variations in content density or task complexity, leading to high computational costs.
Key Innovation: FrameOracle, a lightweight, plug-and-play module that predicts both which frames are most relevant and how many frames are needed. It is trained via a curriculum and a new large-scale VideoQA dataset (FrameOracle-41K) with validated keyframe annotations, improving efficiency and accuracy in video understanding.
207. Latent Domain Prompt Learning for Vision-Language Models
Core Problem: Domain generalization (DG) is crucial for deploying vision-language models (VLMs) in real-world applications, but most existing methods rely on explicit domain labels that may be unavailable or ambiguous, making generalization to unseen domains challenging.
Key Innovation: A strategy for DG without explicit domain labels, representing an unseen target domain as a combination of latent domains automatically discovered from training data. It performs latent domain clustering on image features and fuses domain-specific text features based on image-latent domain similarity, yielding consistent gains and improving robustness under domain shift.
208. Dynamic Reflections: Probing Video Representations with Text Alignment
Core Problem: The temporal nature of video data remains largely unexplored in the context of cross-modal representation alignment, making it difficult to understand the capabilities of modern video and language encoders and their general-purpose utility.
Key Innovation: The first comprehensive study of video-text representation alignment, revealing that cross-modal alignment highly depends on the richness of both visual and text data. It proposes parametric test-time scaling laws and correlates semantic alignment with performance on downstream tasks, introducing video-text alignment as an informative zero-shot way to probe spatio-temporal data representation power.
209. A2GC: Asymmetric Aggregation with Geometric Constraints for Locally Aggregated Descriptors
Core Problem: Optimal transport-based aggregation methods for Visual Place Recognition (VPR) treat source and target marginals symmetrically, limiting effectiveness when image features and cluster centers exhibit substantially different distributions.
Key Innovation: A2GC-VPR, an asymmetric aggregation VPR method with geometric constraints. It employs row-column normalization averaging with separate marginal calibration for asymmetric matching and incorporates learnable coordinate embeddings to fuse compatibility scores with feature similarities, thereby promoting spatially proximal features and enhancing spatial awareness, leading to superior performance.
210. SHAP-Guided Kernel Actor-Critic for Explainable Reinforcement Learning
Core Problem: Actor-critic (AC) methods in reinforcement learning (RL) offer limited interpretability, and existing explainable RL methods often neglect the heterogeneous impacts of individual state dimensions on the reward during training.
Key Innovation: RSA2C (RKHS-SHAP-based Advanced Actor-Critic), an attribution-aware, kernelized, two-timescale AC algorithm that uses state attributions computed from the Value Critic via RKHS-SHAP to modulate Actor gradients and Advantage Critic targets, achieving efficiency, stability, and interpretability.
211. SB-TRPO: Towards Safe Reinforcement Learning with Hard Constraints
Core Problem: In safety-critical domains, reinforcement learning (RL) agents must satisfy strict, zero-cost safety constraints, but existing model-free methods often fail to achieve near-zero violations or become overly conservative.
Key Innovation: Safety-Biased Trust Region Policy Optimisation (SB-TRPO), a principled algorithm for hard-constrained RL that dynamically balances cost reduction with reward improvement by updating via a dynamic convex combination of reward and cost natural policy gradients, providing formal guarantees of local progress on safety while improving reward.
212. Less is More: Label-Guided Summarization of Procedural and Instructional Videos
Core Problem: Turning long procedural and instructional videos into clear, concise representations is challenging, and prior methods often struggle to produce contextually coherent or semantically grounded summaries.
Key Innovation: PRISM (Procedural Representation via Integrated Semantic and Multimodal analysis), a three-stage framework combining adaptive visual sampling, label-driven keyframe anchoring, and contextual validation using a large language model (LLM), which produces semantically grounded video summaries that retain high semantic content with significantly fewer frames and generalize across procedural and domain-specific video tasks.
213. Multivariate Bayesian Last Layer for Regression with Uncertainty Quantification and Decomposition
Core Problem: Standard deep neural networks often lack robust uncertainty quantification, especially in multivariate regression under heteroscedastic noise, making it difficult to disentangle aleatoric and epistemic uncertainty.
Key Innovation: New Bayesian Last Layer neural network models for multivariate regression that provide uncertainty quantification with a single forward pass, capable of disentangling aleatoric and epistemic uncertainty, and can enhance canonically trained deep neural networks with uncertainty-aware capabilities through proposed EM algorithms for parameter learning.
214. On uniqueness in structured model learning
Core Problem: Ensuring uniqueness and convergence in learning physical laws for systems of partial differential equations (PDEs), especially when augmenting existing physical models with components learned from data using neural networks.
Key Innovation: This paper provides theoretical uniqueness and convergence results for structured model learning frameworks, demonstrating that a unique identification of unknown model components is possible under specific conditions (neural network properties, regularization), even with incomplete and noisy measurements.
215. Uncertainty Quantification for Prior-Data Fitted Networks using Martingale Posteriors
Core Problem: Prior-data fitted networks (PFNs), despite their strong predictive performance on tabular data, lack a principled mechanism for providing uncertainty quantification for their predictions.
Key Innovation: This paper proposes a principled and efficient sampling procedure to construct Bayesian posteriors for PFNs based on Martingale posteriors, providing a method for uncertainty quantification for predictive means and quantiles, and proving its convergence.
216. FESTA: Functionally Equivalent Sampling for Trust Assessment of Multimodal LLMs
Core Problem: Accurately assessing the trustworthiness and quantifying uncertainty in predictions generated by multimodal large language models (MLLMs) is challenging due to the diverse nature of multi-modal input paradigms and the black-box nature of these models.
Key Innovation: This paper proposes FESTA (Functionally Equivalent Sampling for Trust Assessment), a novel unsupervised, black-box multimodal input sampling technique that generates an uncertainty measure for MLLMs by probing their consistency and sensitivity through equivalent and complementary input samples, achieving significant improvements in selective prediction performance.
217. Lifelong Learning with Behavior Consolidation for Vehicle Routing
Core Problem: Existing neural solvers for routing problems struggle with lifelong learning, exhibiting poor zero-shot generalization to new tasks and catastrophic forgetting of previously acquired knowledge when fine-tuned on sequential tasks with diverse distributions and scales.
Key Innovation: This paper proposes Lifelong Learning Router with Behavior Consolidation (LLR-BC), a novel framework that enables neural VRP solvers to effectively learn new tasks sequentially while mitigating catastrophic forgetting by consolidating prior knowledge through behavior alignment and confidence-weighted decision-seeking, thereby improving plasticity and zero-shot generalization.
218. Calibrating Decision Robustness via Inverse Conformal Risk Control
Core Problem: Robust optimization methods rely on ad hoc robustness levels, leading to insufficient protection or overly conservative solutions, and conformal prediction still fixes coverage targets a priori, offering little guidance for selection.
Key Innovation: Proposes a framework that provides distribution-free, finite-sample guarantees on both miscoverage and regret for robust predict-then-optimize policies, enabling decision-makers to reliably evaluate and calibrate robustness levels according to their cost-risk preferences.
219. Predicting Atmospheric Effective Sound Speed Using Synthetic Infrasound and Machine Learning
Core Problem: Accurately predicting atmospheric effective sound speed, particularly in the stratosphere, which is crucial for infrasound propagation modeling.
Key Innovation: Development of an ML model trained on synthetic infrasound waveforms that can recover effective sound speeds, showing highest accuracy in the stratosphere.
220. Influences of Summer Northeastern Arctic Sea Ice on September Compound Heatwave and Drought Events in the South China
Core Problem: Understanding the interannual variations and drivers of September compound heatwave and drought events in South China, particularly the influence of Arctic sea ice.
Key Innovation: Evaluation of the influence of summer Northeastern Arctic sea ice on September CHDEs in South China, identifying the meridional position of the westerly jet as a key accounting factor.
221. Deep Learning Atmospheric Models Reliably Simulate Out‐of‐Sample Land Heat and Cold Wave Frequencies
Core Problem: Developing deep learning-based atmospheric models that can reliably simulate out-of-sample land heat and cold wave frequencies.
Key Innovation: Demonstrating that DL-based GCMs can successfully simulate out-of-sample heat and cold wave frequencies with skill comparable to physical models, highlighting the role of surface temperature field autocorrelation.
222. The Critical Role of the Intermittent Two‐Phase Flow Topology on the Solute Dispersion in Partially Saturated Porous Media
Core Problem: Understanding the critical role of intermittent two-phase flow topology on solute dispersion in partially saturated porous media.
Key Innovation: Quantitatively correlating fluctuations of the transient dispersion coefficient with two-phase velocity covariance and using the effective Okubo-Weiss parameter to characterize flow topology.
223. Analytical Model of Velocity Distribution and Penetration Characteristics in Water‐Level Fluctuation Zone With Vegetation
Core Problem: Developing an analytical model for velocity distribution and penetration characteristics in water-level fluctuation zones with vegetation.
Key Innovation: Developing new formulas for vortex invasion at channel-floodplain interfaces, drag coefficient, and a modified von Karman constant.
224. The Impact of System Softness on Haines Jumps and Drainage in Porous Media
Core Problem: Investigating the impact of system softness on Haines jumps and drainage in porous media.
Key Innovation: Developing a pore-scale analytical model showing entrapped gas bubbles alter pressure response during Haines jumps, supported by microfluidic experiments.
225. Prediction of tension leg platform motion responses under extreme conditions based on physics-informed deep learning and uncertainty quantification
Core Problem: Accurate prediction of TLP motion responses for structural safety and dynamic positioning, especially under strong nonlinear hydrodynamic coupling and environmental uncertainty in extreme marine environments.
Key Innovation: Proposing an adaptive spatio-temporal fusion prediction network (ASTF-Net) that integrates multi-resolution wavelet decomposition, adaptive graph learning, and a physical constraint layer for high-precision and reliable prediction of TLP motion responses and early risk warning.
226. Ship hull stress reconstruction with limited sensors: A time-domain coupled modal superposition method and parametric layout optimization
Core Problem: Improving stress reconstruction accuracy and engineering applicability in structural health monitoring of ship hulls under constraints of limited sensor numbers.
Key Innovation: Proposing a time-domain coupled modal superposition (TCMS) method for accurate stress reconstruction under sparse sensing and developing a sensor layout indicator (S) for optimized sensor placement, providing an integrated technical route for full-field stress reconstruction of large ship structures.
227. Echo Flow-Induced Temporal Correlation Learning for Ultrasound Video Object Segmentation
Core Problem: Existing ultrasound video object segmentation methods struggle to capture inter-frame object motion and are suboptimal for ultrasound-specific characteristics, leading to unsatisfactory segmentation in dynamic or low-contrast scenarios.
Key Innovation: Proposed EchoSAM2, a novel method that introduces Echo Flow to capture motion trends and an Echo Modulation Block to enhance temporal relationships and feature representation, achieving state-of-the-art results in ultrasound video object segmentation.
228. Training-Free Breast Ultrasound Image Segmentation With Retrieval-Based SAM2
Core Problem: Supervised medical image segmentation models require large labeled datasets, long training times, and have limited generalization, hindering their application in breast cancer diagnosis.
Key Innovation: Proposed TFSeg, a Training-Free automatic image segmentation framework based on SAM2, which uses image retrieval and mask prompt generation to achieve accurate and efficient breast ultrasound image segmentation without re-training, outperforming most supervised and training-free models.
229. MVICAD2: Multi-View Independent Component Analysis With Delays and Dilations
Core Problem: Existing multi-view ICA methods assume identical sources across subjects, which is too restrictive for neuroscience data where individual variability and temporal dilation effects are common.
Key Innovation: Proposed MVICAD², a Multi-View Independent Component Analysis with Delays and Dilations, which allows sources to differ across subjects in both temporal delays and dilations, outperforming existing methods in simulations and validating its effectiveness with real neuroscience data.
230. Retrieval of Sea Surface Mean-Square Slope From GNSS-R Delay–Doppler Maps Using a Serial CNN-Transformer Model
Core Problem: Accurately deriving sea surface mean-square slope (MSS) from GNSS-R delay–Doppler map (DDM) data is challenging due to the difficulty in obtaining full-band wave spectral data and effectively capturing local and global spatial features in DDM.
Key Innovation: Proposed a serial deep learning framework combining CNN and Transformer structures for retrieving sea surface MSS from DDM data, which effectively captures local and global spatial features and integrates multisource physical auxiliary information, outperforming comparison models with high R2 values against reanalysis and buoy data.
231. Response patterns and mechanisms of weathering crusts on feldspathic metagreywacke under solar radiation at the Helankou petroglyphs site
Core Problem: The severe deterioration of Helankou petroglyphs, carved on feldspathic metagreywacke weathering crusts, is primarily attributed to deformation induced by thermal cycling under solar radiation, but the response patterns and mechanisms are poorly understood.
Key Innovation: Conducts monitoring of solar radiation, atmospheric temperature, and rock temperature/strain, combined with extreme event analyses and finite element numerical simulations, systematically revealing differential thermal expansion-contraction cycles and heterogeneous thermal stresses as the mechanisms causing spatially heterogeneous exfoliation in weathering crusts.
232. Humanitarian Assistance and its Expected Behavior and Land Use Changes: Insights from Bangladesh and Kenya
Core Problem: A critical gap in systematically evaluating longitudinal data to assess the enduring effects of short-term humanitarian interventions on behavior and land-use changes, which are crucial for long-term resilience and development.
Key Innovation: Examined expected behavior and land-use changes from humanitarian programming interventions using a mixed-methods design, highlighting the need for comprehensive assessment of humanitarian investments by monitoring behavior and land-use changes over time to guide strategic resource allocation and enhance understanding of long-term impacts.
233. Is climate risk perception enough? Empirical evidence from Australian farmers
Core Problem: The disconnect between climate risk perception and actual adaptation strategies among farmers, particularly concerning soil health practices, challenging the assumption of a direct link.
Key Innovation: Found that innovation attitude, risk aversion, training attendance, rainfall zone, age, location, and formal education are significant predictors of soil health practice adoption, while climate risk perception is not directly significant, highlighting a “risk-action gap” and the need for targeted interventions beyond awareness-raising.
234. Line importance sampling for reliability analysis with complex failure domain
Core Problem: Existing Line Sampling (LS) methods for reliability analysis are inefficient for highly nonlinear performance functions and complex failure domains, and are sensitive to the chosen important direction.
Key Innovation: Proposes Line Importance Sampling (LIS) and an adaptive Cross-Entropy-based LIS (CE-LIS) framework that integrates LS with importance sampling, deriving optimal sampling densities and adaptively concentrating samples around dominant failure regions, improving accuracy and robustness for complex problems.
235. Multi-object tracking of vehicles and anomalous states in remote sensing videos: Joint learning of historical trajectory guidance and ID prediction
Core Problem: Multi-object tracking (MOT) of vehicles in complex scenarios and their anomalous states after strong deformation interference remains a huge challenge.
Key Innovation: Proposes an end-to-end MOT method integrating a joint learning paradigm of historical trajectory guidance and identity (ID) prediction, featuring a Frame Feature Aggregation Module, Historical Tracklets Flow Encoder, Semantic-Consistent Clustering Module, and Dual-branch Modulation Fusion Unit. Achieves state-of-the-art performance on a new VAS-MOT dataset and an open-source dataset.
236. SuperMapNet for long-range and high-accuracy vectorized HD map construction
Core Problem: Existing methods for vectorized high-definition (HD) map construction suffer from limited perception range, underutilization of cross-modal synergies, spatial disparities, and low accuracy in classifying and localizing map elements.
Key Innovation: Proposes SuperMapNet, a multi-modal framework that tightly couples semantic information from camera images and geometric information from LiDAR point clouds for long-range BEV feature generation. It then uses three-level interactions (Point2Point, Element2Element, Point2Element) for high-accuracy classification and localization, surpassing previous state-of-the-art methods.
237. Roadside lidar-based scene understanding toward intelligent traffic perception: A comprehensive review
Core Problem: Urban transportation systems require robust and precise 3D spatial understanding of dynamic urban scenes, which roadside lidar offers, but challenges remain in deployment optimization, robust registration, generalizable object detection/tracking, and effective utilization of heterogeneous multi-modal data.
Key Innovation: Presents a comprehensive review of roadside lidar-based traffic perception, structured around sensor placement, multi-lidar point cloud fusion, dynamic traffic information extraction, subsequent applications, and benchmark datasets. It identifies current advances, open problems, and future directions in the field.
238. RTPSeg: A multi-modality dataset for LiDAR point cloud semantic segmentation assisted with RGB-thermal images in autonomous driving
Core Problem: LiDAR point cloud semantic segmentation for autonomous driving faces challenges due to the sparse and textureless characteristics of point clouds, and RGB images degrade under poor lighting conditions.
Key Innovation: Introduces RTPSeg, the first multi-modality dataset to simultaneously provide RGB and thermal infrared (TIR) images for point cloud semantic segmentation. It also proposes RTPSegNet, a baseline model that effectively leverages the complementary information between point clouds, RGB, and TIR images, demonstrating enhanced 3D semantic segmentation.
239. Mamba-CNN hybrid Multi-scale ship detection Network driven by a Dual-perception feature of Doppler and Scattering
Core Problem: Existing polarimetric SAR (PolSAR) ship detection approaches face challenges in discriminating targets from clutter, detecting multi-scale objects in complex scenes, and achieving real-time detection.
Key Innovation: Proposes a Mamba-CNN hybrid Multi-scale ship detection Network (MCMN) driven by a Dual-perception feature of Doppler and Scattering (DDS). DDS effectively differentiates ship and clutter pixels, while MCMN uses a Multi-scale Information Perception Module and a Local-Global Feature Enhancement Module for improved multi-scale detection and efficiency, achieving state-of-the-art results.
240. SAR-NanoShipNet: A scale-adaptive network for robust small ship detection in SAR imagery
Core Problem: Small ship target detection in synthetic aperture radar (SAR) imagery faces challenges such as high speckle noise, difficulty in extracting small target features, geometric distortion of ship shapes, and heading dependence.
Key Innovation: Proposes SAR-NanoShipNet, a scale-adaptive network that employs a specialized convolution (DABConv) for ship targets, deformable convolutions and boundary attention mechanisms for adaptive feature capture, and a VerticalCompSPPF module for longitudinal multi-scale convolution and channel attention. The method achieves high precision detection for small targets.
241. Joint analysis and modeling of the hot spot effect from the diurnal reflectance and temperature cycles observed by SEVIRI
Core Problem: Understanding and correcting for the "hot spot effect" (radiance peak when solar and viewing angles coincide) in diurnal reflectance and temperature cycles observed by satellite sensors, as this effect distorts temperature profiles and complicates comparisons across sites/seasons.
Key Innovation: Coupling a four-parameter DTC model with a directional kernel-driven model (KDM) including a hot spot term to create Time-Evolving KDMs. This allows for assessing the directional impact on spectral brightness temperature and LST, revealing significant deviations and the need for angular effect correction.
242. Trampling disturbance affects the stability of soil carbon pools in urban park green spaces by disrupting soil aggregates and altering the composition of organic carbon components
Core Problem: The unclear impact of trampling disturbance on urban park soil carbon cycling and the mechanisms by which varying trampling intensities influence soil carbon pools.
Key Innovation: Elucidation of the mechanisms by which trampling disturbance affects urban park soil carbon pools, demonstrating that heavy trampling significantly compromises soil carbon pool stability and sequestration capacity by disrupting soil aggregates, reducing organic carbon content, altering carbon fraction composition, and suppressing microbial activity.
243. Regime shifts driven by grazing of an inland river wetland ecosystem
Core Problem: Understanding the regime shifts and driving mechanisms of inland river wetland ecosystems under intensive grazing pressure, leading to degradation from a 'moist-healthy' to a 'dry-degraded' state.
Key Innovation: Identified multiple stable states and non-equilibrium threshold intervals for key eco-hydrological variables (e.g., soil bulk density 1.0–1.5 g·cm−3) that drive ecosystem regime shifts, proposing soil moisture content as a potential early-warning signal.
244. Fully coupled physics-informed neural networks for hydro-mechanical analysis of saturated poroelastic media
Core Problem: Traditional methods for fully coupled hydro-mechanical analysis of saturated poroelastic materials often require training data or spatial/time discretization, and struggle to capture strong coupling effects like the Mandel–Cryer effect.
Key Innovation: Presented a continuous-time Physics-Informed Neural Networks (PINNs) framework that simultaneously solves coupled momentum balance and mass conservation equations without training data or discretization. Implemented a simultaneous optimization strategy to accurately capture strong coupling between mechanical and hydraulic fields.
245. Origins, evolutions, and future directions of Landsat science products for advancing global inland water and coastal ocean observations
Core Problem: Developing a standardized global product for Landsat-derived surface water measurements (Aquatic Reflectance) to enable quality-controlled data for emerging aquatic science applications, while addressing challenges in accuracy, scalability, and computational efficiency.
Key Innovation: Introduced the Landsat 8/9 provisional Aquatic Reflectance (AR) product and examined its performance through the SATO framework and quantitative assessment using GLORIA and AERONET-OC datasets. Highlighted that AR retrieval performance is context-dependent, with minimal errors in optically simple waters but increasing errors in complex waters, and discussed key algorithmic considerations for atmospheric correction.
246. Advancing quantum imaging: Electrical tunability enabled by versatile liquid crystals
Core Problem: Current heralded quantum imaging schemes lack dynamic control over imaging functions, limiting their utility for identifying light-sensitive samples under minimal illumination.
Key Innovation: Introduced versatile liquid crystals (bichiral cholesteric, nematic, ferroelectric) to achieve electrically tunable multimode switching and heralded single-photon imaging, enabling dynamic control, specific polarization selection, and ultrafast remote switching for photon-limited scenarios.
247. One-stop strabismus digital diagnosis via AI-integrated skin-like and wearable “Eyelectronics”
Core Problem: Clinical diagnosis of strabismus is limited by low objectivity, poor pediatric compliance, and high cost, requiring multiple instruments and stepwise examinations.
Key Innovation: Proposed an AI-integrated, skin-like, wearable 'Eyelectronics' system (ultrathin, breathable, multidirectional strain-sensing array) that conformally adapts to the eyelid for wireless, mild-restricted measurement of eyelid deformation during eye movements. Combined with biomechanical modeling and a physiology knowledge-driven AI algorithm, it achieves simultaneous measurement of strabismus angle and identification of paretic muscle with high accuracy.