TerraMosaic Daily Digest: Mar 5, 2026
Daily Summary
This March 5, 2026 digest compiles 174 selected papers from 994 analyzed studies. The strongest contributions focus on landslide failure mechanics: reactivation of stepwise-deepening slopes, residual-strength prediction for old loess landslides, deformation and failure of gently inclined bedding slopes, and debris-flow initiation under contrasting hydrologic triggers. Together, these studies move slope analysis toward better-resolved internal weakening, trigger separation, and process-specific mitigation.
A parallel stream extends hazard analysis to coupled risk systems, including cryosphere-driven hazard amplification in the Hindukush-Himalaya, asset-scale coastal adaptation under cyclone forcing, pluvial and debris-flow vulnerability assessment, and remote-sensing-constrained deformation analysis from InSAR, GNSS, and UAV-SLAM. Across the set, the dominant evidence comes from detection-and-monitoring and concepts-and-mechanisms studies, with hazard-modelling papers translating those observations into more decision-ready workflows.
Key Trends
The main trajectory is from descriptive hazard mapping to mechanism-aware, observation-constrained, and intervention-oriented geohazard analysis.
- Failure mechanics are being resolved at the slip-surface scale: studies of deep-seated reactivation, loess residual strength, and bedding-controlled slope failure explicitly quantify how internal structure and post-peak weakening govern renewed movement.
- Trigger attribution is becoming pathway-specific: debris-flow and coastal-hazard studies now distinguish snowmelt, rainfall, storm-wave, and seismic forcing regimes instead of collapsing them into generic hazard indicators.
- Remote sensing products are becoming model constraints: plateau-scale InSAR, trans-continental InSAR-GNSS fields, and UAV-SLAM reconstructions are increasingly used to inform mechanics, vulnerability estimates, and hazard evolution rather than serve only as descriptive overlays.
- Mitigation studies are targeting the governing instability process: bioengineered root reinforcement, electrolysis-induced desaturation, and adaptation measures for exposed infrastructure are being evaluated against the specific mechanisms they are intended to suppress.
- Risk analysis is moving toward asset- and decision-scale application: recent work resolves adaptation value, gridded pluvial susceptibility, debris-flow vulnerability, and perception-resilience coupling in forms that are more usable for operational planning.
Selected Papers
This digest features 174 selected papers from 994 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. Deformation characteristics and mechanism of reactivation for stepwise-deepening landslides: a case study of the Hongzhai Landslide in Guizhou Province, China
Core Problem: Understanding the complex deformation characteristics and reactivation mechanisms of ancient landslides, particularly those exhibiting stepwise-deepening slip surfaces and multiple reactivation events.
Key Innovation: A detailed case study of the Hongzhai landslide, combining field surveys, historical analysis, InSAR-based surface deformation monitoring, and MatDEM numerical simulations to reveal its multi-phase reactivation history, progressive slip zone deepening, and the combined effects of geological factors, providing valuable insights for similar deep-seated landslides.
2. Research on the mechanical properties and residual strength prediction model of an old loess landslide
Core Problem: Lack of comprehensive understanding of the mechanical behavior of sliding zone loess and a reliable predictive model for residual strength, crucial for investigating large deformations and reactivation of old loess landslides.
Key Innovation: Systematic investigation of shear mechanical behavior of loess across full deformation range using ring shear tests, analysis of multi-factor interactive effects on shear strength using Response Surface Methodology (RSM), and establishment of a predictive model for residual strength, providing a scientific basis for identifying and restoring old loess landslides.
3. Numerical simulation study on the deformation and failure of a gently inclined bedding slope: a case study of the Jinhaihu landslide in Guizhou, China
Core Problem: The incomprehensible catastrophic nature and genetic mechanisms of gently inclined bedding landslides, particularly the Jinhaihu landslide, exacerbated by engineering disturbance and rainfall infiltration.
Key Innovation: Detailed field investigation, physical/mechanical tests, and a coupled PFC/FLAC2D numerical model to analyze the genetic mechanism of the Jinhaihu landslide, revealing the four-stage evolution process and the combined effects of excavation and claystone weakening.
4. Vetiver grass for sustainable slope stabilization: a geoslope-based study on laterite and sand-laterite embankments
Core Problem: The significant threat of slope failure to infrastructure, particularly in regions with heterogeneous soil profiles and steep embankments, and the need for sustainable bioengineering solutions.
Key Innovation: Evaluation of vetiver grass as a sustainable bioengineering solution for slope stabilization using GeoSlope's SLOPE/W module, demonstrating its significant improvement in Factor of Safety across various soil types and slope angles, and confirming its potential for mitigating rainfall-induced failures and erosion.
5. A normal classification system and intelligent identification method for slope failure
Core Problem: The significant economic losses and casualties caused by landslide disasters, and the critical need for accurately identifying landslide failure modes to improve monitoring and early warning capabilities.
Key Innovation: Proposal of a general classification system for four slope failure modes and development of an intelligent identification method using five machine learning algorithms (with CNN performing best) based on 219 landslide cases, providing a reference for enhancing landslide monitoring and early warning.
6. Analysis on formation mechanism and hazard of debris flow in small basins under multi-scenario hydrological triggering
Core Problem: Insufficient research on the essential differences in initiation dynamics and disaster patterns between snowmelt and rainfall-triggered debris flows in small basins.
Key Innovation: Development of a numerical simulation framework coupling the Snowmelt Runoff Model (SRM) with the FLO-2D debris flow dynamics model to quantitatively analyze and differentiate the initiation mechanisms, disaster patterns, and spatiotemporal hazard distribution for snowmelt and rainfall-triggered debris flows.
7. Impact-based risk assessment of coastal infrastructure adaptation to climate change: a case study of hurricane risk in Nova Scotia, Canada
Core Problem: Coastal infrastructure faces escalating risks from climate change (intensified tropical cyclones, wind, storm surges), requiring a probabilistic, spatiotemporal framework to quantify physical and economic impacts and evaluate adaptation cost-effectiveness.
Key Innovation: Development of a probabilistic, spatiotemporal framework integrating volunteered geographic information, sectoral damage functions, and hazard intensity to quantify asset-level expected annual damage and evaluate cost-effectiveness of adaptation measures for coastal infrastructure under current and future climate scenarios.
8. Cryospheric changes and resultant hazards: Insights from the Hindukush-Himalaya system
Core Problem: The Hindukush-Himalaya system is experiencing rapid cryospheric changes (warming, glacier mass loss, snow cover loss, glacial lake proliferation, permafrost thaw) at rates faster than the global average, leading to the emergence and intensification of various hazards, with significant research gaps and challenges in adaptation.
Key Innovation: A comprehensive synthesis of nearly 300 studies quantifying ongoing cryospheric changes and their resultant hazards (GLOFs, avalanches, slope destabilization, altered flood risks), providing a foundation to address regional vulnerabilities and global challenges, and highlighting research gaps in the Eastern Himalaya and Pamir.
9. A physics-informed U-Net-LSTM network for nonlinear structural response under seismic excitation
Core Problem: The Finite Element Method (FEM) for nonlinear seismic analysis is computationally demanding, limiting its scalability and real-time applicability, while purely data-driven deep learning models often lack generalization and physical consistency.
Key Innovation: A novel Physics-Informed U-Net-LSTM framework that integrates physical laws (domain-specific constraints) with deep learning to enhance both accuracy and efficiency in predicting the nonlinear seismic response of structures, bridging the gap between data-driven and physics-based modeling.
10. A stochastic multivariate extreme-value model for storm and wave climate off the west coast of India
Core Problem: Quantifying the variability and multivariate extremes of storm and wave variables to assess coastal vulnerability and inform resilient coastal management strategies against increasing erosion from storm-driven sea states, requiring joint probabilistic treatments.
Key Innovation: A reproducible stochastic multivariate extreme-value framework that couples a varying non-homogeneous Poisson process, generalized Pareto distribution tail fits, copula dependence modeling, empirical directional sampling, and Monte Carlo simulation to generate synthetic storm catalogues, providing design-relevant return estimates with uncertainty for six storm and wave variables.
11. Numerical study on seismic dynamics of shallowly-buried immersed tunnel in sandy-clay seabed foundation
Core Problem: Understanding the dynamic response and potential failure modes of immersed tunnels in sandy-clay seabed foundations under seismic activity is crucial for effective aseismic design.
Key Innovation: Employs coupled numerical software FssiCAS with PZIII and Soft Clay models to analyze seismic behavior, identifying liquefaction in the sand stratum, its isolating effect, and three primary instability modes at tunnel joints, providing insights for performance-based design.
12. Spatial Characteristics and Controlling Factors of Permafrost Deformation in the Qinghai–Tibet Plateau Revealed Through InSAR Measurements
Core Problem: The spatial patterns and controlling factors of permafrost deformation across the entire Qinghai–Tibet Plateau (QTP) remain poorly understood, despite accelerated climate warming intensifying permafrost degradation.
Key Innovation: Time-series InSAR (Sentinel-1 SAR imagery) combined with a permafrost deformation model is used to reveal long-term and seasonal deformation characteristics across the QTP, providing the first plateau-wide characterization of seasonal deformation. Dominant controlling factors (solar radiation, precipitation, slope) were identified using the geographic detector method.
13. Debris flow vulnerability assessment in the Eastern Tibetan Plateau using integrated UAV-SLAM 3D reconstruction: Shuihaizigou Gully case study
Core Problem: Quantifying the vulnerability of buildings to debris flow events, specifically establishing the relationship between debris flow parameters (flow depth, velocity, impact pressure) and building damage in the Eastern Tibetan Plateau.
Key Innovation: Integration of UAV-SLAM 3D reconstruction to create high-precision models of damaged buildings, enabling the establishment of a building damage database and derivation of vulnerability curves for debris flows, and a comparative analysis of Torrential Fan models to improve spatial vulnerability assessment accuracy.
14. Modelling the response of Sevastopol bays (Black Sea) to tsunami and meteotsunami impacts
Core Problem: The serious hazard posed by tsunamis and meteotsunamis to coastal zones, leading to inundation and damage, and the need to understand their impact and generated seiche oscillations in complex bay systems.
Key Innovation: Application of nonlinear shallow water equations and a two-stage numerical modeling approach (Black Sea basin model + SWASH model) to study the response of Sevastopol bays to tsunami and meteotsunami impacts, revealing the generation and penetration of intense seiche oscillations (Helmholtz modes) that pose a hazard to vessels and coastal infrastructure.
15. Electrical properties and desaturation effect of horizontal electrolysis in calcareous sand: influence of soil gradation and relative density
Core Problem: The need for innovative induced partial saturation techniques for liquefaction mitigation, and understanding the electrical characteristics and desaturation performance of electrolysis in various sand types.
Key Innovation: Investigation of horizontal electrolysis desaturation in calcareous sands with varying gradations and relative densities, revealing the influence of soil properties on electrical response and desaturation effect, calibrating saturation exponent for resistivity measurements, and analyzing power consumption for practical liquefaction mitigation applications.
16. Scalable Pluvial Flood Risk Assessment: A Data-Driven Framework Integrating Machine Learning (ML) and Discrete Global Grid Systems (DGGS H3)
Core Problem: The escalating frequency of extreme rainfall events necessitates scalable and dynamic tools for urban flood risk assessment, particularly for operationalizing building-level susceptibility indices efficiently across multiple resolutions.
Key Innovation: Development of a data-driven framework integrating machine learning (for building-level susceptibility) and Discrete Global Grid Systems (H3) for scalable, multi-resolution aggregation of pluvial flood susceptibility, significantly improving computational efficiency and supporting consistent multi-scale communication for risk assessment.
17. Connecting perceived flood risk and resilience in Auckland, New Zealand
Core Problem: An existing knowledge gap regarding the interrelationships between perceived flood risk and perceived urban flood resilience among residents, hindering alignment between residents' needs and policymaking.
Key Innovation: Investigation of interrelationships between perceived flood risk and perceived urban flood resilience using survey and network analysis, identifying a 'Flood Resilience Perception cluster' and highlighting cognitive, behavioral, sociocultural, and geographic mediators, offering insights for participatory flood risk governance.
18. Deformation, strains and velocities for the Alpine Himalayan Belt from trans-continental Sentinel-1 InSAR & GNSS
Core Problem: Spatially sparse data from previous GNSS studies limit the characterization of shorter wavelength deformation features across large regions of distributed continental deformation like the Alpine-Himalayan Belt, hindering understanding of tectonic deformation, faulting, and seismic hazard.
Key Innovation: Provides trans-national average surface velocities and time series at 1 km spatial resolution over 11,000 km of the Alpine-Himalayan Belt by combining over 222,000 Sentinel-1 SAR images with a new GNSS compilation, yielding near-continuous horizontal strain rates and revealing details of continental deformation and seismic hazard.
19. 3D analytical modeling of seismic ground motion and energy partitioning in the marine site subjected to obliquely P-SV waves
Core Problem: Understanding seismic wave propagation, ground motion, and energy partitioning in 3D marine porous media, considering factors like pore shape and saturation, to better predict seabed response.
Key Innovation: Developed an analytical solution for seismic wave propagation and ground motion in a 3D marine porous medium, incorporating pore shape and saturation. Systematically analyzed the dependence of seabed displacement, dynamic pore pressure, wave velocity, and energy partitioning on incident wave angle and material properties. Identified key effects of pore shape, porosity, and saturation on stiffness, fluid-solid coupling, and wave reflection/transmission.
20. Upper Mantle Heterogeneity and Weak Subduction Boundaries Control Crustal Stress in the Korean Peninsula
Core Problem: The dominant causes of intraplate earthquakes and the observed seismicity distribution and stress orientations in stable continental interiors like the Korean Peninsula remain elusive, with lithospheric thickness variations alone being insufficient to explain them.
Key Innovation: Three-dimensional numerical models extending to 650km depth, incorporating weak subduction interfaces, slabs extending into the transition zone, and mantle buoyancy, successfully explain the observed seismicity and stress distribution in the Korean Peninsula, highlighting the major role of deep mantle structures and plate boundary effects.
21. Uncertainty-Calibrated Spatiotemporal Field Diffusion with Sparse Supervision
Core Problem: Forecasting and reconstructing physical fields from sparse, time-varying sensor observations is an ill-posed and uncertainty-critical problem, and prior work often relies on dense reanalysis or simulations for training, not directly addressing sparse supervision.
Key Innovation: Presents SOLID, a mask-conditioned diffusion framework that learns spatiotemporal dynamics and provides uncertainty-calibrated posterior sampling of full fields directly from sparse observations alone, using a dual-masking objective to emphasize learning in unobserved regions and upweight reliable anchors, achieving significant improvements in probabilistic error and calibrated uncertainty maps.
22. TianQuan-S2S: A Subseasonal-to-Seasonal Global Weather Model via Incorporate Climatology State
Core Problem: Accurate Subseasonal-to-Seasonal (S2S) weather forecasting is challenging due to the chaotic nature of weather systems, and current data-driven models suffer from inadequate incorporation of climate states and a tendency to produce over-smoothed forecasts.
Key Innovation: Proposes TianQuan-S2S, a global S2S forecasting model that integrates initial weather states with climatological means via patch embedding and enhances variability capture through an uncertainty-augmented Transformer, achieving significant improvements over traditional numerical and advanced data-driven models.
23. Structural safety and energy performance of floating offshore photovoltaic systems with stochastic anchor position deviations
Core Problem: Quantitatively evaluating the impact of stochastic anchor position deviations on the structural safety and energy performance of Floating Offshore Photovoltaic (FOPV) systems under extreme environmental conditions, given that anchors cannot always be installed precisely.
Key Innovation: A methodology using stochastic sampling, time-domain simulations, and statistical modeling to evaluate the effects of anchor position deviations on FOPV systems, revealing significant right-skewed distributions in maximum mooring line tension and horizontal platform displacement, leading to a substantial increase in structural risk, while power generation remains largely insensitive.
24. Numerical and experimental investigation of offshore breakwater with detachable top structures in wave-current environments
Core Problem: Coastal erosion and wave impact require effective mitigation strategies, and traditional breakwaters may have limitations in complex wave-current environments.
Key Innovation: Investigates a novel offshore composite breakwater with a submerged rubble mound base and a detachable superstructure (ballast pontoon and dual porous barriers), demonstrating superior hydrodynamic performance in wave transmission and energy dissipation under combined wave-current conditions.
25. UK-Flow15 Part 1: Development of a coherent national-scale 15-min flow dataset
Core Problem: High-resolution river flow data in the UK are scattered across multiple agencies, lack consistent quality assurance, and are thus limited for large-sample and national-scale hydrological analysis, despite a wealth of records.
Key Innovation: UK-Flow15, a quality-controlled, 15-min, national-scale river flow dataset for the UK, was developed from over 1,300 gauging stations. It includes systematic identification and resolution of data inconsistencies, a comprehensive quality-control framework, and transparent documentation, significantly enhancing capabilities for sub-daily hydrological research and improved flood prediction.
26. An evaluation of near-surface sediments enhancing ground motions excited by shallow local earthquakes near Union City, Oklahoma
Core Problem: Identifying and delineating subsurface structures of near-surface sediments that amplify ground motion excited by shallow local earthquakes, which contribute to damage to the built environment in areas like Union City, Oklahoma.
Key Innovation: Integration of ambient noise data (HVSR) and high-resolution electrical resistivity tomography (ERT) to deduce dominant resonant frequencies and resolve sedimentary structures responsible for ground motion amplification, linking higher frequencies to terrace deposits and lower frequencies to deeper clay layers and impedance contrasts between Permian and Pennsylvanian deposits.
27. Theoretical and experimental study on multiscale coupling shear strength of soil–rock mixture
Core Problem: The complex multiscale coupling shear strength of soil-rock mixtures (S-RM), which are natural geomaterials with hierarchical features, and the need for a model to accurately predict their mechanical behavior.
Key Innovation: Proposal of a multiscale soil-rock cell element model and conducting in situ large-scale direct shear tests to investigate the multiscale coupling shear strength of S-RM, revealing the trans-scale coordinated deformation and microcrack coordination that enhance shear strength, and accurately predicting S-RM shear strength.
28. Time-Dependent System Reliability Analysis of Tunnels Accounting for Degradation of Surrounding Rock and Support Systems
Core Problem: Complex environmental factors lead to time-dependent degradation of tunnel surrounding rock and supporting structures, making it critical to accurately assess the long-term system reliability and identify dominant failure modes for tunnel safety.
Key Innovation: Developed a mechanical model incorporating time-dependent degradation of surrounding rock and supporting structures (considering degradation coefficient, damage coefficient, and carbonation depth) to establish a time-dependent system performance function for three failure modes. Quantified the significant influence of degradation and damage coefficients on failure probability (e.g., 1656.49% increase for insufficient support strength) and identified critical failure modes at different service stages.
29. Experimental investigation of structural additions and removals on the seismic response of building clusters in soft soil
Core Problem: Rapid urbanization involving structural additions and removals within building clusters can significantly alter local site dynamic characteristics and seismic responses, especially in soft soil environments, posing challenges for seismic design and urban planning.
Key Innovation: Conducted physical model tests with eight structure cluster configurations and 54 seismic loading scenarios to analyze dynamic characteristics, site soil magnification, and acceleration responses. Demonstrated that structure clusters intensify seismic amplification in shallow soil, and that the central structure plays a pivotal role in seismic energy transmission, offering critical insights for incorporating site-structure cluster interaction (SSCI) into seismic design.
30. A novel infrared method for segregation frost heaving temperature fields
Core Problem: Accurate, non-invasive, and real-time temperature monitoring in critical zones (segregation cracks, freezing front) during frost heave in frozen soil is challenging due to sensor limitations and displacement.
Key Innovation: A high-precision temperature field reconstruction method combining limited temperature sensors and infrared thermography with dynamic correction, enabling non-destructive, real-time measurement of full cross-sectional temperature distribution in frozen soil, including frost crack edges.
31. Seismic response of disconnected piled raft foundation during a strong earthquake based on dynamic centrifuge tests
Core Problem: Understanding the seismic performance and bending moment distribution in disconnected piled raft (DPR) foundations during strong earthquakes, especially for buildings with high aspect ratios, to prevent underestimation of pile stress.
Key Innovation: Conducted dynamic centrifuge tests on a soil–foundation–superstructure system to analyze DPR seismic response. Distinguished static and dynamic components of pile bending moment, highlighting the impact of cyclic one-sided lateral loads. Demonstrated that DPR significantly reduces axial loads on piles compared to conventional pile foundations, preventing tensile loads and limiting total compressive load.
32. Leveraging large language models for automated knowledge extraction from geological reports
Core Problem: Geological reports contain abundant, unstructured textual data, making it challenging to efficiently extract meaningful information for engineering decision-making and geological risk assessment.
Key Innovation: Benchmarked state-of-the-art LLMs for knowledge graph construction and question answering from geological texts, demonstrating their potential to automate knowledge extraction and support geological risk management, while also identifying effective prompt engineering strategies and developing an interactive platform.
33. MAD-SmaAt-GNet: A Multimodal Advection-Guided Neural Network for Precipitation Nowcasting
Core Problem: Precipitation nowcasting is often computationally expensive with numerical solvers and underutilizes large weather datasets, while deep learning models need to ensure physical consistency and leverage multiple weather variables effectively.
Key Innovation: Proposes MAD-SmaAt-GNet, a multimodal advection-guided neural network for precipitation nowcasting. It extends SmaAt-UNet by incorporating an additional encoder for multiple weather variables and integrating a physics-based advection component to ensure physically consistent predictions, significantly improving rainfall forecasts.
34. Distant Object Localisation from Noisy Image Segmentation Sequences
Core Problem: Localizing distant objects from camera measurements for safety-critical surveillance (e.g., drone-based wildfire monitoring) is challenging when specialized sensor configurations or 3D scene reconstruction are not feasible due to distance or computational limits.
Key Innovation: Demonstrates that 3D object localization can be reliably solved using multi-view triangulation or particle filters (providing shape and uncertainty estimates) from noisy image segmentation sequences, enabling a robust system for drone-based wildfire monitoring.
35. Progressive failure of backfill behind the integral bridge abutment due to seasonal temperature variation
Core Problem: Understanding the progressive soil failure mechanisms in backfills behind integral bridge abutments subjected to cyclic movements caused by seasonal temperature variations.
Key Innovation: Model tests integrating particle image velocimetry and particle tracking velocimetry revealed a three-stage progressive failure process (active, alternating active/passive, and failure surface migration stages) and the underlying mechanisms, including accumulation of residual particle displacements and vortex-like soil flow, providing insights for mitigating earth pressure ratcheting and excessive ground deformation.
36. Numerical investigation of inclined loading and failure envelopes of helical anchors in clay over sand
Core Problem: Understanding the behavior and pullout capacity of helical anchors under inclined loading in stratified soil (clay over sand), as existing approaches mainly focus on vertical loading or homogeneous soil conditions.
Key Innovation: A 3D finite element analysis framework that investigates the influence of various factors on helical anchor performance in clay over sand, revealing significant reductions in pullout capacity compared to single-layer sand, and establishing a practical framework for estimating failure envelopes that integrates combined effects of load inclination and soil stratification.
37. Climate change will increase forest disturbances in Europe throughout the 21st century
Core Problem: The future trajectories and interactions of climate change-induced forest disturbances (wildfires, convective storms, bark beetle outbreaks) and their impact on tree mortality in European forests are largely unknown.
Key Innovation: Projection of increased disturbance rates and tree mortality across Europe in the 21st century under various climate change scenarios, particularly highlighting increased wildfire-induced mortality and its implications for forest structure, carbon sequestration, and biodiversity.
38. Integrating ecosystem adaptability into drought resilience assessment: a case study of the Yellow River Basin, China
Core Problem: Current drought resilience assessment frameworks often overlook ecosystem adaptability, potentially leading to overestimation of resilience loss and misjudgment of ecosystem collapse thresholds.
Key Innovation: Developed an integrated framework that incorporates adaptability as a third core dimension alongside resistance and recovery for drought resilience assessment, providing a novel three-dimensional perspective and detecting early-phase resilience decline in the Yellow River Basin.
39. Resonant effects of seismicity on substance migration in rock fractures
Core Problem: The effects of seismicity-induced oscillatory flow on fluid motion and solute transport across rough rock fractures, and how these disturbances influence macrodispersion and substance migration in groundwater systems.
Key Innovation: Numerical investigation revealing that macrodispersion of fluid-borne solute is significantly enhanced (resonance) when the period of the oscillating flow field approaches the flushing time of the fracture system. Predictive models are developed to quantify these resonance effects, providing a pore-scale perspective on seismic impacts on subsurface flow and substance migration.
40. Dynamic triaxial tests of reinforced silty clay under the combined action of freeze-thaw cycles and train loads
Core Problem: Addressing the engineering challenges posed by non-uniform deformation of high-speed railway roadbeds in seasonally frozen soil regions due to repeated freeze-thaw cycles and train loads, which threaten operational safety.
Key Innovation: Systematically investigated the dynamic response of reinforced silty clay under combined freeze-thaw and train loads using dynamic triaxial tests. Revealed interaction mechanisms of loading frequency, freeze-thaw cycles, and geogrid layers on soil behavior. Demonstrated the dual improvement effect of geogrid reinforcement in suppressing freeze-thaw damage and reducing frequency sensitivity, providing support for railway subgrade design in cold regions.
41. Detecting Morphological Change From DInSAR Data in Ephemeral, Braided, Gravel Bed Rivers. Application to the Ungauged Trionto River, Italy
Core Problem: Characterizing the geometry and morphological change (active width, number of active branches) of ephemeral, braided, gravel bed rivers, especially regarding bedload transport, remains a problem.
Key Innovation: Presents a procedure to estimate morphological change in such rivers from DInSAR data (Sentinel-1 SAR images), demonstrating its potential to capture different channel dynamics related to lateral confinement and flood hydrographs.
42. Machine Learning for Complex Systems Dynamics: Detecting Bifurcations in Dynamical Systems with Deep Neural Networks
Core Problem: Detecting critical transitions and bifurcations in complex dynamical systems typically requires computationally intensive forward simulations or bifurcation analyses, which are often limited by parameter sampling.
Key Innovation: Proposes equilibrium-informed neural networks (EINNs) that infer system parameters from candidate equilibrium states, allowing for the detection of critical thresholds associated with catastrophic regime shifts by analyzing abrupt changes in the learned parameter landscape, offering a flexible alternative to traditional techniques.
43. Fusion and Grouping Strategies in Deep Learning for Local Climate Zone Classification of Multimodal Remote Sensing Data
Core Problem: Improving the accuracy of Local Climate Zone (LCZ) classification from multimodal remote sensing data (SAR and MSI) by comprehensively analyzing different deep learning fusion and grouping strategies.
Key Innovation: Analyzing various deep learning fusion strategies (pixel-, feature-, decision-level) and grouping strategies (band grouping, label merging) for LCZ classification, demonstrating that a hybrid fusion model with band grouping and label merging achieves the highest accuracy and improves prediction for underrepresented classes.
44. When Sensors Fail: Temporal Sequence Models for Robust PPO under Sensor Drift
Core Problem: Real-world reinforcement learning systems, including those for monitoring critical infrastructure or environments, struggle with distributional drift and partial observability caused by temporally persistent sensor failures, leading to degraded performance.
Key Innovation: Augments Proximal Policy Optimization (PPO) with temporal sequence models (Transformers, SSMs) to enable policies to infer missing information from historical data and maintain performance under stochastic sensor failure, demonstrating substantial robustness improvements on continuous-control benchmarks.
45. RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery
Core Problem: Rotated object detection in remote sensing imagery is hindered by non-adaptive receptive field utilization, inadequate long-range multi-scale feature fusion, and discontinuities in angle regression.
Key Innovation: Proposes Rotated Multi-Kernel RetinaNet (RMK RetinaNet), which includes a Multi-Scale Kernel (MSK) Block for adaptive feature extraction, a Multi-Directional Contextual Anchor Attention (MDCAA) for enhanced contextual modeling, a Bottom-up Path for preserving fine-grained spatial details, and an Euler Angle Encoding Module (EAEM) for continuous angle regression, achieving robust and state-of-the-art performance.
46. GloSplat: Joint Pose-Appearance Optimization for Faster and More Accurate 3D Reconstruction
Core Problem: Traditional 3D reconstruction methods treat feature extraction, matching, SfM, and NVS as separate problems, leading to independent optimization objectives and limitations in joint pose-appearance refinement, especially in early stages.
Key Innovation: Introduced GloSplat, a framework for joint pose-appearance optimization during 3D Gaussian Splatting training. It uniquely preserves explicit SfM feature tracks as first-class optimizable entities, providing persistent geometric anchors via a reprojection loss alongside photometric supervision, which prevents early-stage pose drift and enables fine-grained refinement, leading to faster and more accurate 3D reconstruction.
47. Quantifying Salt Precipitation During CO2 Injection: How Flow Rate, Temperature, and Phase State Control Near-Wellbore Crystallization
Core Problem: Salt precipitation near CO2 injection wells reduces permeability and injectivity, but the pore-scale mechanisms coupling multiphase flow, evaporation, and crystallization across variable phase states and flow regimes are not sufficiently quantified.
Key Innovation: Presents high-resolution microfluidic experiments that systematically quantify halite crystallization dynamics during CO2-driven brine evaporation across liquid, gaseous, and supercritical phases, establishing quantitative relationships between dimensionless parameters, kinetic constants, and phase-dependent displacement patterns to predict near-wellbore permeability impairment.
48. SPyCer: Semi-Supervised Physics-Guided Contextual Attention for Near-Surface Air Temperature Estimation from Satellite Imagery
Core Problem: Near-ground sensors provide accurate measurements of near-surface air temperature (NSAT) but are sparse and unevenly distributed, limiting their ability to provide continuous spatial measurements despite detailed surface properties captured by satellites.
Key Innovation: Introduces SPyCer, a semi-supervised physics-guided network that leverages pixel information and physical modeling (surface energy balance, advection-diffusion-reaction PDEs) to guide the learning process for continuous NSAT estimation from satellite imagery. It employs multi-head attention guided by land cover characteristics and modulated with Gaussian distance weighting to produce spatially coherent and physically consistent NSAT estimates.
49. GALACTIC: Global and Local Agnostic Counterfactuals for Time-series Clustering
Core Problem: Existing explainability methods for time-series clustering fail to identify the specific temporal transitions that cause an instance to move across cluster boundaries, and counterfactual explanations have been mostly confined to supervised settings.
Key Innovation: Introduces GALACTIC, the first unified framework for local and global counterfactual explainability in unsupervised time-series clustering. It generates instance-level perturbations via a cluster-aware optimization and formulates a Minimum Description Length (MDL) objective for selecting a non-redundant summary of global explanations, enabling efficient greedy selection with provable guarantees.
50. Dark3R: Learning Structure from Motion in the Dark
Core Problem: Conventional feature- and learning-based structure from motion (SfM) methods fail in extremely low-light conditions (SNR below -4 dB), making 3D reconstruction challenging in such environments.
Key Innovation: Introduces Dark3R, a framework for SfM in the dark that adapts large-scale 3D foundation models to extreme low-light conditions through a teacher-student distillation process. It enables robust feature matching and camera pose estimation directly on raw, noisy images without 3D supervision, and also facilitates state-of-the-art novel view synthesis in the dark.
51. Weather-Related Crash Risk Forecasting: A Deep Learning Approach for Heterogenous Spatiotemporal Data
Core Problem: Forecasting weather-related traffic crash risk is challenging due to the complex, non-linear relationships between crash occurrence and heterogeneous spatiotemporal factors like road characteristics, weather, and traffic conditions.
Key Innovation: Introduces a deep learning framework using an ensemble of Convolutional Long Short-Term Memory (ConvLSTM) models trained over overlapping spatial grids to capture both spatial dependencies and temporal dynamics for weather-related crash risk forecasting, outperforming baselines, especially in high-risk zones.
52. Global versus local internal-external field separation on the sphere: a Hardy-Hodge perspective
Core Problem: Achieving a unique and stable internal-external magnetic field separation when geophysical data is only available in a subdomain of the observation surface, unlike the straightforward procedure for full spherical data.
Key Innovation: Demonstrates that unique separation is possible but highly unstable under the assumption of exterior sources above a source-free spherical shell, providing a theoretical explanation for these intrinsic difficulties based on a Hardy-Hodge decomposition of spherical vector fields.
53. Loop Closure via Maximal Cliques in 3D LiDAR-Based SLAM
Core Problem: Reliable loop closure detection in 3D LiDAR-based SLAM remains a critical challenge under sensor noise, environmental ambiguity, and viewpoint variation, with RANSAC-based approaches often failing.
Key Innovation: A novel deterministic algorithm, CliReg, is introduced for loop closure validation, replacing RANSAC verification with a maximal clique search over a compatibility graph of feature correspondences, demonstrating lower pose error and more reliable loop closures, especially in sparse or ambiguous conditions.
54. RoboPocket: Improve Robot Policies Instantly with Your Phone
Core Problem: Scaling imitation learning is constrained by inefficient data collection, as handheld interfaces operate open-loop, and interactive methods like DAgger require costly physical robot execution.
Key Innovation: RoboPocket is introduced as a portable system for Robot-Free Instant Policy Iteration using smartphones, featuring a Remote Inference framework that visualizes policy trajectories via AR Visual Foresight, allowing collectors to proactively identify failures and focus data collection, coupled with an asynchronous Online Finetuning pipeline for continuous policy updates.
55. FLAIR-HUB: Large-scale Multimodal Dataset for Land Cover and Crop Mapping
Core Problem: The challenge of exploiting the volume and heterogeneity of high-quality Earth Observation data for accurate global land cover and crop type monitoring, requiring large annotated datasets.
Key Innovation: Introduction of FLAIR-HUB, the largest multi-sensor land cover dataset with very-high-resolution annotations (20 cm) covering 2528 km2 of France, combining six aligned modalities and supporting multimodal fusion and deep learning benchmarks.
56. Overtone: Cyclic Patch Modulation for Clean, Efficient, and Flexible Physics Emulators
Core Problem: Transformer-based PDE surrogates suffer from systematic error accumulation at harmonic frequencies due to fixed patch sizes and lack flexibility in computational costs, hindering their efficiency and accuracy in long rollouts.
Key Innovation: Overtone, a unified solution employing cyclic patch modulation (dynamic patch size control) during autoregressive rollouts, which mitigates systematic harmonic error accumulation and enables flexible, compute-adaptive deployment of physics emulators, leading to significantly lower long rollout errors.
57. Gated Differential Linear Attention: A Linear-Time Decoder for High-Fidelity Medical Segmentation
Core Problem: Medical image segmentation requires models that can preserve fine anatomical boundaries efficiently, but existing methods (Transformers, CNNs, linear attention) struggle with either quadratic cost, global reasoning, or training instability/attention dilution.
Key Innovation: Introduces PVT-GDLA, a decoder-centric Transformer featuring Gated Differential Linear Attention (GDLA), which computes two kernelized attention paths with a learnable subtraction and a lightweight gate, combined with a local token-mixing branch, achieving state-of-the-art accuracy and efficiency for high-fidelity medical segmentation.
58. Performance of basic methods for reconstructing the surface displacement from bottom pressure: numerical simulation of irregular directional waves
Core Problem: Accurately assessing statistical surface wave parameters from bottom pressure data using hydrostatic and linear dispersive theories, especially for irregular directional waves and varying water depths, needs validation.
Key Innovation: Tests the accuracy of hydrostatic and linear dispersive theories for reconstructing surface wave parameters from bottom pressure data through direct numerical simulations, showing that hydrostatic theory provides robust estimates in a broad range of conditions and suggesting empirical mappings for accurate reconstruction.
59. LCTNet: Lightweight Convolution-Transformer Network for Hyperspectral Image Classification
Core Problem: Hyperspectral Image (HSI) classification using deep learning models, especially hybrid CNN-Transformer models, often faces challenges with complex structures, high spectral dimensions, and computational complexity, impacting accuracy and efficiency.
Key Innovation: Proposed LCTNet, a lightweight CNN-Transformer network, featuring a spectral extraction-dimension reduction module, a spatial-spectral feature marking module, and a dynamic sparse transformer module, which collectively improve HSI classification accuracy and reduce computational complexity, achieving high accuracy on public datasets.
60. UAVRS-CompNet: An End-to-End Compression Network for UAV Remote Sensing Images
Core Problem: UAV remote sensing images, characterized by high resolution, complex textures, and multiscale features, pose significant challenges for efficient compression and high-fidelity reconstruction compared to natural images.
Key Innovation: Proposed UAVRS-CompNet, an end-to-end deep-learning-based compression network, which uses a multiscale residual bidirectional pyramid backbone and a dual-branch synergistic attention module to enhance feature extraction, reconstruction, and adaptive representation, outperforming state-of-the-art methods in compressing UAV remote sensing imagery.
61. The Potsdam Soil Moisture Observatory: high-coverage reference observations at kilometer scale
Core Problem: The lack of accurate, intermediate-scale, and long-term soil moisture reference datasets for validating remote sensing algorithms and improving hydrological models, due to small-scale heterogeneity and sparse point-scale sensors.
Key Innovation: Establishment of the Potsdam Soil Moisture Observatory (PoSMO), providing a harmonized, high-coverage (approx. 1 km²), and high-density multiscale soil moisture dataset (CRNS, shallow, profile data, manual sampling) for 16 stationary CRNS sensors, serving as a crucial reference for remote sensing products and hydrological models.
62. Assessing seasonal climate predictability using a deep learning application: NN4CAST
Core Problem: Dynamical models often struggle with biases and low signal-to-noise ratios in seasonal climate predictions, and the "black-box" nature of deep learning models makes their application in climate services challenging without explainability.
Key Innovation: Introduction of NN4CAST, a Python deep learning pipeline designed to assess seasonal climate predictability, offering built-in tools for skill evaluation and spatial diagnostics, demonstrating skillful predictions across timescales (lag 0 to longer leads) for both linear and highly non-linear teleconnections, and enabling attribution of predictions to specific input features for improved understanding of climate interactions.
63. Climate research is global — risks and responsibilities should also be distributed
Core Problem: The global nature of climate research and its associated risks and responsibilities are not equitably distributed among nations.
Key Innovation: Advocates for a more distributed approach to climate research, risks, and responsibilities to reflect its global impact and implications for various hazards.
64. Preferential flow reduces overland flow on slopes: insights from a field experiment on the Chinese Loess Plateau
Core Problem: The influence of preferential flow on slope runoff, particularly under different vegetation restoration types and ages on the Chinese Loess Plateau, and its underlying mechanisms, remain unclear.
Key Innovation: Conducted field experiments using artificial rainfall and high-frequency monitoring to quantify the occurrence and contribution of preferential flow patterns under varying vegetation restoration conditions, demonstrating that vegetation restoration significantly increases preferential flow, which in turn reduces slope runoff by enhancing infiltration.
65. Numerical analysis of the influence of electric potential gradient application history on electroosmotic consolidation of soft clay
Core Problem: Optimizing the application of electric potential gradient (EPG) in electroosmotic consolidation of soft clay, which currently relies on empirical staging, by developing a mechanistic understanding of EPG application history.
Key Innovation: Proposed a 'EPG application history' paradigm for mechanically informed design of voltage trajectories. Developed and validated a coupled electric-seepage-stress numerical model based on Biot’s theory, incorporating soil parameter nonlinearity. Systematically investigated drainage, settlement, and pore water pressure under various EPG schemes, demonstrating the importance of application sequence and mechanistic matching between stages for drainage channel stability and optimal consolidation.
66. A New GRIME2: Using an Octagon Calibration Target and Trail Camera to Measure Stream Water Level Over a 2‐Year Period
Core Problem: The need for reliable, easy-to-use, and accurate methods for long-term stream water level measurement using ground-based time-lapse imagery, while accounting for issues like camera movement, biofouling, and environmental conditions.
Key Innovation: Develops a new GRIME2 system integrating an octagon calibration target with open-source software, which greatly eases field setup and image processing, automatically adapts to camera movement, and delivers consistent and accurate water level measurements over long periods.
67. Decorrelating the Future: Joint Frequency Domain Learning for Spatio-temporal Forecasting
Core Problem: Standard direct forecasting models using point-wise objectives fail to capture complex spatio-temporal dependencies in graph-structured signals, and existing frequency-domain methods often overlook spatial and cross spatio-temporal interactions.
Key Innovation: Proposes FreST Loss, a frequency-enhanced spatio-temporal training objective that extends supervision to the joint spatio-temporal spectrum using the Joint Fourier Transform, effectively decorrelating complex dependencies across both space and time and consistently improving state-of-the-art baselines.
68. Flowers: A Warp Drive for Neural PDE Solvers
Core Problem: Traditional neural PDE solvers often rely on Fourier multipliers, dot-product attention, or convolutional mixing, which may not be optimal for capturing complex, adaptive, and global interactions in PDE solutions.
Key Innovation: Introduces 'Flowers,' a neural architecture built entirely from multihead warps that predict displacement fields pointwise, enabling adaptive, global interactions at linear cost without Fourier multipliers or dot-product attention, achieving excellent performance on 2D/3D time-dependent PDEs, particularly flows and waves.
69. Mask-aware inference with State-Space Models
Core Problem: State Space Models (SSMs) lack an inherent mechanism to handle arbitrarily shaped regions of missing or invalid data during inference, which is common in real-world computer vision tasks like depth completion.
Key Innovation: Introduces Partial Vision Mamba (PVM), a novel architectural component that ports the principles of partial operations to the Mamba backbone, enabling mask-aware inference for SSMs and demonstrating efficacy in tasks like depth completion and image inpainting.
70. SGR3 Model: Scene Graph Retrieval-Reasoning Model in 3D
Core Problem: Existing 3D scene graph generation methods often require multi-modal data, rely on heuristic graph construction, and struggle with robust relational reasoning, limiting their applicability and accuracy.
Key Innovation: Introduces SGR3 Model, a training-free framework that leverages multi-modal large language models (MLLMs) with retrieval-augmented generation (RAG) for semantic 3D scene graph generation, enhancing relational reasoning by incorporating semantically aligned scene graphs and a weighted patch-level similarity selection mechanism.
71. Direct Estimation of Tree Volume and Aboveground Biomass Using Deep Regression with Synthetic Lidar Data
Core Problem: Existing methods for forest biomass estimation rely on indirect allometric models, which are limited in accuracy due to measurement uncertainties and the approximate nature of equations, failing to fully account for tree and forest variability.
Key Innovation: Proposes a direct approach leveraging synthetic 3D forest plot point cloud data (generated via a lidar simulator) to train deep regression networks (e.g., PointNet, PointNet++, DGCNN, PointConv). These networks are then applied to real lidar data for plot-level wood volume and Aboveground Biomass (AGB) estimation, showing significantly lower discrepancies compared to indirect allometric and FullCAM approaches.
72. A Benchmark Study of Neural Network Compression Methods for Hyperspectral Image Classification
Core Problem: Deep neural networks for hyperspectral image classification have high computational and memory requirements, limiting their deployment on resource-constrained remote sensing devices and edge systems.
Key Innovation: Conducts a systematic evaluation of neural network compression methods (pruning, quantization, knowledge distillation) for hyperspectral land cover classification. Demonstrates that compressed models significantly reduce model size and computational cost while maintaining competitive classification performance, providing insights into trade-offs for efficient deep learning deployment in remote sensing.
73. Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset
Core Problem: Real-world infrared image super-resolution (IISR) is challenging due to coupled optical and sensing degradations that jointly deteriorate structural sharpness and thermal fidelity, and existing works often rely on simulated datasets.
Key Innovation: Proposes Real-IISR, a unified autoregressive framework that progressively reconstructs fine-grained thermal structures and clear backgrounds using thermal-structural guided visual autoregression, a Condition-Adaptive Codebook, and a Thermal Order Consistency Loss, along with a new real-world FLIR-IISR dataset.
74. SURE: Semi-dense Uncertainty-REfined Feature Matching
Core Problem: Existing image correspondence methods struggle in challenging scenarios (large viewpoint changes, textureless regions) because they rely solely on feature similarity, leading to overconfident errors and a lack of explicit reliability estimation for predicted matches.
Key Innovation: Proposed SURE, a Semi-dense Uncertainty-REfined matching framework, which jointly predicts correspondences and their confidence by explicitly modeling both aleatoric and epistemic uncertainties. It introduces a novel evidential head for trustworthy coordinate regression and a lightweight spatial fusion module, significantly outperforming state-of-the-art semi-dense matching models in accuracy and efficiency.
75. TAPFormer: Robust Arbitrary Point Tracking via Transient Asynchronous Fusion of Frames and Events
Core Problem: Combining RGB frames and event streams for arbitrary point tracking is challenged by temporal misalignment and severe degradation when one modality fails, due to synchronous or non-adaptive fusion strategies.
Key Innovation: Introduces TAPFormer, a transformer-based framework for robust and high-frequency arbitrary point tracking. Its key innovation is a Transient Asynchronous Fusion (TAF) mechanism that explicitly models temporal evolution between discrete frames via continuous event updates, and a Cross-modal Locally Weighted Fusion (CLWF) module that adaptively adjusts spatial attention based on modality reliability, outperforming existing trackers and introducing a new real-world frame-event TAP dataset.
76. MI-DETR: A Strong Baseline for Moving Infrared Small Target Detection with Bio-Inspired Motion Integration
Core Problem: Infrared Small Target Detection (ISTD) is challenging because tiny, low-contrast targets are easily obscured by complex and dynamic backgrounds, and conventional multi-frame approaches often require additional motion supervision or explicit alignment.
Key Innovation: MI-DETR, a bio-inspired dual-pathway detector, explicitly models motion by converting infrared frame sequences into a motion map using a retina-inspired cellular automaton (RCA) and facilitates feature interaction via a Parvocellular-Magnocellular Interconnection (PMI) Block, achieving strong performance on ISTD benchmarks.
77. Generic Camera Calibration using Blurry Images
Core Problem: Calibrating generic camera models accurately requires many images, making motion blur practically unavoidable for individual users, which complicates feature estimation.
Key Innovation: Presents a method for generic camera calibration using blurry images that simultaneously estimates feature locations and spatially varying point spread functions, while resolving translational ambiguity, by leveraging geometric constraints and a local parametric illumination model.
78. Mario: Multimodal Graph Reasoning with Large Language Models
Core Problem: Existing LLM-based multimodal reasoning methods encode image-text pairs in isolation, ignoring the relational structure of real-world multimodal data (multimodal graphs), leading to weak cross-modal consistency and issues with heterogeneous modality preference.
Key Innovation: Proposes Mario, a unified framework with a graph-conditioned VLM design for joint refinement of textual and visual features through fine-grained cross-modal contrastive learning, and a modality-adaptive graph instruction tuning mechanism that uses a learnable router to surface the most informative modality configuration to the LLM for each node and its neighborhood.
79. Towards a data-scale independent regulariser for robust sparse identification of non-linear dynamics
Core Problem: Data normalization, a common preprocessing step, can severely distort the discovery of governing equations by magnitude-based sparse regression methods like SINDy, undermining the sparsity assumption and leading to dense, uninterpretable, and physically incorrect models, especially with measurement noise.
Key Innovation: Introduces the Sequential Thresholding of Coefficient of Variation (STCV), a novel, computationally efficient sparse regression algorithm that is inherently robust to data scaling. STCV replaces conventional magnitude-based thresholding with a dimensionless statistical metric, the Coefficient Presence (CP), making the discovery process invariant to arbitrary data scaling and robustly identifying correct, sparse physical laws.
80. A Behaviour-Aware Federated Forecasting Framework for Distributed Stand-Alone Wind Turbines
Core Problem: Accurate short-term wind power forecasting is essential, but centralizing turbine data raises privacy, cost, and heterogeneity concerns for distributed stand-alone wind turbines.
Key Innovation: A two-stage federated learning framework that first clusters turbines by long-term behavioural statistics using Double Roulette Selection (DRS) with recursive Auto-split refinement, and then trains cluster-specific LSTM models via FedAvg, achieving competitive forecasting accuracy while preserving data locality.
81. Fusion4CA: Boosting 3D Object Detection via Comprehensive Image Exploitation
Core Problem: Existing LiDAR-RGB fusion methods for 3D object detection in autonomous driving suffer from over-reliance on the LiDAR branch, insufficiently exploiting the rich information available in RGB images.
Key Innovation: Proposes Fusion4CA, which enhances 3D object detection by comprehensively exploiting visual input. It introduces a contrastive alignment module to calibrate image features with 3D geometry, a camera auxiliary branch for sufficient RGB information mining during training, leverages a cognitive adapter for pretrained image weights, and integrates a coordinate attention module for further boost.
82. Learning Causal Structure of Time Series using Best Order Score Search
Core Problem: Learning causal structure from multivariate time series data is challenging due to temporal dependence and scalability issues, especially in high auto-correlation regimes.
Key Innovation: Introduction of TS-BOSS, a scalable permutation-based search method for dynamic Bayesian network structures, leveraging grow-shrink trees for efficient score computation, and providing theoretical guarantees for effective causal discovery in time series, particularly in high auto-correlation regimes.
83. Harnessing Synthetic Data from Generative AI for Statistical Inference
Core Problem: The widespread use of synthetic data from generative AI models raises fundamental statistical questions about its valid, reliable, and principled application for downstream discovery, inference, and prediction.
Key Innovation: The paper reviews the current landscape of synthetic data generation and use from a statistical perspective, clarifying assumptions for meaningful support of downstream tasks, surveying models, benefits, limitations, and pitfalls, and discussing emerging frameworks for principled use.
84. FPGA-Enabled Machine Learning Applications in Earth Observation: A Systematic Review
Core Problem: The challenge of processing large volumes of Earth Observation data from small platforms (UAVs, NewSpace) in real-time with limited bandwidth, requiring onboard decision-making.
Key Innovation: A systematic review and analysis of 68 experiments deploying ML models on FPGAs for Remote Sensing, introducing taxonomies for model architectures and FPGA implementation strategies to guide future onboard processing.
85. RadarVLM: A Vision-Language Model Approach for Radar Scene Understanding
Core Problem: Existing machine learning approaches for radar sensors are fragmented and task-specific, lacking unified scene-level representations, despite radar's reliability in adverse conditions.
Key Innovation: RadarVLM, a vision-language framework that learns unified scene-level representations for radar through structured spatial language supervision. It introduces a structured caption framework encoding vehicle distributions in radar's native coordinate system and a Spatially-Grounded CLIP (SG-CLIP) objective that uses continuous scene similarity for fine-grained spatial reasoning, demonstrating improved spatial accuracy.
86. DiffusionHarmonizer: Bridging Neural Reconstruction and Photorealistic Simulation with Online Diffusion Enhancer
Core Problem: Neural reconstruction methods for autonomous driving simulation often produce rendering artifacts and struggle to realistically integrate dynamic objects, limiting simulation fidelity.
Key Innovation: Introduces DiffusionHarmonizer, an online generative enhancement framework that transforms imperfect neural renderings into temporally consistent and realistic outputs using a single-step, temporally-conditioned diffusion enhancer, significantly elevating simulation fidelity.
87. Dr. Seg: Revisiting GRPO Training for Visual Large Language Models through Perception-Oriented Design
Core Problem: The assumption that GRPO training paradigms for language reasoning can seamlessly transfer to visual perception tasks (like segmentation) in VLLMs is invalid, leading to suboptimal performance due to differences in output space and reward requirements.
Key Innovation: Proposes Dr. Seg, a plug-and-play GRPO-based framework for VLLMs that incorporates a Look-to-Confirm mechanism and a Distribution-Ranked Reward module, addressing the unique needs of perception-oriented tasks and significantly improving performance in complex visual scenarios.
88. Quadrotor Navigation using Reinforcement Learning with Privileged Information
Core Problem: Prior learning-based quadrotor navigation methods struggle to navigate effectively around large obstacles that block the goal location.
Key Innovation: Introduces a reinforcement learning-based quadrotor navigation method that leverages time-of-arrival (ToA) maps as privileged information and a yaw alignment loss to guide the robot around large obstacles, achieving higher success rates and robust real-world deployment.
89. Auto-Adaptive PINNs with Applications to Phase Transitions
Core Problem: Accurately resolving characteristic interfacial regions in PDE solutions using Physics Informed Neural Networks (PINNs) without post-hoc resampling, which is challenging for standard PINN training.
Key Innovation: An auto-adaptive sampling method for PINNs that allows for sampling based on arbitrary problem-specific heuristics, demonstrated to be effective in accurately resolving interfacial regions in Allen-Cahn equations.
90. FMint-SDE: A Multimodal Foundation Model for Accelerating Numerical Simulation of SDEs via Error Correction
Core Problem: The inherent trade-off between accuracy and computational efficiency in traditional numerical integrators for dynamical systems, and the need for a universal neural network-based approach that doesn't require retraining for each specific case.
Key Innovation: FMint-SDE, a multimodal foundation model based on a decoder-only transformer, which learns a universal error-correction scheme for accelerating numerical simulations of stochastic differential equations (SDEs) across diverse systems.
91. Improving long-term forecast stability of OceanCastNet: Efficient fine-tuning via a temporal change loss function
Core Problem: Deep learning wave forecast models suffer from data dependency and long-term drift, limiting their stability and accuracy, especially under data-limited conditions.
Key Innovation: Introduces an efficient fine-tuning method for OceanCastNet using a Temporal Change Loss function, which guides training to learn robust physical dynamics, improving long-term stability and accuracy even with reduced datasets, and effectively suppressing error accumulation.
92. A maritime panoramic visual perception framework with joint video stitching and target detection to enhance autonomous surface vehicles navigation
Core Problem: Autonomous Surface Vehicles (ASVs) require accurate panoramic visual perception for reliable navigation, but existing vision-based methods are limited by field of view and performance degradation in complex maritime environments.
Key Innovation: A maritime panoramic visual perception framework was proposed, integrating real-time panoramic video stitching (cylindrical projection, hash mapping) and a panoramic vision-oriented detection model (Context-Gaussian Hybrid Module, Feature Modulation and Upsampling unit) to achieve robust 360° perception and enhanced target detection for ASV navigation.
93. Reinforcement learning-based path planning for underwater gliders using ocean current forecasts
Core Problem: Path planning for underwater gliders in complex dynamic ocean environments is challenging, requiring robust and efficient navigation without relying on complex task modeling.
Key Innovation: The RL-OCF-PP framework was proposed, integrating reinforcement learning (an improved SEN-SAC-LrD algorithm with SENet and learning rate decay) with ocean current forecast data and a lightweight simulation environment, achieving superior performance in task integrity, safety, and efficiency for autonomous underwater glider navigation.
94. Predicting Bacterial Diversity in European Croplands Using Earth Observation and Meteorological Data
Core Problem: Predicting soil bacterial diversity in croplands is crucial for environmental monitoring, but traditional methods can be resource-intensive and lack broad spatial coverage.
Key Innovation: Developed models using random-forest and Cubist regressors, combining spectral indices from Sentinel-2 and meteorological variables, to predict alpha and beta diversity indices of soil bacteria with high skill, particularly for beta diversity, offering a cost-effective monitoring solution.
95. Cross-Modal Attention-Modulated Feature Enhancement Network for Visible-Infrared Ship Detection
Core Problem: Multimodal visible-infrared ship detection lacks paired datasets and faces challenges with insufficient consistency, complementarity, and redundant information during cross-modal semantic fusion.
Key Innovation: Proposed a Cross-Modal Attention-Modulated Feature Enhancement Network (CAMFEN) that establishes two pseudoinfrared datasets and utilizes a collaborative-differential enhancement feature fusion module to effectively fuse multimodal information, outperforming existing ship detection methods.
96. Perception-Driven Generative Model for High-Precision and Low-Cost Remote Sensing Image Generation in Resource-Constrained Cotton Health Monitoring
Core Problem: High-precision crop health monitoring requires expensive multispectral sensors for key bands (NIR, RE), limiting applications in resource-constrained scenarios, and existing cross-band generation methods often lack semantic utility and are susceptible to background interference.
Key Innovation: A perception-driven generative framework that purifies vegetation data, uses selective reconstruction prioritizing vegetation regions, and employs a transformer-based multichannel spectral fusion model to generate high-quality NIR and RE bands from RGB images, significantly improving cotton health classification accuracy.
97. A Coupled Cellular Automata-Deep Learning-XGBoost for Land Use and Land Cover Change Prediction in Yancheng
Core Problem: Existing land use and land cover change (LUCC) models struggle to simultaneously capture complex spatiotemporal dependencies and nonlinear socioeconomic drivers, often failing when historical trends are disrupted by sudden policy interventions (structural breaks).
Key Innovation: A parallel mechanism-coupled cellular automata–deep learning–XGBoost (CDX) framework is proposed, integrating CNN-LSTM for spatiotemporal inertia, XGBoost for driver suitability, and CA for spatial diffusion. This framework successfully characterized policy-driven regime shifts, achieving superior prediction accuracy and adaptability for LUCC.
98. Bathymetry Inversion Using Full Tensor Gravity Gradients: A Case Study in the Bay of Bengal
Core Problem: Conventional bathymetry inversion methods using gravity field data typically only use gravity anomaly and vertical gravity gradient, neglecting the potential contribution of the other five components of the gravity gradient tensor (GGT).
Key Innovation: This study investigates using all six components of the gravity gradient tensor (GGT) for bathymetry inversion, deriving formulas based on Parker's formula. Combining inversion results from all six GGT components yielded a new bathymetric model with higher accuracy than single-component methods.
99. Elevation change of the Greenland Ice Sheet and its peripheral glaciers: 1992–2023
Core Problem: Quantifying consistent and reliable long-term elevation and mass changes of the Greenland Ice Sheet and its peripheral glaciers, considering short-term variability and improving data integration and bias reduction.
Key Innovation: Integration of multiple satellite altimetry missions and improved methodology to produce a consistent, reliable, and publicly available dataset (ITS_LIVE) covering 1992–2023, revealing significant and accelerating mass loss patterns and providing a comprehensive understanding of Greenland's contribution to global sea-level rise.
100. A Consolidated Database of Mercury Observations for Permafrost Regions
Core Problem: Quantifying and modeling the large terrestrial mercury (Hg) pool in permafrost soils is hindered by a lack of harmonized, spatially resolved observations, despite the threat of climate-driven thaw releasing this legacy Hg.
Key Innovation: Compilation of a consolidated database of 117,802 harmonized mercury observations (1988-2022) from 59 studies across Northern Hemisphere permafrost regions, enabling cross-site synthesis, model calibration, and environmental assessments, while highlighting spatial and observation-type biases.
101. Touching with torque enables human-level robotic dexterity
Core Problem: Achieving human-like forceful manipulation in robotics is challenging due to the lack of critical environmental interaction cues such as collisions, balance, and resistance.
Key Innovation: Developed a torque-angle-pressure (TAP) tactile sensor leveraging magnetic flux density gradients for ultrasensitive, high-linearity, bidirectional torque sensing, enabling a robot to perform vision-free stable object placement, balance beam stacking, and adaptive slicing, surpassing human performance in some metrics.
102. Integrating machine learning with numerical simulation for 3D mineral prospectivity modeling in the Sanshandao-Haiyu gold belt, Eastern China
Core Problem: Existing numerical simulation software for deep geodynamic processes has limitations (e.g., FIDAP lacks rock deformation, FLAC2D/3D lacks chemical reactions), and 3D mineral prospectivity modeling often suffers from insufficient deep data.
Key Innovation: Integrating chemical reactions into FLAC3D via a custom program to calculate mineralization rates, and then employing machine learning to combine simulation outcomes (volumetric strain, mineralization rate) with fault morphology to construct and validate predictive models for 3D mineral prospectivity, delineating two prospective targets.
103. Algorithm for improving drill core sizing accuracy using image segmentation
Core Problem: Inaccurate and inefficient manual or unreliable algorithmic measurement of drill core sample dimensions (Length of Intact Rock Core Pieces - LIRCP), leading to operational delays and imprecise quantitative information for calculating key geotechnical parameters like Rock Quality Designation (RQD).
Key Innovation: Development and implementation of an integrated system using the YOLOv11 segmentation model for accurate, real-time calculation of actual length and volume of drill core samples (98% accuracy), significantly improving operational efficiency and providing precise data for geological decision-making.
104. Nano-Scale Study of the Mechanical Behavior of Aragonite-Dominant Coral Reef Limestones
Core Problem: Ensuring the safety and durability of infrastructure built with Shallow Coral Reef Limestone (CRL) requires a deep understanding of its long-term mechanical behavior under sustained loading, particularly the influence of its aragonite-dominant structures.
Key Innovation: Utilized large-scale nanoindentation mapping and the Mori–Tanaka homogenization model to reveal variations in Young's modulus (25.5-43.77 GPa) and quantified the critical threshold of mineral content. Applied the three-element Voigt model to demonstrate consistent creep behavior predominantly controlled by dislocation motion, identifying peak load as the dominant factor for long-term deformation.
105. Experimental Investigation on Rock-Breaking Mechanism by Scaled Button Cutters Assisted by Vibration with Different Amplitudes
Core Problem: Abstract not provided.
Key Innovation: Abstract not provided.
106. Deep Generative Modeling of Digital Rock CT Images Within a Conditional Diffusion Framework
Core Problem: The high cost and limited accessibility of pore-scale X-ray computed tomography (CT) imaging constrain digital rock physics (DRP) studies, limiting the ability to generate diverse rock microstructures for analysis.
Key Innovation: Proposes a conditional diffusion model (CDM) with classifier-free guidance and a residual U-Net noise predictor to generate pore-scale rock CT images under multi-level conditions (porosity, phase composition maps). Achieves high structural similarity (mean SSIM of 0.9393) and demonstrates effective generation for digital rock physics.
107. Rock Salt Properties for Gas Storage in Mined Caverns: Part I–Short-Term Deformation and Yield, Long-Term Creep, and Gas Sealing Capacity
Core Problem: Ensuring the long-term safety and integrity of deep rock salt caverns for energy storage requires a thorough understanding of the short-term deformation, yield, long-term creep, and gas sealing capacity of different rock salt lithofacies under various stress conditions.
Key Innovation: Conducted a comprehensive laboratory appraisal of two contrasting rock salt lithofacies from the Canning Basin, including mineralogical, microstructural, petrophysical characterization, and long-term four-stage triaxial tests. Monitored strains, ultrasonic velocities, and conducted long-term gas transmissivity and creep tests, providing data on gas permeability (e.g., 5.10–5 mD for clean salt) and creep rates, identifying dislocation glide as the likely dominant creep mechanism.
108. Seawater intrusion in karst aquifers: A critical review of factors, methods, and mitigation actions
Core Problem: Seawater intrusion (SWI) in karst aquifers is often oversimplified through porous-media assumptions, neglecting karst architecture as a governing control, leading to biases in investigative design and model inference.
Key Innovation: A critical review that foregrounds karst architecture as a primary control on SWI mechanisms, organizes drivers, structures investigative and predictive methods (hydrochemical/isotopic tracers, hydrogeophysical imaging, numerical modeling), and provides decision-ready guidance for management and mitigation actions tailored to karst systems.
109. Pulse fragmentation-induced uncertainty in forest LAI mapping using UAV LiDAR
Core Problem: Accurate estimation of Leaf Area Index (LAI) in complex forests using UAV LiDAR is challenged by LiDAR detection rates and the need to account for the wood component, which introduce bias.
Key Innovation: Uses simulated UAV LiDAR data to evaluate how pulse fragmentation leads to incomplete detection and biases LAI estimates, and assesses the effectiveness of wood masks, concluding that small footprint LiDAR systems are better for LAI mapping due to reduced beam fragmentation and improved leaf/wood classification.
110. Bayesian updating of LRFD resistance factor for CPT-based pile design methods using combined static and dynamic pile load tests
Core Problem: Improving the reliability of Load and Resistance Factor Design (LRFD) for driven piles by providing a robust framework for calibrating resistance factors (ϕ) using CPT-based methods, integrating various site-specific test data.
Key Innovation: Developed a hybrid Bayesian updating framework to calibrate LRFD resistance factors for driven piles, combining prior regional distributions with site-specific static and dynamic pile load test data. Demonstrated how early evidence and bias factors influence ϕ updates and the need for extensive calibration in highly variable sites. Provides a probabilistic basis for site-specific LRFD calibration.
111. InverseNet: Benchmarking Operator Mismatch and Calibration Across Compressive Imaging Modalities
Core Problem: Existing benchmarks for compressive imaging systems do not quantify operator mismatch, which is a common issue in deployed systems and significantly degrades the performance of deep learning methods.
Key Innovation: Introducing InverseNet, the first cross-modality benchmark for operator mismatch in compressive imaging, demonstrating that deep learning methods lose significant performance under mismatch and that operator-conditioned methods with blind calibration can effectively recover these losses.
112. PinPoint: Evaluation of Composed Image Retrieval with Explicit Negatives, Multi-Image Queries, and Paraphrase Testing
Core Problem: Current Composed Image Retrieval (CIR) benchmarks are limited, lacking multiple ground-truth answers, explicit negative examples, robustness testing, and support for multi-image queries, hindering comprehensive evaluation of CIR methods.
Key Innovation: Presents PinPoint, a comprehensive real-world benchmark for CIR with multiple correct answers, explicit hard negatives, six instruction paraphrases for robustness, multi-image composition support, and demographic metadata, along with a training-free reranking method to address identified drawbacks.
113. ConTSG-Bench: A Unified Benchmark for Conditional Time Series Generation
Core Problem: The field of conditional time series generation lacks a standardized and systematic benchmarking framework to evaluate generative models across diverse conditions and modalities.
Key Innovation: Introduces ConTSG-Bench, a unified benchmark comprising a large-scale, well-aligned dataset and a comprehensive suite of metrics, enabling systematic evaluation of conditional time series generation methods and highlighting challenges in structural controllability and downstream task utility.
114. Revisiting Shape from Polarization in the Era of Vision Foundation Models
Core Problem: Existing Shape from Polarization (SfP) methods underperform RGB-only vision foundation models in surface normal estimation due to domain gaps from limited/unrealistic synthetic datasets and unmodeled sensor noise, despite the strong physical relationship between polarization and surface geometry.
Key Innovation: Developed a high-quality polarization dataset using 1,954 3D-scanned real-world objects, incorporated pretrained DINOv3 priors, and introduced polarization sensor-aware data augmentation. This allows a lightweight model with polarization cues to significantly outperform state-of-the-art SfP and RGB-only VFMs in surface normal estimation, reducing training data or model parameters.
115. Locality-Attending Vision Transformer
Core Problem: Vision Transformers (ViTs), while excellent for image-level classification due to global self-attention, often obscure fine-grained spatial details crucial for segmentation tasks, limiting their performance in such applications.
Key Innovation: A simple yet effective add-on that modulates ViT's self-attention with a learnable Gaussian kernel to bias attention toward neighboring patches and refines patch representations, thereby enhancing segmentation performance by encouraging local focus while preserving global information, without altering the training regime or sacrificing classification performance.
116. FC-VFI: Faithful and Consistent Video Frame Interpolation for High-FPS Slow Motion Video Generation
Core Problem: Large pre-trained video diffusion models struggle to generate high-fidelity frames in video frame interpolation due to reliance on intrinsic generative priors, and existing methods often lack structural context or rely on error-prone optical flow for temporal consistency.
Key Innovation: FC-VFI, a novel framework for faithful and consistent video frame interpolation that employs a temporal modeling strategy on latent sequences to inherit fidelity cues, leverages semantic matching lines for structure-aware motion guidance, and introduces a temporal difference loss to mitigate temporal inconsistencies, achieving high performance and structural integrity for high-FPS video generation.
117. BiEvLight: Bi-level Learning of Task-Aware Event Refinement for Low-Light Image Enhancement
Core Problem: Low-light image enhancement using event cameras suffers from dual degradation: intrinsic background activity (BA) noise in events and low signal-to-noise ratio (SNR) in images, leading to severe noise coupling during modal fusion and a critical performance bottleneck.
Key Innovation: Proposes BiEvLight, a hierarchical and task-aware framework that collaboratively optimizes enhancement and denoising through bi-level optimization. It uses a gradient-guided event denoising prior and recasts event denoising as a bilevel problem constrained by the enhancement task, significantly outperforming state-of-the-art approaches.
118. 3D-RFT: Reinforcement Fine-Tuning for Video-based 3D Scene Understanding
Core Problem: Existing approaches for 3D scene understanding largely rely on Supervised Fine-Tuning (SFT), where the token-level cross-entropy loss acts as an indirect proxy, causing misalignment between training objectives and actual task performance.
Key Innovation: Introduces 3D-RFT, the first framework to extend Reinforcement Learning with Verifiable Rewards (RLVR) to video-based 3D perception and reasoning. It directly optimizes models towards evaluation metrics (e.g., 3D IoU, F1-Score) using Group Relative Policy Optimization (GRPO) after SFT, achieving state-of-the-art performance on various video-based 3D scene understanding tasks.
119. Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding
Core Problem: Long video understanding is challenging due to dense visual redundancy, long-range temporal dependencies, and the tendency of chain-of-thought and retrieval-based agents to accumulate semantic drift and correlation-driven errors.
Key Innovation: Proposes VideoHV-Agent, a multi-agent framework that reformulates video question answering as a structured hypothesis-verification process. It uses a Thinker, Judge, Verifier, and Answer agent to articulate, derive clues, ground evidence, and integrate validated information, achieving state-of-the-art accuracy with enhanced interpretability and lower computational cost.
120. How far have we gone in Generative Image Restoration? A study on its capability, limitations and evaluation practices
Core Problem: Despite impressive perceptual realism, the practical capabilities and limitations of Generative Image Restoration (GIR) models, and their evaluation practices, are not systematically understood compared to previous methods.
Key Innovation: Presents a large-scale study using a new multi-dimensional evaluation pipeline to assess GIR models on detail, sharpness, semantic correctness, and overall quality. The study reveals performance disparities, identifies a paradigm shift in failure modes (from under-generation to over-generation), and trains a new IQA model aligned with human perception, providing insights for future GIR development.
121. Tell2Adapt: A Unified Framework for Source Free Unsupervised Domain Adaptation via Vision Foundation Model
Core Problem: Existing Source Free Unsupervised Domain Adaptation (SFUDA) methods are typically designed for low-gap, specific domain shifts and lack a unified, multi-modal, multi-target framework, hindering real-world application, especially in diverse clinical settings.
Key Innovation: Introduces Tell2Adapt, a novel SFUDA framework leveraging Vision Foundation Models (VFMs). It ensures high-fidelity VFM prompts via Context-Aware Prompts Regularization (CAPR) for pseudo-label generation and incorporates Visual Plausibility Refinement (VPR) using VFM's anatomical knowledge to re-ground predictions, achieving state-of-the-art performance in medical image segmentation across diverse domains and targets.
122. Generalizable Multiscale Segmentation of Heterogeneous Map Collections
Core Problem: Most map recognition work focuses on specialist models for homogeneous map series, failing to address the diversity of historical map collections in style, scale, and geographic focus, thus limiting their integration into historical geographic studies.
Key Innovation: Develops a generalizable semantic segmentation framework and ontology for heterogeneous map collections. It introduces Semap, a new benchmark dataset, and combines procedural data synthesis with multiscale integration to improve robustness and transferability, achieving state-of-the-art performance and enabling the integration of diverse cartographic archives.
123. Trainable Bitwise Soft Quantization for Input Feature Compression
Core Problem: Machine learning applications in IoT are constrained by limited compute/memory resources and the need to transmit large amounts of data from edge devices to remote servers, leading to bandwidth, latency, or energy issues.
Key Innovation: Proposes a task-specific, trainable feature quantization layer that compresses input features of a neural network by allowing each feature to be quantized to a user-defined number of bits, using trainable sigmoid-approximated step functions for bitwise soft quantization, achieving significant compression without major performance loss.
124. CATNet: Collaborative Alignment and Transformation Network for Cooperative Perception
Core Problem: Existing cooperative perception research often overlooks critical challenges in real-world multi-source data integration, specifically high temporal latency and multi-source noise, which degrade scene understanding.
Key Innovation: CATNet, an adaptive compensation framework that resolves temporal latency and noise interference in multi-agent systems through Spatio-Temporal Recurrent Synchronization, a Dual-Branch Wavelet Enhanced Denoiser, and an Adaptive Feature Selector, demonstrating superior robustness and adaptability in complex traffic conditions.
125. Fusion-CAM: Integrating Gradient and Region-Based Class Activation Maps for Robust Visual Explanations
Core Problem: Existing Class Activation Map (CAM) methods for explaining deep neural networks either provide fine-grained but noisy/incomplete gradient-based explanations or broad but over-smoothed region-based explanations, lacking robustness and comprehensive coverage.
Key Innovation: Fusion-CAM, a novel framework that integrates denoised gradient-based maps with region-based maps using adaptive similarity-based pixel-level fusion, producing robust, discriminative, and context-aware visual explanations that outperform existing CAM variants.
126. EdgeDAM: Real-time Object Tracking for Mobile Devices
Core Problem: Single-object tracking (SOT) on edge devices requires accurate and continuous target localization under challenging conditions (occlusion, distractors, fast motion), but existing state-of-the-art methods are computationally intensive or prone to drift.
Key Innovation: EdgeDAM, a lightweight detection-guided tracking framework that reformulates distractor-aware memory for bounding-box tracking under edge constraints, introducing Dual-Buffer Distractor-Aware Memory and Confidence-Driven Switching with Held-Box Stabilization, achieving improved robustness and real-time performance on mobile devices.
127. Towards 3D Scene Understanding of Gas Plumes in LWIR Hyperspectral Images Using Neural Radiance Fields
Core Problem: Combining information from multiple LWIR HSI images for 3D scene understanding and gas plume detection is challenging, especially with limited images, hindering comprehensive environmental analysis.
Key Innovation: Proposes using Neural Radiance Fields (NeRFs) with a novel adaptive weighted MSE loss to create 3D scene reconstructions from sparse LWIR HSI, enabling effective gas plume detection with significantly fewer training images.
128. Dictionary Based Pattern Entropy for Causal Direction Discovery
Core Problem: Discovering causal direction from symbolic temporal observational data is challenging due to the lack of functional models and noise assumptions.
Key Innovation: Proposes Dictionary Based Pattern Entropy (DPE), a framework integrating Algorithmic Information Theory and Shannon Information Theory, to infer causal direction and identify specific subpatterns driving changes by constructing direction-specific dictionaries and quantifying their influence using entropy-based measures.
129. The Volterra signature
Core Problem: Modern approaches for learning from non-Markovian time series often rely on implicit memory mechanisms that are difficult to interpret or train over long horizons.
Key Innovation: Proposes the Volterra signature (VSig) as a principled, explicit feature representation for history-dependent systems, leveraging the Volterra-Chen identity to derive rigorous learning-theoretic guarantees and demonstrating its efficacy in dynamic learning tasks.
130. Improving the accuracy of physics-informed neural networks via last-layer retraining
Core Problem: Physics-informed neural networks (PINNs) typically yield only moderately accurate solutions for solving partial differential equations (PDEs) due to challenges in determining suitable training strategies.
Key Innovation: Proposing a post-processing method that couples PINNs with last-layer retraining, which seeks the best approximation in a function space, leading to significantly lower errors (four to five orders of magnitude) and enabling transfer learning for complex problems.
131. On the Statistical Optimality of Optimal Decision Trees
Core Problem: Despite their empirical success, globally optimal empirical risk minimization (ERM) decision trees lack rigorous theoretical guarantees for their statistical performance.
Key Innovation: The work develops a comprehensive statistical theory for ERM trees, establishing sharp oracle inequalities and deriving minimax optimal rates over a novel function class (PSHAB space) that captures sparsity, anisotropic smoothness, and spatial heterogeneity, providing a principled foundation for ERM tree optimality.
132. ms-Mamba: Multi-scale Mamba for Time-Series Forecasting
Core Problem: Existing time-series forecasting architectures process input at a single temporal scale, which is sub-optimal for tasks where information changes over multiple time scales.
Key Innovation: Introduces Multi-scale Mamba (ms-Mamba), a novel architecture that incorporates multiple temporal scales by using multiple Mamba blocks with different sampling rates, outperforming state-of-the-art time-series forecasting models on various benchmarks.
133. Collaborative Learning of Local 3D Occupancy Prediction and Versatile Global Occupancy Mapping
Core Problem: Autonomous driving systems require robust 3D semantic occupancy prediction, especially in challenging conditions (occlusion, poor illumination), and need long-term memory priors for enhanced local perception and continuous global map updating.
Key Innovation: LMPOcc, a plug-and-play framework that integrates global occupancy priors to boost local prediction and simultaneously updates global maps. It features an efficient Current-Prior Fusion module and a model-agnostic prior format for continuous global information updating and large-scale map building.
134. RESAR-BEV: An Explainable Progressive Residual Autoregressive Approach for Camera-Radar Fusion in BEV Segmentation
Core Problem: Bird's-Eye-View (BEV) semantic segmentation for autonomous driving suffers from multi-modal misalignment and sensor noise, and existing approaches lack progressive refinement and robustness in challenging scenarios.
Key Innovation: RESAR-BEV, a progressive refinement framework for camera-radar fusion in BEV segmentation. It uses residual autoregressive learning with a cascaded Drive-Transformer and Modifier-Transformer, robust BEV representation combining ground-proximity voxels and dual-path encoding, and decoupled supervision for improved performance and robustness in adverse conditions.
135. Structured Kolmogorov-Arnold Neural ODEs for Interpretable Learning and Symbolic Discovery of Nonlinear Dynamics
Core Problem: The fundamental challenge of understanding and modeling nonlinear dynamical systems, particularly achieving models that are both accurate and physically interpretable using deep learning.
Key Innovation: Introduction of Structured Kolmogorov-Arnold Neural ODEs (SKANODEs), a framework that integrates structured state-space modeling with KANs within a Neural ODE architecture to perform virtual sensing, recover interpretable latent states (e.g., displacements, velocities), and extract compact, symbolic expressions for the system's governing dynamics, demonstrated on canonical oscillators and real-world F-16 data.
136. Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology
Core Problem: The absence of a comprehensive benchmark to holistically evaluate visual grounded reasoning models, particularly regarding focused visual perception, traceable evidence, and second-order reasoning, hindering the development of models with explainable reasoning pathways.
Key Innovation: Introduction of TreeBench, a diagnostic benchmark for visual grounded reasoning, and TreeVGR, a training paradigm that uses reinforcement learning to jointly supervise localization and reasoning, significantly improving model accuracy and traceability in visual question-answering tasks.
137. Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models
Core Problem: Existing diffusion models (like EDM) are limited by their reliance on fixed Gaussian noise, which restricts their ability to explain arbitrary-noise-based flow methods and can negatively impact image restoration tasks by corrupting degraded images and increasing complexity.
Key Innovation: EDA (Elucidates the Design space of Arbitrary-noise diffusion models), a unified theoretical framework that expands noise pattern flexibility for diffusion models, minimizes restoration distance, and introduces no additional computational overhead, demonstrating strong generalization capability across diverse medical and natural image restoration tasks.
138. Optimizing Multi-Modality Trackers via Significance-Regularized Tuning
Core Problem: Existing fine-tuning methods for multi-modality trackers struggle with the plasticity-stability trade-off, leading to suboptimal adaptation of pre-trained models for RGB data.
Key Innovation: Proposes a novel significance-regularized fine-tuning framework that incorporates intrinsic parameter significance (prior and transfer) to delicately refine the learning process, enhancing transferability across modalities and outperforming state-of-the-art techniques.
139. FLoC: Facility Location-Based Efficient Visual Token Compression for Long Video Understanding
Core Problem: The scalability of video-LMMs for long video understanding is severely limited by the overwhelming volume of visual tokens generated from extended video sequences.
Key Innovation: Proposes FLoC, an efficient, training-free, model-agnostic, and query-agnostic visual token compression framework based on the facility location function. It swiftly selects a compact, representative, and diverse subset of visual tokens, achieving significant efficiency gains and near-optimal performance in long video understanding benchmarks.
140. SPOT: Single-Shot Positioning via Trainable Near-Field Rainbow Beamforming
Core Problem: Existing analytical and learning-based schemes for frequency-dependent rainbow beamforming and user positioning incur high overhead and may not maximize localization accuracy.
Key Innovation: Proposes SPOT, an end-to-end deep learning-based scheme that simultaneously designs rainbow beams and estimates user positions. It treats PS and TTD coefficients as trainable variables to synthesize task-oriented beams for maximized localization accuracy and uses a lightweight fully connected module to recover user coordinates from single downlink transmission feedback, reducing overhead and delivering lower positioning error.
141. ReCast: Reliability-aware Codebook Assisted Lightweight Time Series Forecasting
Core Problem: Conventional time series forecasting methods struggle with local, complex, and highly dynamic patterns, and have high model complexity, limiting real-time or resource-constrained applications.
Key Innovation: ReCast, a reliability-aware codebook-assisted framework that uses patch-wise quantization and a dual-path architecture with a reliability-aware codebook update strategy (fusing multiple reliability factors via DRO) for lightweight and robust prediction of recurring local shapes and irregular fluctuations.
142. NeuralRemaster: Phase-Preserving Diffusion for Structure-Aligned Generation
Core Problem: Standard diffusion models destroy spatial structure by corrupting phase components, making them unsuitable for tasks requiring geometric consistency (e.g., re-rendering, sim-to-real transfer).
Key Innovation: Phase-Preserving Diffusion (phi-PD), a model-agnostic reformulation of the diffusion process that preserves input phase while randomizing magnitude, enabling structure-aligned generation for tasks like re-rendering and sim-to-real enhancement, and improving planner transfer performance in simulators like CARLA.
143. Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement
Core Problem: Existing approaches for learning general-purpose representations for Points-of-Interest (POIs) primarily focus on place identity from static textual metadata or trajectory context, neglecting the crucial signal of POI function (how places are actually used).
Key Innovation: Mobility-Embedded POIs (ME-POIs) is a framework that augments POI embeddings with large-scale human mobility data to learn POI-centric, context-independent representations grounded in real-world usage, improving map enrichment tasks and outperforming text-only and mobility-only baselines.
144. Learning to Drive is a Free Gift: Large-Scale Label-Free Autonomy Pretraining from Unposed In-The-Wild Videos
Core Problem: Ego-centric driving videos available online lack annotations, making it difficult to learn representations that capture both semantic structure and 3D geometry for autonomous driving perception.
Key Innovation: Proposes LFG, a label-free, teacher-guided framework using a feedforward architecture with an autoregressive module to learn unified pseudo-4D representations (point maps, camera poses, semantic segmentation, motion masks) from raw unposed videos, outperforming baselines on autonomous driving planning and perception tasks.
145. Seeing Through Uncertainty: A Free-Energy Approach for Real-Time Perceptual Adaptation in Robust Visual Navigation
Core Problem: Enabling robots to maintain robust, goal-directed visual navigation despite noisy, incomplete, or abruptly shifting sensory inputs, a challenge for standard Deep Neural Networks.
Key Innovation: Introduced FEP-Nav, a biologically-inspired framework implementing real-time perceptual adaptation by decomposing Variational Free Energy (VFE) into prediction error and Bayesian surprise, using a Top-down Decoder and Adaptive Normalisation to significantly improve navigation performance under diverse visual corruptions.
146. Distilling Privileged Information for Dubins Traveling Salesman Problems with Neighborhoods
Core Problem: Efficiently generating optimal tours for non-holonomic vehicles in Dubins Traveling Salesman Problems with Neighborhoods (DTSPN), which is computationally intensive for traditional methods.
Key Innovation: Developed a novel two-phase learning approach that leverages privileged information from expert trajectories via reinforcement learning, followed by supervised learning, to produce solutions 50 times faster than LKH and substantially outperform other learning schemes.
147. Learning Physical Systems: Symplectification via Gauge Fixing in Dirac Structures
Core Problem: Physics-informed deep learning struggles with systems involving dissipation and holonomic constraints (common in robotics), where the canonical symplectic form degenerates, undermining stability and long-term prediction.
Key Innovation: Introduces Presymplectification Networks (PSNs), the first framework to learn the symplectification lift via Dirac structures, restoring non-degenerate symplectic geometry by embedding constrained systems into a higher-dimensional manifold, demonstrated on quadruped robot dynamics.
148. CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions
Core Problem: Reinforcement learning (RL) often prioritizes performance over safety, leading to catastrophic outcomes, and traditional online safety filters using Control Barrier Functions (CBFs) can result in conservative behaviors because the RL policy lacks CBF knowledge.
Key Innovation: Proposes CBF-RL, a framework that enforces CBFs during RL training by minimally modifying the policy and safety filtering rollouts, enabling the learned policy to internalize safety constraints for safer exploration, faster convergence, and robust performance without an online safety filter, demonstrated on navigation and humanoid robot tasks.
149. Bayesian Inference for PDE-based Inverse Problems using the Optimization of a Discrete Loss
Core Problem: Accurately solving PDE-based inverse problems, especially when measurements are incomplete or indirect, and quantifying the uncertainty of the inferred parameters.
Key Innovation: Introduction of B-ODIL, a Bayesian extension of ODIL, which integrates PDE loss as prior knowledge with a likelihood to infer solutions with quantified uncertainties.
150. Testing Most Influential Sets
Core Problem: The lack of a formal, principled method to determine when the influence of small, influential data subsets on model conclusions is statistically excessive rather than expected under natural random sampling variation.
Key Innovation: A principled framework for testing most influential sets, including deriving exact influence formulas and identifying extreme value distributions of maximal influence (Fréchet or Gumbel), enabling rigorous hypothesis tests for excessive influence.
151. Symmetric Aggregation of Conformity Scores for Efficient Uncertainty Sets
Core Problem: Aggregating predictive uncertainties from multiple models to produce reliable and efficient uncertainty quantification within the conformal prediction framework remains an underexplored challenge, as combining individual prediction sets is difficult.
Key Innovation: SACP (Symmetric Aggregated Conformal Prediction), a novel method that aggregates nonconformity scores from multiple predictors by transforming them into e-values and combining them using any symmetric aggregation function, leading to sharper prediction sets and improved efficiency.
152. Development and applicability of a novel fully coupled framework for floating wind energy systems
Core Problem: The limitations of existing hybrid frameworks in accurately capturing tower dynamics and providing seamless two-way coupling between hydrodynamics and aero-servo-elastic responses for floating offshore wind turbines (FOWTs).
Key Innovation: OpenF2A, a novel open-source multi-physics framework that integrates OpenFAST into AQWA via a modified DLL, enabling seamless two-way coupling and more accurate representation of platform-tower interaction by incorporating platform restoring forces and added mass effects directly into the tower's equation of motion.
153. Predicting remaining fatigue life in the presence of HCF or LCF pre-damage
Core Problem: Existing fatigue life models for offshore structures under variable-amplitude loading (VAL) are often calibrated under constant-amplitude loading and neglect prior high-cycle (HCF) or low-cycle (LCF) fatigue damage, leading to inaccurate remaining-life estimates and potentially unsafe decisions.
Key Innovation: A unified damage-mechanics framework that derives explicit relationships between damage evolution equations (for HCF, LCF, MCF) and fatigue life equations (S–N, ε–N curves), incorporating load sequence via a stress-ratio parameter, demonstrating superior prediction accuracy (91.59% within 2x error band) across various materials and sequential VAL conditions.
154. A hybrid data-physics model for real-time prediction of riser dynamics using particle filtering
Core Problem: Accurately acquiring the full-field dynamic response of risers is challenging with limited sensor data, and standalone data-driven models may lack physical consistency or robustness to noise.
Key Innovation: A novel hybrid data-physics fusion method combining an Euler-Bernoulli beam theory-based physics model with an LSTM network-based data-driven model, utilizing particle filtering to optimally predict full-field riser dynamic response from limited sensor data with higher accuracy and exceptional resistance to measurement noise compared to standalone LSTM.
155. Prediction and Image Formation for Serendipitous Resident Space Object Underflights of Landsat 8
Core Problem: Serendipitous imaging of Resident Space Objects (RSOs) by Earth-viewing satellites like Landsat 8 often results in indecipherable RSO appearances in traditional image products due to drastically different relative motion, hindering Space Domain Awareness (SDA).
Key Innovation: A novel methodology for predicting RSO underflights for nadir-viewing platforms, calculating the RSO's apparent velocity, and resampling raw sensor image data specifically for underflight scenarios to produce more exploitable RSO images, advancing SDA capabilities.
156. Geocoding Refinement via Planar Block Adjustment for Spaceborne SAR Imagery in Low-Relief Terrain: A Case Study of LuTan-1
Core Problem: Conventional SAR geocoding refinement methods struggle in low-relief terrains due to insufficient and unevenly distributed control points, as they rely on slope variations in DEMs. Existing planar block adjustment (PBA) techniques are also ineffective as they assume prerefined single scenes.
Key Innovation: A robust planar block adjustment (PBA) method is proposed that directly achieves holistic geocoding refinement, bypassing the need for error compensation. This approach consistently attained subpixel geolocation accuracy and improved DEM elevation accuracy (better than 2m RMSE) in low-relief areas using LuTan-1 InSAR data.
157. DSS-Mamba: Deformable Spatial–Spectral State-Space Model for Hyperspectral Land Cover Classification
Core Problem: Hyperspectral image (HSI) data's high spectral dimensionality and dense spatial information pose a challenge for state-space models like Mamba to efficiently construct sequences that align with these structural characteristics for accurate land cover classification.
Key Innovation: An efficient deformable spatial–spectral Mamba network (DSS-Mamba) is proposed, featuring a two-branch architecture (spatial deformable Mamba and spectral bidirectional Mamba) and a spatial–spectral complementary fusion module. This approach efficiently mines valuable spatial–spectral features, leading to superior land cover classification accuracy.
158. GlocalDualNet: Disentangling Scale and Representation for Few-Shot Remote Sensing Segmentation
Core Problem: Few-shot semantic segmentation (FSS) in remote sensing imagery is challenged by extreme scale variations among target objects and significant intraclass heterogeneity, which existing FSS methods (often relying on single, global prototypes) fail to address effectively.
Key Innovation: GlocalDualNet, an FSS framework, is proposed with a multiscale support prototype extraction module (generating heterogeneous local prototypes) and a dual-branch segmentation network (explicitly disentangling feature learning for large- and small-scale targets). This improves segmentation accuracy across disparate scales.
159. Contrastive Prototype Clustering for Multimodal Remote Sensing Data Based on Spectral–Spatial Cross Mamba
Core Problem: Joint clustering of multimodal remote sensing (RS) data is limited by inadequate exploration of complex cross-modal interactions and long-range dependencies, and insufficient capability in handling large-scale datasets.
Key Innovation: Contrastive Prototype Clustering for Multimodal RS data based on Spectral–Spatial Cross Mamba (CPCM) is proposed. It features a multimodal spectral–spatial cross Mamba (S2CM) for global contextual modeling and deep semantic fusion, and an end-to-end joint optimization framework integrating contrastive learning with clustering learning through prototype learning for scalability.
160. A Transformer-based agent model of GEOS-Chem v14.2.2 for informative prediction of PM2.5 and O3 levels to future emission scenarios: TGEOS v1.0
Core Problem: Traditional chemical transport models are computationally expensive for efficient air quality prediction, and existing emulators lack comprehensive estimates and struggle with varying emissions and regional transport.
Key Innovation: Developed TGEOS v1.0, a Transformer-based agent model of GEOS-Chem, for efficient and informative prediction of PM2.5 and O3 levels under future emission scenarios, capable of capturing extreme pollution events and accounting for sectoral emissions, regional emissions, and meteorology.
161. Climate change and geopolitics threaten water supplies — but disaster is not inevitable
Core Problem: Global water supplies are increasingly threatened by climate change (e.g., glacier loss) and geopolitical tensions.
Key Innovation: The Indus Waters Treaty serves as a successful model for managing transboundary water resources amidst environmental and political challenges, offering a blueprint for other regions.
162. Support communities to conserve the Third Pole
Core Problem: Communities in the Third Pole region require support for conservation efforts to address environmental challenges and their potential impacts.
Key Innovation: Emphasizes the importance of supporting communities for conservation in the Third Pole, contributing to environmental stability and resilience.
163. AI and quantum now drive NSF grantmaking, officials say
Core Problem: The need for strategic direction in scientific funding to advance key technological areas like AI and quantum computing.
Key Innovation: The National Science Foundation's shift to prioritize AI and quantum computing in its grantmaking.
164. Global majority countries must embed critical minerals into AI governance
Core Problem: National AI strategies in Global Majority countries often overlook the strategic importance of critical minerals, limiting their leverage in the global AI value chain.
Key Innovation: Advocating for the integration of critical mineral considerations into AI governance frameworks in Global Majority countries to enhance their geopolitical position.
165. The influence of gas injection and production rates on pore-scale gas-water movement in low-permeability heterogeneous gas storage based on microfluidic technique
Core Problem: Clarifying the microscopic gas-water movement laws at the pore-scale in low-permeability heterogeneous gas storage is crucial for improving operational efficiency, as the impact of injection/production rates is not well understood.
Key Innovation: Conducting microfluidics experiments to investigate the impact of gas injection and production rates on pore-scale gas-water movement, demonstrating that slow injection combined with fast production significantly improves gas-bearing pore space utilization by promoting uniform displacement and suppressing water-locking, and discussing pore-scale transport mechanisms.
166. Understanding the role of large wood in restoring river systems across scales and structures: A systematic review
Core Problem: Gaps in understanding how large wood (LW) structural complexity, geomorphic context, and multi-scale hydraulic and ecological responses interact to shape river restoration outcomes.
Key Innovation: A systematic review synthesizing research on natural LW and engineered log structures, providing a process-based framework that classifies LW interventions by their dominant hydro-geomorphic effects and evaluates both benefits and risks to guide evidence-driven restoration design.
167. A framework for integrating spatiotemporal deep learning methods with landsat for annual land cover and impervious surface mapping
Core Problem: There is a need to extend the temporal resolution and reduce production latency of the National Land Cover Database (NLCD) for consistent U.S. land cover mapping.
Key Innovation: Develops the Land Cover Artificial Mapping System (LCAMS), a prototype spatiotemporal deep learning framework using a two-stage architecture (spatial and temporal models) to produce annual 30m resolution land cover and impervious surface outputs from Landsat data, achieving comparable accuracy with higher thematic resolution and automated production.
168. Organic fertilization promotes soil organic carbon sequestration by offsetting gaseous losses in sloping purple soil croplands
Core Problem: Limited understanding of soil organic carbon (SOC) loss through both gaseous emissions and runoff-related pathways in sloping agroecosystems, despite considerable research on soil respiration and erosion.
Key Innovation: Demonstrated that organic fertilization in sloping purple soil croplands leads to trade-offs among SOC loss pathways, increasing gaseous and dissolved organic carbon losses while reducing losses through sediment runoff, suggesting potential for enhanced net ecosystem carbon balance under suitable management.
169. Image-based ice shape and accretion process prediction on wind turbine blades via deep learning with curvature consistency constraint
Core Problem: Blade icing poses a major threat to wind turbine operation, and existing prediction methods are computationally demanding and complex.
Key Innovation: An image-based deep learning model (SegFormer with a novel curvature consistency-based loss function) for efficient and geometrically accurate prediction of ice shape and accretion processes on wind turbine blades.
170. Wind-tunnel study of uneven snow distribution on stepped flat roofs considering the effects of snow fences
Core Problem: Uneven snow distribution on stepped flat roofs under wind action leads to highly non-uniform snow loads, increasing the risk of local overloading and structural failures.
Key Innovation: A wind-tunnel study investigating the evolution of snow distribution patterns on stepped flat roofs with increasing wind speed and quantitatively evaluating the effectiveness of different snow fence parameters (height, spacing) in mitigating uneven snow distribution.
171. A two-stage approach to evaluate parameter equifinality in build-up/wash-off stormwater quality models
Core Problem: Addressing the issue of parameter equifinality in build-up/wash-off stormwater quality models, which compromises the robustness and physical interpretability of model outputs, and understanding how land use, pollutant type, and temporal calibration scale influence this equifinality.
Key Innovation: Introduced a novel two-stage calibration-evaluation framework integrating seasonal parameter pooling and sequential conditional/joint likelihood reweighting (using GLUE) to explicitly evaluate parameter generalizability and predictive reliability under equifinality, providing a flexible alternative to deterministic approaches.
172. Response and adaptation of global terrestrial vegetation production to extreme precipitation
Core Problem: Quantitative assessments of how ecosystem carbon sequestration adapts to extreme precipitation events (EPEs) at the global scale and over the long term are lacking.
Key Innovation: Systematically quantified the response, adaptation, and recovery of gross primary productivity (GPP) to EPEs using FLUXNET data, demonstrating that EPre significantly reduces ecosystem carbon sequestration and identifying key drivers influencing GPP recovery.
173. Macro- and micro-scale mechanical properties of anisotropic binary mixtures: A DEM study
Core Problem: Systematically investigating the macroscopic and microscopic mechanical behavior of anisotropic binary mixtures, considering coupled effects of deposition angle, coarse particle content, and particle shape, to understand stress-force-fabric relationships.
Key Innovation: Used DEM to validate stress-force-fabric relationships, proposed a novel descriptor (P s eff) for load-bearing network efficiency, and identified consistent linear correlations between stress ratio and anisotropy descriptors. Explored microscale coordination and interlocking, providing a theoretical foundation for constitutive models.
174. Influence of fines content on shear strength and fabric evolution of unbound aggregate base: An experiment and simulation study
Core Problem: Understanding how fines content (Fc) affects the micromechanical load transfer, fabric evolution, and ultimately the shear strength and performance of unbound aggregate bases, which is critical for their design.
Key Innovation: Conducted laboratory compaction, CBR, and triaxial compression tests, complemented by DEM simulations, to quantify coordination, force–fabric anisotropy, and particle kinematics. Found a non-monotonic dependence of peak deviator stress and CBR strength on Fc, peaking at 2.5%. Revealed micromechanical insights into coordination, force chain evolution, and particle rotation, providing a basis for fines control in aggregate base design.