TerraMosaic Daily Digest: Mar 3, 2026
Daily Summary
This March 3, 2026 digest compiles 184 selected papers from 1094 analyzed studies. The clearest high-impact contributions remain process-resolved and observation-anchored: integrated GNSS-seismic-hydrologic diagnostics for rainfall-driven slope acceleration in Taiwan, substrate-controlled runout mechanics for a large Tibetan Plateau rock avalanche, and a new central Apennines inventory that constrains the orogen-scale distribution of large rock-slope deformation.
Across the full set, the dominant paper mix is Concepts & Mechanisms plus Detection and Monitoring, with a second tier of hazard-modeling studies. This is reflected in coupled workflows that connect failure initiation to consequence and decision support, including Etna flank thermomechanical deformation modeling, SPH-FEM simulation of landslide-surge interaction with concrete dams, dynamic Bayesian slope reliability updating, and climate-conditioned flood, avalanche, and drought risk analyses.
Key Trends
The dominant trajectory is from isolated hazard indicators toward integrated, uncertainty-aware geohazard workflows.
- Mechanism-aware slope diagnostics are strengthening: integrated GNSS, ambient-noise seismic velocity, rainfall, and groundwater observations are being used to resolve trigger timing and acceleration phases rather than only mapping deformation magnitude.
- Failure and runout analyses are increasingly material-conditioned: substrate properties, hydro-mechanical coupling, and structural setting are treated as first-order controls for rock-avalanche emplacement and broader slope-instability behavior.
- Cascading hazard chains are being modeled end-to-end: recent studies explicitly link initiation and downstream effects, including landslide-surge-dam interaction and levee-failure cascade dynamics.
- Climate-risk work is moving from trend reporting to actionable assessment: flood adaptation portfolios, avalanche-frequency drivers, and drought-transition intensification are increasingly quantified in forms usable for planning.
- Methodological contributions are converging on operational geohazard tasks: large volumes of AI and geospatial research now emphasize monitoring, forecasting, and reliability updating pipelines that can be integrated with physical hazard models.
Selected Papers
This digest features 184 selected papers from 1094 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. Unraveling rainfall-induced accelerated slope deformation through the integration of GNSS, seismic, and hydrologic data in Lantai, Taiwan
Core Problem: Challenges in studying deep-seated catastrophic landslides due to complex conditions, slow movements, limited real-time monitoring, and difficulties in accurately assessing sliding behavior, particularly rainfall-induced acceleration.
Key Innovation: Developed an integrated monitoring-modeling framework combining GNSS, seismic velocity changes (from ambient noise interferometry), rainfall, and groundwater-level records to unravel rainfall-induced accelerated slope deformation. Identified specific rainfall thresholds for acceleration, consistent temporal sequences of displacement and groundwater response, and successfully simulated groundwater responses with a 1D hydrological model, improving hazard assessment and early warning.
2. Influence of substrate on emplacement of a large rock avalanche in the west-central Tibetan Plateau, China: Insights from geomorphology and sedimentology
Core Problem: The fundamental mechanisms governing the interaction between rock avalanche masses and different substrates, and their resultant effects on emplacement behaviors and hypermobility, are not fully elucidated.
Key Innovation: Conducted a comprehensive geomorphological and sedimentological investigation of the Dzarang Tso Rock Avalanche, demonstrating that variable substrate nature (gravel-dominated vs. soft saturated lacustrine sediments) critically controls emplacement behaviors and final morphologies (e.g., flowbands vs. hummocks), and showing that soft saturated substrates can significantly enhance runout by providing a lubricated layer.
3. Large rock slope deformations: Evidence of orogen-scale distribution from an original inventory in central Apennines (Italy)
Core Problem: A comprehensive understanding of the occurrence and distribution of Large Rock Slope Deformations (DSGSDs) in mountain regions, specifically the central Apennines, remains complex.
Key Innovation: Presents an original inventory of 337 DSGSDs in the central Apennines, documenting 260 new sites. Uses statistical and geospatial analysis to identify morphometric parameters, linear clustering along tectonic lineaments, and the influence of geolithological conditions, providing a basis for understanding their morphogenetic significance in landscape evolution.
4. Modeling the Deformation Response to Mt. Etna Sliding Flank
Core Problem: Understanding the processes (gravitational spreading, tectonic forces, magmatic forcing) underlying the slow sliding of Mt. Etna's southeastern flank and explaining its deformation and seismic patterns.
Key Innovation: Presented new 3D thermomechanical rheological models incorporating geological features, demonstrating that ductile rheology of rocks and the presence of supercritical fluids horizontally constrained by pre-existing volcanic structures are essential to explain deformation patterns and act as a lubricant for gravitational sliding.
5. Assessing the Effectiveness of Nature‐Based Solutions and Building‐Level Flood Risk Reduction Measures: An Open‐Source Coupled Model
Core Problem: The increasing frequency and severity of floods due to climate change necessitate additional adaptation measures, but a comprehensive framework is needed to compare the effectiveness of diverse nature-based solutions and building-level measures across a catchment.
Key Innovation: Extending the Geographical, Environmental, and Behavioral (GEB) model by coupling it to hydrodynamic and flood risk models to provide an open-source, catchment-wide assessment of the impacts and cost-effectiveness of various nature-based solutions (reforestation, natural grassland, retention ponds) and building-level flood adaptation measures (dry/wet-proofing).
6. Probabilistic and Alarm-Based Evaluation of a b-Value-Driven Deep Learning Earthquake Forecast
Core Problem: Improving short-term earthquake forecasting using deep learning models, specifically by leveraging spatiotemporal evolution of seismic b-values, and rigorously evaluating their performance.
Key Innovation: Evaluates a b-value-driven deep learning model for short-term earthquake forecasting, demonstrating consistently positive Brier Skill Scores and elevated event capture rates at low alarm fractions, indicating that spatiotemporal b-value variations contain a persistent, relevant signal for probabilistic earthquake forecasting.
7. Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents
Core Problem: Current MLLMs in Earth Observation (EO) lack the capability to tackle complex tasks requiring multi-step reasoning and the use of domain-specific tools, limiting their utility for comprehensive spatiotemporal analysis.
Key Innovation: Introduces Earth-Agent, the first agentic framework that unifies RGB and spectral EO data within an MCP-based tool ecosystem, enabling cross-modal, multi-step, and quantitative spatiotemporal reasoning for complex scientific tasks, supported by a new comprehensive benchmark (Earth-Bench).
8. A review of measured rogue waves: Spatio-temporal characteristics, occurrence probability, physical mechanisms, and predictability
Core Problem: Understanding the statistical characteristics, physical mechanisms, and predictive potential of rogue waves under realistic sea states, given their hazardous nature and observed occurrence exceeding classical predictions.
Key Innovation: Synthesizing measurement-based findings on rogue waves, clarifying their spatio-temporal clustering, the central role of group-scale energy focusing, the dominance of transient linear focusing and weakly nonlinear interactions, and the fundamental constraints on their deterministic predictability.
9. Changing drivers of regional large magnitude avalanche frequency throughout Colorado, USA
Core Problem: Understanding the historical frequency and changing climate drivers of large magnitude snow avalanches in Colorado to improve forecasting and mitigation strategies, especially given recent extreme events.
Key Innovation: Paired tree disturbance data (dendrochronology, 1698-2020) with meteorological data (1901-2020) to reconstruct over three centuries of avalanche frequency and identify shifts in dominant avalanche regimes and drivers, revealing that extreme events are still possible despite declining snowpack.
10. A thorough review of the 5 May 1990 Potenza (Southern Italy) earthquake: constraints from macroseismology and insights from hydrology
Core Problem: Previous macroseismic studies of the 1990 Potenza earthquake contained inconsistent and exaggerated intensity values, hindering accurate characterization of the event and its effects.
Key Innovation: Re-evaluated all available macroseismic data from original sources, compiling a new robust dataset of 1393 macroseismic data points, and integrated extensive observations of seismically-induced hydrological changes (liquefaction, streamflow) to provide independent magnitude estimates and significantly enhance the accuracy of the event's characterization.
11. Numerical modeling-based deformation prediction for tunnel portal talus slopes
Core Problem: Predicting the deformation behavior and assessing the stability of talus slopes at tunnel portals under the combined influence of varying reservoir water levels and dynamic train loads.
Key Innovation: Developing and calibrating a numerical model to simulate talus slope deformation, demonstrating that deformation is greater during reservoir filling, and slope stability is substantially reduced when train loading coincides with elevated water levels, providing a robust framework for similar engineering projects.
12. Coupled SPH-FEM modelling from landslide surges to dynamic response of concrete dams
Core Problem: The serious threats posed to dam safety by surge waves generated by landslides (triggered by earthquakes or heavy rainfall), requiring a robust method to simulate these multi-hazard interactions and their dynamic impact on dam structures.
Key Innovation: Developed and validated a coupled GPU-accelerated SPH-FEM framework to simulate landslide-induced surge dynamics, bidirectional fluid-structure interactions, and their subsequent impact on concrete dams. Revealed that landslide characteristics (volume, fragmentation, proximity) predominantly govern surge dynamics and pressure, with moderate fragmentation amplifying secondary/tertiary waves, underscoring its importance in hazard assessments.
13. Seismic risk analysis of high-speed railway bridges based on new seismic hazard model
Core Problem: Improving seismic risk analysis for high-speed railway bridges requires a more accurate seismic hazard model and a robust method for calculating engineering demand parameter (EDP) hazard and seismic risk probabilities.
Key Innovation: This study proposed a new seismic hazard model based on quadratic polynomials and derived a semi-analytical solution for EDP demand hazard and seismic risk probabilities. A case study demonstrated its higher computational accuracy compared to classical models, enhancing seismic risk analysis for critical infrastructure like high-speed railway bridges.
14. Time-variant deformation prediction and reliability assessment of slopes based on dynamic Bayesian model averaging
Core Problem: Conventional slope reliability assessment often fails to accurately predict time-varying deformation due to neglecting material strength degradation, rainfall effects, and structural uncertainty in computational models.
Key Innovation: Proposes a dynamic Bayesian model averaging (DBMA) framework that integrates limit equilibrium and finite element methods, accounts for material deterioration and rainfall, and uses real-time monitoring data with a recursive Bayesian scheme to dynamically update model weights, leading to superior accuracy and robustness in time-variant slope deformation prediction and reliability assessment.
15. Ice-dammed lakes, outburst floods, and formation of the Gaccetávži ravine in Finnmark, NE Norway
Core Problem: Understanding the formation, evolution, and eventual drainage mechanisms of ice-dammed lakes that led to the formation of the Gaccetávži ravine in NE Norway.
Key Innovation: Used LiDAR-derived DTMs, field investigations, and remote sensing to propose a model for the formation, evolution, and drainage of ice-dammed lakes, demonstrating how initial smaller lakes coalesced and subsequently drained through ice-dam breakthroughs, triggering outburst floods and initiating ravine formation, with successive drainage events recorded in glaciofluvial terraces.
16. PLINIVS-based seismic vulnerability mapping during volcanic unrest in Campi Flegrei: a replicable DRM model
Core Problem: The need for a probabilistic and scalable methodology for areal seismic vulnerability mapping of buildings during volcanic unrest to support mitigation actions and strategic risk planning.
Key Innovation: Develops a PLINIVS-based methodology for seismic vulnerability mapping, classifying over 8,600 buildings and computing absolute and relative vulnerability indices on a grid. Integrates GIS, post-event validation, and probabilistic damage modeling, quantifying classification uncertainty, and demonstrating effectiveness for anticipatory Disaster Risk Management in multi-hazard volcanic environments.
17. A Bayesian network framework for quantitative analysis of education continuity under earthquake risk: integrating physical and social vulnerabilities
Core Problem: Challenges in quantitatively assessing education disruption risks under earthquake scenarios, requiring the integration of both physical and social vulnerabilities and handling associated uncertainties.
Key Innovation: Proposes a Bayesian network (BN) framework to quantitatively analyze education continuity under earthquake risk, integrating advanced system reliability tools for physical vulnerability and a Delphi survey for social vulnerability. This framework reveals varying risks across schools and provides insights for equity-informed risk assessment and policy simulations.
18. Review on major changes in sediment yields and sources in the Chinese Loess Plateau over the past 60 years
Core Problem: A lack of systematic assessments of sediment yield and source changes following large-scale revegetation initiatives in severely eroded regions, specifically the Chinese Loess Plateau.
Key Innovation: Provides a comprehensive review and analysis of sediment yields and sources in the Chinese Loess Plateau, demonstrating a significant decline (16% to 65%) in annual specific sediment yields after the 1999 Grain for Green Programme, primarily due to reduced gully sediment (88% of total change), with vegetation cover identified as the dominant controlling factor.
19. Characteristics of mud and water inrush disasters in high water pressure filled karst tunnels and the evolution law of surrounding rock response
Core Problem: High water pressure filled karst cavities ahead of tunnel faces pose a significant threat of mud and water inrush, and the disaster characteristics and evolution law of surrounding rock response are not fully understood, hindering effective prevention.
Key Innovation: Developed a 3D model test system to reveal that surrounding rock instability in high water pressure filled karst tunnels progresses through three stages (seepage channel expansion, local failure, overall instability), and that the anti-outburst rock mass thickness dictates failure mechanisms, shifting from overall compressive deformation to bending and tensile cracking near critical safety thickness.
20. Three-dimensional site characterization for unsaturated soil nail wall analysis using integration of multivariate co-kriging and finite element method
Core Problem: Conventional 2D and 3D deterministic analyses of soil nail walls fail to capture the inherent heterogeneity of soil properties, particularly out-of-plane spatial variability, leading to potential underestimation of failure risk.
Key Innovation: Introduction of a three-dimensional estimation approach using multivariate co-kriging to interpolate spatial variability of eight soil parameters, integrated with a finite element method for unsaturated soil nail wall analysis, revealing significant variations in stability responses and a more accurate assessment of failure risk compared to 2D stochastic analyses.
21. Cascading failure in levee systems: A two-stage material point method framework
Core Problem: Conventional flood risk assessments treat levee slope sliding and overtopping failures separately, failing to account for their interactive and cascading effects, leading to an incomplete understanding of post-sliding levee geometry's influence on subsequent overtopping-induced flood risk.
Key Innovation: Development of a novel two-stage Material Point Method (MPM) framework that sequentially simulates levee slope sliding and subsequent overtopping, enabling a unified analysis of cascading failures and demonstrating how post-sliding geometry (e.g., remaining crest width) governs overtopping-induced flood risk under various hydrological scenarios.
22. Differential Craton Destruction Controlled by Fossil Structures in the Central North China Craton
Core Problem: Understanding the mechanism of craton destruction in the Shanxi Rift Zone within the North China Craton, which is characterized by frequent volcanoes and earthquakes, by investigating its anisotropic lithospheric structure.
Key Innovation: Used surface-wave Eikonal tomography to reveal strong north-south contrasts in lithospheric structure, attributing variations (especially thermal structure) to Paleozoic subduction and later magmatic underplating, and proposing that different fossil structures control the differential craton destruction and associated geohazards.
23. Meteorological to Agricultural Drought Transitions Compounded by Heat Waves in Historical and Future Climates
Core Problem: The intensity of agricultural drought for a given precipitation deficit is not fully understood, particularly how warm temperature anomalies (heat waves) affect the transition from meteorological to agricultural droughts and how this will evolve in future climate scenarios.
Key Innovation: Quantifying how warm temperature anomalies affect the evolution from meteorological to agricultural drought using Earth System Model data, predicting that today's meteorological droughts would propagate into agricultural droughts roughly one classification more severe in future SSP3-7.0 scenarios, leading to major increases in moderate to extreme drought events globally.
24. Characterizing and Predicting Wildfire Evacuation Behavior: A Dual-Stage ML Approach
Core Problem: Wildfire evacuation behavior is highly variable and influenced by complex interactions, making it difficult to characterize and predict, which hinders targeted preparedness strategies and equitable emergency planning.
Key Innovation: Integrates unsupervised and supervised machine learning methods to uncover latent behavioral typologies (based on household resources, preparedness, etc.) and predict key evacuation outcomes, advancing data-driven understanding for targeted preparedness and resource allocation.
25. Deep Learning Based Wildfire Detection for Peatland Fires Using Transfer Learning
Core Problem: Conventional wildfire detectors struggle with peatland fires due to their distinct visual characteristics (smoldering, low flame, persistent smoke) and limited labeled data.
Key Innovation: A transfer learning-based approach that initializes a deep learning peatland fire detector with pretrained weights from a general wildfire model and fine-tunes it on peatland-specific data, significantly improving detection accuracy and robustness.
26. Towards Accurate and Interpretable Time-series Forecasting: A Polynomial Learning Approach
Core Problem: Existing interpretable time-series forecasting methods struggle with insufficient modeling of temporal dependencies, lack feature-level interpretability for early warning, and difficulty in simultaneously achieving accuracy and interpretability.
Key Innovation: Proposes Interpretable Polynomial Learning (IPL) which explicitly models original features and their interactions via polynomial representations, preserving temporal dependencies, providing feature-level interpretability, and offering a flexible trade-off between accuracy and interpretability, enabling simpler and more efficient early warning mechanisms.
27. Stabilized Adaptive Loss and Residual-Based Collocation for Physics-Informed Neural Networks
Core Problem: Traditional Physics-Informed Neural Networks (PINNs) struggle with problems characterized by high stiffness or shock-dominated dynamics, exhibiting unbalanced training and inaccuracy even with small physics residuals.
Key Innovation: Develops a new adaptive loss balancing scheme using smoothed gradient norms to ensure satisfaction of initial and boundary conditions, and an adaptive residual-based collocation scheme to improve solution accuracy in high-residual regions. This significantly reduces L2 errors for challenging PDEs like Burgers' and Allen-Cahn equations.
28. COP-GEN: Latent Diffusion Transformer for Copernicus Earth Observation Data -- Generation Stochastic by Design
Core Problem: Earth observation applications rely on multimodal sensor data, but the non-injective relationships between modalities mean deterministic models fail to represent uncertainty and variability, leading to collapsed conditional means and limitations in tasks like data completion and cross-sensor translation.
Key Innovation: Introduces COP-GEN, a multimodal latent diffusion transformer that models the joint distribution of heterogeneous Earth Observation modalities, enabling flexible any-to-any conditional generation (including zero-shot modality translation, spectral band infilling, and generation under partial/missing inputs) while generating diverse, physically consistent, and high-fidelity realisations.
29. StreamSplat: Towards Online Dynamic 3D Reconstruction from Uncalibrated Video Streams
Core Problem: Real-time, online reconstruction of dynamic 3D scenes from uncalibrated video streams is challenging due to the need for robust methods under strict latency and memory constraints, with most existing methods relying on extensive per-scene optimization.
Key Innovation: Introduces StreamSplat, a fully feed-forward framework for instant, online dynamic 3D Gaussian Splatting reconstruction from uncalibrated video streams of arbitrary length. It uses a probabilistic sampling mechanism, a bidirectional deformation field to mitigate error accumulation, and an adaptive Gaussian fusion operation, achieving state-of-the-art quality with significant speedup.
30. Closure to “Cyclic Behavior of Kaolin Clay under Undrained Conditions: Role of Microfabric”
Core Problem: This item is an author closure responding to technical comments on a previously published cyclic clay behavior study; it does not present a new primary experimental dataset.
Key Innovation: Provides formal clarifications and boundary conditions for interpreting microfabric-controlled cyclic soil response, improving how the original findings are applied in seismic geotechnical practice.
31. Discussion of “Cyclic Behavior of Kaolin Clay under Undrained Conditions: Role of Microfabric”
Core Problem: This item is a technical discussion of a previously published cyclic clay behavior study; it does not present a new standalone experimental campaign.
Key Innovation: Contributes critical interpretation and methodological scrutiny that sharpen uncertainty bounds and practical implications of microfabric effects under cyclic undrained loading.
32. Numerical modelling of wave-current-seabed-pipeline system: pore pressure dynamics and pipeline stability in liquefiable seabed
Core Problem: Existing studies on soil-structure interaction (SSI) in wave-current-seabed-pipeline (WCSP) systems use simplified wave-current solutions, leading to deviations in seabed boundary conditions and response predictions for liquefaction and pipeline stability.
Key Innovation: Established a numerical solution modifying wave parameters under wave-current interactions based on conservation laws, employing this for enhanced boundary conditions. Integrated this with Biot's consolidation theory and CycLiqCPSP constitutive model within a coupled numerical system to investigate WCSP system response, revealing critical links between pipeline stability and spatiotemporal liquefaction evolution.
33. Experimental study on the fluid-solid coupling effect of geosynthetic-reinforced calcareous sand revetment-breakwater
Core Problem: Understanding the dynamic response mechanism and fluid-solid coupling effects in geosynthetic-reinforced calcareous sand revetment-breakwaters under wave loading to improve their performance against marine hazards.
Key Innovation: Experimental study revealing a multi-level energy dissipation system within GRCS structures, identifying wave height as a key driver, and demonstrating how reinforcements enhance stability against wave impact and suction effects.
34. Harnessing multi-source hydro-meteorological data for high flows modelling in a partially glacierized Himalayan basin
Core Problem: Accurately modeling streamflow and simulating floods in data-scarce, complex, partially glacierized Himalayan basins, which are highly susceptible to floods during the monsoon.
Key Innovation: Developed a conceptual, semi-distributed hydrological model enhanced with static and dynamic glacier modules, calibrated using multi-variable data (including satellite-based glacier water loss and actual evapotranspiration) to achieve plausible streamflow reproduction and internally consistent water balance estimates, demonstrating a transferable parsimonious modeling strategy for data-limited mountain catchments.
35. EMO-1: an improved version of the high-resolution multi-variable gridded meteorological dataset for Europe
Core Problem: The need for high-quality, high-resolution gridded meteorological datasets for continental-scale hydrological modeling, particularly for operational flood forecasting systems like EFAS.
Key Innovation: Introduces EMO-1, an advanced 1 arc-minute (~1.5 km) daily and 6-hourly multi-variable gridded meteorological dataset for Europe (1990-2024), significantly upgrading its predecessor by integrating observations from 47 data providers, using "virtual stations" to minimize data gaps, and employing a rigorous quality control and an efficient Angular Distance Weighting (ADW) interpolation scheme.
36. Leveraging the Fine-Control MICP framework for cross-scale geoenvironmental applications: a roadmap of innovations and challenges
Core Problem: Despite numerous reviews, a comprehensive synthesis addressing the cross-scale challenges in Microbial-induced calcium carbonate precipitation (MICP) for geotechnical engineering and geo-disaster mitigation is missing.
Key Innovation: This review developed a unified 'Fine-Control' framework for MICP, spanning strain enhancement, environmental modulation, process control, and geological adaptation, providing a pathway for predictive biomineralization and identifying critical challenges for field-scale implementation in geo-disaster mitigation.
37. Ridge formation in deposits of rapid granular flows over slope–horizontal surfaces: An experimental study
Core Problem: The underlying formation mechanisms of intricate transverse ridges commonly produced by rapid granular flows, and the key dynamic controls on their development, remain poorly understood.
Key Innovation: Conducted systematic flume experiments using high-speed imaging and PIV to identify two recurring ridge morphologies (arc-shaped and V-shaped) in rapid granular flows, demonstrating their primary association with the inflow Froude number and summarizing conditions in a depositional phase diagram, providing a physically grounded basis for interpreting ridge-bearing deposits in granular flows and landslides.
38. Revisiting slope-area thresholds for gully initiation: a systematic review and meta-analysis
Core Problem: Gully erosion is a significant global soil erosion process, but there is a need for a systematic review and evaluation of existing methods for deriving slope-area (S-A) thresholds for gully initiation and assessing the effects of various environmental conditions on these thresholds.
Key Innovation: Systematically reviews and evaluates S-A threshold derivation methods and data collection, identifying orthogonal regression as the optimal approach and generating a global threshold (SA0.269 = 0.014) based on 127 datasets, while also assessing the influence of climate, lithology, soil texture, and land use on threshold parameters.
39. Responses of hydrology-driven soil carbon and nitrogen losses to rock fragment content under different rainfall scenarios
Core Problem: The specific response of soil carbon and nitrogen losses to rock fragment content through hydrological processes (leaching, gas emission) under different rainfall scenarios is unclear.
Key Innovation: Provides new mechanistic insights into how rock fragment content influences soil water retention, hydraulic conductivity, and consequently, the leaching and gaseous emissions of carbon and nitrogen under varying rainfall intensities, highlighting the interplay of flushing, dilution, and pore connectivity effects on terrestrial C and N losses in stony soils.
40. Intelligent constitutive modeling of frozen soil under true triaxial conditions using liquid neural networks and analysis of its mechanical response
Core Problem: Conventional constitutive models are inadequate for reproducing the full mechanical evolution of frozen soil under complex multi-principal stress loading, considering coupled influences of temperature, water content, intermediate principal stress coefficient, and confining pressure, which is crucial for safety in cold regions engineering and ground reinforcement.
Key Innovation: Developed an intelligent constitutive modeling framework integrating liquid neural networks (LNNs), LSTM, and an Attention mechanism. This framework dynamically modulates mechanical responses, captures long-term dependencies, and identifies critical deformation stages, achieving high-precision fitting of deviatoric stress-strain curves and strong generalization capability for predicting frozen soil behavior.
41. Analysis of microseismic evolution law and damage zone detection in oil storage caverns during operational period
Core Problem: Ensuring the safety and stability of water-sealed oil caverns during their operational period is a critical rock mechanics issue, requiring effective methods for identifying surrounding rock damage and understanding its evolution.
Key Innovation: Utilized microseismic monitoring technology to identify surrounding rock damage in an oil cavern during operation, analyzing the evolution of microseismic parameters (e.g., b-value, energy, Schmidt number) to detect local damage areas and provide a reference for early warning of potential instability.
42. An explicit method for in-situ determination of undrained elastic shear modulus using ROBOCONE torsional and vertical modules
Core Problem: Efficient and accurate in-situ determination of undrained elastic shear modulus, especially in offshore projects, without the need for extensive sampling and laboratory testing, and overcoming the difficulty of determining interface stiffness.
Key Innovation: Development of an explicit method using the ROBOCONE's coupled torsional and vertical modules (t–z and τ–θ) to determine undrained elastic shear modulus, effectively eliminating the influence of interface stiffness and enhancing site investigation efficiency.
43. Understanding compression behavior and structural transitions of clay-rich media with varying moisture contents through large-scale molecular dynamics simulations
Core Problem: Existing molecular dynamics studies on clay-rich media are limited in scope (simplified geometries, fixed stress ranges), preventing them from capturing stress-driven consolidation pathways, pore collapse, and fabric anisotropy development from dispersed to densely compacted states.
Key Innovation: Development of a large-scale molecular dynamics simulation framework (inspired by oedometer tests) to investigate the compression behavior and structural transitions of a nanoscale clay platelet assembly across a broad range of physical states and stress conditions, revealing three distinct structural transition stages and deriving closed-form equations for void ratio and S-order evolution.
44. Transverse differential settlement in lateritic subgrade of high-speed Railway under cyclic wetting
Core Problem: Transverse differential settlement in high-speed railway lateritic subgrades, induced by cyclic wetting (rainfall and water level fluctuations), poses a severe threat to train operational safety, yet a comprehensive analysis method for predicting and quantifying this settlement under varying environmental factors is lacking.
Key Innovation: Development of a 3-D finite element analysis method for predicting transverse differential settlement in high-speed railway lateritic subgrades under cyclic wetting, incorporating a cumulative deformation model and load model, and systematically quantifying the influence of slope angle, rainfall frequency/intensity, and water level on settlement and unevenness, providing recommendations for monitoring and prevention.
45. Hydro-thermo-mechanical evolution of foundation pits during horizontal freezing and thawing: From <em>in situ</em> tests to prediction methods
Core Problem: The hydro-thermo-mechanical evolution mechanisms of horizontal freezing and thawing in foundation pits, which differ significantly from vertical processes, are not well understood, leading to a lack of reliable prediction methods for freeze-thaw characteristics and prevention of freezing damage.
Key Innovation: A combined in situ test and numerical simulation study that investigates the hydro-thermo-mechanical evolution of foundation pits during horizontal freezing and thawing, clarifying the intrinsic relationships among temperature, water, and stress fields, and proposing a validated predictive equation for the maximum horizontal frost heaving force on pit side walls.
46. Effects of water immersion on the mechanical properties and hydrochemical characteristics of cemented calcareous soil
Core Problem: The stability of geoengineering structures and geological phenomena (like landslides) in areas with cemented calcareous soils is critically influenced by the complex response of these soils to water immersion, particularly regarding their mechanical and hydrochemical properties.
Key Innovation: Demonstrated that prolonged water immersion significantly alters the hydraulic conductivity and hydrochemical properties of cemented calcareous soil, leading to initial strength augmentation followed by a progressive decline due to dissolution of cementing material and undermining of interparticle bonds, ultimately causing structural transformations and potential soil collapse.
47. Contrasting Crustal Rheology and Seismicity in Northeast China: Far‐Field Responses to the Pacific Plate Subduction
Core Problem: Understanding how intracontinental deformation and lithospheric architecture in NE China respond to the westward subduction of the Pacific plate, particularly concerning seismicity patterns.
Key Innovation: Employed an integrated geophysical-petrological approach to reveal significant lithospheric thinning and a contrast between cold, dry, strong crust west of the North-South Gravity Lineament (NSGL) and hot, wet, weak crust to the east, which reconciles observed seismicity patterns and suggests critical control by differential crustal rheology.
48. Neural Electromagnetic Fields for High-Resolution Material Parameter Reconstruction
Core Problem: Current methods for creating Digital Twins are visually rich but functionally incomplete, lacking underlying material properties (e.g., permittivity, conductivity) which are difficult to acquire non-contact and non-invasively due to the ill-posed nature of physical inversion.
Key Innovation: Introduces NEMF, a novel framework for dense, non-invasive physical inversion that systematically disentangles geometry, ambient field, and target materials. It leverages high-fidelity geometry from images and ambient RF signals with a differentiable layer to explicitly output a continuous, spatially-varying field of underlying material parameters, enabling functional digital twins.
49. Improving Anomaly Detection with Foundation-Model Synthesis and Wavelet-Domain Attention
Core Problem: Industrial anomaly detection faces significant challenges due to the scarcity of anomalous samples and the complexity of real-world anomalies.
Key Innovation: Proposes a foundation model-based anomaly synthesis pipeline (FMAS) that generates highly realistic anomalous samples without fine-tuning. Introduces a Wavelet Domain Attention Module (WDAM) which exploits adaptive sub-band processing to enhance anomaly feature extraction, leveraging distinct frequency-domain characteristics. The combination significantly improves anomaly detection sensitivity and computational efficiency.
50. TinyIceNet: Low-Power SAR Sea Ice Segmentation for On-Board FPGA Inference
Core Problem: Timely and reliable sea ice mapping is crucial for safe maritime navigation, but conventional ground-based processing of Sentinel-1 SAR data is limited by downlink bandwidth, latency, and energy costs.
Key Innovation: Presents TinyIceNet, a compact semantic segmentation network co-designed for low-power on-board FPGA inference of sea ice from Sentinel-1 SAR imagery, achieving high F1 score with significantly reduced energy consumption compared to GPU baselines.
51. Discussion on “performance of sustainable drainage capillary barrier systems for climate change adaptation in temperate climates”
Core Problem: This item is a technical discussion piece on a previously published capillary-barrier drainage study rather than a new primary field or laboratory investigation.
Key Innovation: Adds focused critique and interpretive clarification on design assumptions and performance claims, helping refine engineering use of capillary-barrier drainage for climate adaptation contexts.
52. Coastal FRP bar anti-floating anchor rods bearing performance field tests and anchorage mechanism analysis
Core Problem: Understanding the actual engineering performance, internal force distribution, and anchorage mechanism of FRP bars as substitutes for steel bars in coastal anti-floating projects, given their unclear behavior under real-world conditions.
Key Innovation: Conducts field tests and mesoscopic/macroscopic analysis of FRP bar anchor rods in coastal anti-floating projects, demonstrating superior bearing capacity but larger displacement compared to steel bars, and clarifying the time-varying displacement, axial force transmission, and shear stress mechanisms at the anchor interface.
53. Explainable and probabilistic nearshore wave modeling along living shorelines using a CASGP-X framework
Core Problem: Accurate, reliable, interpretable, and uncertainty-quantified characterization of nearshore wave conditions is needed for evaluating living shoreline performance and resilience against coastal hazards.
Key Innovation: Developed CASGP-X, an interpretable probabilistic framework integrating cascaded Sparse Gaussian Process Regression with uncertainty quantification and SHAP, to jointly predict water depth, significant wave height, and peak period, offering a scalable, uncertainty-aware, and explainable alternative to traditional models for coastal engineering applications.
54. Optimization design research of vibration control parameters for Submerged Floating Tunnel based on GA-AIF-PSO algorithm
Core Problem: Ensuring the safety of Submerged Floating Tunnels (SFTs) during operation, particularly under seismic conditions, by optimizing vibration control measures and their parameters.
Key Innovation: Proposes a passive control scheme (Revetment Isolator-Multiple Tuned Mass Damper - RIS-MTMD) for SFTs under seismic conditions, introduces a hybrid optimization algorithm (GA-AIF-PSO) to optimize control parameters, and develops a two-step optimization method, demonstrating improved vibration control performance and enhanced safety.
55. How earthquakes and lightning help explain squeaky sneakers
Core Problem: Understanding the fundamental physics of friction and sound generation, drawing parallels between seemingly disparate phenomena like squeaky sneakers, lightning, and earthquakes.
Key Innovation: High-speed footage revealed that shoe squeaks can initiate with a tiny bolt of lightning, and the study also provided evidence for a debated brain phenomenon, using analogies to earthquakes to explain physical processes.
56. Service performance evolution of squeezing rock tunnels: A case study
Core Problem: Squeezing rock tunnels are highly susceptible to long-term degradation (lining cracking, pavement cracking, water leakage) due to rock creep and ambient temperature variations, which impairs lining performance and degrades tunnel safety and durability.
Key Innovation: Utilized automated remote monitoring and finite element analysis to demonstrate that tunnel lining deformation and stress fluctuate sinusoidally with seasonal temperature, with thermal stress as a primary factor, and that rock creep drives a two-stage evolution of displacement and stress, with over 69% of vertical displacement occurring within the first 20 years.
57. Numerical simulations of hydrogen seepage in “fairy circles”
Core Problem: Evaluating the hypothesis that 'fairy circles' are surface expressions of natural hydrogen accumulations, specifically understanding the coupled fluid-mechanical processes that lead to their formation and associated surface subsidence.
Key Innovation: Numerical simulations quantifying how capillarity, buoyancy, and overpressure control fairy circle radius and gas flux, and identifying gas-induced volumetric compaction as a mechanism for surface subsidence consistent with field observations.
58. Operator learning for consolidation: An architectural comparison for DeepONet variants
Core Problem: The limited application of Deep Operator Networks (DeepONets) in geotechnical engineering, specifically for the consolidation problem, and the need for robust architectures that can handle significant variations in target solutions and accelerate uncertainty quantification.
Key Innovation: Systematic evaluation of DeepONet architectures for the consolidation problem, proposing a Trunknet Fourier feature-enhanced DeepONet (Model 4) that significantly outperforms standard configurations, captures rapidly varying functions, and achieves substantial computational speedup (e.g., 1000x in 3D) for geotechnical applications.
59. Synthesis of composite seismic intensity measure based on Ridge Regression: Application to steel frame structures
Core Problem: Individual scalar seismic intensity measures (IMs) often fail to fully capture ground motion characteristics, while composite IMs face challenges of parameter selection and information redundancy due to multicollinearity among variables.
Key Innovation: Development of a composite seismic intensity measure (IM) using the Ridge Regression (RR) method, which effectively addresses multicollinearity and information redundancy among individual IMs, demonstrating superior performance in characterizing ground motion information for performance-based earthquake engineering applications in steel frame structures.
60. Joint Influence of Supportive and Unsupportive Environmental Conditions on Tropical Cyclone Rapid Intensification in Two HighResMIP Simulations
Core Problem: Understanding how multiple environmental factors jointly influence tropical cyclone rapid intensification (RI), particularly the interplay between supportive and unsupportive conditions, and improving its prediction in climate models.
Key Innovation: Demonstrated that while TC RI probability increases with supportive conditions, it is strongly suppressed by even a few unsupportive ones, and that current climate models (HighResMIP) show weaker sensitivity to these conditions due to unrealistic representation of vertical wind shear and low free-troposphere humidity, emphasizing the need to explicitly consider unsupportive conditions for better RI prediction.
61. Advancing Earth Observation Through Machine Learning: A TorchGeo Tutorial
Core Problem: Earth observation machine learning pipelines differ fundamentally from standard computer vision workflows (e.g., large georeferenced scenes, varied label formats, spatially aware sampling), making it challenging to use geospatial data in ML pipelines.
Key Innovation: TorchGeo, a PyTorch-based domain library providing datasets, samplers, transforms, and pre-trained models specifically designed to simplify and enable the use of geospatial data in machine learning pipelines, demonstrated through an end-to-end multispectral water segmentation case study.
62. SGMA: Semantic-Guided Modality-Aware Segmentation for Remote Sensing with Incomplete Multimodal Data
Core Problem: Multimodal semantic segmentation for remote sensing faces challenges with incomplete data (missing modalities), including multimodal imbalance, intra-class variation, and cross-modal heterogeneity, which existing methods often fail to address comprehensively.
Key Innovation: The Semantic-Guided Modality-Aware (SGMA) framework, which uses a Semantic-Guided Fusion (SGF) module to extract class-wise semantic prototypes for adaptive fusion and a Modality-Aware Sampling (MAS) module to dynamically reweight training samples, ensuring balanced multimodal learning and robust segmentation with incomplete data.
63. NeighborMAE: Exploiting Spatial Dependencies between Neighboring Earth Observation Images in Masked Autoencoders Pretraining
Core Problem: Existing Masked Image Modeling approaches for Earth Observation (EO) data largely overlook the rich contextual information provided by spatial dependencies between neighboring images.
Key Innovation: NeighborMAE, a framework that learns spatial dependencies by joint reconstruction of neighboring EO images, dynamically adjusting mask ratio and loss weight, leading to significantly improved performance across various downstream tasks relevant to EO data analysis.
64. SemGS: Feed-Forward Semantic 3D Gaussian Splatting from Sparse Views for Generalizable Scene Understanding
Core Problem: Existing methods for semantic scene reconstruction and semantic-aware novel view synthesis often rely on dense multi-view inputs and require scene-specific optimization, limiting their practicality and scalability.
Key Innovation: SemGS, a feed-forward framework that reconstructs generalizable semantic 3D fields from sparse images using a dual-branch architecture with camera-aware attention and dual-Gaussians, achieving state-of-the-art performance with rapid inference and strong generalization capabilities.
65. Generalizable Knowledge Distillation from Vision Foundation Models for Semantic Segmentation
Core Problem: Conventional knowledge distillation for semantic segmentation primarily preserves in-domain accuracy but compromises out-of-domain generalization, a limitation exacerbated when distilling from robust Vision Foundation Models (VFMs).
Key Innovation: Generalizable Knowledge Distillation (GKD), a multi-stage framework that decouples representation and task learning, uses selective feature distillation and a query-based soft distillation mechanism to explicitly enhance generalization, consistently outperforming existing KD methods on domain generalization benchmarks.
66. Track4World: Feedforward World-centric Dense 3D Tracking of All Pixels
Core Problem: Existing monocular 3D tracking methods are limited to sparse point tracking or slow optimization-based frameworks for dense tracking, hindering efficient holistic 3D understanding of video dynamics.
Key Innovation: Proposes Track4World, a feedforward model that enables efficient, holistic 3D tracking of every pixel in a world-centric coordinate system from monocular video. It uses a novel 3D correlation scheme to estimate pixel-wise 2D and 3D dense flow, enabling subsequent efficient 3D tracking.
67. Seeing Clearly without Training: Mitigating Hallucinations in Multimodal LLMs for Remote Sensing
Core Problem: Multimodal large language models (MLLMs) suffer from hallucinations in remote sensing visual question-answering (RS-VQA) due to visual grounding failures in large-scale scenes or misinterpretation of fine-grained small targets.
Key Innovation: RSHBench, a benchmark for diagnosing hallucinations, and Relative Attention-Driven Actively Reasoning (RADAR), a training-free inference method, leverage intrinsic attention to guide progressive localization and fine-grained local reasoning, mitigating hallucinations in RS-VQA.
68. Scale-invariant Gaussian derivative residual networks
Core Problem: Deep networks often fail to generalize across different image scales, leading to poor performance on out-of-distribution data when images are presented at scales not seen during training.
Key Innovation: Introducing GaussDerResNets, provably scale-invariant Gaussian derivative residual networks that achieve strong scale generalization and scale selection properties on rescaled datasets, significantly improving accuracy while maintaining good scale generalization compared to previous Gaussian derivative layers.
69. Multimodal-Prior-Guided Importance Sampling for Hierarchical Gaussian Splatting in Sparse-View Novel View Synthesis
Core Problem: Achieving high-quality 3D reconstruction and novel view synthesis from sparse-view imagery often suffers from overfitting texture-induced errors and noise from pose/appearance inconsistencies, especially in underconstrained regions.
Key Innovation: Introducing a multimodal-prior-guided importance sampling mechanism for hierarchical 3D Gaussian Splatting that fuses photometric residuals, semantic priors, and geometric priors to robustly estimate local recoverability, enabling selective injection of fine Gaussians and achieving state-of-the-art reconstructions in sparse-view novel view synthesis.
70. Disentangling regional impacts of joint teleconnections using causal representation learning
Core Problem: Existing methods for understanding climate teleconnections either lack causal interpretability (deep learning) or are based on simple indices, risking loss of signal and subject to model biases, hindering seasonal predictability and understanding of climatic changes.
Key Innovation: Introduces DAG-VAE, a causal representation learning approach that embeds a physics-informed directed acyclic graph in the latent space of a variational autoencoder. It jointly learns nonlinear reduced representations of large-scale modes of variability and their causal interactions, enabling data-driven counterfactuals for extreme rain seasons.
71. Distributed Dynamic Invariant Causal Prediction in Environmental Time Series
Core Problem: Existing methods for causal analysis in environmental time series either neglect environmental contexts in dynamic analysis or focus on static invariant causal inference, leaving a gap in distributed temporal settings for robust decision-making and stable causal predictor recovery.
Key Innovation: Proposes DisDy-ICPT, a novel framework for Distributed Dynamic Invariant Causal Prediction in Time-series. It learns dynamic causal relationships over time while mitigating spatial confounding variables without requiring data communication, and theoretically proves recovery of stable causal predictors, with applications in carbon monitoring and weather forecasting.
72. Any Resolution Any Geometry: From Multi-View To Multi-Patch
Core Problem: High-resolution joint estimation of surface normals and depth is difficult due to the trade-off between preserving fine local detail and maintaining global consistency.
Key Innovation: Proposing the Ultra Resolution Geometry Transformer (URGT), a multi-patch transformer that processes image patches augmented with coarse priors in a single forward pass, enforcing global coherence through cross-patch attention, and introducing a GridMix patch sampling strategy for spatial robustness, achieving state-of-the-art high-resolution geometry estimation.
73. From Complex Dynamics to DynFormer: Rethinking Transformers for PDEs
Core Problem: Conventional Transformer-based neural operators for Partial Differential Equations (PDEs) treat all discretized spatial points uniformly, leading to computationally prohibitive global attention that inefficiently mixes smooth large-scale dynamics with high-frequency fluctuations.
Key Innovation: Proposes DynFormer, a dynamics-informed neural operator that assigns specialized network modules to distinct physical scales. It leverages Spectral Embedding for low-frequency modes with Kronecker-structured attention and a Local-Global-Mixing transformation for small-scale turbulent cascades, integrated into a hybrid evolutionary architecture for robust long-term temporal stability.
74. I-CAM-UV: Integrating Causal Graphs over Non-Identical Variable Sets Using Causal Additive Models with Unobserved Variables
Core Problem: Existing causal discovery approaches are typically designed for single datasets, making it challenging to integrate causal graphs from multiple datasets with non-identical variable sets, especially when unobserved variables act as confounders.
Key Innovation: Proposes I-CAM-UV, an approach to integrate causal graphs derived from Causal Additive Models with Unobserved Variables (CAM-UV) across multiple datasets. It leverages the structural consistency between the ground truth causal graph and CAM-UV information to enumerate all consistent causal graphs, providing a method for robust causal discovery from diverse data sources.
75. Utonia: Toward One Encoder for All Point Clouds
Core Problem: Point clouds from diverse domains (remote sensing, LiDAR, etc.) have distinct sensing geometries, densities, and priors, making it challenging to train a single model that benefits all.
Key Innovation: Utonia, a self-supervised point transformer encoder trained across diverse point cloud domains, learns a consistent representation space that improves perception, benefits embodied/multimodal reasoning, and transfers across domains, serving as a step toward foundation models for sparse 3D data.
76. ALARM: Automated MLLM-Based Anomaly Detection in Complex-EnviRonment Monitoring with Uncertainty Quantification
Core Problem: Deploying MLLM-based visual anomaly detection (VAD) in complex environments is challenging because anomalies are often highly contextual and ambiguous, necessitating robust uncertainty quantification.
Key Innovation: ALARM, a UQ-supported MLLM-based VAD framework, integrates uncertainty quantification with quality-assurance techniques (reasoning chain, self-reflection, MLLM ensemble) for robust and accurate anomaly detection in complex environments, demonstrating generic applicability.
77. Channel-Adaptive Edge AI: Maximizing Inference Throughput by Adapting Computational Complexity to Channel States
Core Problem: The lack of a tractable theoretical framework for characterizing end-to-end inference performance hinders the efficient design of integrated communication and computation (IC^2) technologies for edge AI in 6G networks.
Key Innovation: A tractable analytical model for end-to-end inference accuracy was developed using a Mixture of von Mises distribution, leading to a channel-adaptive AI algorithm. This algorithm jointly optimizes transmit-side feature compression and receive-side model complexity based on channel conditions to maximize inference throughput under latency and accuracy constraints, showing superior performance over fixed-complexity methods.
78. Partial Weakly-Supervised Oriented Object Detection
Core Problem: The high cost of dataset annotation for oriented object detection (OOD) hinders its widespread application, as current methods require extensive or specific annotations.
Key Innovation: The first Partial Weakly-Supervised Oriented Object Detection (PWOOD) framework, which efficiently leverages large amounts of unlabeled data with partially weak annotations (horizontal boxes or single points), outperforming weakly supervised algorithms and achieving comparable results to semi-supervised methods. It includes an Orientation-and-Scale-aware Student (OS-Student) model and a Class-Agnostic Pseudo-Label Filtering (CPF) strategy.
79. Inpainting the Red Planet: Diffusion Models for the Reconstruction of Martian Environments in Virtual Reality
Core Problem: Extraterrestrial heightmaps, particularly for Mars, often contain missing values due to acquisition constraints, and current interpolation techniques fail to preserve geometric coherence, hindering accurate 3D terrain reconstruction for VR applications and scientific analysis.
Key Innovation: A novel method for reconstructing the Martian surface using an unconditional diffusion model, trained on augmented HiRISE heightmaps, which significantly outperforms established void-filling and inpainting techniques in reconstruction accuracy and perceptual similarity.
80. Online Data Curation for Object Detection via Marginal Contributions to Dataset-level Average Precision
Core Problem: Existing online data sampling strategies for machine learning are not well-suited for object detection due to its structural complexity and domain gaps, hindering efficient data curation for improved model performance.
Key Innovation: DetGain, an online data curation method for object detection that estimates the marginal perturbation of each image to dataset-level Average Precision (AP) based on prediction quality, enabling efficient selection of informative samples and consistent accuracy improvements on COCO with various detectors, especially under low-quality data.
81. Prediction of Multiscale Features Using Deep Learning-based Preconditioner-Solver Architecture for Darcy Equation in High-Contrast Media
Core Problem: Accurately and efficiently modeling subsurface fluid flow in highly heterogeneous and multi-scale porous media, which is crucial for applications like oil and gas exploration but challenging for existing methods.
Key Innovation: Fourier Preconditioner-based Hierarchical Multiscale Net (FP-HMsNet), an efficient deep learning architecture combining Fourier Neural Operators with multi-scale neural networks, capable of accurately reconstructing multi-scale basis functions for high-dimensional subsurface fluid flow, demonstrating superior accuracy, generalization, robustness, and computational efficiency.
82. Nonparametric Reaction Coordinate Optimization with Histories: A Framework for Rare Event Dynamics
Core Problem: Identifying an optimal reaction coordinate (RC) for rare but critical events in complex, high-dimensional stochastic systems (e.g., extreme weather) is notoriously difficult due to challenges like absence of ground truth, lack of a general loss function, overfitting, irregular/incomplete data, limited sampling, and extreme data imbalance.
Key Innovation: A nonparametric reaction coordinate optimization framework that incorporates trajectory histories, circumventing common methodological challenges. This framework enables robust analysis of irregular or incomplete data without requiring extensive sampling, accurately characterizing rare event dynamics in diverse complex systems, including conceptual ocean circulation models relevant to extreme weather.
83. Shipborne monitoring of significant wave height from BeiDou reflected signal considering satellite elevation and azimuth angles
Core Problem: Achieving high-resolution, dynamic monitoring of significant wave height (SWH) from shipborne platforms using BeiDou reflected signals, while effectively accounting for the coupling among observables induced by variations in observation geometry.
Key Innovation: Proposes a bivariate correction (BVC) model for retrieving SWH, utilizing average signal to noise ratio (ASNR) and time delay window (TDW) extracted from delay-Doppler maps, considering satellite elevation and azimuth angles, achieving high-resolution SWH monitoring with low RMSE.
84. CPT-based spatial variability of marine clay: a case study from Gwangyang Bay of Korea
Core Problem: Quantifying the spatial variability of marine clay properties, specifically vertical correlation length, which is crucial for reliable random field modeling in geotechnical engineering but often lacks sufficient empirical data.
Key Innovation: Investigating the spatial variability of Gwangyang marine clay using 51 CPT profiles, evaluating vertical correlation length based on the autocorrelation function (ACF) approach, and contributing to the growing database of spatial correlation lengths for clays.
85. Installation characteristics of a novel group-drag-anchor system in soft clay seabed
Core Problem: Providing sufficient bearing capacity for offshore wind turbine mooring systems, especially in challenging soft clay seabeds, and understanding the installation characteristics of novel anchoring systems.
Key Innovation: Develops a novel group-drag-anchor system (GDAS) and investigates its installation characteristics in soft clay seabeds using large deformation finite element analyses (CEL technique), quantifying the influence of key factors on diving performance and deriving an explicit expression for predicting ultimate embedment depth.
86. 20 years of trials and insights: bridging legacy and next generation in ParFlow and Land Surface Model Coupling
Core Problem: Groundwater is oversimplified in most Earth system models (ESMs), limiting their ability to accurately represent key land-atmosphere interactions, including evapotranspiration partitioning, drought propagation, and boundary layer development.
Key Innovation: Reviewing two decades of ParFlow-land/atmosphere coupled modeling efforts, highlighting the critical role of physically based groundwater in ESMs, demonstrating improved performance with renewed coupling efforts, and proposing a modular framework and a community-led model intercomparison project (PLCMIP) for future development.
87. Reloading experimental research on the mechanical characteristics and failure mechanism of sandstone samples considering pre-peak cyclic loading
Core Problem: Mitigating surrounding rock stability issues caused by cyclic stress disturbance during tunnel excavation, requiring a better understanding of sandstone's mechanical characteristics and failure mechanisms under such conditions.
Key Innovation: Conducted pre-peak pre-damaged cyclic disturbance and true triaxial reloading tests on sandstone, analyzing mechanical behavior, energy conversion, and deformation mechanisms under varying intermediate principal stresses and disturbance frequencies. Established a segmented softening damage constitutive model and provided theoretical foundation for assessing rock mass damage and predicting stability.
88. How do international boundary river dynamics affect riparian land? Insights from the China-Russia Ussuri River
Core Problem: There is a lack of quantitative assessment of the cross-border impacts of international boundary river dynamics on riparian land, especially considering past and future changes.
Key Innovation: Develops a coupled bankline forecasting (DSAS) and land use and land cover (ANN-MLP) prediction model to simulate the dynamic interactions of the Ussuri River. Reveals significant asymmetric effects and distinct national vulnerabilities to erosion, providing a scientific basis for establishing bilateral risk warning and joint governance frameworks.
89. Projected Global Diversity of Marine Heatwaves in the 21st Century
Core Problem: While the diversity of marine heatwaves (MHWs) has been highlighted, future changes in their six representative types under different climate change scenarios are unknown, posing serious threats to marine ecosystems.
Key Innovation: Analyzed projected changes in MHW diversity for the late 21st century using multi-model ensembles from CMIP6. Results show that bimodal MHWs are projected to increasingly dominate, with nearly all ocean regions governed by bimodal types by mid-century under both moderate and high-emission scenarios, providing critical information for marine conservation and climate adaptation.
90. Asymmetry in ENSO Prediction Skill Linked to Consecutive La Niña Events Within the IRI Real‐Time Forecast System
Core Problem: Accurately predicting the El Niño–Southern Oscillation (ENSO) remains a key challenge, with an observed asymmetry in prediction skill, particularly for consecutive La Niña events.
Key Innovation: Evaluated the International Research Institute for Climate and Society (IRI) real-time ENSO forecast system, revealing consistently high skill for El Niño and first-year La Niña events, but significantly lower predictability for consecutive La Niña events. This asymmetry is linked to deviations from the linear recharge–discharge oscillator framework due to enhanced nonlinear processes, suggesting improved representation of these processes could enhance prediction skill.
91. Forecasting as Rendering: A 2D Gaussian Splatting Framework for Time Series Forecasting
Core Problem: Time series forecasting (TSF) is challenged by intricate intraperiod-fluctuations and interperiod-trends, and existing 2D period-phase representations suffer from topological mismatch and inefficient, fixed-size modeling capacity.
Key Innovation: Introduces TimeGS, a novel framework that shifts TSF from regression to 2D generative rendering, reconceptualizing the future sequence as a continuous latent surface and using Gaussian splatting with Multi-Basis Gaussian Kernel Generation and Multi-Period Chronologically Continuous Rasterization for adaptive, chronologically continuous modeling.
92. Efficient Sparse Selective-Update RNNs for Long-Range Sequence Modeling
Core Problem: Recurrent Neural Networks (RNNs) suffer from 'memory decay' in long-range sequence modeling due to a rigid update schedule that constantly overwrites internal state, making it difficult to learn from distant past events.
Key Innovation: Introduces Selective-Update RNNs (suRNNs), a non-linear architecture that learns to preserve memory by using a neuron-level binary switch to update only for informative events, decoupling recurrent updates from raw sequence length and enabling efficient long-term storage and Transformer-level performance.
93. Physics-Informed Neural Networks with Architectural Physics Embedding for Large-Scale Wave Field Reconstruction
Core Problem: Large-scale wave field reconstruction faces challenges with computational efficiency and accuracy using traditional numerical methods or pure data-driven approaches, and standard Physics-Informed Neural Networks (PINNs) suffer from slow convergence, optimization instability, and spectral bias.
Key Innovation: Introduces Architecture Physics Embedded (PE)-PINN, which integrates physical guidance directly into the neural network architecture via a new envelope transformation layer, achieving over 10 times speedup in convergence and orders of magnitude reduction in memory usage for high-fidelity large-scale wave field reconstruction.
94. A Comparative Study of UMAP and Other Dimensionality Reduction Methods
Core Problem: A lack of systematic evaluation and understanding of the performance and limitations of UMAP, especially its supervised extensions for regression settings, compared to other dimensionality reduction methods.
Key Innovation: Provides a comprehensive comparative analysis of UMAP and other dimensionality reduction methods, showing supervised UMAP's strong performance for classification but identifying limitations in effectively incorporating response information for regression, highlighting future development directions.
95. Using the SEKF to Transfer NN Models of Dynamical Systems with Limited Data
Core Problem: Data-driven models of dynamical systems require extensive training data, which is often impractical to gather for many applications due to cost or safety concerns.
Key Innovation: Proposing the Subset Extended Kalman Filter (SEKF) to adapt pre-trained neural network models to new, similar dynamical systems with limited data, demonstrating significant data reduction (as little as 1% of original training data) and improved generalization.
96. ModalPatch: A Plug-and-Play Module for Robust Multi-Modal 3D Object Detection under Modality Drop
Core Problem: Multi-modal 3D object detection systems in autonomous driving are vulnerable to transient data interruptions and missing modalities (modality drop), posing critical safety risks.
Key Innovation: ModalPatch, a plug-and-play module that leverages temporal sensor data for perceptual continuity, using a history-based module to predict and compensate for unavailable features and an uncertainty-guided cross-modality fusion strategy to enhance robustness and accuracy under diverse modality-drop conditions.
97. WTHaar-Net: a Hybrid Quantum-Classical Approach
Core Problem: Enhancing deep learning models by leveraging quantum computing for efficient linear filtering operations in convolutional neural networks, specifically by finding a quantum-realizable transform that provides spatially localized, multi-resolution representations.
Key Innovation: WTHaar-Net, a hybrid quantum-classical convolutional neural network that replaces the Hadamard Transform with the Haar Wavelet Transform (HWT), which is quantum-realizable and provides better inductive biases for vision tasks, leading to substantial parameter reduction and competitive accuracy.
98. CAWM-Mamba: A unified model for infrared-visible image fusion and compound adverse weather restoration
Core Problem: Existing multimodal image fusion methods under adverse weather generally only handle single types of degradation (haze, rain, snow) and fail when multiple degradations coexist, limiting their utility in real-world applications like autonomous driving and UAV monitoring.
Key Innovation: Proposes CAWM-Mamba, the first end-to-end framework for joint image fusion and compound weather restoration with unified shared weights. It includes a Weather-Aware Preprocess Module, a Cross-modal Feature Interaction Module, and a Wavelet Space State Block with Freq-SSM to handle multi-frequency degradations and improve generalization.
99. ATD: Improved Transformer with Adaptive Token Dictionary for Image Restoration
Core Problem: Transformer-based image restoration models struggle to balance performance and computational burden, often restricting attention to local windows, leading to limited receptive fields and suboptimal performance.
Key Innovation: Proposes Adaptive Token Dictionary (ATD), a transformer-based architecture that enables global dependency modeling with linear complexity. It uses a learnable token dictionary and a token dictionary cross-attention (TDCA) mechanism to enhance features and improve performance on various image restoration tasks.
100. Addressing Missing and Noisy Modalities in One Solution: Unified Modality-Quality Framework for Low-quality Multimodal Data
Core Problem: Multimodal data in real-world scenarios often suffers from low quality, including noisy and missing modalities, which are typically handled separately, severely hindering model performance and robustness.
Key Innovation: A Unified Modality-Quality (UMQ) framework that jointly addresses missing and noisy modalities by training a quality estimator with a rank-guided strategy, constructing modality-specific quality enhancers, and employing a quality-aware mixture-of-experts module for robust multimodal affective computing.
101. Cross-view geo-localization, Image retrieval, Multiscale geometric modeling, Frequency domain enhancement
Core Problem: Cross-view geo-localization (CVGL) is challenging due to severe geometric asymmetry, texture inconsistency, and degradation of local information, with existing methods being sensitive to large viewpoint variations.
Key Innovation: The Spatial and Frequency Domain Enhancement Network (SFDE) proposes a three-branch parallel architecture leveraging complementary spatial and frequency domain representations to characterize consistency across domains, improving CVGL performance.
102. Rethinking Time Series Domain Generalization via Structure-Stratified Calibration
Core Problem: Existing cross-domain generalization methods for time series assume comparable samples, but real-world datasets from structurally heterogeneous dynamical systems lead to spurious correspondences and negative transfer.
Key Innovation: A structurally stratified calibration framework explicitly distinguishes structurally consistent samples and performs amplitude calibration exclusively within structurally compatible sample clusters, alleviating generalization failures caused by structural incompatibility in time series data.
103. Adapting Time Series Foundation Models through Data Mixtures
Core Problem: Time Series Foundation Models (TSFMs) suffer performance degradation in zero-shot forecasting for new domains not adequately represented in their pretraining data, and existing fine-tuning methods may not optimally capture sub-domain specific distributions.
Key Innovation: Proposing MixFT, a method that re-partitions time series data into more homogeneous sub-domains using Bayesian mixtures and fine-tunes separate modules for each, leading to improved specialization of TSFMs and better zero-shot forecasting performance.
104. Eliciting Numerical Predictive Distributions of LLMs Without Autoregression
Core Problem: Autoregressive decoding in Large Language Models (LLMs) for continuous-valued outputs is computationally expensive for obtaining predictive distributions over numerical targets.
Key Innovation: Demonstrates that training regression probes to predict statistical functionals (mean, median, quantiles) of LLM's numerical output distribution directly from its internal representations can recover distributional properties and numerical uncertainty without explicit autoregressive generation.
105. HDINO: A Concise and Efficient Open-Vocabulary Detector
Core Problem: Most open-vocabulary object detection methods rely on manually curated fine-grained training datasets and resource-intensive cross-modal feature extraction.
Key Innovation: Proposes HDINO, a concise and efficient open-vocabulary object detector using a two-stage training strategy with a One-to-Many Semantic Alignment Mechanism (O2M) and a Difficulty Weighted Classification Loss (DWCL) to improve semantic alignment and mine hard examples, achieving state-of-the-art performance without extensive data curation.
106. Enhancing Physics-Informed Neural Networks with Domain-aware Fourier Features: Towards Improved Performance and Interpretable Results
Core Problem: Physics-Informed Neural Networks (PINNs) are difficult to train and interpret, despite their ability to embed partial differential equations (PDEs) into their loss function.
Key Innovation: Proposes a novel modeling approach using Domain-aware Fourier Features (DaFFs) for positional encoding in PINNs. DaFFs encapsulate domain-specific characteristics, eliminating the need for explicit boundary condition loss terms and loss balancing, simplifying optimization, and reducing computational cost. It also develops an LRP-based explainability framework, demonstrating that PINN-DaFFs achieve lower errors, faster convergence, and more physically consistent feature attributions.
107. Leveraging Label Proportion Prior for Class-Imbalanced Semi-Supervised Learning
Core Problem: Semi-supervised learning (SSL) often suffers under class imbalance, where pseudo-labeling amplifies majority bias and suppresses minority performance.
Key Innovation: Introduces Proportion Loss from learning from label proportions (LLP) as a regularization term into SSL. This loss aligns model predictions with the global class distribution, mitigating bias across both majority and minority classes. A stochastic variant is also formulated. Integrating Proportion Loss into existing SSL methods consistently improves performance, particularly under scarce-label conditions.
108. Semi-Supervised Few-Shot Adaptation of Vision-Language Models
Core Problem: Few-shot adaptation of Vision-Language Models (VLMs) struggles in extremely low-shot regimes, especially with inherent class imbalances in tasks like medical imaging, leading to underrepresented categories and penalizing overall model performance.
Key Innovation: Proposes an efficient semi-supervised solver that propagates text-informed pseudo-labels during few-shot adaptation of VLMs. This method leverages unlabeled data, enabling lower-budget annotation pipelines and reducing labeling effort by over 50% in low-shot regimes.
109. LAGO: A Local-Global Optimization Framework Combining Trust Region Methods and Bayesian Optimization
Core Problem: Standard non-linear local optimization algorithms can struggle with exploring the full design space, while global optimization methods might be slow to converge locally.
Key Innovation: Proposes LAGO (LocAl-Global Optimization), an algorithm that combines gradient-enhanced Bayesian Optimization (BO) with gradient-based trust region local refinement through an adaptive competition mechanism. It separates global exploration from local refinement and incorporates local evaluations into the global Gaussian process under specific criteria, leading to improved exploration of the design space while maintaining fast local convergence.
110. Spatial Autoregressive Modeling of DINOv3 Embeddings for Unsupervised Anomaly Detection
Core Problem: Existing unsupervised anomaly detection methods using DINO models ignore spatial relationships between patches and incur substantial memory and computational overhead for modeling normative distributions.
Key Innovation: Proposing a simple and efficient framework that explicitly models spatial and contextual dependencies between patch embeddings using a 2D autoregressive model, learning a compact parametric normative distribution for faster and more memory-efficient inference.
111. MoECLIP: Patch-Specialized Experts for Zero-shot Anomaly Detection
Core Problem: Existing Zero-Shot Anomaly Detection (ZSAD) methods based on CLIP use a patch-agnostic design, limiting their ability to specialize for anomaly detection tasks while preserving CLIP's powerful generalization.
Key Innovation: Proposes MoECLIP, a Mixture-of-Experts architecture that achieves patch-level adaptation by dynamically routing each image patch to a specialized Low-Rank Adaptation (LoRA) expert. It also introduces Frozen Orthogonal Feature Separation (FOFS) and a simplex equiangular tight frame (ETF) loss to prevent expert redundancy.
112. VL-KGE: Vision-Language Models Meet Knowledge Graph Embeddings
Core Problem: Traditional knowledge graph embedding (KGE) methods struggle with heterogeneous multimodal knowledge graphs, processing modalities in isolation and lacking strong cross-modal alignment for unified representations.
Key Innovation: VL-KGE, a framework that integrates cross-modal alignment from Vision-Language Models (VLMs) with structured relational modeling to learn unified multimodal representations of knowledge graphs, improving link prediction on multimodal datasets.
113. Conformal Graph Prediction with Z-Gromov Wasserstein Distances
Core Problem: Principled uncertainty quantification for supervised graph prediction problems, especially with distribution-free coverage guarantees in structured output spaces, remains limited.
Key Innovation: A conformal prediction framework for graph-valued outputs, defining nonconformity via the Z-Gromov-Wasserstein distance and introducing Score Conformalized Quantile Regression (SCQR) to provide adaptive prediction sets with distribution-free coverage guarantees.
114. Exact Functional ANOVA Decomposition for Categorical Inputs Models
Core Problem: The lack of an explicit closed-form expression for Functional ANOVA decomposition for general dependent distributions has forced practitioners to rely on costly sampling-based approximations, especially for categorical inputs.
Key Innovation: Derived a computationally efficient, closed-form functional ANOVA decomposition for categorical inputs without any independence assumptions, seamlessly recovering the classical independent case and extending to arbitrary dependence structures, and providing a natural generalization of SHAP values.
115. Safe and Robust Domains of Attraction for Discrete-Time Systems: A Set-Based Characterization and Certifiable Neural Network Estimation
Core Problem: Accurately characterizing safe and robust domains of attraction (DOAs) for general nonlinear uncertain discrete-time systems, especially with state constraints, remains challenging due to theoretical and computational limitations.
Key Innovation: A novel framework was proposed for DOA estimation using value functions defined on metric spaces of compact sets, deriving associated Bellman-type functional equations. This was coupled with a physics-informed neural network framework and a formal verification procedure to learn and certify these estimates, demonstrating effectiveness for nonlinear uncertain systems.
116. Scalable Uncertainty Quantification for Black-Box Density-Based Clustering
Core Problem: There is a need for a scalable framework to quantify uncertainty in clustering, especially for high-dimensional and irregularly shaped data, and when leveraging modern black-box neural density estimators.
Key Innovation: A novel framework was introduced that combines the martingale posterior paradigm with density-based clustering, effectively propagating uncertainty from density estimation to the clustering structure. This approach scales to high-dimensional data by leveraging neural density estimators and parallel computation, with established frequentist consistency guarantees.
117. RealOSR: Latent Guidance Boosts Diffusion-based Real-world Omnidirectional Image Super-Resolutions
Core Problem: Existing Omnidirectional Image Super-Resolution (ODISR) methods are limited by simplified degradation assumptions and suffer from slow inference in latent-based diffusion approaches.
Key Innovation: Proposes RealOSR, a diffusion-based framework for real-world ODISR, featuring efficient latent-based condition guidance within a one-step denoising paradigm, enabled by the Latent Gradient Alignment Routing (LaGAR) module, achieving significant visual quality improvements and inference acceleration.
118. DMTrack: Spatio-Temporal Multimodal Tracking via Dual-Adapter
Core Problem: The challenge of effective spatio-temporal multimodal tracking, especially with efficient parameter usage and robust cross-modality fusion.
Key Innovation: DMTrack, a novel dual-adapter architecture for spatio-temporal multimodal tracking, featuring a spatio-temporal modality adapter (STMA) for adjusting features and a progressive modality complementary adapter (PMCA) for facilitating cross-modality prompting, achieving state-of-the-art performance with minimal trainable parameters.
119. MIRAGE: Runtime Scheduling for Multi-Vector Image Retrieval with Hierarchical Decomposition
Core Problem: Conventional multi-vector image retrieval (MVR) methods for RAG in MLLMs suffer from sub-optimal accuracy and efficiency due to overlooking alignment between queries and varying image objects, and redundant fine-grained image segments.
Key Innovation: MIRAGE, an efficient scheduling framework for image retrieval that uses a novel hierarchical decomposition paradigm with multiple intermediate granularities for better alignment, and minimizes redundancy by leveraging cross-hierarchy similarity consistency and sparsity, significantly improving accuracy and reducing computation.
120. Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment
Core Problem: While reasoning-based Image Quality Assessment (IQA) models trained with reinforcement learning (RL) show exceptional generalization, their underlying mechanisms are underexplored, and they incur high inference energy usage and latency, limiting deployment.
Key Innovation: Verification that RL-trained MLLMs achieve generalization by converting redundant visual representations into compact, cross-domain aligned text representations; and RALI, a novel contrastive learning algorithm that directly aligns images with these generalizable text representations, achieving comparable generalization with significantly fewer parameters and inference time.
121. Map the Flow: Revealing Hidden Pathways of Information in VideoLLMs
Core Problem: Despite recent advances in Video Large Language Models (VideoLLMs), their internal mechanisms for extracting and propagating video and textual information, especially for spatiotemporal reasoning, remain less explored.
Key Innovation: A mechanistic interpretability study revealing consistent patterns of information flow in VideoLLMs, including early cross-frame interactions for temporal reasoning, progressive video-language integration, and the model's ability to retain performance by selecting effective information pathways.
122. Self-Aug: Query and Entropy Adaptive Decoding for Large Vision-Language Models
Core Problem: Large Vision-Language Models (LVLMs) inherit a tendency to hallucinate from their underlying language models, and existing visual contrastive decoding methods often use generic augmentations, limiting their effectiveness in mitigating this issue.
Key Innovation: Self-Aug, a novel training-free decoding strategy that mitigates hallucination in LVLMs by employing a self-augmentation prompting strategy for query-dependent visual augmentation and an adaptive thresholding algorithm that adjusts next token candidate size based on output sparsity, significantly enhancing factual consistency.
123. Boosted Trees on a Diet: Compact Models for Resource-Constrained Devices
Core Problem: Deploying effective machine learning models on resource-constrained IoT devices is challenging due to their memory and computational limitations.
Key Innovation: A compression scheme for boosted decision trees that trains compact ensembles with reduced memory footprint (4-16x compression) by rewarding feature and threshold reuse, enabling autonomous operation on minimal-power IoT devices for applications like remote monitoring.
124. Discrete Solution Operator Learning for Geometry-Dependent PDEs
Core Problem: Traditional neural operator learning struggles with geometry-dependent Partial Differential Equations (PDEs) where discrete structural changes (e.g., topological changes, boundary conditions) break the smooth-variation premise.
Key Innovation: Introduction of Discrete Solution Operator Learning (DiSOL), a paradigm that learns discrete solution procedures by factorizing the solver into learnable stages, enabling stable and accurate predictions under diverse and out-of-distribution geometries.
125. CLEAR: Calibrated Learning for Epistemic and Aleatoric Risk
Core Problem: Existing uncertainty quantification methods typically address either aleatoric (measurement noise) or epistemic (limited data) uncertainty, but not both in a balanced and calibrated manner, leading to suboptimal predictive intervals for regression tasks.
Key Innovation: CLEAR, a novel calibration method with two distinct parameters (gamma_1 and gamma_2) that effectively combines aleatoric and epistemic uncertainty components. It improves the conditional coverage and reduces the width of predictive intervals across diverse datasets while maintaining nominal coverage, and is compatible with various uncertainty estimators.
126. PROFusion: Robust and Accurate Dense Reconstruction via Camera Pose Regression and Optimization
Core Problem: Current RGB-D SLAM systems fail to achieve robust and accurate real-time dense scene reconstruction during unstable camera motions (large viewpoint changes, fast motions, or sudden shaking).
Key Innovation: Combines a learning-based camera pose regression network for robust initialization with a randomized optimization algorithm for accurate refinement, achieving both robustness for unstable motions and accuracy for dense reconstruction in real-time.
127. Impact-induced damage mechanisms and seabed-stiffness-based design correction for pipe-in-pipe systems resting on flexible clay seabeds
Core Problem: Most studies on dropped-object impacts on pipe-in-pipe (PIP) systems idealize the seabed as rigid, neglecting the deformable effects of clay seabeds, leading to inaccurate damage assessment and design guidance.
Key Innovation: Developed a 3D nonlinear explicit finite element model with large deformation and coupled pipe-soil interaction to investigate PIP impact on saturated clay seabeds, introducing a seabed-strength-based correction factor to refine conventional rigid-seabed impact assessment methods.
128. Dynamic responses and performance predictions of two semi-submersible floating offshore wind turbine platforms under wave–current condition
Core Problem: The dynamic responses of semi-submersible floating offshore wind turbine (FOWT) platforms under viscous effects, especially wave-current interaction, have not been thoroughly investigated, hindering optimization of stability and power generation.
Key Innovation: Developed a fully coupled computational fluid dynamics (CFD) model to investigate wave height and wave-current interaction effects on two platforms, and a CNN-LSTM deep learning model to predict dynamic responses, providing insights for optimizing platform stability and power generation.
129. Multitemporal Scale Fusion Transformers for Interpretable Sea Surface Temperature Prediction
Core Problem: The inherent complexity and nonlinearity of Sea Surface Temperature (SST) dynamics present major challenges for achieving accurate and interpretable forecasting, hindering marine risk assessment and climate-informed decision-making.
Key Innovation: A novel interpretable framework named Multitemporal Scale Fusion Transformers (MTSFT) that incorporates Enhanced Multitemporal Scale Periodic Features and an improved Temporal Fusion Transformer to achieve reliable SST prediction performance and provide multilevel interpretability for marine risk assessment.
130. A 1 km hourly high-resolution 3D wind field dataset over the Yangtze River Delta incorporating dynamical downscaling, observational assimilation, and land use updates
Core Problem: The coarse spatial resolution of existing reanalysis data (e.g., ERA5) limits the ability to capture local-scale atmospheric processes, impacting applications like extreme weather forecasting.
Key Innovation: Developed YRD1km, a 1 km hourly high-resolution 3D wind field dataset for the Yangtze River Delta, generated through dynamical downscaling of ERA5 using a customized WRF model, integrating multi-source observational nudging and high-resolution land use parameterization, demonstrating superior performance in capturing fine-scale wind gradients and terrain-induced circulations.
131. Paleo-erosion rates of the Shanxi Rift since the late Miocene: Insights from the <sup>10</sup>Be and <sup>26</sup>Al records obtained from core SG-1 in the Yuncheng Basin
Core Problem: Quantifying long-term erosion rates in the Shanxi Rift to understand the tectonic and climatic drivers of surface processes, particularly how they differ from other regions.
Key Innovation: Derived paleo-erosion rates for the Shanxi Rift over the past ~6.6 Ma using cosmogenic nuclides (10Be and 26Al) from core SG-1, revealing two distinct erosion pulses primarily governed by regional tectonic activities (uplift of Luliang Mountain and regional activities during Late Pliocene to Early Pleistocene), contrasting with climate-driven erosion in western China.
132. Monitoring water level dynamics in neotropical peatlands with earth observation data
Core Problem: Poor understanding of peatland hydrology due to limited accessibility and sparse field measurements, hindering effective monitoring of water level dynamics.
Key Innovation: Demonstrates the potential of combining L-band SAR backscatter with vegetation structure metrics from GEDI and other remote sensing data to monitor and predict above-ground water level variations across diverse neotropical peatland ecosystems with high accuracy (R2 = 0.8–0.94), providing a methodological pathway for regional hydrological monitoring.
133. Numerical Simulations of Sulfate Formation via the Diffusive Reaction Between Ca2+ in Volcanic Ash and SO2 Gas: Evaluation of SO2 Losses From Eruption Clouds During the 1991 Pinatubo Eruption
Core Problem: Quantifying the amount and efficiency of SO2 scavenging by volcanic ash in eruption clouds to better predict the impact of volcanic eruptions on global climate.
Key Innovation: Developed a 3D numerical simulation combining volcanic clouds and diffusion modeling to estimate SO2 scavenging, showing that pyroclastic flows are efficient at high-temperature SO2 scavenging and that 8.0%–90% of SO2 emitted during the 1991 Pinatubo eruption could be scavenged.
134. Constraints on Past CO2 and Climate Sensitivity From Global Temperature and Sea Level Reconstructions Across the Plio‐Pleistocene
Core Problem: Previous reconstructions of global mean sea level (GMSL) showed less variability throughout the Pleistocene, leading to uncertainties in constraining past CO2 concentrations and equilibrium climate sensitivity from these records.
Key Innovation: Assessed new reconstructions of global mean surface temperature and GMSL (which include large Pleistocene variability) using an energy balance model. Results suggest a high CO2 scenario (300–450 ppm before 2 Ma) is most consistent with the new climate reconstructions, yielding an equilibrium climate sensitivity of 1.8–1.9 K, which is slightly lower than some recent estimates.
135. Surface Observations From Atmospheric Radiation Measurement Sites Constrain the Anthropogenic Contribution to Cloud Droplet Number
Core Problem: Uncertainty in anthropogenic forcing driven by aerosol-cloud interactions (aci), specifically the change in cloud droplet number concentration (Nd) due to changes in aerosols, limits the ability to infer Earth system sensitivity from historical records.
Key Innovation: Combined a perturbed parameter ensemble from a global Earth system model with surface observations of cloud condensation nuclei (CCN) and single-layer-cloud Nd to constrain the anthropogenic contribution to present-day Nd. This observational evidence constrains the preindustrial to present-day change in Nd to be between 11 and 43 cm-3, suggesting a stronger historical aerosol cooling and, in turn, larger future warming.
136. MERG3R: A Divide-and-Conquer Approach to Large-Scale Neural Visual Geometry
Core Problem: Existing neural visual geometry models are fundamentally limited by GPU memory, preventing them from scaling to large, unordered image collections for high-quality 3D reconstruction.
Key Innovation: MERG3R, a training-free divide-and-conquer framework that reorders and partitions images into overlapping subsets, reconstructs them independently, and then merges the local reconstructions through global alignment and bundle adjustment, enabling large-scale 3D modeling beyond native memory limits.
137. VLMFusionOcc3D: VLM Assisted Multi-Modal 3D Semantic Occupancy Prediction
Core Problem: Current voxel-based occupancy models in autonomous driving struggle with semantic ambiguity in sparse geometric grids and performance degradation under adverse weather conditions.
Key Innovation: VLMFusionOcc3D, a multimodal framework for 3D semantic occupancy prediction, which uses VLM linguistic priors (InstVLM) to anchor ambiguous features, dynamically re-weights sensor contributions based on weather (WeathFusion), and aligns geometry (DAGA loss), significantly enhancing performance in challenging weather scenarios for autonomous driving.
138. Towards an Incremental Unified Multimodal Anomaly Detection: Augmenting Multimodal Denoising From an Information Bottleneck Perspective
Core Problem: Existing incremental unified multimodal anomaly detection models suffer from catastrophic forgetting, exacerbated by spurious and redundant features, especially in naive aggregations of unimodal architectures.
Key Innovation: Introduction of IB-IUMAD, a novel denoising framework that uses a Mamba decoder to disentangle inter-object feature coupling and an information bottleneck fusion module to filter redundant features, thereby preserving discriminative information and mitigating catastrophic forgetting.
139. SEP-YOLO: Fourier-Domain Feature Representation for Transparent Object Instance Segmentation
Core Problem: Transparent object instance segmentation is challenging due to boundary blur, low contrast, and high dependence on background context, which existing methods struggle with due to their reliance on strong appearance cues.
Key Innovation: SEP-YOLO, a novel framework integrating a dual-domain collaborative mechanism. It includes a Frequency Domain Detail Enhancement Module to enhance weak high-frequency boundary components and a multi-scale spatial refinement stream (Content-Aware Alignment Neck and Multi-scale Gated Refinement Block) for precise feature alignment and boundary localization.
140. Causal Learning Should Embrace the Wisdom of the Crowd
Core Problem: Learning complex causal structures (DAGs) from observational data is challenging due to combinatorial explosion and inherent ambiguities, especially when knowledge is fragmented and imperfect across different sources.
Key Innovation: A systematic framework integrating scalable crowdsourcing, interactive expert knowledge elicitation, robust aggregation techniques, and LLM-based simulation to synthesize fragmented insights and recover a global causal structure unachievable by any individual agent alone.
141. FiDeSR: High-Fidelity and Detail-Preserving One-Step Diffusion Super-Resolution
Core Problem: Existing diffusion-based image super-resolution methods struggle to simultaneously preserve fine details and ensure high-fidelity reconstruction, often resulting in suboptimal visual quality.
Key Innovation: FiDeSR, a high-fidelity and detail-preserving one-step diffusion super-resolution framework that introduces a detail-aware weighting strategy during training, low- and high-frequency adaptive enhancers during inference, and residual-in-residual noise refinement for improved accuracy and detail recovery.
142. MiM-DiT: MoE in MoE with Diffusion Transformers for All-in-One Image Restoration
Core Problem: All-in-one image restoration is challenging because different degradation types (e.g., haze, blur, noise, low-light) impose diverse requirements on restoration strategies, making it difficult for a single model to handle them effectively.
Key Innovation: MiM-DiT, a unified image restoration framework that integrates a dual-level Mixture-of-Experts (MoE) architecture with a pretrained diffusion model, enabling coarse-grained adaptation across major degradation types (Inter-MoE) and fine-grained modulation for intra-class variations (Intra-MoE).
143. R3GW: Relightable 3D Gaussians for Outdoor Scenes in the Wild
Core Problem: Existing 3D Gaussian Splatting (3DGS) methods do not explicitly model scene illumination, making them unsuitable for relighting tasks and struggling to reconstruct scenes captured in the wild with unconstrained photo collections featuring changing lighting conditions.
Key Innovation: R3GW, a novel method that learns a relightable 3DGS representation of an outdoor scene captured in the wild by separating the scene into a relightable foreground and a non-reflective background, and modeling view-dependent lighting effects using Physically Based Rendering.
144. Harmonic Beltrami Signature Network: a Shape Prior Module in Deep Learning Framework
Core Problem: Efficiently extracting and utilizing shape prior information in deep learning frameworks for 2D simply connected shapes, while maintaining invariance to transformations, is challenging.
Key Innovation: Introduces the Harmonic Beltrami Signature Network (HBSN), a novel deep learning architecture for computing the Harmonic Beltrami Signature (HBS) from binary-like images, providing a robust shape representation and improving segmentation models by embedding geometric shape information.
145. Contextual Latent World Models for Offline Meta Reinforcement Learning
Core Problem: Learning effective task representations without supervision remains a challenge in context-based offline meta-reinforcement learning, leading to limited generalization to unseen tasks.
Key Innovation: Introduces contextual latent world models, which condition latent world models on inferred task representations and train them jointly with the context encoder, enforcing task-conditioned temporal consistency to yield more expressive task representations and significantly improve generalization to unseen tasks.
146. Breaking the Prototype Bias Loop: Confidence-Aware Federated Contrastive Learning for Highly Imbalanced Clients
Core Problem: Prototype bias in federated contrastive learning caused by local class imbalance and data heterogeneity, leading to accumulation of errors across communication rounds.
Key Innovation: Proposing Confidence-Aware Federated Contrastive Learning (CAFedCL) which uses a confidence-aware aggregation mechanism to downweight high-variance local prototypes, integrates generative augmentation for minority classes, and applies geometric consistency regularization to stabilize inter-class structure.
147. SEHFS: Structural Entropy-Guided High-Order Correlation Learning for Multi-View Multi-Label Feature Selection
Core Problem: Existing information-theoretic methods for multi-view multi-label feature selection struggle to learn high-order structural correlations and are prone to converging to local optima.
Key Innovation: Proposing SEHFS, which converts the feature graph into a structural-entropy-minimizing encoding tree to quantify and learn high-order feature correlations, and uses a framework fusing information theory and matrix methods to balance global and local optimization.
148. cPNN: Continuous Progressive Neural Networks for Evolving Streaming Time Series
Core Problem: Existing solutions for unbounded data streams with temporal dependencies and concept drift address these issues separately and struggle with catastrophic forgetting.
Key Innovation: Proposes Continuous Progressive Neural Networks (cPNN), a continuous version of Progressive Neural Networks, designed to jointly handle concept drifts, temporal dependencies, and bypass catastrophic forgetting in streaming time series data.
149. IoUCert: Robustness Verification for Anchor-based Object Detectors
Core Problem: Scaling formal robustness verification to object detection is difficult due to complex non-linear coordinate transformations and Intersection-over-Union (IoU) metrics.
Key Innovation: Introduces IoUCert, a novel formal verification framework for anchor-based object detectors that uses a coordinate transformation and Interval Bound Propagation to derive optimal IoU bounds, enabling robustness verification for models like SSD and YOLO.
150. Reinforcement Learning with Symbolic Reward Machines
Core Problem: Traditional Reward Machines in Reinforcement Learning require manual user input for creating labeling functions, limiting their applicability.
Key Innovation: Proposes Symbolic Reward Machines (SRMs) that consume standard environment output and process observations directly through guards represented by symbolic formulas, eliminating the need for manual labeling functions while providing interpretable task representations.
151. Context Adaptive Extended Chain Coding for Semantic Map Compression
Core Problem: Efficient compression of semantic maps is needed for applications in robotics, autonomous systems, and extended reality, while preserving structured semantic information.
Key Innovation: Introduces a novel chain-coding-based framework for lossless semantic map compression, utilizing an extended chain code (ECC), a context-adaptive entropy coding scheme, and a skip-coding mechanism to exploit contour topology and shared boundaries, achieving significant bitrate reduction and runtime efficiency.
152. ProSMA-UNet: Decoder Conditioning for Proximal-Sparse Skip Feature Selection
Core Problem: U-shaped encoder-decoder architectures like U-Net propagate low-level textures, background clutter, and acquisition noise through skip connections, allowing irrelevant information to bypass deeper semantic filtering, especially in low-contrast imaging.
Key Innovation: Proposes ProSMA-UNet, which reformulates skip gating as a decoder-conditioned sparse feature selection problem. It constructs a multi-scale compatibility field and enforces explicit sparsity via an L1 proximal operator with learnable thresholds, combined with decoder-conditioned channel gating, to remove noisy and semantically irrelevant activations.
153. Specificity-aware reinforcement learning for fine-grained open-world classification
Core Problem: Reasoning Large Multimodal Models (LMMs) tend to produce overly generic predictions when performing fine-grained image classification in open-world settings, despite possessing intrinsic fine-grained domain knowledge.
Key Innovation: Proposes SpeciaRL, a novel specificity-aware reinforcement learning framework to fine-tune reasoning LMMs. It introduces a dynamic, verifier-based reward signal anchored to the best predictions within online rollouts, promoting specificity while respecting the model's capabilities to prevent incorrect predictions.
154. LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory
Core Problem: Scaling dense 3D reconstruction from feedforward geometric foundation models to extremely long video sequences (minutes-long) is bottlenecked by quadratic attention complexity or limited effective memory in recurrent designs, leading to issues like scale drift and lack of global consistency.
Key Innovation: Introduces LoGeR, a novel architecture for long-context geometric reconstruction that uses a learning-based hybrid memory module (parametric Test-Time Training memory for global frame anchoring and non-parametric Sliding Window Attention for uncompressed context) to achieve robust, globally consistent 3D reconstruction over unprecedented horizons from thousands of frames.
155. Combinatorial Sparse PCA Beyond the Spiked Identity Model
Core Problem: Existing combinatorial sparse PCA algorithms are typically only provably successful under the restrictive 'spiked identity model,' failing for general covariance matrices where SDP-based algorithms are required.
Key Innovation: Developed the first combinatorial method for sparse PCA that provably succeeds for general covariance matrices, using a variant of the truncated power method, with specific sample and time complexities.
156. Neural quantum support vector data description for one-class classification
Core Problem: Existing one-class classification (OCC) techniques struggle with the increasing complexity and dimensionality of modern datasets, lacking sufficient expressivity and efficiency for advanced anomaly detection.
Key Innovation: Introduces Neural Quantum Support Vector Data Description (NQSVDD), a classical-quantum hybrid framework for OCC that performs end-to-end optimized hierarchical representation learning, achieving competitive or superior AUC performance on benchmark datasets.
157. A Covering Framework for Offline POMDPs Learning using Belief Space Metric
Core Problem: Offline policy evaluation (OPE) methods for partially observable Markov decision processes (POMDPs) suffer from the curse of horizon and memory due to the necessity of inferring hidden states from past observations.
Key Innovation: A novel covering analysis framework was introduced that exploits the intrinsic metric structure of the belief space to relax traditional coverage assumptions. This framework yields tighter error bounds and coverage requirements in terms of belief space metrics, mitigating exponential blow-ups in horizon and memory length and illustrating improved sample efficiency for OPE algorithms.
158. Infinite dimensional generative sensing
Core Problem: Existing theoretical guarantees for deep generative models in inverse problems are confined to finite-dimensional vector spaces, creating a gap when physical signals are modeled as functions in Hilbert spaces.
Key Innovation: Presents a rigorous framework for generative compressed sensing in Hilbert spaces, extending notions like local coherence and Restricted Isometry Property to infinite dimensions, and deriving optimal, resolution-independent sampling distributions for stable recovery.
159. Physics-informed post-processing of stabilized finite element solutions for transient convection-dominated problems
Core Problem: Numerical simulation of convection-dominated transient transport phenomena poses significant computational challenges due to sharp gradients and propagating fronts, leading to spurious oscillations in classical methods and difficulties for standalone PINNs to capture sharp structures.
Key Innovation: Presents a hybrid computational framework extending PASSC to unsteady problems, combining a semi-discrete stabilized finite element method with a PINN-based correction strategy applied selectively near the terminal time, significantly improving accuracy for transient convection-diffusion-reaction equations.
160. Learning Lagrangian Interaction Dynamics with Sampling-Based Model Order Reduction
Core Problem: High computational expense of simulating physical systems governed by Lagrangian dynamics, especially for localized, highly dynamic behaviors like fluids and granular media.
Key Innovation: Proposes GIOROM (Geometry-Informed Reduced-Order Modeling), a sampling-based reduction framework that evolves Lagrangian systems in physical space using data-driven neural PDE operators and a learnable kernel parameterization, achieving significant dimensionality reduction and high-fidelity evaluations across diverse Lagrangian regimes.
161. Language-guided Open-world Video Anomaly Detection under Weak Supervision
Core Problem: Existing video anomaly detection (VAD) methods assume invariable anomaly definitions, making them unsuitable for open-world scenarios where expected events and anomaly definitions can change based on user requirements.
Key Innovation: Proposes a novel open-world VAD paradigm with variable definitions guided by natural language at inference time, and introduces LaGoVAD, a model that dynamically adapts anomaly definitions under weak supervision using dynamic video synthesis and contrastive learning. Also collects PreVAD, the largest and most diverse video anomaly dataset to date.
162. Strictly Constrained Generative Modeling via Split Augmented Langevin Sampling
Core Problem: The limitation of deep generative models in scientific and engineering applications due to the lack of guarantees on physical plausibility and the inability to rigorously enforce known physical constraints on generated outputs.
Key Innovation: CASAL (Constrained Alternated Split Augmented Langevin), a novel primal-dual sampling algorithm that rigorously enforces mathematical constraints in generative models (specifically diffusion models) through variable splitting, improving forecast accuracy and preservation of critical conserved quantities in complex physical systems and optimal control.
163. An Explainable and Interpretable Composite Indicator Based on Decision Rules
Core Problem: The need for explainable, interpretable, and transparent composite indicators in multi-criteria decision aiding, beyond just producing a final score or classification.
Key Innovation: A novel framework for constructing explainable and interpretable composite indicators using if-then decision rules, based on the Dominance-based Rough Set Approach, which can handle various scenarios, missing values, and efficiently induces minimal rules.
164. TTT3R: 3D Reconstruction as Test-Time Training
Core Problem: Modern Recurrent Neural Networks for 3D reconstruction suffer from significant performance degradation and limited length generalization when applied beyond their training context length.
Key Innovation: Proposes TTT3R, a training-free intervention that frames 3D reconstruction as an online learning problem and derives a closed-form learning rate for memory updates based on alignment confidence, substantially improving length generalization (2x in global pose estimation) while maintaining efficiency.
165. Human3R: Everyone Everywhere All at Once
Core Problem: Existing 4D human-scene reconstruction methods rely on multi-stage pipelines, iterative refinement, and heavy dependencies (e.g., human detection, depth estimation, SLAM), making them inefficient and complex.
Key Innovation: Introduces Human3R, a unified, feed-forward framework that jointly recovers global multi-person SMPL-X bodies, dense 3D scenes, and camera trajectories in a single pass, achieving real-time performance (15 FPS) with low memory footprint and superior performance across multiple reconstruction tasks.
166. Efficient Resource-Constrained Training of Transformers via Subspace Optimization
Core Problem: The expanding scale of modern neural networks, particularly transformers, creates significant memory and computational obstacles for efficient on-device training, hindering energy conservation and data privacy efforts.
Key Innovation: Weight-Activation Subspace Iteration (WASI), a subspace-based training method for transformers that restricts training to a fixed subspace, mitigating memory bottlenecks and boosting inference efficiency, achieving comparable accuracy with substantial reductions in memory usage and computational cost.
167. Graph Homomorphism Distortion: A Metric to Distinguish Them All and in the Latent Space Bind Them
Core Problem: Existing expressivity measures for graph neural networks primarily focus on graph structure, neglecting node features, making it difficult to assess similarity between graphs based on both structural and feature proximity.
Key Innovation: Introduces 'graph homomorphism distortion,' a new (pseudo-)metric that measures the minimal worst-case distortion of node features when mapping one graph to another, thereby complementing existing expressivity measures and improving GNN predictive capabilities by defining structural encodings.
168. SURFACEBENCH: A Geometry-Aware Benchmark for Symbolic Surface Discovery
Core Problem: Existing benchmarks for symbolic regression primarily evaluate low-dimensional scalar functions and lack geometry-aware metrics, failing to capture structural and geometric equivalence crucial for discovering complex 3D surface equations governing physical phenomena.
Key Innovation: SURFACEBENCH, the first geometry-aware benchmark for symbolic discovery of 3D surfaces, comprising 183 science-inspired equations across various representations, incorporating symbolic equivalence checks and geometric metrics (Chamfer, Hausdorff distances) to evaluate functional fidelity beyond algebraic syntax, revealing limitations of current methods in surface-level reasoning.
169. Multi-Scenario Highway Lane-Change Intention Prediction: A Temporal Physics-Informed Multi-Modal Framework
Core Problem: Lane-change intention prediction for autonomous driving is difficult in naturalistic traffic due to noisy kinematics, severe class imbalance, and limited generalization across heterogeneous highway scenarios.
Key Innovation: Proposes TPI-AI, a hybrid framework fusing deep temporal representations (Bi-LSTM) with physics-inspired interaction cues for multi-scenario highway lane-change intention prediction. It incorporates imbalance-aware optimization, achieving robust prediction across different highway scenarios and outperforming baselines.
170. UniDrive-WM: Unified Understanding, Planning and Generation World Model For Autonomous Driving
Core Problem: Existing autonomous driving approaches typically treat perception, prediction, and planning as separate modules, and current Vision-Language Model (VLM)-based planning methods lack a unified framework for these tasks.
Key Innovation: Proposes UniDrive-WM, a unified VLM-based world model that jointly performs driving-scene understanding, trajectory planning, and trajectory-conditioned future image generation within a single architecture. It uses future image predictions as supervisory signals to enhance scene understanding and iteratively refine trajectory generation, improving planning performance by 5.9% in L2 trajectory error and 9.2% in collision rate.
171. Graph Recognition via Subgraph Prediction
Core Problem: Visual relationship recognition, modeled as extracting a graph from an image, remains challenging due to the lack of a canonical, transferable approach, with most existing solutions being problem-specific.
Key Innovation: Development of GraSP (Graph Recognition via Subgraph Prediction), a general, broadly applicable, and transferable method for recognizing diverse types of graphs in images without task-specific modifications.
172. SceneStreamer: Continuous Scenario Generation as Next Token Group Prediction
Core Problem: Most existing data-driven traffic simulation methods for autonomous driving rely on static initialization or log-replay, limiting their ability to model dynamic, long-horizon scenarios with evolving agent populations.
Key Innovation: Proposes SceneStreamer, a unified autoregressive framework for continuous scenario generation that represents the entire scene as a sequence of tokens, enabling realistic, diverse, and adaptive traffic behaviors over an unbounded horizon for autonomous driving simulation.
173. Stochastic Control Methods for Optimization
Core Problem: The need for robust global optimization methods for non-convex and/or non-differentiable objective functions in both Euclidean and Wasserstein spaces.
Key Innovation: Investigates a stochastic control framework for global optimization, approximating the problem with regularized stochastic control problems and proposing Monte Carlo-based derivative-free numerical schemes with theoretical convergence rates.
174. A Boundary Integral-based Neural Operator for Mesh Deformation
Core Problem: Traditional finite element methods for mesh deformation are computationally expensive, and existing neural operators struggle with Dirichlet boundary conditions for vector fields in linear elasticity boundary value problems.
Key Innovation: The Boundary-Integral-based Neural Operator (BINO) is introduced, which uses a direct boundary integral representation with a Dirichlet-type Green's tensor to express internal displacement solely from boundary displacements. This method learns the geometry- and material-aware Green's traction kernel, offering high accuracy, computational efficiency, and potential for cross-geometry adaptation in mesh deformation.
175. Integrating machine learning algorithms and game theory for optimized shipyard site selection in Istanbul
Core Problem: Overcoming limitations of traditional Multi-Criteria Decision-Making (MCDM) methods in shipyard site selection by integrating big data and dynamic optimization, considering multiple complex factors like environment, logistics, security, and energy infrastructure.
Key Innovation: Integrates Machine Learning (ML) algorithms (Random Forest, Categorical Boosting) and Game Theory (Nash Equilibrium) with GIS-based spatial analysis to optimize shipyard site selection, identifying highly suitable zones and specific candidate regions in Istanbul.
176. MFCF: A multimodal cascade fusion system for efficient pipe damage detection
Core Problem: Unimodal detection approaches often fail to comprehensively capture complex characteristics of pipeline damage, leading to missed detections or false assessments, compromising structural integrity and safety.
Key Innovation: Proposes a Multimodal Feature Cascade Fusion System (MFCF) based on strain and acceleration signals, employing Dendrite Networks (DD) for polynomial-based relationships and enhanced learning efficiency, demonstrating superior accuracy and robustness in fiber-reinforced pipe damage detection.
177. Reliability analysis of monopile-supported offshore wind turbines considering load correlations and multi-source uncertainties under data scarcity conditions
Core Problem: Accurately assessing the dynamic reliability of monopile-supported offshore wind turbines (MOWTs) requires considering wind-wave load correlations and multi-source uncertainties, especially under data scarcity.
Key Innovation: Proposes a novel reliability analysis framework fusing Bayesian Bootstrap, Copula theory, and the Probability Density Evolution Method (PDEM) to correct load distributions, quantify epistemic uncertainty, construct a joint wind-wave probability model, and systematically analyze the impact of multi-source uncertainties and load correlations on MOWT dynamic reliability.
178. Synergistic Frequency-Spatial Enhancement With Temporal Correlation for Robust Satellite Video Tracking
Core Problem: Satellite video single object tracking (SOT) is challenged by the adjacency effect, which introduces blurry targets and similar-object interference, leading to severe ambiguity in template matching and error accumulation.
Key Innovation: A Synergistic Perception Tracker (SPTrack) that integrates frequency, spatial, and temporal reasoning through a frequency-spatial separate-interact (FSSI) block and an interframe temporal correlation (IFTC) module, achieving robust object tracking in satellite videos under challenging conditions like blur and similar objects.
179. Edge-Aware Dual-Stream Dynamic Fusion Network for Building Extraction in Remote Sensing Images
Core Problem: Extracting buildings from high-resolution remote sensing images faces significant challenges due to complex urban scenes, spectral ambiguities, geometric diversity, and boundary uncertainties, making it difficult to balance global context with local details and leverage edge priors effectively.
Key Innovation: An Edge-Aware Dual-Stream Dynamic Fusion Network (EA-DSFNet) that uses a dual-path encoder (CNN and Transformer) with a feature reuse and fusion module, a feature selection gate, and a multitask edge-directed decoder to achieve comprehensive multiscale feature representation and improve building extraction accuracy, especially at boundaries.
180. Entanglement-assisted non-local optical interferometry in a quantum network
Core Problem: The sensitivity of non-local optical measurements at low light intensities, crucial for applications like long-baseline telescope arrays, is fundamentally limited by quantum noise and photon losses.
Key Innovation: Demonstrated entanglement-assisted differential phase measurement of weak incident light between two spatially separate stations using entangled quantum memories in a quantum network, overcoming quantum noise limitations and offering potential for new quantum-enhanced optical imaging methods.
181. Understanding Stress-Induced Fragmentation of Ultra-hard Rocks in Abrasive Water Jet Cutting
Core Problem: Efficient fragmentation of ultra-hard rocks in deep excavation environments remains a challenge, limiting operational efficiency and safety in mining and tunneling, and the influence of in situ stress on abrasive water jet cutting is not fully understood.
Key Innovation: This study conducted abrasive water jet cutting experiments under various stress conditions, identifying a 'Jet Stress Threshold' (e.g., 20.88 MPa for granite) that characterizes the transition from inhibition to promotion of fragmentation. It revealed stress-dependent fracture modes (intergranular to transgranular), offering guidance for optimizing water jet technology in stress-laden environments.
182. Morphologic adjustment and sediment balance of the Nile River in Egypt under cascade dam regulation
Core Problem: Insufficient understanding of the geomorphologic adjustments and cumulative channel responses of intensively regulated river systems, such as the Nile River, to cascade dam regulation.
Key Innovation: Developed an integrated GIS-based DEM differencing and 2D hydrodynamic modeling framework to quantify morphologic change (sediment budget, zonal classification) and hydraulic responses (water surface elevation, shear stress) in the cascade-regulated Nile River, revealing spatially heterogeneous aggradation/degradation patterns and demonstrating that each reach responds uniquely based on geomorphic pattern, sediment inputs, and human pressures.
183. Exploring the potential of Harmonized Landsat-Sentinel-2 in predicting boreal forest structure from UAV-LiDAR data in Northwestern America
Core Problem: Boreal forest structural attributes, particularly in complex treeline ecosystems, remain poorly characterized, and satellite-derived products may have inaccuracies.
Key Innovation: Explores the potential of Harmonized Landsat-Sentinel-2 multispectral data to predict UAV-LiDAR-derived boreal forest structure (Canopy Height and Crown Cover) using Random Forest models, highlighting the value of fine-scale LiDAR data for algorithm building and assessments of satellite-derived products.
184. A hybrid multi-task 1D CNN and XGBoost approach for ballast layer characterization using GPR data
Core Problem: Current Ground-Penetrating Radar (GPR) practices for railway ballast layer characterization rely on manual interpretation and heuristic thresholding, which are inefficient and prone to subjectivity, limiting automated subsurface condition monitoring.
Key Innovation: Development of a hybrid multi-task 1D CNN and XGBoost framework that automates ballast layer characterization from raw GPR A-scan signals, jointly estimating ballast thickness, relative dielectric permittivity, and fouling level with high accuracy, thereby advancing intelligent inspection and maintenance management of railway infrastructure.