TerraMosaic Daily Digest: Feb 25, 2026
Daily Summary
This February 25, 2026 digest captures a mature shift from single-hazard description to coupled hazard-process analysis. The strongest papers link seismic loading, hydro-mechanical response, and slope or infrastructure performance in one causal chain: examples include earthquake magnitude classification for operational warning, fault-coupled paleo-landslide evolution, co-seismic tailings instability, and energy-based retaining-wall response in liquefiable ground.
Flood and climate-facing research is likewise moving toward decision-ready products rather than static maps: crop-specific flood damage functions, probabilistic road-function loss under inundation, social-media-derived urban flood intelligence, and uncertainty-aware precipitation correction all directly improve emergency allocation and planning. Across the long tail of method papers, relevance is highest when transferability to geohazard workflows is explicit and uncertainty is quantified.
Key Trends
The main technical direction is integration: mechanism, uncertainty, and operations are being treated as a single design problem.
- Seismo-geotechnical coupling is now central, not peripheral: top studies jointly model earthquake forcing, material degradation, and slope/embankment response, yielding physically interpretable indicators for stability assessment and warning design.
- Landslide science is becoming structure-aware and scale-aware: susceptibility studies increasingly test model-data alignment (e.g., geo-tabular structure effects) and combine mapping with kinematic or reliability constraints, improving transfer beyond single catchments.
- Flood analytics is shifting to actionable risk operations: current work emphasizes transport functionality, crop loss, social sensing, and emergency accessibility under uncertainty, rather than only hazard extent mapping.
- Cryosphere and mountain hazards are entering data-rich monitoring mode: glacier-lake inventory expansion and transferable avalanche forecasting signal stronger foundations for regional, cross-border early action.
- Methodological AI remains high-volume, but quality is increasingly judged by uncertainty control and domain fit: generic models are valuable only when they demonstrably improve geohazard detection, forecasting, or decision support under real data constraints.
Selected Papers
This digest features 176 selected papers from 1034 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. Multi-class earthquake magnitude classification in Kahramanmaraş region using feature engineering: artificial intelligence approaches
Core Problem: Accurately classifying earthquake magnitudes, which is crucial for effective disaster management, early warning, and seismic risk assessment, is challenging due to the complex nature of seismic data.
Key Innovation: Develops a robust multi-class earthquake magnitude classification framework for the Kahramanmaraş region, leveraging extensive feature engineering (integrating seismic energy indicators, temporal dependency, geophysical parameters) and advanced AI models (Gradient Boosting achieving 98.70% accuracy), offering potential for disaster management, early warning, and seismic risk assessment.
2. Landslide susceptibility mapping and kinematic analysis in Bachchangad catchment, Uttarakhand, India: a comparative study using machine learning benchmark classifiers
Core Problem: The need for accurate landslide susceptibility mapping and a holistic geotechnical assessment in complex, landslide-prone terrains like the Uttarakhand Himalaya.
Key Innovation: A comparative study of Logistic Regression (LR) and Random Forest (RF) algorithms for landslide susceptibility mapping using 13 causative factors and a comprehensive landslide inventory. The RF model demonstrated superior performance (AUC 78% vs 71.3% for LR). The study integrated kinematic analysis at key sites to enhance the accuracy and robustness of the susceptibility mapping, providing a framework for landslide risk assessment and disaster management.
3. Structure-driven evaluation of GBDT and neural networks for Geo-Tabular landslide hazard assessment on the Tibetan plateau
Core Problem: Despite the widespread application of machine learning in landslide hazard assessment, there is limited understanding of how the structural properties of geospatial data (termed 'Geo-Tabular Data') influence model adaptability and performance.
Key Innovation: Introduction of the concept of 'Geo-Tabular Data' (characterized by low dimensionality, skewed feature distributions, and weak spatial continuity) and a structure-aware framework to evaluate model adaptability for landslide hazard assessment. The study benchmarked 14 models, revealing that GBDT models (LightGBM, XGBoost, CatBoost) significantly outperformed deep tabular neural networks, attributing this to structural misalignment between model architectures and the intrinsic properties of Geo-Tabular Data. This reframes landslide susceptibility modeling as a Geo-Tabular Learning task, providing practical guidance for large-scale geospatial risk mapping.
4. Field investigations on large-scale instability triggered by the Chenghai-Binchuan fault zone, northwestern Yunnan, China
Core Problem: Systematically investigating the formation mechanisms and chronological evolution of large-scale paleo-landslides triggered by the tectonically active Chenghai-Binchuan fault zone in northwestern Yunnan, China.
Key Innovation: Through field investigations, OSL/ESR dating, and analysis of 284 landslides, it revealed that strong fault–slope coupling, involving long-term creep weakening and abrupt seismic acceleration, is the primary control on the initiation, evolution, and spatial clustering of major paleo-landslides, providing new insights into fault-controlled landslide mechanisms.
5. Variance reduction function modeling for a potential circular slip surface in spatially variable slopes
Core Problem: The need for accurate and efficient probabilistic modeling of geotechnical systems, particularly for spatially variable slopes, where existing variance reduction function (VRF) methods for circular slip surfaces are often oversimplified or limited to specific autocorrelation function types.
Key Innovation: Proposing a one-dimensional arc length-based proxy autocorrelation function (ACF) to systematically evaluate multiple 2D ACF types for VRF modeling along circular slip surfaces, leading to a Variance Reduction Stochastic Response Surface Method (VRSRSM) for efficient probabilistic slope stability analysis, validated with a real-world earthquake-induced landslide.
6. Failure mechanism of closed fine-grained tailing deposits due to co-seismic damage accumulation: Insight from shaking table tests
Core Problem: Understanding the failure mechanism of closed, compacted fine-grained tailings deposits due to co-seismic damage accumulation from multi-stage earthquakes, a process that is understudied.
Key Innovation: Used shaking table tests and PIV to simulate and monitor progressive failure of saturated dense tailings, identifying cyclic mobility as the cause of slope failure in three stages, proposing a plastic displacement coefficient (D), and highlighting increased instability risk from damage accumulation.
7. New Insights Into Basal Slip Processes and Kinematics of a Giant Pleistocene Submarine Mass Transport Complex, West of New Zealand's North Island
Core Problem: Basal slip processes controlling submarine landslide runout are poorly understood due to limited observations, hindering accurate assessment of sediment mobilization impacts and tsunamigenic potential.
Key Innovation: Uses 2D/3D seismic data to examine the evolution and kinematics of a giant Pleistocene MTC, identifying distinct failure sectors with varying runout. Proposes a new conceptual model where remobilized pre-existing MTC material behaves like viscous mud, creating high-friction zones that arrest movement, influencing subsequent sliding processes and improving tsunamigenic hazard assessments.
8. Disaster Question Answering with LoRA Efficiency and Accurate End Position
Core Problem: Individuals in natural disaster situations often lack domain-specific knowledge and experience, and existing LLM-based Q&A systems may provide irrelevant or hallucinated information, exacerbating confusion and hindering appropriate responses.
Key Innovation: A disaster-focused question answering system for Japanese disaster scenarios, utilizing a BERT-based architecture with Bi-LSTM and LoRA efficiency optimization, achieving high End Position accuracy (70.4%) and a 0.885 Span F1 score with significantly fewer parameters, suitable for real disaster response.
9. Instrumented 74-mm Dynamic Cone Penetration and Shear Wave Velocity Tests for Assessing Liquefaction in Gravelly Soils Subjected to the 2020 Petrinja, Croatia, Earthquake
Core Problem: Evaluating liquefaction in gravelly soils is challenging due to interference with traditional in situ tests, and there's a need for improved methods and case history data for probabilistic triggering relationships.
Key Innovation: Application and evaluation of large-diameter (74-mm) instrumented DPT (iDPT, cDPT) and shear wave velocity (Vs) profiling (MASW, HVSR) at sites affected by the 2020 Petrinja earthquake, demonstrating that energy-corrected DPT and Vs-based methods correctly predict liquefaction manifestation in gravelly soils and provide valuable case history data for refining triggering relationships.
10. Flood damage functions for rice: synthesizing evidence and building data-driven models
Core Problem: Flood damage models for crops, particularly rice, are scarce, often lack validation, uncertainty estimates, and assessments of their performance when transferred to new regions.
Key Innovation: A comprehensive review of 20 flood damage models for rice and the development of a suite of empirical data-driven models (deterministic/probabilistic stage-damage functions, Bayesian regression, Random Forest) using data from Thailand and Myanmar, assessing their predictive performance and spatial transferability.
11. Rapidly changing lake-terminating glaciers in High Mountain Asia: a dataset from 1990 to 2022
Core Problem: A lack of a comprehensive inventory for lake-terminating glaciers (LTGs) and their associated proglacial lakes across High Mountain Asia (HMA) limits the understanding of their spatial heterogeneity in glacier change and associated hazards.
Key Innovation: Construction of a comprehensive, semi-automated inventory dataset of LTGs and proglacial lakes in HMA from 1990 to 2022, quantifying their rapid changes (e.g., 81.7% increase in proglacial lake area) and highlighting regions with high concentrations of glacier-lake systems, to improve understanding of glacier-related hazards.
12. Spatiotemporal vulnerability analysis of China’s high-speed railway network to seismic hazards
Core Problem: Assessing the vulnerability of complex infrastructure like high-speed railway networks to seismic hazards, considering both spatial and temporal aspects.
Key Innovation: Integrates complex network theory with spatiotemporal hotspot analysis to assess the vulnerability of China's high-speed railway network to seismic hazards.
13. Better understanding flood resilience in rural china: a review integrating regional disaster systems and 4R theory
Core Problem: A lack of systematic conceptual research, integrated evaluation methods, and comprehensive understanding of influencing factors and improvement strategies for rural community flood disaster resilience in China.
Key Innovation: Develops an analytical framework for rural community flood disaster resilience by integrating regional disaster system elements and 4R crisis management theory, identifying key research gaps, core influencing factors, and proposing a strategy system for improvement, while also advancing the theoretical resilience framework by integrating static and dynamic perspectives.
14. Intensification of typhoon extremes along the South China coast: a standard deviation perspective
Core Problem: Traditional analyses of typhoon intensification often focus on mean shifts, potentially missing critical changes in the distribution of extreme events, which are crucial for assessing destructive potential and secondary hazards.
Key Innovation: Demonstrates that typhoon extreme intensification along the South China coast is more consistently expressed through an increase in the standard deviation of maximum wind speed and minimum central pressure, rather than a mean shift, providing a standard-deviation-driven interpretation for increased probability of extreme outcomes and a quantitative basis for non-stationary extreme value analyses in hazard assessment.
15. Evaluating and enhancing the performance of satellite precipitation products by considering uncertainty in rain gauge observations
Core Problem: Accurate evaluation and bias correction of Satellite Precipitation Products (SPPs) are challenging, especially when accounting for the inherent uncertainty in ground-based rain gauge observations, which limits their utility for applications like flood forecasting.
Key Innovation: Develops a machine-learning-driven hierarchical framework that accounts for rain gauge observation uncertainty (modeling them as interval-valued data) to evaluate and correct SPP biases, introducing a novel distance-based evaluation index and a neural network-based correction model that significantly reduces bias, particularly for heavy precipitation events, thereby enhancing SPP utility for flood forecasting.
16. Optimizing surveillance efficiency with deep learning-driven flood segmentation
Core Problem: Accurate identification of flood-affected regions in ground-surveillance images is crucial for rapid assessment and disaster response, but existing methods often lack accuracy and generalization, especially for local or regional scenes.
Key Innovation: Two deep learning-based segmentation architectures for ground-level flood detection, including a novel framework called Flood-X (utilizing an Xception-based encoder with a lightweight decoder). The proposed model achieved state-of-the-art performance, producing accurate segmentation masks on unseen real-world flood images with an mIoU of 94%, surpassing existing approaches in both efficiency and accuracy.
17. Experimental study of self-healing effect of slip zone soil in red-bed translational landslide
Core Problem: The conventional view of residual strength in slip zone soil as a static parameter overlooks the potential for self-healing, which influences the reactivation potential of landslides, particularly in red-bed translational landslides.
Key Innovation: Experimentally demonstrated the self-healing effect of red-bed slip zone soil, quantifying how strength recovery is influenced by consolidation time and normal stress, and identified the underlying mechanisms (consolidation, mineral characteristics), providing insights into delaying landslide reoccurrence.
18. Taxonomy-driven exposure mapping for earthquake risk assessment in the active tectonic setting of Eastern Indonesia
Core Problem: The limited sector-specific exposure and loss assessments incorporating building taxonomy in Eastern Indonesia, a highly tectonically active and earthquake-vulnerable region.
Key Innovation: Developed a taxonomy-based exposure and seismic loss assessment framework by integrating spatial datasets with construction typologies (GED4ALL), revealing spatial and sectoral variability in structural vulnerability and economic losses under different earthquake scenarios, thus enabling precise identification of high-risk areas for mitigation.
19. Pseudo-static Stability Analysis of Jointed Rock Slopes Reinforced by Passive Bolts
Core Problem: Accurately assessing the stability of jointed rock slopes reinforced by passive bolts under combined gravity and seismic loading, considering complex material and joint properties.
Key Innovation: A limit analysis kinematic approach using mixed modeling (3D continuum rock, beam-like bolts, 2D cohesive-frictional joints, Hoek-Brown criterion) to derive rigorous lower bound estimates of required reinforcement strength for pseudo-static seismic conditions, validated against finite element solutions.
20. Fluid-driven particle erosion promotes the frictional instability of Granular Shear-Zone gouge
Core Problem: How fluid-driven particle erosion reorganizes grain-scale structures to modulate sliding stability in Earth's shear zones, influencing the initiation of earthquakes, landslides, and related failures, remains poorly understood.
Key Innovation: Systematically investigated the frictional responses to varying intensities of fluid-driven particle erosion in granular shear-zone gouge using direct shear experiments and AE monitoring. Developed a force-chain-based micromechanical framework to elucidate the weakening mechanisms and how stress/geometry modulate this effect, advancing understanding of fluid-driven instability in landslides and faults.
21. Layered or Simple: Making Impact Chains more accessible and useful to stakeholders
Core Problem: The challenge of designing models and tools that accurately convey the intricacies of multi-hazards, compounded impacts, vulnerabilities, and mitigation actions to diverse stakeholders without overwhelming them, thus hindering effective decision-making in DRR/M.
Key Innovation: Introducing Layered Impact Chains and Simplified Impact Chains as frameworks to streamline stakeholders’ understanding of complex multi-hazard models, validated through a case study involving an earthquake-triggered cascade of secondary hazards (dam-break flood, fires, liquefaction), thereby empowering stakeholders to craft integrated and more effective DRR/M solutions.
22. Human health implications of cascading effects in chemical plants triggered by seismic NaTech events
Core Problem: Assessing the human health impacts of cascading effects (toxic substance release) in industrial chemical plants triggered by seismic events (NaTech incidents), particularly considering the vulnerabilities of steel tanks and the lack of holistic assessment frameworks.
Key Innovation: This research presents fragility curves for various steel tank typologies under seismic loads and introduces a novel holistic methodological framework to quantitatively assess environmental contamination, human exposure, and potential health effects on workers and communities following hazardous substance releases.
23. From the Swiss Alps to the Pyrenees: Evaluating the transferability of machine learning models for avalanche forecasting
Core Problem: The challenge of addressing spatial variation in predictive performance and verifying the capability to transfer machine learning avalanche forecasting models, originally trained in Switzerland, to other mountain regions with different snow climate conditions.
Key Innovation: Demonstrated that machine learning models for avalanche danger and wet-snow avalanche activity, trained on Swiss data, can be transferred to the Spanish Central Pyrenees with only a moderate loss in performance (60%-70% agreement for danger level), suggesting their potential for integration into forecasting in new regions.
24. Seismic performance of T-shaped retaining walls in liquefiable sites based on Arias Intensity
Core Problem: Assessing the seismic performance of T-shaped retaining walls in liquefiable sites, particularly the limitations of relying solely on Peak Ground Acceleration (PGA) for dynamic displacement response.
Key Innovation: Investigated seismic performance using large-scale shaking table tests, correlating excess pore water pressure and wall displacement with Arias Intensity (I_a), and developing empirical models that account for total seismic energy, revealing a two-stage displacement behavior and an energy threshold.
25. A Boundary Element Model for Assessing Large‐Scale Pressurization in Faulted Geological Storage Systems
Core Problem: Computational challenges in efficiently assessing large-scale pressurization in complex, faulted geological storage systems for applications like wastewater injection and CO2 sequestration.
Key Innovation: Development of a computationally efficient boundary element model integrating single-phase semi-analytical solutions to simulate pressure propagation in multilayered 3D systems with vertical faults, enabling efficient basin-scale pressurization assessment and risk evaluation.
26. VCDF: A Validated Consensus-Driven Framework for Time Series Causal Discovery
Core Problem: Existing time series causal discovery methods are often sensitive to noise, non-stationarity, and sampling variability, limiting their robustness and reliability in understanding dynamic systems.
Key Innovation: Proposes the Validated Consensus-Driven Framework (VCDF), a method-agnostic layer that improves robustness of causal discovery by evaluating the stability of causal relations across blocked temporal subsets, demonstrating significant improvements in F1 scores on synthetic and simulated real-world data.
27. Proportion optimization and performance-oriented design method of fluidized solidified soil for offshore wind turbine scour repair
Core Problem: Local scour poses a critical threat to the structural safety of offshore wind turbine foundations, and effective materials for scour remediation require optimized mix proportions.
Key Innovation: A multi-objective optimization framework is developed for the mix proportioning of cement-based fluidized solidified soil (FSS) for scour repair, utilizing least squares regression and NSGA-II/NSGA-III algorithms to establish a performance-oriented design method, experimentally validated for accuracy.
28. Squeaking at soft–rigid frictional interfaces
Core Problem: Understanding the mechanisms behind squeaking at soft-rigid frictional interfaces, particularly at velocities that produce squeaking, and how these relate to rupture dynamics.
Key Innovation: Experimental investigation using high-speed imaging and acoustic analysis reveals that squeaking at soft-rigid interfaces arises from intersonic opening slip pulses propagating at the shear wave speed, analogous to earthquake ruptures, and that geometric confinement can stabilize rupture and transform irregular dynamics into coherent pulse trains.
29. Risk communication in Nepal: a scoping review of trends, gaps and future directions
Core Problem: Risk communication in Nepal, a disaster-prone country, is fragmented, largely non-proactive, and lacks academic research applying established theories across various hazards.
Key Innovation: Provides a novel synthesis of fragmented evidence on risk communication in Nepal over 4 decades, using a '5 Ws and H' framework to identify trends and gaps across multiple hazards, and offers policy recommendations for institutionalizing government-led communication and tailored tools.
30. Bridging the gender data gap: measuring welfare losses from climate-related disasters for men and women
Core Problem: The significant gender data gap in disaster contexts hinders equitable climate action and effective, inclusive disaster response and recovery policies.
Key Innovation: Develops and applies a novel rapid assessment survey tool to collect individual-level, sex-disaggregated welfare loss data across multiple domains (food security, sanitation, healthcare, work, gender-based violence) immediately after a climate-related disaster (Cyclone Remal), providing an empirically grounded basis for equitable response and resilience planning.
31. Urban flood information extraction and spatial-temporal analysis of “7.20” Zhengzhou flood risk from social media perspective
Core Problem: Traditional urban flood monitoring methods suffer from insufficient spatio-temporal coverage, response delays, and potential failure during extreme weather events.
Key Innovation: An integrated framework combining multimodal social media data collection, natural language processing for flood event location and context extraction, Mask R-CNN for water depth identification from images, and GIS-based spatial correlation analysis. This framework provides high spatio-temporal resolution flood information, accurate flood location extraction, and insights into public sentiment and spatial clustering patterns, enhancing urban disaster prevention and resilience.
32. Probabilistic functionality assessment of road networks for medical emergency vehicles during flooding
Core Problem: Current analyses of road network functionality loss during floods are often based on specific flood scenarios, lacking a probabilistic assessment of stability loss for various vehicle types and its impact on emergency services.
Key Innovation: A method to probabilistically assess flood risk to road networks by evaluating the probability of stability loss for different vehicle types (SUVs/emergency vehicles, cars). The study generated flood risk maps and measured reduced accessibility via isochrones for emergency vehicles, offering strategies to mitigate flood impacts on road networks and prepare emergency medical services.
33. Robust and efficient iterative algorithm for copula-based FORM and its application to slope reliability analysis
Core Problem: Accurately and efficiently performing slope reliability analysis, especially under incomplete probability information and considering the dependence structure of shear strength parameters, remains a challenge for conventional FORM algorithms.
Key Innovation: Developed a robust and efficient iterative algorithm (iHLRF-BFGS) for copula-based FORM, which improves accuracy and efficiency in finding the design point for slope reliability analysis, and demonstrated its ability to comprehensively investigate the impact of shear strength parameter dependence, highlighting the importance of proper copula selection.
34. Study on Mechanical Response and Energy Dissipation Mechanisms of Pre-Damaged Rocks Under Confining Pressure Unloading Conditions
Core Problem: Insufficient understanding of the coupled evolution mechanisms of mechanical energy-seepage interactions in pre-damaged rocks under confining pressure unloading, which frequently induces rock damage deterioration and water/mud inrush disasters.
Key Innovation: Systematically investigated the mechanical responses, energy evolution, and permeability variation patterns of pre-damaged red sandstone under unloading conditions, and developed a damage constitutive model incorporating confining pressure unloading–seepage coupling, providing theoretical foundations for early warning of seepage-induced disasters.
35. Drained Residual Strength of Texcoco Clays Under Different Testing Conditions
Core Problem: Characterizing the drained residual strength of high plastic remoulded clays from the Valley of Mexico under various testing conditions, which is a critical parameter for assessing long-term slope stability and landslide susceptibility.
Key Innovation: Investigated the influence of displacement rate, consolidation time, and OCR on the drained residual strength of Texcoco clays, showing that increased displacement rate causes an increase in residual friction angle, while increased consolidation time and OCR primarily affect peak strength, with residual strength remaining largely unchanged.
36. Investigating the Anti-suffusion Capacity of Granular Soils with Different Consolidation Pressures Using Coupled CFD-DEM
Core Problem: Understanding how in-situ stress conditions (consolidation pressure) influence the development and resistance to suffusion in granular soils, which is critical for the stability of geotechnical structures.
Key Innovation: Using coupled CFD-DEM simulations to investigate macroscopic and microscopic suffusion mechanisms, revealing three distinct stages of erosion, the formation of preferential flow paths, and the role of higher confining stress in promoting clogging effects and enhancing anti-suffusion resistance.
37. Velocity-dependent friction of granular ice: insights from temperature-controlled ring shear tests
Core Problem: Limited experimental data on the frictional behavior of granular ice under high-stress conditions, which is crucial for understanding glacier dynamics and ice-related geohazards.
Key Innovation: Experimental investigation using a novel temperature-controlled ring shear apparatus revealing a transition from velocity strengthening to velocity weakening friction in granular ice, substantial frictional strengthening (healing) with hold time, and providing robust constitutive parameters for refining predictive models for deep glacial shear zones and engineering geological processes.
38. Correlation between seismic signal wavelength and valley significant response
Core Problem: Traditional research on site effects in earthquake-induced geological disasters has insufficiently examined the conditions under which topographical site effects should be considered, complicating disaster assessment, especially in river valleys.
Key Innovation: This study uses finite element method to reveal a significant relationship between seismic signal wavelength and valley span, affecting seismic response. It provides a criterion (wavelength > 60 times valley depth) for when topographical effects can be neglected, offering a basis for rapid regional earthquake risk evaluation.
39. Exploring the potential of Sentinel-3 and Sentinel-6 SAR altimetry measurements for discharge estimation: Case studies from the Rhine and Po rivers
Core Problem: The relative paucity of in situ river discharge data necessitates the development of accurate satellite-based methods for discharge estimation, especially for Essential Climate Variables.
Key Innovation: Explored the potential of Sentinel-3A & B and Sentinel-6 SAR altimetry for discharge estimation on the Rhine and Po Rivers. Demonstrated that enhanced high-resolution 80 Hz processing and rigorous data selection improve accuracy. Achieved 4–16% NRMSE with empirical rating curves, 11–26% with Manning equation (improving to 8–22% with variable roughness), and potentially 10% with a semi-empirical Bjerklie equation, indicating high-quality SAR altimetry-based discharge products can detect projected discharge changes.
40. Shaking table tests and seismic response mechanisms of group pile foundations for nuclear power plants
Core Problem: Investigating the seismic response mechanisms of group pile foundations supporting nuclear island structures on soil sites, particularly the influence of pile number and spacing.
Key Innovation: Conducted scaled shaking table tests and validated 3D finite element models to analyze the seismic response of pile group foundations, demonstrating how increasing pile number (reducing spacing) affects system stiffness, damping, displacement, acceleration, and bending moments, and identifying key governing parameters.
41. Model test and numerical simulation on secant pile-diaphragm wall caisson foundation during excavation
Core Problem: Conventional caisson construction involves time-consuming staged excavation and secondary lining processes, and understanding ground deformation during deep excavation is critical for foundation stability.
Key Innovation: An innovative secant pile-diaphragm wall caisson is proposed, and its performance during excavation is analyzed through physical model tests and 3D finite element simulations, revealing the influence of pile parameters on deformation, the mobilization of soil arching, and providing practical guidance for design and optimization.
42. Assessment of lightning mortality in the United States from 1979 to 2023
Core Problem: Despite a potential increase in hazardous lightning activity due to climate change, understanding the long-term trends and underlying reasons for lightning-related mortality in the US is crucial for effective risk management.
Key Innovation: Provides a comprehensive assessment of lightning-related mortality in the United States from 1979 to 2023, demonstrating a steady decline in incidence and rates, and attributes this success to a multitude of risk reduction strategies, highlighting the effectiveness of public health and safety interventions against a potentially increasing natural hazard.
43. Forecasting heavy precipitation in Portugal: AROME numerical prediction model performance
Core Problem: Accurate forecasting of heavy precipitation remains challenging due to complex driving mechanisms and model limitations, despite its critical role in public safety and economic loss minimization during extreme events.
Key Innovation: The first assessment of the AROME model's precipitation forecast skill in Portugal, comparing its output with observed data from 105 weather stations. The study analyzed performance across various precipitation thresholds and accumulation periods, and identified skillful convective predictors (K-index, Total-Totals index, Potential instability, Bulk-Shear) for timely early warning.
44. Quantifying situation-dependent uncertainty in tropical cyclone track forecasts with a recurrent neural network approach
Core Problem: The need for reliable and situation-dependent forecasts of tropical cyclone tracks to accurately anticipate their environmental and societal impacts, as existing methods may not fully capture spatiotemporal correlations of forecast errors.
Key Innovation: Application of a Long Short-Term Memory (LSTM) model to better represent spatiotemporal correlations of typhoon track forecast errors and provide a situation-dependent Cone of Uncertainty (CoU). LSTM-calibrated forecasts generally outperformed official forecasts, and the estimated CoU reliably covered observed tracks. Incorporating global numerical model information further reduced forecast uncertainty while maintaining reliable coverage.
45. Advanced evaluation of pre-trained CNN models for accurate forest fire detection
Core Problem: Timely and accurate detection of forest fires is crucial for mitigating their impact, but achieving high accuracy remains a significant challenge due to the inherent complexity of the task.
Key Innovation: An advanced evaluation of six pre-trained Convolutional Neural Network (CNN) models (AlexNet, GoogLeNet, SqueezeNet, ResNet-18, MobileNetV2, and EfficientNetB0) for forest fire detection. The models were fine-tuned using two public datasets, achieving high accuracies (e.g., AlexNet 99.74%, ResNet-18 98%), demonstrating a robust and reliable solution that outperforms existing state-of-the-art methods for wildfire detection.
46. Tensile strength characteristics and predictive modeling of expansive soil containing termite nests
Core Problem: The stability of canal embankments is threatened by the deterioration of expansive soil's tensile strength, particularly exacerbated by the presence of termite nests and varying water content, making it difficult to predict failure.
Key Innovation: Conducted splitting tests with PIV to quantify the reduction in tensile strength of expansive soil due to termite nest pores and water content, and developed a predictive model for this deterioration, providing a scientific basis for evaluating embankment stability.
47. Dynamic Mechanical Behavior and Fractal Characteristic of NPR Cable-Anchored Sandstone Specimens in SHPB Tests
Core Problem: Enhancing the impact resistance of rock masses and mitigating rock failure under dynamic loading conditions, which is crucial for preventing rockfall and rockbursts.
Key Innovation: Demonstrated that NPR (Negative Poisson's Ratio) cable-anchored specimens exhibit superior impact resistance (higher dynamic peak stress, reduced strain, less fragmentation, improved integrity) compared to unanchored or PR-anchored specimens, providing a quantitative basis for their role in mitigating rock failure.
48. Integrating geological mapping and modelling for informed resource utilization of groundwater and aggregates in Voss, Norway
Core Problem: Underutilization of in-situ resource mapping and investigation in planning, hindering sustainable development and resilience, and the need to integrate geological conditions and geohazard assessment into early-stage area planning.
Key Innovation: An integrated approach combining high-density geophysical data, borehole data, and Quaternary geological interpretations to develop a 3D geological model for mapping resource potential (groundwater, geothermal energy, aggregates) and ensuring geohazard-safe development, providing a scalable framework for geoscience integration into urban planning.
49. How microstructural evolution governs the rheological behavior of deep-sea sediments: Evidence from shear plane imaging and rheological tests
Core Problem: The microstructural mechanisms underlying shear-induced failure and rheological behavior of deep-sea sediments, which pose significant risks to seabed engineering, are poorly understood.
Key Innovation: Developed an innovative workflow combining drainage plasticization and vane shearing to non-destructively expose shear planes. Correlated microstructural evolution with rheological responses, revealing distinct mechanisms for different sediment types, thereby deepening the understanding of deep-sea sediment rheology for stability assessment.
50. Temporal transferability of spatially derived Manning’s roughness across flood regimes in the Mississippi River
Core Problem: The need to balance using consistent roughness values for model reliability with adjustments needed to reflect changing conditions and improve model performance when evaluating the temporal transferability of spatially derived Manning's roughness across different flood regimes.
Key Innovation: Applied a high-resolution 2D unstructured grid hydrodynamic model with temporally invariant, spatially variable Manning's roughness to simulate flood events over a decade (2011, 2019, 2022) in the Middle Mississippi River. Demonstrated excellent model performance (Nash–Sutcliffe Efficiency values of 0.98–0.99 for water surface elevations and 0.92–0.99 for discharge) and strong agreement with observed flood inundation, suggesting that performance reductions are more likely due to hydrologic forcing rather than temporal changes in roughness.
51. Fault Slip and Fluid Flow: Seismic Source Analysis to Assess Role of Multiple Slip Patches in Fault Permeability
Core Problem: The physical processes underlying the relationship between fault reactivation, microearthquakes, and permeability evolution during fluid injection are not fully understood, particularly how cumulative seismic moments relate to permeability.
Key Innovation: Conducts fault reactivation experiments and analyzes acoustic emission signals, revealing that permeability enhancement is driven by the sequential and interacting activation of multiple millimeter-scale slip patches (not a single large event), which cumulatively cover the fault multiple times, creating a continuous flow pathway and offering AEs as a diagnostic for constraining crustal permeability changes.
52. ECHOSAT: Estimating Canopy Height Over Space And Time
Core Problem: Existing global tree height maps provide only static snapshots, failing to capture temporal forest dynamics (growth, loss events) essential for accurate carbon accounting and climate change mitigation.
Key Innovation: Introduces ECHOSAT, a global and temporally consistent 10m resolution tree height map spanning multiple years, derived from multi-sensor satellite data using a specialized vision transformer. It employs a self-supervised growth loss to regularize predictions, accurately quantifying tree growth and disturbances like fires over time.
53. Deep Unfolding Real-Time Super-Resolution Using Subpixel-Shift Twin Image and Convex Self-Similarity Prior
Core Problem: Twin-image super-resolution (TISR), a challenging multi-image super-resolution (MISR) scenario crucial for satellite remote sensing (e.g., SPOT-5 supermode imaging), is less investigated and requires robust, real-time solutions, especially when subpixel shifts are not spatially uniform in real-world data.
Key Innovation: Formulates TISR using a convex criterion implemented via a novel deep unfolding network (COSUP). It uses an embedded simple shift operator to address coupled data-fitting terms and a transformer with a convex self-similarity loss for regularization, achieving state-of-the-art performance with millisecond-level computational time, even on real-world data with non-uniform subpixel shifts.
54. Generalizing Visual Geometry Priors to Sparse Gaussian Occupancy Prediction
Core Problem: Existing 3D occupancy prediction methods for embodied intelligence primarily rely on depth priors and make limited use of 3D cues, restricting performance and generalization, especially for volumetric interiors rather than just visible surfaces.
Key Innovation: GPOcc leverages generalizable visual geometry priors by extending surface points inward to generate volumetric samples represented as Gaussian primitives for probabilistic occupancy inference, and includes a training-free incremental update strategy, significantly improving mIoU and efficiency for 3D scene understanding.
55. ABM-UDE: Developing Surrogates for Epidemic Agent-Based Models via Scientific Machine Learning
Core Problem: Agent-based epidemic models (ABMs) are too slow for real-time planning despite encoding crucial heterogeneity, and existing Universal Differential Equation (UDE) methods for identification can be unstable or overconfident.
Key Innovation: ABM-UDE develops fast, trustworthy surrogates for ABMs using UDEs with a neural-parameterized contact rate, adapting multiple shooting and an observer-based prediction-error method for stable identification, enforcing physical constraints, and achieving significantly faster, more accurate, and calibrated forecasts with broad applicability to other scientific domains.
56. SEF-MAP: Subspace-Decomposed Expert Fusion for Robust Multimodal HD Map Prediction
Core Problem: Multimodal fusion for High-Definition (HD) map prediction in autonomous driving often suffers from inconsistencies between camera and LiDAR modalities, leading to performance degradation under challenging conditions like low-light, occlusions, or sparse point clouds.
Key Innovation: SEFMAP, a Subspace-Expert Fusion framework, explicitly disentangles BEV features into four semantic subspaces (LiDAR-private, Image-private, Shared, Interaction), each with a dedicated expert, combined with an uncertainty-aware gating mechanism and distribution-aware masking to enhance robustness and promote role specialization, achieving state-of-the-art performance.
57. Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping
Core Problem: Forecasting dynamic scenes remains a fundamental challenge in computer vision, as limited observations make it difficult to capture coherent object-level motion and long-term temporal evolution.
Key Innovation: MoGaF (Motion Group-aware Gaussian Forecasting), a framework for long-term scene extrapolation using 4D Gaussian Splatting, introduces motion-aware Gaussian grouping and group-wise optimization to enforce physically consistent motion across rigid and non-rigid regions, enabling realistic and temporally stable scene evolution via a lightweight forecasting module.
58. SF3D-RGB: Scene Flow Estimation from Monocular Camera and Sparse LiDAR
Core Problem: Existing learning-based scene flow estimation methods tend to focus on single modalities (image or LiDAR), limiting robustness and accuracy in perceiving dynamic scene changes.
Key Innovation: Presents SF3D-RGB, an end-to-end deep learning architecture that fuses 2D monocular images and 3D sparse LiDAR point clouds to achieve more accurate and robust sparse scene flow estimation with fewer parameters compared to single-modality or other fusion methods.
59. Assessing airborne laser scanning and aerial photogrammetry for deep learning-based stand delineation
Core Problem: Accurate forest stand delineation, essential for forest management, remains largely manual and subjective, and existing deep learning methods face operational scalability issues due to temporal misalignment between ALS and aerial imagery, raising questions about the reliability of DAP-derived CHMs and the utility of DTMs.
Key Innovation: Assessed a U-Net-based semantic segmentation framework for forest stand delineation, demonstrating that DAP-derived CHMs can reliably replace ALS-derived CHMs and that the inclusion of a DTM does not significantly improve performance, indicating the framework's resilience to input data variations and enabling the assembly of large, temporally aligned datasets for deep learning.
60. Global-Aware Edge Prioritization for Pose Graph Initialization
Core Problem: Existing Structure-from-Motion (SfM) pose graph initialization methods rely on independent image retrieval, ignoring global consistency and leading to less reliable and accurate 3D reconstructions, especially in sparse or ambiguous scenes.
Key Innovation: Introduces global-aware edge prioritization for pose graph initialization, using a GNN trained with SfM-derived supervision to predict globally consistent edge reliability, guiding multi-minimal-spanning-tree-based pose graph construction, and employing connectivity-aware score modulation to improve reconstruction accuracy and create more reliable/compact pose graphs.
61. Olbedo: An Albedo and Shading Aerial Dataset for Large-Scale Outdoor Environments
Core Problem: Progress in intrinsic image decomposition (IID) for large-scale outdoor scenes is limited by the lack of real-world datasets with reliable albedo and shading supervision, crucial for environmental understanding and change analysis.
Key Innovation: Olbedo, a large-scale aerial dataset containing 5,664 UAV images with multi-view consistent albedo and shading maps, metric depth, surface normals, and illumination data, derived from an inverse-rendering refinement pipeline, enabling state-of-the-art IID models to generalize to real outdoor imagery and supporting applications like relighting and scene change analysis.
62. DualWeaver: Synergistic Feature Weaving Surrogates for Multivariate Forecasting with Univariate Time Series Foundation Models
Core Problem: Effectively extending the strong univariate forecasting success of Time-Series Foundation Models (TSFMs) to multivariate forecasting, particularly in capturing cross-variable dependencies, remains challenging.
Key Innovation: Proposes DualWeaver, a novel framework that adapts univariate TSFMs for multivariate forecasting by using a pair of learnable, structurally symmetric surrogate series generated by a shared auxiliary feature-fusion module, enabling parameter-free reconstruction and enhanced robustness.
63. Lumosaic: Hyperspectral Video via Active Illumination and Coded-Exposure Pixels
Core Problem: Existing passive snapshot hyperspectral systems struggle with motion during capture and inefficient photon utilization, limiting real-time capture of dynamic scenes with spectral fidelity.
Key Innovation: Lumosaic, a compact active hyperspectral video system combining a narrowband LED array with a coded-exposure-pixel camera for joint encoding of scene information across space, time, and wavelength, enabling robust 31-channel hyperspectral video at 30 fps with improved fidelity and temporal stability, highly relevant for remote sensing in geohazard monitoring.
64. Surrogate models for Rock-Fluid Interaction: A Grid-Size-Invariant Approach
Core Problem: High-fidelity numerical models for rock-fluid interaction are computationally expensive due to high resolution requirements, limiting their applicability for multi-query problems like uncertainty quantification and optimization.
Key Innovation: Develops eight surrogate models, including a novel grid-size-invariant framework for single neural networks (UNet and UNet++), to predict fluid flow in porous media with rock dissolution, reducing memory consumption and outperforming reduced-order models.
65. Conditional neural control variates for variance reduction in Bayesian inverse problems
Core Problem: Accurate Monte Carlo estimation for Bayesian inverse problems, particularly those constrained by partial differential equations, demands a prohibitive number of samples due to significant posterior variability, making it computationally expensive.
Key Innovation: Introduces conditional neural control variates, a modular method that learns amortized control variates from joint model-data samples to reduce the variance of Monte Carlo estimators. It scales to high-dimensional problems using Stein's identity and hierarchical coupling layers, demonstrating substantial variance reduction on Darcy flow inverse problems even with learned surrogate scores.
66. Active operator learning with predictive uncertainty quantification for partial differential equations
Core Problem: Deploying neural operators as surrogate models for Partial Differential Equations (PDEs) in scientific applications requires understanding prediction accuracy and associated error levels, but existing uncertainty quantification (UQ) methods are often computationally expensive.
Key Innovation: Proposes a lightweight predictive UQ method for DeepONets (generalizable to other operator networks) that provides unbiased and accurate out-of-distribution uncertainty estimates with fast inference, demonstrating its utility in Bayesian optimization and active learning for improved accuracy and data-efficiency in solving PDEs.
67. Modern Neural Networks for Small Tabular Datasets: The New Default for Field-Scale Digital Soil Mapping?
Core Problem: Traditional machine learning methods have long been the default for field-scale Digital Soil Mapping (DSM) due to challenges posed by small training sample sizes and high feature-to-sample ratios for deep learning.
Key Innovation: A comprehensive benchmark demonstrating that modern Artificial Neural Networks (ANNs) consistently outperform classical methods for field-scale DSM, particularly TabPFN, suggesting ANNs as the new default for predicting critical soil properties.
68. Characteristic Root Analysis and Regularization for Linear Time Series Forecasting
Core Problem: The effectiveness of complex models in time series forecasting varies unpredictably, while simple linear models, despite their robustness, lack deep theoretical investigation, particularly regarding the role of characteristic roots in temporal dynamics and the production of spurious roots in noisy regimes.
Key Innovation: A systematic study of linear models for time series forecasting, analyzing characteristic roots in noise-free and noisy settings, and proposing two robust root restructuring strategies: rank reduction techniques (Reduced-Rank Regression, Direct Weight Rank Reduction) and a novel adaptive method called Root Purge, which collectively improve forecasting accuracy and interpretability.
69. Characterization and Learning of Causal Graphs with Latent Confounders and Post-treatment Selection from Interventional Data
Core Problem: Identifying causal relations from interventional data is challenged not only by latent confounders but also by post-treatment selection, a common yet overlooked issue that introduces spurious dependencies and distorts causal discovery results.
Key Innovation: A novel causal formulation that explicitly models post-treatment selection, characterizes its Markov properties, introduces a Fine-grained Interventional equivalence class (FI-Markov equivalence) represented by F-PAG, and develops a provably sound and complete algorithm (F-FCI) to identify causal relations, latent confounders, and post-treatment selection.
70. SciTS: Scientific Time Series Understanding and Generation with LLMs
Core Problem: Current multimodal Large Language Models (LLMs) are insufficient for comprehensive scientific time series understanding and generation due to limitations in encoding numerical sequences as text or images, leading to loss of precision and handling excessively long sequences; existing unified time series models also lack generalizability for non-periodic, heterogeneous scientific signals.
Key Innovation: The introduction of SciTS, a comprehensive benchmark spanning 12 scientific domains and 43 tasks for scientific time series, and TimeOmni, a framework that equips LLMs with the ability to understand and generate time series while remaining compatible with general-purpose LLM training, demonstrating improved generalizability over specialized models.
71. Aerial Vision-Language Navigation with a Unified Framework for Spatial, Temporal and Embodied Reasoning
Core Problem: Existing Aerial Vision-and-Language Navigation (VLN) methods for UAVs rely on costly and complex inputs (panoramic images, depth, odometry), hindering practical deployment for lightweight UAVs in complex urban environments.
Key Innovation: Presents a unified aerial VLN framework that operates solely on egocentric monocular RGB observations and natural language instructions, formulating navigation as a next-token prediction problem, and incorporating keyframe selection and action merging/label reweighting to achieve strong results on AerialVLN and OpenFly benchmarks, narrowing the performance gap with state-of-the-art panoramic RGB-D counterparts.
72. OGD4All: A Framework for Accessible Interaction with Geospatial Open Government Data Based on Large Language Models
Core Problem: Citizens struggle with accessible, transparent, and reproducible interaction with geospatial Open Government Data (OGD).
Key Innovation: Presents OGD4All, an LLM-based framework that combines semantic data retrieval, agentic reasoning, and sandboxed execution to provide explainable, multimodal, and verifiable access to geospatial OGD, minimizing hallucination risks.
73. From Pairs to Sequences: Track-Aware Policy Gradients for Keypoint Detection
Core Problem: Existing learning-based keypoint detection methods are trained on image pairs, failing to explicitly optimize for the long-term trackability of keypoints across challenging image sequences, which is critical for robust 3D vision systems.
Key Innovation: TraqPoint, an end-to-end Reinforcement Learning framework that redefines keypoint detection as a sequential decision-making problem, optimizing keypoint track-quality directly on image sequences using a track-aware reward mechanism.
74. Multi-agent deep reinforcement learning with centralized training and decentralized execution for transportation infrastructure management
Core Problem: Life-cycle management of large-scale transportation systems (pavement, bridges) is challenging due to high-dimensional state/action spaces, multiple uncertainties, and critical operational constraints, leading to limitations in optimality, scalability, and proper uncertainty handling in traditional approaches.
Key Innovation: Proposing DDMAC-CTDE, a Deep Decentralized Multi-Agent Actor-Critic reinforcement learning architecture with Centralized Training and Decentralized Execution, to solve constrained POMDPs for transportation infrastructure management, demonstrating superior performance over baselines on a new comprehensive benchmark environment that incorporates practical constraints and nonstationary degradation.
75. A dynamic data driven application system for ocean environment prediction using reduced-order modeling and adaptive glider sensing networks
Core Problem: Real-time, accurate prediction of regional three-dimensional ocean environments is challenging due to sparse, noisy observations and complex spatiotemporal dynamics.
Key Innovation: A novel Dynamic Data-Driven Application System (DDDAS) framework is proposed, integrating hybrid reduced-order modeling, sparse-aware data assimilation (Kriged Ensemble Kalman Filter with Group Lasso), and Deep Reinforcement Learning for adaptive glider sensing optimization, enabling real-time and accurate ocean environmental forecasting.
76. We need a global assessment of avoidable climate-change risks
Core Problem: Policymakers and citizens lack a full analysis of what is at stake regarding climate-change risks, hindering understanding of the urgency for emissions reductions.
Key Innovation: Calls for a global assessment of avoidable climate-change risks to inform policymakers and citizens about the urgency of emissions reductions.
77. The scientific evidence of the applications of social media to climate change adaptation and disaster risk reduction: current status, implications and way forward
Core Problem: The understanding of social media's role in climate change adaptation (CCA) and disaster risk reduction (DRR) is fragmented, with key aspects like resilience and preparedness often neglected.
Key Innovation: A systematic review providing scientific evidence on social media applications for CCA/DRR, identifying current status, implications, and neglected areas. It proposes a social media-integrated CCA-DRR governance framework and suggests future research directions, including tracking mis/disinformation and examining policy shifts on social media platforms.
78. A proposed gravity model application for community resilience assessments
Core Problem: Existing community resilience indices often rely on static indicators and fail to represent the dynamic interactions, accessibility, and spatial/structural processes that shape adaptive capacity.
Key Innovation: A proposed integrated framework combining resilience metrics with a production-constrained gravity model, constrained by community networks. This framework dynamically represents how accessibility, connectivity, and spatial structure contribute to resilience outcomes by modeling interactions between households (origins) and resources (destinations), shifting emphasis from static capital-based measures to dynamic processes of interaction.
79. A single-stress elasto-plastic triaxial model for saturated and unsaturated soils
Core Problem: Developing a unified constitutive model to accurately predict the complex elasto-plastic behavior of both saturated and unsaturated soils under general triaxial loading conditions.
Key Innovation: Proposed a single-stress elasto-plastic triaxial model that extends the scaled stress concept, incorporating a capillary bonding function to define a unified normal compression line (UNCL) and an alternative yield function, validated against triaxial tests on fine-grained soils.
80. An Intelligent Prediction Method for Rock Core Integrity Based on Deep Semantic Segmentation
Core Problem: Addressing the inefficiency and high subjectivity of manual rock core integrity assessment and Rock Quality Designation (RQD) calculation.
Key Innovation: Developed DSS-RCI, a deep semantic segmentation algorithm with a position-aware circular convolution feature extraction network, multi-level feature enhancement, and a dynamic up-sampler, achieving high-precision intelligent prediction of rock core integrity and RQD with an average error of only 2.89%.
81. Fine-Tuned SAM Adaptation with Multi-scale Guidance for Automated Detection Toward Image-Based Core Length and RQD Measurement
Core Problem: The labor-intensive nature of traditional manual measurement of Rock Quality Designation (RQD) from boreholes and the insufficient detection accuracy and generalizability of mainstream deep learning methods for rock core logging.
Key Innovation: Proposed MG-FSAM4Seg, a fine-tuned Segment Anything Model (SAM) with multi-scale guidance for rock core instance segmentation, combined with a fine-grained RQD analytics method, achieving high accuracy (88.07% AP) and strong generalization across diverse datasets, with RQD estimates showing high agreement (R^2 > 0.95) with manual measurements.
82. The Influence of Microstructures on the Coal Fragment’s Kinetic Energy Generated by the True Triaxial Unloading Impact Experiments
Core Problem: Explaining the anomalous, significantly higher kinetic energy of coal fragments observed in some true triaxial unloading impact experiments, despite identical macroscopic conditions, which contributes to dynamic impact disasters in deep mining.
Key Innovation: Used small-angle neutron scattering (SANS) to reveal nanoscale pore characteristics (higher porosity in the 1-100 nm range and a distinct pore size distribution) as a microstructural explanation for the sharp increase in fragment kinetic energy during impact experiments.
83. Sustainability-Oriented Grouting Techniques for Soil Improvement and Retaining Pile Installation in Urban MRT Construction
Core Problem: Ensuring structural stability and groundwater control for urban MRT construction in challenging ground conditions (soft silty clay, shallow groundwater, dense utilities) where conventional methods are infeasible.
Key Innovation: Field-validated evidence on the performance of integrated multi-scale jet grouting systems (standard high-pressure and large-diameter super high-pressure) for retaining pile installation and soil improvement, providing context-specific performance benchmarks.
84. 3D SPT-seismic full-waveform inversion of S- and P-wave fields with correlation to SPT-N in karstic terrain of Florida, USA
Core Problem: Conventional site characterization methods provide only point measurements, leaving much of the subsurface uncharacterized, especially in complex geological settings like karstic terrains, hindering effective geotechnical assessments.
Key Innovation: Developed and validated a 3D SPT-seismic method that integrates seismic wave analysis from SPT hammer blows with 3D full-waveform inversion to produce high-resolution 3D shear wave velocity profiles. This transforms conventional SPT into a volumetric imaging tool, enabling reliable 3D estimation of SPT-N values and identification of subsurface anomalies like voids in karstic terrain.
85. Exploring the influence of team characteristics on the quality of scenarios in disaster management
Core Problem: Ensuring high-quality scenario development in disaster management teams, particularly understanding how team characteristics like social intelligence and creativity influence scenario quality, and addressing gaps in assessing and fostering these traits.
Key Innovation: Investigating the rapid development of plausible future scenarios in disaster management, finding that teams with higher social intelligence and creativity developed higher quality scenarios, and discussing the implications for practice, training, and cultural barriers in encouraging creative thinking for novel disaster situations.
86. Interpretable Budyko-constrained machine learning framework for monthly runoff attribution in U.S. CAMELS basins
Core Problem: Many traditional Budyko equations overlook key factors like snowmelt and soil water storage, limiting the accuracy and interpretability of runoff change attribution under climate change and human activity, which is critical for adaptive basin management.
Key Innovation: Developed a Budyko-constrained machine learning framework (Budyko-ML) that integrates physical consistency with data-driven flexibility by extending the Budyko equation to account for changes in snow water equivalent (ΔSWE) and soil water storage (ΔS). Improved attribution stability (NSE increased by 30%) compared to purely data-driven models. Used SHAP to quantify climate and human driver contributions, identifying potential evapotranspiration (PET) as the dominant factor for declining runoff in most basins, and revealed nonlinear thresholds via GAM analysis.
87. Correlation of pedotransfer function residuals with input variables and the effect of database similarity on predictive performance
Core Problem: Insufficient understanding of how specific input data characteristics drive pedotransfer function (PTF) prediction performance and challenges in assessing PTF transferability beyond their development datasets, limiting their robustness and broad application for estimating soil hydraulic properties.
Key Innovation: Employed the hierarchical Rosetta3 as a development dataset and evaluated its PTF performance using two independent application datasets (NCSS, HYBRAS-V2) comprising over 51,900 samples. Demonstrated that incorporating additional inputs moderately reduces residual-input correlations (higher correlations linked to inferior performance) and that lower Chamfer Distance (better dataset resemblance) leads to better PTF performance. Showed that increasing input complexity in hierarchical Rosetta3 models mitigates the effect of resemblance, enhancing robustness.
88. Emerging Effective Radiative Forcing in the Radiative Imbalance Since 2010
Core Problem: Explaining the substantial and rapidly increasing Earth's top-of-atmosphere (TOA) radiative imbalance observed since 2010, which exceeds current climate model projections.
Key Innovation: Estimated trends in effective radiative forcing (ERF) from 2010-2024, showing they are significantly larger than projections from state-of-the-art models (exceeding 1.0 Wm-2 per decade for net and shortwave fluxes), highlighting a widening discrepancy between observations and models.
89. Pseudo-View Enhancement via Confidence Fusion for Unposed Sparse-View Reconstruction
Core Problem: 3D scene reconstruction under unposed sparse viewpoints, especially in outdoor scenes with complex lighting and scale variation, is highly challenging. Extremely limited input views lead to unreasonable geometry and harm reconstruction quality when directly using diffusion models to synthesize pseudo frames.
Key Innovation: Proposes a novel framework for sparse-view outdoor reconstruction that achieves high-quality results through bidirectional pseudo frame restoration and scene perception Gaussian management. It introduces a diffusion-based synthesis guided by adjacent frames with a lightweight pseudo-view deblur model and confidence mask inference, alongside a strategy to optimize Gaussians based on joint depth-density information.
90. Lie Flow: Video Dynamic Fields Modeling and Predicting with Lie Algebra as Geometric Physics Principle
Core Problem: Modeling 4D scenes requires capturing both spatial structure and temporal motion, which is challenging due to the need for physically consistent representations of complex rigid and non-rigid motions, often leading to spatial inconsistency and physically implausible motion.
Key Innovation: LieFlow, a dynamic radiance representation framework, explicitly models motion within the SE(3) Lie group, enabling coherent learning of translation and rotation in a unified geometric space and enforcing physically inspired constraints to maintain motion continuity and geometric consistency.
91. HybridINR-PCGC: Hybrid Lossless Point Cloud Geometry Compression Bridging Pretrained Model and Implicit Neural Representation
Core Problem: Learning-based point cloud compression methods suffer from training data dependency (pretrained models) or time-consuming online training and bitstream overhead (Implicit Neural Representation-based methods).
Key Innovation: HybridINR-PCGC, a novel hybrid framework, bridges pretrained models and INR by using a Pretrained Prior Network (PPN) to generate a robust prior for accelerating the convergence of a Distribution Agnostic Refiner (DAR), which is decomposed into base and enhancement layers, with a supervised model compression module to minimize bitrate.
92. Send Less, Perceive More: Masked Quantized Point Cloud Communication for Loss-Tolerant Collaborative Perception
Core Problem: Existing collaborative perception approaches struggle to achieve high accuracy under strict bandwidth constraints and remain highly vulnerable to random transmission packet loss when sharing sensory information among connected vehicles.
Key Innovation: QPoint2Comm, a quantized point-cloud communication framework, reduces bandwidth by directly communicating quantized point-cloud indices using a shared codebook, ensures robustness to packet loss via a masked training strategy, and enhances multi-vehicle information integration with a cascade attention fusion module.
93. TiMi: Empower Time Series Transformers with Multimodal Mixture of Experts
Core Problem: Existing multimodal time series forecasting methods struggle to effectively incorporate diverse multimodal data, especially textual information with causal influence, due to challenges in modality alignment.
Key Innovation: Proposes TiMi, a Time Series Transformer with a Multimodal Mixture-of-Experts (MMoE) module, which leverages LLMs for causal reasoning from textual information to guide numerical time series forecasting, achieving state-of-the-art performance without explicit representation-level alignment.
94. LiREC-Net: A Target-Free and Learning-Based Network for LiDAR, RGB, and Event Calibration
Core Problem: Existing learning-based multi-sensor calibration methods are typically designed for only bi-modal setups, lacking a unified framework for jointly calibrating multiple sensor modalities like LiDAR, RGB, and event data in a target-free manner.
Key Innovation: Proposes LiREC-Net, a target-free, learning-based network that jointly calibrates LiDAR, RGB, and event data within a unified framework, introducing a shared LiDAR representation to reduce computation and improve efficiency, achieving competitive performance for tri-modal calibration.
95. TIRAuxCloud: A Thermal Infrared Dataset for Day and Night Cloud Detection
Core Problem: Clouds significantly hinder Earth observation, limiting critical remote sensing applications (e.g., fire disaster response) 24/7. While thermal infrared (TIR) imagery is crucial for nighttime cloud detection, it faces challenges due to limited spectral information and lower spatial resolution, and there's a scarcity of comprehensive datasets.
Key Innovation: Introduces TIRAuxCloud, a multi-modal dataset centered around thermal spectral data (TIR, optical, NIR from Landsat and VIIRS) combined with auxiliary information (elevation, land cover, meteorological variables, cloud-free references) to facilitate day and night cloud segmentation and reduce surface-cloud ambiguity, including automated and manually annotated cloud masks.
96. NESTOR: A Nested MOE-based Neural Operator for Large-Scale PDE Pre-Training
Core Problem: Existing neural operators for solving PDEs typically rely on a single network architecture, limiting their capacity to fully capture heterogeneous features and complex system dependencies, which bottlenecks large-scale PDE pre-training.
Key Innovation: Proposes NESTOR, a large-scale PDE pre-trained neural operator based on a nested Mixture-of-Experts (MoE) framework (image-level MoE for global, token-level Sub-MoE for local dependencies) to selectively activate suitable expert networks, enhancing generalization and transferability across diverse PDE systems.
97. RAMSeS: Robust and Adaptive Model Selection for Time-Series Anomaly Detection Algorithms
Core Problem: Time-series data vary widely across domains, making a universal anomaly detector impractical, as methods that perform well on one dataset often fail to transfer due to context-dependent anomaly definitions.
Key Innovation: RAMSeS, a Robust and Adaptive Model Selection framework, combines a stacking ensemble optimized with a genetic algorithm and an adaptive model-selection branch (using Thompson sampling, GANs, and Monte Carlo simulations) to exploit collective model strength and adapt to dataset-specific characteristics, outperforming prior methods on F1 score.
98. A Comprehensive Survey on Underwater Image Enhancement Based on Deep Learning
Core Problem: Underwater image enhancement (UIE) is a significant challenge in computer vision research, and despite numerous algorithms, a thorough and systematic review is still absent, hindering future advancements.
Key Innovation: Provides a comprehensive survey of UIE based on deep learning, introducing physical models, data construction, evaluation metrics, and loss functions, categorizing algorithms by contribution, performing quantitative and qualitative evaluations of state-of-the-art methods, and identifying key areas for future research.
99. Modular Deep Learning for Multivariate Time-Series: Decoupling Imputation and Downstream Tasks
Core Problem: Existing end-to-end deep learning models for multivariate time-series with missing values tightly couple imputation and downstream tasks, limiting reusability, interpretability, and quality assessment.
Key Innovation: Proposes a modular deep learning approach that decouples imputation and downstream tasks, demonstrating maintained high performance while enhancing flexibility and reusability across various datasets and models, which is crucial for real-world applications.
100. Enhancing Spatio-Temporal Forecasting with Spatial Neighbourhood Fusion:A Case Study on COVID-19 Mobility in Peru
Core Problem: Accurate spatio-temporal forecasting of human mobility is challenging due to the spatial sparsity of hourly mobility counts across grid cells, which limits the predictive power of conventional time series models.
Key Innovation: A lightweight and model-agnostic Spatial Neighbourhood Fusion (SPN) technique is proposed that augments each cell's features with aggregated signals from its immediate neighbors, consistently improving spatio-temporal forecasting performance (e.g., up to 9.85% reduction in test MSE) for sparse mobility data.
101. The Curious Case of In-Training Compression of State Space Models
Core Problem: State Space Models (SSMs) face a design challenge in balancing expressivity with computational burden, as update costs scale with state dimension, and training efficient models directly at smaller dimensions often leads to a loss of task-critical structure.
Key Innovation: CompreSSM, an approach that applies Hankel singular value analysis from control theory to SSMs *during training* to identify and preserve only dimensions of high influence, significantly accelerating optimization and preserving expressivity by allowing models to begin large and shrink, outperforming models trained directly at smaller dimensions.
102. WeirNet: A Large-Scale 3D CFD Benchmark for Geometric Surrogate Modeling of Piano Key Weirs
Core Problem: Reliable prediction of hydraulic performance for Piano Key Weir (PKW) design is challenging due to complex 3D geometry and operating conditions, and progress in surrogate modeling is limited by scarce large, well-documented datasets.
Key Innovation: WeirNet, a large 3D CFD benchmark dataset (71,387 simulations) for geometric surrogate modeling of PKWs, enabling faster and more reliable hydraulic-structure design by providing data for training efficient surrogate models.
103. Wave forces of subsea cables laid on meso-scale rough seabeds
Core Problem: The stability design of subsea cables on rough seabeds is challenged by complex wave forces, particularly the influence of Reynolds number and cable-seabed interaction.
Key Innovation: An experimental investigation reveals that wave forces on subsea cables vary significantly with the cable's relative position to meso-scale roughness elements, and recommends incorporating velocity overshoot and equivalent gap ratio effects into stability design, especially under severe storm waves.
104. Assessment of gap-filling techniques applied to satellite phytoplankton composition products for the Atlantic Ocean
Core Problem: Satellite-derived ocean color products suffer from limited temporal and spatial coverage due to non-optimal observing conditions (e.g., clouds, sun glint), hindering comprehensive spatiotemporal assessments.
Key Innovation: Evaluation of two robust gap-filling methods (DINEOF and DINCAE) for satellite ocean color products (Chl a and PFTs), demonstrating their ability to restore spatial structure and increase data availability while preserving statistical trends, with potential transferability to geohazard monitoring.
105. Upper mantle low-velocity layer tied to volatile-charged carbonate melts
Core Problem: Understanding how minute melt fractions in Earth’s deep interior migrate through solid rocks and generate large-scale geophysical anomalies.
Key Innovation: High-pressure experiments demonstrating that volatile-charged carbonate-rich melts act as super-spreaders, completely wetting mantle mineral surfaces and forming fully interconnected networks at trace amounts (0.02-0.08 vol%), facilitating efficient melt migration and global material recycling.
106. Coupled geomorphic and climate-driven biogeochemical processes regulate soil organic carbon stocks in agricultural terraces
Core Problem: The extent and drivers of agricultural terraces' impact on the carbon cycle and soil organic carbon (SOC) stocks remain highly uncertain, despite their widespread presence and influence on soil landscapes.
Key Innovation: Established a framework showing that SOC stock changes in terraces are governed by coupled C turnover-geomorphic processes (topsoil C replacement at eroding positions and buried SOC stabilization at depositional positions), with climate strongly modulating these processes, leading to region-specific outcomes for C sequestration.
107. A Freezing front advancing method for measuring hydraulic conductivity of frozen soil
Core Problem: Traditional steady-state methods for measuring hydraulic conductivity in frozen soils are time-consuming and yield only limited, discrete data points, hindering efficient parameterization of seepage behavior in permafrost engineering.
Key Innovation: Proposed a novel method based on freezing front advancing to measure hydraulic conductivity in frozen soils over a continuous temperature range, significantly reducing experimental duration. This provides an efficient tool for parameterizing seepage issues in cold region engineering.
108. SpecAware: a spectral-content aware foundation model for unifying multi-sensor learning in hyperspectral remote sensing mapping
Core Problem: The inherent heterogeneity of hyperspectral imaging (HSI) data, particularly spectral channel variations across sensors, limits model generalization and adaptability for cross-sensor joint learning in LULC mapping.
Key Innovation: Proposal of SpecAware, a spectral-content aware foundation model that uses a hypernetwork-driven unified image embedding process and a meta-content aware module to dynamically generate a unified hyperspectral token representation, enabling adaptive multi-sensor joint pre-training for diverse HSI mapping tasks.
109. Preparation and properties of cross-linked polymer/bentonite nanocomposite for containment of chemically aggressive liquids
Core Problem: Conventional bentonite and even sodium-activated calcium bentonite (NCB) are ineffective at containing chemically aggressive liquids, especially when utilizing abundant but less hydrophilic calcium bentonite.
Key Innovation: Developed a polymerization method to transform sodium-activated calcium bentonite (NCB) into a polymer-modified bentonite (PMB) nanocomposite, achieving significantly low hydraulic conductivity (<10−11 m/s) to aggressive liquids by forming a stable polymer network that increases swelling, decreases pore size, and creates tortuous flow pathways.
110. Spatial and vertical wind patterns in a path‐induced blowout in opposing wind conditions
Core Problem: Fine-scale airflow patterns and their interaction with blowout morphology under bi-directional wind regimes are poorly characterized, especially considering anthropogenic modifications.
Key Innovation: Investigates airflow dynamics in a small trough blowout using extensive anemometer data, revealing distinct internal flow structures, flow bifurcation, and near-surface disturbance patterns influenced by blowout morphology, wind regime, and anthropogenic alteration.
111. On Utilizing Spatial Gradients to Discover Functional Relationships in Hydrology
Core Problem: Difficulty in discovering functional relationships in hydrology, especially given the static nature of some variables and reliance on natural gradients.
Key Innovation: Proposes utilizing spatial gradients across different systems to discover and test functional relationships in hydrology, integrating theory, data, and models.
112. A Stacking Ensemble Model for Specific Yield Prediction: Framework Development and Application to Groundwater Storage Change Estimation in the North China Plain
Core Problem: Accurately predicting specific yield (SY) in unconfined aquifers for groundwater storage change estimation, especially given its variability with groundwater-level fluctuations and unsaturated-zone moisture fluxes.
Key Innovation: Development of a stacking ensemble model (using random forest and light gradient boosting machine) for SY prediction, which improves the accuracy of groundwater storage change assessment and provides support for sustainable groundwater resource management.
113. Generative Bayesian Computation as a Scalable Alternative to Gaussian Process Surrogates
Core Problem: Gaussian Process (GP) surrogates, while default for emulating expensive computer experiments, suffer from cubic computational cost, restrictive stationarity assumptions, and limitations to Gaussian predictive distributions.
Key Innovation: Proposes Generative Bayesian Computation (GBC) via Implicit Quantile Networks (IQNs) as a scalable surrogate framework. GBC learns the full conditional quantile function, offering linear scaling, improved performance on non-Gaussian and jump-process benchmarks, and better active learning capabilities compared to GPs.
114. D-Flow SGLD: Source-Space Posterior Sampling for Scientific Inverse Problems with Flow Matching
Core Problem: The need for robust, uncertainty-aware posterior sampling strategies for scientific inverse problems using Flow Matching (FM) priors, especially for complex scientific benchmarks where fidelity beyond measurement misfit is crucial, and the limitations of existing inference-time strategies.
Key Innovation: Proposing D-Flow SGLD, a source-space posterior sampling method that augments differentiable source inference with preconditioned stochastic gradient Langevin dynamics. This enables scalable exploration of the source posterior induced by new measurement operators without retraining the prior or modifying the learned FM dynamics, demonstrating strong performance on various scientific inverse problems.
115. Learning Recursive Multi-Scale Representations for Irregular Multivariate Time Series Forecasting
Core Problem: Existing multi-scale methods for Irregular Multivariate Time Series (IMTS) forecasting use resampling, which alters original timestamps and disrupts valuable sampling pattern information, making it difficult to capture diverse dependencies across multiple time scales.
Key Innovation: Proposes ReIMTS, a recursive multi-scale modeling approach that keeps timestamps unchanged and recursively splits samples into subsamples with progressively shorter time periods. It uses an irregularity-aware representation fusion mechanism to capture global-to-local dependencies for accurate forecasting.
116. From Basis to Basis: Gaussian Particle Representation for Interpretable PDE Operators
Core Problem: Existing neural operators and Transformer-based models for learning PDE dynamics lack interpretability, struggle with localized high-frequency structures, and incur quadratic computational costs.
Key Innovation: A Gaussian basis representation for fields and a Gaussian Particle Operator are introduced, acting in modal space with learned Gaussian modal windows and PG Gaussian Attention, achieving resolution-agnostic, near-linear complexity, and intrinsic interpretability for PDE dynamics.
117. Tokenizing Semantic Segmentation with RLE
Core Problem: Semantic segmentation, especially for videos and panoptic segmentation, can be computationally intensive and lack a unified approach that efficiently handles mask representation and instance information.
Key Innovation: Presents a new unified approach for semantic segmentation in images and videos by tokenizing masks using Run Length Encoding (RLE) and training a modified Pix2Seq model for autoregressive output, proposing novel tokenization strategies to compress sequence length and incorporate instance information for panoptic segmentation.
118. CCCaption: Dual-Reward Reinforcement Learning for Complete and Correct Image Captioning
Core Problem: Ground-truth human annotations for image captioning are often incomplete or incorrect, limiting caption models, and current quality assessment overlooks objective completeness and correctness.
Key Innovation: CCCaption, a dual-reward reinforcement learning framework with a dedicated fine-tuning corpus, explicitly optimizes for completeness (rewarding captions that answer more visual queries) and correctness (penalizing hallucinations) to generate more objective and accurate captions.
119. Hierarchical Lead Critic based Multi-Agent Reinforcement Learning
Core Problem: Cooperative Multi-Agent Reinforcement Learning (MARL) is often limited to either local (independent learning) or global (centralized learning) perspectives, hindering its ability to learn from multiple perspectives on different hierarchy levels for complex tasks requiring coordination.
Key Innovation: Hierarchical Lead Critic (HLC), a novel sequential training scheme and MARL architecture, learns from multiple perspectives on different hierarchy levels, leveraging local and global insights to achieve improved performance, high sample efficiency, and robust policies in cooperative, non-communicative, and partially observable MARL benchmarks.
120. SAPNet++: Evolving Point-Prompted Instance Segmentation with Semantic and Spatial Awareness
Core Problem: Point-prompted instance segmentation (PPIS) faces challenges like granularity ambiguity (distinguishing whole objects vs. parts) and boundary uncertainty due to single-point annotations, and existing methods relying on proposal selection often fail to resolve these issues.
Key Innovation: Proposes SAPNet++, which integrates Point Distance Guidance and Box Mining Strategy to address granularity ambiguity, incorporates completeness scores for spatial granularity awareness (S-MIL), and uses Multi-level Affinity Refinement to narrow boundary uncertainty, significantly improving segmentation performance on challenging datasets.
121. XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression
Core Problem: Learning-based 3D visual geometry models like StreamVGGT suffer from unbounded growth in the Key-Value (KV) cache due to massive influx of vision tokens, leading to increased memory consumption and inference latency, which limits scalability for long-horizon applications.
Key Innovation: Proposes XStreamVGGT, a tuning-free approach that seamlessly integrates pruning and quantization to systematically compress the KV cache, enabling extremely memory-efficient streaming inference for 3D applications with mostly negligible performance degradation and substantial reductions in memory usage and inference time.
122. Scan Clusters, Not Pixels: A Cluster-Centric Paradigm for Efficient Ultra-high-definition Image Restoration
Core Problem: Ultra-High-Definition (UHD) image restoration models are computationally unsustainable due to pixel-wise operations, and even state space models (SSMs) like Mamba retain a fundamental bottleneck with pixel-serial scanning for millions of pixels.
Key Innovation: Introduces C^2SSM, a visual state space model that shifts from pixel-serial to cluster-serial scanning. It distills UHD image features into sparse semantic centroids, enabling a dual-path process that scans and reasons over clusters, then diffuses global context back to pixels, achieving significant efficiency gains and state-of-the-art results across five UHD restoration tasks.
123. Learning Unknown Interdependencies for Decentralized Root Cause Analysis in Nonlinear Dynamical Systems
Core Problem: Root cause analysis (RCA) in networked industrial systems is difficult due to unknown, dynamically evolving interdependencies among heterogeneous, geographically distributed clients, and existing federated learning (FL) methods often assume homogeneous feature spaces or retrainable client models.
Key Innovation: Presents a federated cross-client interdependency learning methodology for feature-partitioned, nonlinear time-series data. It augments proprietary local client models with ML models encoding cross-client interdependencies, coordinated via a global server with calibrated differential privacy noise, and establishes theoretical convergence guarantees.
124. Bayesian Generative Adversarial Networks via Gaussian Approximation for Tabular Data Synthesis
Core Problem: Conditional tabular GANs (CTGAN) struggle to effectively balance the risk-utility trade-off in synthesizing mixed tabular data, while traditional Bayesian GANs using MCMC are computationally intensive.
Key Innovation: Gaussian Approximation of CTGAN (GACTGAN) integrates Stochastic Weight Averaging-Gaussian (SWAG) within the CTGAN generator to synthesize tabular data, reducing computational overhead and yielding better synthetic data with improved preservation of tabular structure and inferential statistics, and less privacy risk.
125. Robustness in sparse artificial neural networks trained with adaptive topology
Core Problem: Developing efficient deep learning models often comes at the cost of robustness, and understanding the robustness of sparse neural networks, particularly with adaptive topologies, is crucial for reliable applications.
Key Innovation: Demonstrates that sparse artificial neural networks (99% sparsity) trained with adaptive topology (updating topology between epochs) not only achieve competitive accuracy but also maintain robustness under various perturbations, including random link removal, adversarial attacks, and link weight shuffling.
126. PanoEnv: Exploring 3D Spatial Intelligence in Panoramic Environments with Reinforcement Learning
Core Problem: Current Vision-Language Models (VLMs) struggle with 3D spatial reasoning on panoramic images due to geometric distortion and limited 3D supervision, hindering holistic scene understanding.
Key Innovation: Introduction of PanoEnv, a large-scale VQA benchmark for 3D spatial intelligence in panoramic environments, and a reinforcement learning post-training framework (GRPO with ground-truth-guided reward and a two-stage curriculum) that significantly enhances VLMs' 3D spatial reasoning capabilities.
127. RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations
Core Problem: Existing Vision-based Retrieval-Augmented Generation (VisRAG) models degrade significantly when visual inputs are affected by distortions (e.g., blur, noise), as semantic and degradation factors become entangled in visual encoders.
Key Innovation: RobustVisRAG, a causality-guided dual-path framework that separates degradation signals from purified semantics, along with Non-Causal Distortion Modeling and Causal Semantic Alignment objectives, leading to improved retrieval and generation performance under various real-world visual degradations. Also introduces the Distortion-VisRAG dataset.
128. RT-RMOT: A Dataset and Framework for RGB-Thermal Referring Multi-Object Tracking
Core Problem: Existing Referring Multi-Object Tracking (RMOT) methods exhibit limitations in low-visibility conditions such as nighttime or smoke.
Key Innovation: Proposes the RT-RMOT task, the RefRT RGB-Thermal RMOT dataset, and the RTrack framework (built upon an MLLM with GSPO and CAS strategies) to enable all-day referring multi-object tracking by fusing RGB and thermal modalities.
129. AdaSpot: Spend Resolution Where It Matters for Precise Event Spotting
Core Problem: Existing Precise Event Spotting methods process all video frames uniformly, leading to redundant computation on non-informative regions and loss of fine-grained details due to spatial downsampling.
Key Innovation: Proposes AdaSpot, a framework that processes low-resolution videos for global features while adaptively selecting and processing high-resolution regions-of-interest in each frame using an unsupervised, task-aware strategy, achieving state-of-the-art performance with high efficiency.
130. On Imbalanced Regression with Hoeffding Trees
Core Problem: Effectively performing imbalanced regression tasks in continuous data streams using Hoeffding trees, specifically integrating kernel density estimation (KDE) and hierarchical shrinkage (HS) for improved performance.
Key Innovation: Extends KDE to streaming environments and implements HS for incremental decision tree models, showing KDE is beneficial in early stream parts for imbalanced regression, potentially applicable to geohazard prediction with imbalanced streaming data.
131. Don't stop me now: Rethinking Validation Criteria for Model Parameter Selection
Core Problem: The evaluation of generalization on the validation set for neural classifiers, particularly how validation criteria for model selection (e.g., early stopping based on accuracy vs. loss) affect test performance.
Key Innovation: A systematic empirical and statistical study showing that early stopping based on validation accuracy performs worst, while loss-based validation criteria yield comparable and more stable test accuracy, suggesting avoiding validation accuracy for parameter selection in neural classifiers, which could improve geohazard model reliability.
132. SigmaQuant: Hardware-Aware Heterogeneous Quantization Method for Edge DNN Inference
Core Problem: Deploying deep neural networks (DNNs) on resource-constrained edge devices is challenging due to memory, energy, and computational limits, and existing heterogeneous quantization methods lack adaptability to varied hardware conditions without exhaustive search.
Key Innovation: SigmaQuant, an adaptive layer-wise heterogeneous quantization framework that efficiently balances accuracy and resource usage for varied edge environments without exhaustive search, addressing the limitations of uniform and existing heterogeneous quantization methods, potentially enabling more efficient deployment of geohazard models on edge devices.
133. WeaveTime: Stream from Earlier Frames into Emergent Memory in VideoLLMs
Core Problem: Current Video-LLMs suffer from 'Time-Agnosticism' in streaming settings, treating videos as unordered evidence, leading to failures in temporal order reasoning and distinguishing present observations from history.
Key Innovation: WeaveTime, a simple, efficient, and model-agnostic framework that teaches order to Video-LLMs via a Temporal Reconstruction objective and uses a Past-Current Dynamic Focus Cache for uncertainty-triggered retrieval, improving accuracy and reducing latency in streaming benchmarks, potentially useful for real-time geohazard video monitoring.
134. NoLan: Mitigating Object Hallucinations in Large Vision-Language Models via Dynamic Suppression of Language Priors
Core Problem: Object hallucination in Large Vision-Language Models (LVLMs), where models generate objects not present in the input image, primarily due to strong priors from the language decoder.
Key Innovation: NoLan (No-Language-Hallucination Decoding), a training-free framework that mitigates object hallucinations by dynamically suppressing language priors, modulated by the output distribution difference between multimodal and text-only inputs, significantly improving accuracy on benchmarks like POPE, which could enhance reliability of LVLMs in geohazard interpretation.
135. CASR: A Robust Cyclic Framework for Arbitrary Large-Scale Super-Resolution with Distribution Alignment and Self-Similarity Awareness
Core Problem: Arbitrary-Scale Super-Resolution (ASISR) is fundamentally limited by cross-scale distribution shift, causing noise, blur, and artifacts to accumulate sharply when inference scales deviate from the training range.
Key Innovation: Proposes CASR, a cyclic SR framework that reformulates ultra-magnification as a sequence of in-distribution scale transitions, using SDAM for structural distribution alignment and SARM for high-frequency texture restoration, ensuring stable inference and superior generalization at arbitrary scales with a single model.
136. Neu-PiG: Neural Preconditioned Grids for Fast Dynamic Surface Reconstruction on Long Sequences
Core Problem: Temporally consistent, high-fidelity, and drift-free surface reconstruction of dynamic 3D objects from unstructured point cloud data remains challenging for very long sequences, with existing methods being slow or requiring category-specific training.
Key Innovation: Neu-PiG, a fast deformation optimization method using a novel preconditioned latent-grid encoding, achieves superior accuracy, scalability, and significantly faster reconstruction (at least 60x faster than existing training-free methods) for dynamic 3D surfaces without explicit correspondences or priors.
137. Global Sequential Testing for Multi-Stream Auditing
Core Problem: Continuously auditing machine learning systems across multiple data streams to quickly detect unusual behavior is critical, but standard global tests (e.g., Bonferroni) can be inefficient under certain alternative hypotheses.
Key Innovation: Constructs new sequential tests by merging test martingales, achieving improved expected stopping times under both sparse and dense alternative hypotheses, and demonstrating effectiveness on synthetic and real-world data.
138. Dream-SLAM: Dreaming the Unseen for Active SLAM in Dynamic Environments
Core Problem: Existing active SLAM methods are limited by underlying SLAM modules, short-sighted motion planning, and struggles with dynamic scenes.
Key Innovation: Dream-SLAM, a novel monocular active SLAM method, addresses these limitations by 'dreaming' cross-spatio-temporal images and semantically plausible structures of partially observed dynamic environments to improve camera pose estimation, 3D scene representation, and long-horizon planning.
139. Coarsening Bias from Variable Discretization in Causal Functionals
Core Problem: Discretizing continuous variables in causal effect functionals introduces a first-order coarsening bias, distinct from statistical estimation error, which can lead to non-negligible approximation errors.
Key Innovation: A simple bias-reduced functional that evaluates the outcome regression at within-bin conditional means, eliminating the leading term of the coarsening bias and yielding a second-order approximation error, demonstrated to reduce bias and improve confidence interval coverage.
140. Slice and Explain: Logic-Based Explanations for Neural Networks through Domain Slicing
Core Problem: Logic-based approaches for explaining neural network predictions, while offering correctness guarantees, suffer from scalability concerns due to the complexity of logical constraints.
Key Innovation: An approach leveraging domain slicing to reduce the complexity of logical constraints, thereby facilitating logic-based explanation generation for neural networks and decreasing explanation time by up to 40%.
141. Parallel Split Learning with Global Sampling
Core Problem: Distributed deep learning in resource-constrained environments faces scalability and generalization challenges due to large effective batch sizes and non-identically distributed client data.
Key Innovation: Introduces a server-driven sampling strategy that maintains a fixed global batch size by dynamically adjusting client-side batch sizes, decoupling the effective batch size from the number of participating devices and ensuring global batches better reflect the overall data distribution, improving accuracy, training efficiency, and convergence stability.
142. Grounding-IQA: Grounding Multimodal Language Model for Image Quality Assessment
Core Problem: Existing MLLM-based Image Quality Assessment (IQA) methods primarily rely on general contextual descriptions, limiting their ability to perform fine-grained quality assessment.
Key Innovation: Introduces 'grounding-IQA,' a new paradigm that integrates multimodal referring and grounding with IQA to enable more fine-grained quality perception, supported by a new dataset (GIQA-160K) and benchmark (GIQA-Bench).
143. Training speedups via batching for geometric learning: an analysis of static and dynamic algorithms
Core Problem: The effect of different batching algorithms (static vs. dynamic) on training time and model performance for Graph Neural Networks (GNNs) has not been thoroughly explored, despite GNNs' growing use in various domains.
Key Innovation: Analyzes static and dynamic batching algorithms for GNNs, demonstrating that batching can provide significant training speedups (up to 2.7x) and that the optimal algorithm depends on data, model, batch size, hardware, and training steps, with some combinations showing significant differences in learning metrics.
144. FFINO: Factorized Fourier Improved Neural Operator for Modeling Multiphase Flow in Underground Hydrogen Storage
Core Problem: Fast and accurate modeling of hydrogen plume migration and pressure field evolution is crucial for Underground Hydrogen Storage (UHS) field management, but traditional numerical simulators are computationally expensive and slow.
Key Innovation: FFINO, a factorized Fourier improved neural operator, is proposed as a fast, accurate, and stable surrogate model for multiphase flow problems in UHS, achieving significant reductions in parameters, training time, and GPU memory while improving prediction accuracy and being 7,850 times faster than numerical simulators.
145. Voxel Densification for Serialized 3D Object Detection: Mitigating Sparsity via Pre-serialization Expansion
Core Problem: Serialized 3D object detection frameworks lack voxel expansion capabilities, hindering performance, especially for sparse foreground objects, as they strictly maintain input/output voxel dimension consistency.
Key Innovation: Proposes a novel Voxel Densification Module (VDM) that promotes pre-serialization spatial expansion by leveraging sparse 3D convolutions to densify feature representations and aggregate local context, consistently improving detection accuracy across multiple 3D object detection benchmarks.
146. Rethinking Consistent Multi-Label Classification Under Inexact Supervision
Core Problem: Existing consistent approaches for partial and complementary multi-label learning under inexact supervision often require accurate estimation of the label generation process or assume uniform distributions, conditions that are difficult to satisfy in real-world scenarios.
Key Innovation: Novel consistent approaches that do not rely on specific assumptions about label generation processes, utilizing two risk estimators based on first- and second-order strategies, with theoretical proofs of consistency and derived convergence rates, demonstrating effectiveness on both real-world and synthetic datasets.
147. MNO: Multiscale Neural Operator for 3D Computational Fluid Dynamics
Core Problem: Existing neural operators for solving partial differential equations (PDEs) in 3D Computational Fluid Dynamics (CFD) on irregular domains are limited in accuracy and scalability, particularly in capturing multiscale fluid flow structures.
Key Innovation: Introduces the Multiscale Neural Operator (MNO), an architecture that explicitly decomposes information across global, local, and micro scales using attention modules, significantly outperforming state-of-the-art baselines in 3D CFD tasks.
148. Domain Adaptation for Camera-Specific Image Characteristics using Shallow Discriminators
Core Problem: Camera-specific image characteristics degrade image quality and create a domain gap in learning-based perception algorithms, impeding performance when training data lacks these characteristics.
Key Innovation: Proposes shallow discriminator architectures with smaller receptive fields to improve learning of unknown local image distortions, achieving better instance segmentation performance and efficiency compared to previous pixel-level domain adaptation methods.
149. Improving Variational Autoencoder using Random Fourier Transformation: An Aviation Safety Anomaly Detection Case-Study
Core Problem: Improving the training process and inference of Autoencoders (AEs) and Variational Autoencoders (VAEs) for reconstruction-based anomaly detection, particularly regarding their ability to learn high-frequency features.
Key Innovation: Introduces Random Fourier Transformation (RFT) to AEs and VAEs, showing that models with RFT learn low and high-frequency features simultaneously, leading to superior performance in anomaly detection, demonstrated with an aviation safety dataset. A trainable RFT variant is also introduced.
150. Aligned Stable Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency
Core Problem: Generative image inpainting methods often produce unnatural results due to unwanted object insertion and noticeable color shifts in the inpainted regions.
Key Innovation: Introduces Aligned Stable inpainting with UnKnown Areas prior (ASUKA), a framework that uses reconstruction-based priors to suppress hallucinated objects and a specialized VAE decoder for local harmonization to reduce color shifts, leading to more natural and color-consistent inpainted images.
151. Pay Attention to Where You Looked
Core Problem: In few-shot novel view synthesis, existing methods often assume equal importance for all input views relative to the target, leading to suboptimal results.
Key Innovation: Introduces a camera-weighting mechanism that adaptively adjusts the importance of source views based on their relevance to the target (using geometric properties or a cross-attention-based learning scheme), enhancing accuracy and realism in novel view synthesis.
152. Extracting and Analyzing Rail Crossing Behavior Signatures from Videos using Tensor Methods
Core Problem: Traditional approaches to railway crossing safety analyze crossings individually, limiting the ability to identify shared behavioral patterns across locations and inform targeted interventions.
Key Innovation: Proposed a multi-view tensor decomposition framework using TimeSformer embeddings from railway crossing videos to capture behavioral similarities across temporal phases. Revealed latent behavioral components, showing location as a stronger determinant than time of day, and approach-phase behavior as discriminative.
153. RAYNOVA: Scale-Temporal Autoregressive World Modeling in Ray Space
Core Problem: World foundation models struggle to simulate the evolution of the real world with physically plausible behavior, especially in handling spatial and temporal correlations across multiple views and scales.
Key Innovation: RAYNOVA, a geometry-agnostic multiview world model that employs a dual-causal autoregressive framework for scale-wise and temporal reasoning in ray space, enabling robust generalization to diverse camera setups and long-horizon video generation without explicit 3D scene representation.
154. MPC of Uncertain Nonlinear Systems with Meta-Learning for Fast Adaptation of Neural Predictive Models
Core Problem: Designing optimal controllers for reference tracking in uncertain nonlinear systems, especially when limited data is available for the specific target system, making it challenging to quickly train accurate neural predictive models.
Key Innovation: Proposing an iMAML (implicit model-agnostic meta-learning) framework that leverages data from source systems to expedite training and enhance control performance of neural State-Space Models (NSSMs) for MPC in target uncertain nonlinear systems, enabling fast adaptation with limited online data and reduced computational complexity.
155. InsightX Agent: An LMM-based Agentic Framework with Integrated Tools for Reliable X-ray NDT Analysis
Core Problem: Existing deep-learning-based X-ray NDT approaches lack interactivity, interpretability, and critical self-assessment, limiting reliability and operator trust in industrial quality assurance.
Key Innovation: Proposes InsightX Agent, an LMM-based agentic framework that orchestrates a Sparse Deformable Multi-Scale Detector (SDMSD) and an Evidence-Grounded Reflection (EGR) tool to provide reliable, interpretable, and interactive X-ray NDT analysis, enhancing diagnostic reliability and trustworthiness through active reasoning.
156. Synthetic vs. Real Training Data for Visual Navigation
Core Problem: Visual navigation policies trained in simulation often suffer significant performance degradation (the sim-to-real gap) when evaluated in the real world.
Key Innovation: A novel navigation policy architecture that leverages pretrained visual representations to bridge the sim-to-real appearance gap, enabling simulator-trained policies to match or outperform real-world-trained counterparts and prior state-of-the-art methods in navigation success rate on mobile robots and drones.
157. MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile Manipulation
Core Problem: Collecting large-scale, diverse human demonstrations for multi-step bimanual mobile manipulation is costly and time-consuming, especially considering robot base placement for reachability and camera positioning for visibility.
Key Innovation: MoMaGen, a framework that formulates data generation as a constrained optimization problem to satisfy hard (e.g., reachability) and soft (e.g., visibility) constraints, enabling the generation of much more diverse datasets for bimanual mobile manipulation and training successful imitation learning policies from a single source demo.
158. Multimodal Datasets with Controllable Mutual Information
Core Problem: There is a lack of benchmark datasets that allow for systematic studies of mutual information estimators and multimodal self-supervised learning techniques with known and controllable mutual information between modalities.
Key Innovation: A framework for generating highly multimodal datasets where mutual information between modalities is explicitly calculable, achieved using a flow-based generative model and a structured causal framework for correlated latent variables, providing a novel testbed for MI estimators and multimodal SSL.
159. HetroD: A High-Fidelity Drone Dataset and Benchmark for Autonomous Driving in Heterogeneous Traffic
Core Problem: Autonomous driving systems struggle to navigate real-world heterogeneous traffic, especially with vulnerable road users (VRUs) exhibiting complex, unstructured behaviors (e.g., hook turns, lane splitting), which are underrepresented in existing datasets focused on structured traffic.
Key Innovation: Presents HetroD, a large-scale drone-based dataset and benchmark providing holistic observations of heterogeneous traffic with centimeter-accurate annotations, HD maps, and traffic signal states. It comprises over 65.4k high-fidelity agent trajectories (70% VRUs) and supports modeling VRU behaviors, revealing that state-of-the-art models struggle with its challenges.
160. Ductile-to-brittle transitions with decreasing temperature in BCC transition metals: fractographic observations and issues with conventional hypotheses
Core Problem: Understanding the abrupt ductile-to-brittle transitions in BCC transition metals and alloys at decreasing temperatures, and challenging existing hypotheses.
Key Innovation: High-resolution fractographic observations providing new insights into the characteristics of ductile-to-brittle transitions in BCC metals, questioning conventional explanations.
161. A new model for the ductile-brittle transition in BCC metals
Core Problem: Developing an improved model for the ductile-brittle transition in body-centred cubic (BCC) metals, informed by recent high-resolution fractographic data.
Key Innovation: Presentation of a new model for the ductile-brittle transition in BCC metals, incorporating insights from high-resolution images of brittle fracture surfaces.
162. Effect of Mg17Al12 on brittle and ductile crack extension behaviours in magnesium
Core Problem: Addressing the low fracture toughness of magnesium and its alloys by investigating how the common precipitation phase Mg17Al12 influences crack extension behaviors.
Key Innovation: Analysis of the effect of Mg17Al12 on both brittle and ductile crack extension in magnesium, providing insights into improving the fracture toughness of these alloys.
163. A surface-physics-based hypothesis for ductile-to-brittle transitions in a variety of pure metals and crystal structures
Core Problem: Explaining the similar characteristics of 'brittle' fractures in various pure metals and crystal structures at low temperatures, which challenges existing theories.
Key Innovation: Proposal of a surface-physics-based hypothesis to account for ductile-to-brittle transitions observed across a range of pure metals and crystal structures, based on fractographic observations.
164. A hybrid POD-LSTM framework for efficient vortex-induced vibration prediction in marine risers via multi-fidelity data fusion
Core Problem: Accurate and computationally efficient prediction of vortex-induced vibrations (VIV) in marine risers is challenging due to the high cost of high-fidelity models and the inaccuracy of low-fidelity approximations.
Key Innovation: A novel reduced-order modeling framework integrating Proper Orthogonal Decomposition (POD) and Long Short-Term Memory (LSTM) networks with multi-fidelity data fusion, enabling efficient and accurate VIV prediction in marine risers by leveraging sparse high-resolution data and abundant low-fidelity simulations.
165. Spatial-frequency fusion object detection in low-quality forward-looking sonar images based on improved YOLOv12
Core Problem: Object detection in Forward-Looking Sonar (FLS) images is severely challenged by low signal-to-noise ratio (SNR), leading to blurred target edges and structural discontinuities, and existing image enhancement techniques often introduce perturbations that degrade deep neural network performance.
Key Innovation: Proposed a novel spatial-frequency fusion object detection method based on improved YOLOv12, incorporating a Color-Adaptive Multi-scale Retinex (CAMRetinex) module for image enhancement and a SpectC2f module for feature extraction, achieving superior accuracy and efficiency for real-time underwater sonar object detection.
166. Transformer-based multimodal target detection and projection-based pose estimation for autonomous deployment and recovery of surface-state UUVs
Core Problem: Significant challenges in target detection and pose estimation for autonomous UUV deployment and recovery, including sparse point clouds, occlusion, insufficient feature fusion, and strict real-time constraints.
Key Innovation: Proposing a robust multimodal perception framework featuring MultiFormer Net (Transformer-based cross-attention and point-channel dual attention for detection) and two projection-based pose estimation algorithms (DensE-Pose and Dual-PCA) to achieve high accuracy and real-time performance in complex operating scenarios.
167. Hydrometeorological and hydrological data from Baker Creek Research Watershed, Northwest Territories, Canada Release V.3
Core Problem: Lack of comprehensive, long-term hydrometeorological and hydrological data for subarctic Canadian Shield terrain to understand hydrological responses to climate change.
Key Innovation: Release of a unique, long-term (2003-2025) dataset of half-hourly hydrometeorological and daily streamflow data for the Baker Creek Research Watershed, enabling better understanding of subarctic hydrological responses to warming trends and precipitation cycles.
168. Enhancing process interpretation with isotopes: potential discharge-isotope trade-offs in ecohydrological modelling of heavily managed lowland catchments
Core Problem: Quantifying ecohydrological processes and constraining model equifinality in heavily managed catchments using tracer-aided modeling (TAM), especially with sparse isotope data.
Key Innovation: Demonstrated that even sparse isotope data can provide informative insights in TAM for complex, heavily managed catchments, helping to constrain process representation and identify epistemic errors (e.g., un-represented water withdrawals).
169. Uncertainties in long-term ensemble estimates of contextual evapotranspiration over southern France
Core Problem: Accurately estimating evapotranspiration (ET) at large spatial scales using remote sensing, and quantifying the uncertainties associated with different model inputs and formulations.
Key Innovation: Developed an ensemble-based contextual tool (EVASPA) for long-term ET estimation, demonstrating its ability to provide reliable flux estimates and meaningful uncertainty spreads, identifying LST inputs and EF formulations as dominant uncertainty sources.
170. Century-long coral records of the South China Sea throughflow slowdown
Core Problem: Long-term changes of the South China Sea throughflow (SCSTF) under a warming climate and its underlying dynamics are unclear due to limited observations.
Key Innovation: Reconstructed centennial-scale variability of the SCSTF (1894–2022) using coral oxygen isotope and satellite data, revealing a marked long-term decreasing trend (−0.14 ± 0.02 Sv per decade) driven by intensified trade winds in the tropical western Pacific.
171. Intelligent lithology identification method based on spectral curve images of advanced horizontal drilling cuttings and engineering application
Core Problem: Efficient and accurate lithology identification from drilling cuttings is crucial for tunnel engineering and resource exploration, but traditional methods can be time-consuming or less precise.
Key Innovation: Developed an intelligent lithology identification method combining object detection (YOLOv5s) with VNIR spectroscopy on drilling cuttings, achieving high accuracy in both laboratory and field settings, and demonstrating its practical application in engineering projects.
172. Unraveling bofedal change and degradation: Multidimensional analysis of pastoral management, local knowledge, and image analysis in Sajama National Park, Bolivia
Core Problem: Understanding the multidimensional drivers and processes of bofedal change and degradation in Sajama National Park, integrating ecological, local knowledge, and remote sensing data.
Key Innovation: Identified a decline in cushion-forming species and contraction of dense bofedales, revealing simultaneous processes of decline, stability, and recovery influenced by climate change, environmental stressors, and seismic activity, while highlighting the role of water management.
173. Application of digital image correlation and strain inversion to determine the five elastic constants of transversely isotropic rock from a single rock core by a strip load
Core Problem: Efficiently and accurately determining the five elastic constants of transversely isotropic rock from a single rock core, which are fundamental parameters for rock mechanics and geohazard assessment.
Key Innovation: Developed and validated a method combining digital image correlation (DIC) and strain inversion to determine the five elastic constants of transversely isotropic rock from a single rock core under a strip load, offering an efficient and accurate approach.
174. A novel ductile fuse for seismic-resilient precast concrete frames accounting simplified fabrication
Core Problem: Addressing the difficulty and cost of post-earthquake repair and replacement in conventional RC moment-resisting frames and the challenges of achieving MRF connections in precast construction.
Key Innovation: A novel precast concrete frame system incorporating ductile steel link beams (replaceable steel fuses) that confine seismic damage to easily replaceable components, demonstrating stable, degradation-free hysteretic behavior and improved energy dissipation.
175. Seismic fragility assessment of prefabricated concrete frame structures with grouted sleeve connections
Core Problem: Limited understanding of the overall seismic performance of prefabricated concrete frame structures incorporating grouted sleeve connections, despite extensive research on local connection behavior.
Key Innovation: A hybrid numerical approach combining detailed finite element simulation of beam-column joints with global structural modeling to perform seismic fragility assessment of multi-story prefabricated concrete frames with grouted sleeve connections, confirming reliable seismic performance.
176. Development of granular road assessment and asset management tool for quantifying superload impacts
Core Problem: Granular roads are highly vulnerable to damage from superload vehicles, leading to accelerated deterioration, and there is a need for tools to evaluate the structural and economic effects of these superloads.
Key Innovation: Development of RISAT (Road Infrastructure-Superload Analysis Tool), a data-driven, user-friendly tool integrating mechanistic analysis with ANN models to predict critical road responses, rutting damage, costs, and service life reduction under superload conditions, validated against field data.