TerraMosaic Daily Digest: Feb 26, 2026
Daily Summary
This February 26, 2026 digest shows a stronger coupling of process physics and operational analytics than in previous cycles. High-priority studies do not treat landslides, earthquakes, floods, and coastal hazards as isolated topics; they resolve trigger-to-impact chains, from multifactor InSAR deformation attribution in karst terrain to reservoir-rainfall coupling in slope failure, seismic source characterization, and physically constrained convective nowcasting for flash-flood-prone systems.
A second clear signal is methodological selectivity: AI-heavy contributions are increasingly valuable only when they add physical consistency, uncertainty control, or direct decision utility. The most actionable advances include tsunami wavefield mapping from SWOT, probabilistic large-deformation slope analysis with spatial variability, cold-region thaw susceptibility under scarce data, and multi-risk coastal frameworks that integrate exposure, vulnerability, and long-horizon adaptation.
Key Trends
The technical frontier is shifting from model-centric performance claims to evidence-backed, deployable geohazard intelligence.
- Remote sensing is moving from detection to attribution: current work combines time-series InSAR, multi-satellite observations, and DEM/DEM-like mechanics to separate hydroclimatic, tectonic, and anthropogenic drivers of deformation.
- Hydro-mechanical coupling is becoming standard in slope analysis: studies now incorporate hysteresis, storage effects, and spatial hydraulic variability to improve both failure timing and post-failure behavior prediction.
- Seismic hazard research is emphasizing physically interpretable response metrics: new models link crustal structure, inelastic response spectra, and fault-zone materials to hazard estimates that are more transferable to engineering design.
- Flood risk work is increasingly behavior-aware and systems-aware: research combines precipitation nowcasting, crowd decision dynamics, and network-scale forecasting to improve warning relevance and emergency response timing.
- Generic AI papers are being filtered by transfer value: methods without explicit hazard grounding remain secondary unless they demonstrably improve geohazard monitoring, forecasting, or resilience workflows under realistic data constraints.
Selected Papers
This digest features 185 selected papers from 1030 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. Coupled Multifactor Analysis of Surface Deformation in the Karst Regions of Guangxi Using Time-Series InSAR Monitoring
Core Problem: The Guangxi karst region is highly susceptible to collapse and surface deformation, posing significant risks, but lacks sufficient long-term, region-scale monitoring and comprehensive multidriver coupling analyses to understand the complex causes of deformation.
Key Innovation: The first autonomous-region-scale, systematic InSAR deformation dataset for Guangxi (derived from 691 Sentinel-1A SAR images), coupled with a multifactor analysis identifying diverse drivers (drought, irrigation, groundwater extraction, earthquakes, fault zones, climate, industrial activity, mining), providing a robust scientific foundation for spatial planning and disaster-risk mitigation.
2. A comparative study on spatial patterns and dominant factors of landslide susceptibility under geomorphological differentiation
Core Problem: Spatial heterogeneity significantly challenges large-scale landslide susceptibility assessment, often leading to generalized models that overlook local variations.
Key Innovation: Conducted a comparative study focusing on southeastern [region] to analyze spatial patterns and dominant factors of landslide susceptibility under geomorphological differentiation, aiming to overcome the limitations of generalized models by accounting for local variations.
3. Tracking advances in landslide numerical simulation: A comprehensive bibliometric study
Core Problem: The need for a systematic and comprehensive evaluation of the evolution and current state of landslide numerical simulation research to identify key trends, hotspots, influential contributors, and inform future research directions.
Key Innovation: Provides a comprehensive bibliometric analysis of landslide numerical simulation research from 1995-2024, characterizing its evolution through distinct phases, identifying research hotspots (triggering mechanisms, stability analysis, dynamics, methodologies), and highlighting leading countries and journals.
4. Coupled Effects of Reservoir Level Fluctuations and Rainfall on Landslide Deformation Using InSAR and DEM Analysis
Core Problem: Understanding the deformation characteristics and failure evolution mechanism of the Taping 1 # landslide under the coupled effects of reservoir level fluctuations and rainfall.
Key Innovation: InSAR technology and Discrete Element Method (DEM) revealed three distinct landslide deformation zones, quantified movement rates, and demonstrated that coupled reservoir water drop and rainfall infiltration cause a multi-stage retrogression-type failure, initiating at the toe and progressing to the middle.
5. Ocean Bottom Seismometers Provide Direct Measurements of Pulsed‐Structure and Turbulence of Turbidity Currents Overspilling From a Submarine Channel
Core Problem: The overspill of turbidity currents onto channel-levees and abyssal mixing remain poorly constrained due to a lack of direct observations.
Key Innovation: Ocean Bottom Seismometers (OBS) deployed on Congo Canyon-Channel levees captured the structure and turbulence of overspill during a large canyon-flushing event, revealing that overspill is long-lasting, highly pulsed, and generates strong turbulence, providing insights into levee growth and deep-ocean mixing.
6. Advancing Convective Precipitation Nowcasting via 3D Polarimetric Radar Data and Physics‐Constrained Deep Learning Model
Core Problem: Accurate nowcasting of severe convective precipitation remains challenging, and most deep learning models lack physical constraints, limiting their consistency with atmospheric processes.
Key Innovation: This study introduces FURECast, a physics-constrained deep learning model that leverages 3D polarimetric radar data (ZH, ZDR, KDP) and embeds their intrinsic self-consistency relation as a physical constraint, achieving a 14.1% improvement in 90-min critical success index and reducing physical inconsistency by two orders of magnitude.
7. Distinct Crustal Structure Across the Alpine Fault, New Zealand: Seismic Imaging of a Through‐Going Vertical Fault Beneath Its Central Section
Core Problem: The unclear subsurface accommodation of fault geometry changes (from near-vertical to dipping) along the Alpine Fault, which impacts understanding fault motion and future rupture scenarios.
Key Innovation: Using receiver functions from seismic stations to image the subsurface, revealing distinct crustal structures and suggesting the co-existence of vertical and dipping fault structures beyond a section boundary, consistent with microseismicity and previous tectonic studies.
8. Distributed Lower‐Crustal Flow Beneath the Central Xianshuihe–Xiaojiang Fault System: Reconciling Geodesy and Geology
Core Problem: The long-standing mismatch between geodetic and geologic slip-rate estimates across the Zemuhe-Daliangshan fault zone when using conventional elastic dislocation models.
Key Innovation: Incorporating a viscoelastic rheology for the lower crust into models, which reconciles geodetic and geologic slip rates (7mm/yr on Zemuhe, 2mm/yr on Daliangshan) and yields locking depths consistent with microseismicity, indicating distributed lower-crustal flow in this immature system.
9. A Synergistic Approach: Dynamics-AI Ensemble in Tropical Cyclone Forecasting
Core Problem: AI-based weather forecasting, particularly for tropical cyclones, faces a critical challenge in bridging the gap between computational efficiency and dynamic consistency, as conventional ensembles are computationally expensive and AI ensembles often lack adequate perturbation methods.
Key Innovation: The study develops an AI-driven optimized ensemble forecast system using Orthogonal Conditional Nonlinear Optimal Perturbations (O-CNOPs) that generate dynamically optimized, orthogonal perturbations respecting nonlinear dynamics, demonstrating superior deterministic and probabilistic skills in tropical cyclone track forecasting.
10. First mapping of a tsunami wavefield by SWOT satellite: observation data and preliminary numerical simulation of the 19 May 2023 tsunami near the Loyalty Islands
Core Problem: The challenge of dynamically and accurately monitoring tsunami wavefields and assessing the capability of new satellite technology for this purpose.
Key Innovation: SWOT satellite's unprecedented 2D observation of a tsunami wavefield, demonstrating its ability to record and inform tsunami propagation and modelling, offering a breakthrough for better predictions.
11. Seismic source location using deep learning and a metaheuristic-optimized grid search algorithm under a sparse network in underground coal mines
Core Problem: Achieving high-accuracy seismic source location in underground coal mines, especially with sparse monitoring networks, is challenging, hindering effective assessment of rock stability and mitigation of dynamic hazards.
Key Innovation: Developed a method combining deep learning and a metaheuristic-optimized grid search algorithm for accurate seismic source location in underground coal mines, even with sparse networks, to improve rock stability assessment and hazard mitigation.
12. Analysis of the structural characteristics and deformation mechanisms of unstable rock masses in the red beds along the Renchi Expressway
Core Problem: Understanding the structural characteristics and deformation mechanisms of unstable rock masses in red-bed strata along the Renhuai-Chishui Expressway to assess and manage associated hazards.
Key Innovation: Integrated UAV tilt photogrammetry and PFC simulation to analyze unstable rock masses in red-bed strata along the Renchi Expressway, revealing their development patterns and deformation mechanisms.
13. Influence of the basal shear zone on long-term earthflow evolution
Core Problem: Understanding the long-term evolution of earthflows, particularly the role of the basal shear zone in their intermittent movement, including slowing down and surging stages, is crucial.
Key Innovation: Describes a hydro-mechanical scenario where a thick, highly remoulded basal shear zone plays a key role in earthflow evolution. Based on site investigations, this scenario explains the generally high thickness of the shear zone and the relatively high frictional shear strength mobilized by the soil mass during movement.
14. Probabilistic large deformation analysis of rainfall-induced slope failures considering spatial variability of saturated hydraulic conductivity
Core Problem: Accurately predicting post-failure characteristics, landslide types, and probability of rainfall-induced slope failures requires considering the spatial variability of soil properties like saturated hydraulic conductivity, which is computationally challenging with existing methods.
Key Innovation: The study proposes a computationally efficient coupled hydro-mechanical framework, Random Limit Equilibrium and Material Point Methods (RLE-MPM), for probabilistic large deformation analysis of rainfall-induced slope failures, demonstrating its effectiveness in evaluating post-failure characteristics and probability of failure while highlighting the significant impact of spatial variability of saturated hydraulic conductivity.
15. Jigsaw-fit blocks: A tale of segregation and disaggregation
Core Problem: The mechanisms behind the constrained disaggregation and matrix infilling of jigsaw-fit blocks within volcanic debris avalanches and debris flows remain unclear, challenging existing theories of granular flow rheology.
Key Innovation: Experimental studies demonstrated that disaggregation of jigsaw-fit blocks occurs regardless of fragmentation pattern, is conditioned by fragment density (faster in lighter fragments), and results from fragment rotation and particle infilling, providing a new basis for interpreting debris avalanche deposits and inferring kinematic features.
16. Farmers’ social networks and adoption of disaster risk reduction measures: An experimental study in Uganda
Core Problem: Strengthening local community resilience to hazards like landslides and flash floods requires improving farmers' knowledge and adoption of farm-based Disaster Risk Reduction (DRR) measures, but the effectiveness of informal, local social networks in information transfer for DRR remains unclear.
Key Innovation: An experimental study demonstrating that knowledge transfer through Citizen Scientists (an example of local social networks) is more effective in enhancing the adoption of tree planting (and in group sessions, diversion channels) as DRR measures compared to conventional outreach by formal extension workers in disaster-prone Western Uganda.
17. A method for better mapping of susceptibility to thaw hazards in data-scarce cold regions
Core Problem: Accurate assessment of thaw hazard susceptibility in data-scarce permafrost regions (e.g., Qinghai–Tibet Plateau) is challenging, especially with limited data.
Key Innovation: Developed an inventory of thaw-hazard sites and demonstrated that the novel TabPFN machine learning model significantly outperforms conventional models for thaw-hazard susceptibility mapping under small-sample constraints, providing robust, interpretable predictions and aiding risk mitigation.
18. Importance of hydraulic hysteresis and storage coefficient of soils on slope stability analysis under rainfalls
Core Problem: Accurately evaluating slope stability under rainfalls, specifically considering the influences of hydraulic hysteresis and the storage coefficient of soils.
Key Innovation: Investigated the significant influences of hydraulic hysteresis (drying/wetting soil-water retention curves) and a newly derived storage coefficient on slope safety factors and failure times under various rainfall conditions for different soil types, providing a more feasible scheme for prediction.
19. Ground motion model for inelastic response spectra in the Indian subcontinent using a transfer learning approach
Core Problem: Accurate prediction of inelastic response spectra (IRS) and seismic hazard assessment for the Indian subcontinent is challenging due to its unique seismotectonic characteristics, and current code provisions may underestimate seismic hazard.
Key Innovation: Developed a neural network-based ground-motion model (GMM) for inelastic response spectra in the Indian region using a transfer learning approach. The model effectively predicts IRS for acceleration, velocity, and displacement, facilitates improved probabilistic seismic hazard assessment (PSHA), and demonstrates potential underestimation by the Indian Standard Code.
20. Size‐Dependent Melting Behavior of Ultrafine Fault Rocks
Core Problem: The thermal and mechanical effects of extreme grain-size reduction in principal slip zones (PSZs) on dynamic weakening via melting during earthquakes remain poorly constrained.
Key Innovation: This study examines the melting behavior of dry granitoid fault rocks ball-milled to ultrafine grain sizes, demonstrating that extreme grain-size reduction significantly lowers the melting temperature (from 1,270°C to 1,090°C), contributing to fault weakening.
21. Global River Forecasting with a Topology-Informed AI Foundation Model
Core Problem: Widespread hydrology data scarcity restricts data-driven river forecasting to isolated predictions, hindering systemic simulation of inherently interconnected river networks, especially in 'ColdStart' scenarios without historical river states.
Key Innovation: GraphRiverCast (GRC), a topology-informed AI foundation model, simulates multivariate river hydrodynamics in global river systems, operating in 'ColdStart' mode with high Nash-Sutcliffe Efficiency (0.82) by leveraging topological encoding and physics-based pre-training.
22. TerraCodec: Compressing Optical Earth Observations
Core Problem: Massive streams of multispectral image time series from Earth observation (EO) satellites pose significant challenges for storage and transmission, with existing learned EO compression methods being fragmented, lacking pretrained codecs, and underexploring temporal redundancy.
Key Innovation: Introduces TerraCodec (TEC), a family of learned codecs pretrained on Sentinel-2 EO data, including a Temporal Transformer model (TEC-TT) that leverages temporal dependencies and Latent Repacking for flexible-rate models. TerraCodec achieves 3-10x higher compression at equivalent image quality and enables zero-shot cloud inpainting, establishing neural codecs as a promising direction for Earth observation.
23. Nature-based solutions for coastal protection based on intuitionistic fuzzy group decision making
Core Problem: Existing frameworks for selecting Nature-based Solutions (NbS) for coastal protection are limited in handling complex interactions among multiple indicators and inherent uncertainty in expert evaluations.
Key Innovation: An integrated assessment framework combining Fuzzy Delphi Method (FDM) and entropy-based intuitionistic fuzzy TOPSIS (IF-TOPSIS) is developed to facilitate expert-driven multi-criteria decision analysis for selecting NbS in coastal protection, demonstrated through a case study in Taiwan.
24. Analysis of L-Band Radar Signatures of Surface Topography, Soil Moisture, and Vegetation Water Content
Core Problem: A comprehensive understanding of the individual and combined impacts of surface topography, soil moisture, and vegetation water content (VWC) on L-band SAR backscatter is needed for improved remote sensing applications.
Key Innovation: A comprehensive analysis using a two-dimensional binned framework to isolate the impacts of soil moisture, VWC, and surface topography (STD slope) on L-band SAR backscatter, revealing strong backscatter response to STD slope and soil moisture under low VWC, and modeling these relationships with piecewise linear and exponential functions.
25. Human decision-making in crowds in a virtual flood scenario
Core Problem: Lack of empirical data on individual human decision-making during natural hazards, specifically how social cues (crowd behavior) interact with other factors to influence flood evacuation choices and delays.
Key Innovation: Uses Virtual Reality to demonstrate that crowd behavior strongly determines route choice and evacuation latency in flood scenarios, often outweighing other factors, and highlights the need to integrate realistic human behavior into flood risk models and evacuation planning.
26. AGILE v0.1: The Open Global Glacier Data Assimilation Framework
Core Problem: Glacier models face challenges in integrating heterogeneous observations in a dynamically consistent way, and estimates of current glacier volume and area remain uncertain due to outdated inventories.
Key Innovation: Presents AGILE, an Open Global Glacier Data Assimilation Framework, a time-dependent variational method inspired by 4D-Var data assimilation. Built on a PyTorch reimplementation of OGGM, it uses automatic differentiation to efficiently optimize control variables (e.g., glacier bed topography, ice volume) for transient calibration.
27. Multi-risk assessment in the NW coast of the Island of Malta (central Mediterranean Sea)
Core Problem: Coastal communities are increasingly vulnerable to multiple, interacting weather- and climate-related hazards (erosion, flooding, slope instability, sea level rise), necessitating comprehensive multi-risk assessment frameworks for effective coastal management.
Key Innovation: The research applies a multi-risk framework using an index-based approach and GIS tools to assess and map exposure, vulnerability, and susceptibility to temporary/permanent inundation, shoreline erosion, and rock fall along the NW coast of Malta, projecting multi-risk scenarios for 2100 to identify critical zones for adaptation.
28. Risk awareness characterisation in multi-hazard, high-tourist-interest sites: the case of Stromboli volcano during the 2019 eruptive crisis (Aeolian islands UNESCO site, Italy)
Core Problem: In multi-hazard, high-tourism volcanic regions, increasing tourist flows amplify volcanic risk, and there's a need to understand the contrasting risk awareness profiles between local inhabitants and tourists regarding various threats (volcanic, seismic, landslide, tsunami) to improve disaster risk management.
Key Innovation: The study employs a mixed-methods approach (interviews, surveys) to characterize risk awareness among inhabitants and tourists on Stromboli during a volcanic crisis, revealing distinct profiles (place-based knowledge vs. fragmented awareness) and highlighting limited awareness of non-volcanic hazards (landslides, tsunamis) among tourists, providing insights for integrated disaster risk management in dynamic demographic settings.
29. A novel approach for physical modelling of seismic actions on a retaining wall supported by the rock-socketed pile foundation
Core Problem: Physical modeling of scaled-down retaining walls and rock-socketed pile foundations under seismic actions is complicated, making it challenging to accurately evaluate their earthquake response and inform seismic design.
Key Innovation: Developed a novel physical modeling approach using 1g shaking table experiments on a scaled-down retaining wall model with a rotational spring assembly. This method successfully replicated earthquake response, investigated backfill inertial forces, and highlighted the critical importance of considering the ductility of the rock-socketed pile foundation in seismic design.
30. Retrieval and Attribution of Tropical Cyclone Vertical Tilt From SAR and Infrared Satellite Imagery
Core Problem: The need for a systematic understanding and attribution of tropical cyclone vertical tilt, which plays a key role in storm structure and intensity changes, but has been challenging to observe globally.
Key Innovation: Constructing a global satellite-based tilt dataset from 1,024 paired SAR and infrared images for 264 TCs (2016–2024), revealing systematic links between TC tilt and intensity, motion, and latitude, and using interpretable regression and deep learning to attribute tilt to vertical wind shear, tropospheric temperature differences, TC intensity, and cloud-related indices.
31. Reflectance Multispectral Imaging for Soil Composition Estimation and USDA Texture Classification
Core Problem: Traditional soil texture determination methods are slow, labor-intensive, costly, or too coarse for routine field deployment, hindering efficient geotechnical screening and agricultural management.
Key Innovation: A robust, cost-effective, and field-deployable multispectral imaging (MSI) system combined with machine learning to accurately and non-destructively estimate soil composition (clay, silt, and sand percentages) and classify USDA texture, achieving high R^2 and accuracy, suitable for geotechnical screening.
32. Statistical Characteristics of a Natural Marine Sand under Constant Volume Cyclic Direct Simple Shear
Core Problem: Quantifying the uncertainty and variability in the cyclic resistance of sandy soil under dynamic loads, which hinders reliable design and risk assessment for foundation systems in seismic regions.
Key Innovation: A robust statistical dataset from 30 repeated cyclic direct simple shear (CDSS) tests, providing insights into the variability of cyclic soil properties, and enabling a probabilistic framework for cyclic design and confidence-based correction of design parameters.
33. High Spatial Resolution of GRACE-Derived Ice Mass Change Reveals Glacier-Scale Mass Loss in Greenland Ice Sheet
Core Problem: The coarse spatial resolution (∼300 km) of GRACE/GRACE-FO data limits detailed understanding of ice mass change response mechanisms to climate change at finer spatial scales (below 200,000 km²), particularly in the Greenland Ice Sheet.
Key Innovation: The first application of high-resolution downscaling (from 0.25° × 0.25° to 5 km × 5 km) to continuous reconstructed GRACE/GRACE-FO data over the Greenland Ice Sheet using geographically weighted regression (GWR), enabling glacier-scale analysis of ice mass loss and climatic forcing mechanisms.
34. Lessons learned in institutional preparedness and response during the 2022 European drought
Core Problem: The increasing frequency and severity of droughts in Europe necessitate understanding how forecasting systems and Drought Management Plans (DMPs) impact response timing and effectiveness.
Key Innovation: Quantifies that organizations with forecasting systems or DMPs implemented drought response measures earlier and rated their effectiveness higher, emphasizing the necessity of standardized, continent-wide drought risk management coordination and a Drought Directive.
35. How well do hydrological models simulate streamflow extremes and drought-to-flood transitions?
Core Problem: It is unclear how well hydrological models capture compound extreme events like drought-to-flood transitions and which modeling decisions are most important for model performance in simulating these events.
Key Innovation: Showed that general model performance (KGE) does not guarantee good detection of streamflow extremes and transitions, and that performance for extremes primarily depends on capturing streamflow timing. Highlighted that model structure, catchment characteristics, and meteorological forcings play a key role, with generally poor representation in semi-arid and high-mountain catchments.
36. Mechanism of time-lapse electrical resistivity tomography (ERT) response to mining-induced fracture evolution in shallow coal seams: a coupled DEM-FEM study
Core Problem: Dynamically monitoring the Water Flowing Fractured Zone (WCFZ) and mapping the complex spatiotemporal evolution of mining-induced fractures to electrical signals is challenging, yet critical for preventing water hazards in shallow coal seams.
Key Innovation: Proposes a time-lapse electrical forward modeling strategy coupling Discrete Element Method (DEM) and Finite Element Method (FEM) via Digital Image Processing (DIP). This approach reveals the 'vertical initiation-penetration-compaction recovery' mechanism and its distinct electrical signatures, validating ERT for quantitatively evaluating the 'damage-recovery' state of the goaf.
37. Integrating rainfall duration to urban flood exposure assessment in Bornova Watershed, Izmir, Turkey
Core Problem: Traditional urban flood exposure assessments often overlook the critical impact of varying rainfall durations, leading to underestimation of flood extent and population/building exposure, especially in the context of urbanization and climate change.
Key Innovation: The study uses high-resolution terrain data and HEC-RAS 2D simulations under 12h and 24h rainfall duration scenarios to quantify the increased flood extent and exposure (population, buildings, vulnerable groups) due to longer rainfall, emphasizing the need for duration-inclusive risk assessments and localized adaptation strategies.
38. Evaluating flood risk in the Yangtze River Delta region using explainable machine learning
Core Problem: Assessing comprehensive flood risk across large urbanized regions is challenging due to complex interactions of multiple geospatial indicators and the need for interpretable models to inform context-specific governance.
Key Innovation: The study develops an explainable machine learning framework (Random Forest with SHAP values) integrating 14 geospatial indicators to assess and map flood risk across 41 cities, identifying key drivers (typhoon frequency, GDP, runoff depth, precipitation, DEM) and their varying contributions, providing a scientific basis for regional flood risk governance.
39. The impact of antecedent topography on tsunami deposition
Core Problem: Despite the importance of sedimentary records for extending long-term tsunami data, the specific depositional mechanisms of tsunami waves and how antecedent topography influences these deposits are poorly understood, limiting the interpretation of palaeotsunami evidence.
Key Innovation: The study uses flume experiments, photogrammetry, and video analysis to examine the depositional mechanisms of bore-type tsunami waves under differing antecedent topography, demonstrating that sediment transport is primarily near-bed and deposition occurs in depressions and onshore areas, providing insights to better interpret palaeotsunami deposits.
40. A semi-empirical approach to estimate the future frequency of extreme sea level events: the case study of Trieste (North Adriatic)
Core Problem: Accurately estimating the future frequency of extreme sea level events, considering both climate model outputs and observed statistical distributions, is crucial for urban adaptation but presents methodological challenges, especially for long-term projections.
Key Innovation: The study proposes a semi-empirical methodology combining climate model outputs, IPCC mean sea level projections, and observed extreme event statistics to estimate the future frequency of extreme sea level events (daily mean and Highest High Waters) for Trieste, revealing a projected increase of more than one order of magnitude by 2100, primarily due to mean sea level rise, highlighting the urgency for urban adaptation.
41. Similarity study of soil spatial variability between the longitudinal sections of new and old subgrade foundations based on the autocorrelation function of CPTU test
Core Problem: Reliably characterizing the spatial variability, especially anisotropy and non-stationary variation, of soil parameters (OCR, Su, Es) in new and old subgrade foundations to address the challenge of uneven settlement in infrastructure expansion projects, where conventional methods are limited.
Key Innovation: Proposes a soil spatial variation quantization method combining two-dimensional Bayesian compressed sensing (BCS) and discrete cosine transform (DCT) basis functions, applied to CPTU data. This framework efficiently reconstructs and analyzes parameter fields, characterizes autocorrelation, and provides high-precision support for foundation performance evaluation and differential settlement control.
42. Towards global mapping of dynamic surface water extents using Sentinel-1 SAR data
Core Problem: The need for a fully automated, scalable, and accurate method for mapping dynamic surface water extents from single-acquisition Sentinel-1 SAR imagery to support near-real-time monitoring of floods, droughts, and water-resource management.
Key Innovation: A novel, fully automated, and scalable method integrating adaptive thresholding, fuzzy-logic classification, region growing, dark land estimation, and a bimodality test to map surface water extents from Sentinel-1 SAR data with high accuracy (exceeding 85%), demonstrating robust performance across diverse environments and for monitoring floods and droughts.
43. Competitive equilibrium between carbon loss and sequestration driven by erosion: Stratified responses of microbial metabolism and mineral protection
Core Problem: The underlying mechanisms by which soil erosion shapes carbon source/sink patterns on slopes remain inadequately understood, particularly the synergistic mechanisms between physical protection and biological metabolic processes beyond just physical translocation.
Key Innovation: This study reveals that the redistribution of particulate organic carbon (POC) and mineral-associated organic carbon (MAOC) across eroded landscapes, coupled with microbial metabolic adaptation, governs carbon source/sink dynamics along slopes, emphasizing the need to quantify both physical translocation losses and mineral-biological synergy-driven sequestration.
44. A new matrix propagation method for theoretical analysis of wave propagation in low-temperature gradient layered rock mass with parallel ice-filled joints
Core Problem: The mechanisms of stress wave propagation through layered rock masses containing ice-filled joints in high-altitude cold regions, particularly during blasting operations, are poorly understood due to distinct stratified ground temperature distributions.
Key Innovation: A novel theoretical model combining the matrix propagation method (MPM) with a nonlinear displacement discontinuity model (DDM) is developed to analyze stress wave propagation in layered rock with ice-filled joints. It reveals that low-temperature gradients and ice-filled joints inhibit wave propagation, but multiple reflections enhance total transmitted wave energy, offering theoretical guidance for blasting operations in cold regions.
45. Strength of a cement-improved clay and a machine learning evaluation of interactions between observed mechanical behaviours
Core Problem: Gaining a thorough understanding of the improvement process in cement-improved soils, specifically how various factors (cement content, density, saturation, confining pressure) interact to influence mechanical properties.
Key Innovation: Uses machine learning (correlation analysis, PCA, tree-based regression) to systematically investigate cement-improved clay, revealing that specimen density critically governs strength properties in addition to cement content, and demonstrating the superior modeling capability of ensemble tree-based algorithms for soil improvement.
46. Physics-informed geological characterization for enhanced cross-domain generalization of tunneling-induced surface settlement prediction
Core Problem: Data-driven models for tunneling-induced settlement prediction often suffer from poor cross-domain generalization, limiting their applicability in diverse geological settings.
Key Innovation: Established a Physics-Informed Geological Characterization (PIGC) framework by integrating mechanism-based parameters (unloading–reloading modulus, Poisson’s ratio) and a normalized weighting strategy for multi-layered strata, implemented with a Bayesian-optimized Bidirectional LSTM and custom WMSE loss function, significantly enhancing cross-domain generalization and accuracy for large-magnitude settlements.
47. Multiscale characterization of pore damage evolution and mechanical response of rocks: Real-time computed tomography scanning and discrete element simulation
Core Problem: Understanding rock damage evolution, especially multiscale behaviors in sedimentary rocks like sandstone and mudstone, is essential for deep resource extraction and geological hazard prevention.
Key Innovation: Investigation of rock damage evolution under uniaxial loading using in situ real-time CT scanning and mineral-based discrete element simulations (CT-GBM), deriving new parameters (Kpec, Kpnsc, Dk, Dt) to evaluate mesoscopic damage, and offering methodological insights for predicting rock mass safety.
48. Morphological instability in restored intertidal flats: How anthropogenic structures drive early‐stage evolution
Core Problem: The early-stage evolution and morphological instability of restored intertidal flats, particularly how anthropogenic structures influence them, are poorly understood.
Key Innovation: This study examines sediment deposition, vertical accretion, and morphological evolution in three human-induced intertidal flats using sediment traps and digital elevation models, revealing rapid accretion in unusual locations and high erosion of artificial structures, leading to self-cannibalization.
49. Modeling Transient Groundwater Flow in Unconfined Aquifers Under Dynamic Conditions Using Physics‐Informed Neural Networks
Core Problem: The challenge of training Physics-Informed Neural Networks (PINNs) for unconfined aquifers and under dynamic conditions due to nonlinearity and the large number of required collocation points, limiting their applicability beyond steady-state confined aquifers.
Key Innovation: Adapting and comparing three PINNs approaches (continuous time, discrete time, and time decomposition) for modeling transient groundwater flow in unconfined aquifers, demonstrating the superior accuracy and training efficiency of the discrete-time approach, which can be 10 times more efficient than standard PINNs.
50. Deep Accurate Solver for the Geodesic Problem
Core Problem: Traditional methods for computing geodesic distances on continuous surfaces using discretized polygonal meshes are limited to at most second-order accuracy, and existing local solvers are not sufficiently accurate.
Key Innovation: A higher-order accurate deep learning method for computing geodesic distances on surfaces, which uses a neural network-based local solver to implicitly approximate the continuous surface structure, achieving third-order accuracy and providing a bootstrapping recipe for further improvement.
51. AviaSafe: A Physics-Informed Data-Driven Model for Aviation Safety-Critical Cloud Forecasts
Core Problem: Current AI weather forecasting models predict conventional atmospheric variables but cannot distinguish between cloud microphysical species critical for aviation safety, which are characterized by extreme sparsity, discontinuous distributions, and complex interactions.
Key Innovation: AviaSafe, a hierarchical, physics-informed neural forecaster, produces global, six-hourly predictions of four hydrometeor species for up to 7 days, integrating the Icing Condition (IC) index as a physics-based constraint and outperforming operational numerical models for certain variables.
52. AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction
Core Problem: Existing 4D scene reconstruction methods struggle with monocular aerial conditions, leading to severe depth ambiguity and unstable motion estimation for dynamic objects, limiting their application in complex aerial environments.
Key Innovation: Introduces AeroDGS, a physics-guided 4D Gaussian splatting framework for monocular UAV videos, which uses a Monocular Geometry Lifting module and a Physics-Guided Optimization module to achieve superior and physically consistent reconstruction fidelity in dynamic aerial environments.
53. GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views
Core Problem: Existing feed-forward 3D reconstruction methods from sparse views struggle with out-of-domain data, lack inference-time refinement, and become slow when integrating generative priors due to their one-shot prediction paradigm.
Key Innovation: GIFSplat, a purely feed-forward iterative refinement framework for 3D Gaussian Splatting from sparse unposed views, which uses forward-only residual updates and distills a frozen diffusion prior into Gaussian-level cues to achieve high quality and efficiency without test-time gradient optimization.
54. Spectrally Distilled Representations Aligned with Instruction-Augmented LLMs for Satellite Imagery
Core Problem: The adoption of Vision-language foundation models (VLFMs) for satellite imagery is hindered by the difficulty of consistently exploiting multi-spectral inputs (due to redundancy/misalignment) and the limited semantic expressiveness of CLIP-style text encoders.
Key Innovation: SATtxt, a spectrum-aware VLFM that operates with RGB inputs only at inference while retaining spectral cues learned during training, achieved through Spectral Representation Distillation and Spectrally Grounded Alignment with Instruction-Augmented LLMs, improving zero-shot classification, retrieval, and linear probing on Earth observation datasets.
55. UFO-DETR: Frequency-Guided End-to-End Detector for UAV Tiny Objects
Core Problem: Small target detection in UAV imagery faces significant challenges such as scale variations, dense distribution, and the dominance of small targets, making it difficult for existing algorithms to balance accuracy and complexity.
Key Innovation: Proposed UFO-DETR, an end-to-end object detection framework for UAV tiny objects, integrating an LSKNet-based backbone, multi-scale spatial relationship modeling, and a DynFreq-C3 module for frequency feature enhancement, achieving superior performance and efficiency.
56. GSTurb: Gaussian Splatting for Atmospheric Turbulence Mitigation
Core Problem: Atmospheric turbulence causes significant image degradation (pixel displacement and blur) in long-range imaging, hindering effective remote sensing and monitoring applications.
Key Innovation: Proposes GSTurb, a novel framework that integrates optical flow-guided tilt correction and Gaussian splatting for modeling non-isoplanatic blur, optimizing Gaussian parameters across multiple frames to enhance image restoration under both synthetic and real-world turbulence conditions.
57. Scaling Laws of Global Weather Models
Core Problem: Optimizing training efficiency and model performance in data-driven weather forecasting requires understanding the empirical scaling laws between model performance, model size, dataset size, and compute budget.
Key Innovation: Empirical scaling laws were analyzed, revealing that Aurora exhibits strong data-scaling behavior, GraphCast has high parameter efficiency, and compute-optimal analysis suggests prioritizing longer training durations over increasing model size. Weather models consistently favor increased width over depth, guiding future model architecture design.
58. Learning Physical Operators using Neural Operators
Core Problem: Neural operators, while promising for solving partial differential equations (PDEs), struggle to generalize beyond training distributions and are often constrained to a fixed temporal discretisation, limiting their applicability to novel physical regimes.
Key Innovation: Introduces a physics-informed training framework that decomposes PDEs using operator splitting methods, training separate neural operators to learn individual non-linear physical operators while approximating linear operators with fixed finite-difference convolutions. This modular architecture, formulated as a neural ordinary differential equation (ODE), enables generalization to unseen physics, continuous-in-time predictions, and temporal extrapolation.
59. Partial recovery of meter-scale surface weather
Core Problem: Current weather analyses and forecasts lack meter-scale variability in near-surface atmospheric conditions, despite its importance due to land cover and topography, making it unclear if such variability is predictable.
Key Innovation: A computationally feasible approach to infer spatially continuous fields of near-surface wind, temperature, and humidity at 10m resolution across the contiguous United States, by conditioning coarse atmospheric state on sparse surface station measurements and high-resolution Earth observation data, demonstrating the recoverability of meter-scale weather variability.
60. UniScale: Unified Scale-Aware 3D Reconstruction for Multi-View Understanding via Prior Injection for Robotic Perception
Core Problem: Accurate extraction of environmental structure from raw image sequences, including metric scale, is critical for vision-based robotic navigation, but existing methods often lack unified scale-awareness and flexible prior integration.
Key Innovation: UniScale, a unified, scale-aware multi-view 3D reconstruction framework that jointly estimates camera parameters, scale-invariant depth/point maps, and metric scene scale from multi-view images, flexibly integrating geometric priors for robust, metric-aware 3D reconstruction in robotic applications.
61. LineGraph2Road: Structural Graph Reasoning on Line Graphs for Road Network Extraction
Core Problem: Existing methods for automatic road extraction from satellite imagery struggle to capture long-range dependencies and complex topologies, limiting accuracy for applications like navigation and urban planning.
Key Innovation: Proposes LineGraph2Road, a framework that formulates connectedness prediction as binary classification on a global sparse Euclidean graph. It transforms the graph into its line graph and applies a Graph Transformer for structural link representation and relational reasoning, achieving state-of-the-art results on road network extraction benchmarks and resolving multi-level crossings.
62. VGG-T$^3$: Offline Feed-Forward 3D Reconstruction at Scale
Core Problem: Offline feed-forward 3D reconstruction methods suffer from computational and memory requirements that grow quadratically with the number of input images, due to the varying-length Key-Value space representation of scene geometry.
Key Innovation: VGG-T$^3$ (Visual Geometry Grounded Test Time Training) scales linearly with input views by distilling the varying-length KV space into a fixed-size Multi-Layer Perceptron (MLP) via test-time training, achieving significant speed-up and improved point map reconstruction error for 3D reconstruction.
63. Sapling-NeRF: Geo-Localised Sapling Reconstruction in Forests for Ecological Monitoring
Core Problem: Existing 3D sensing methods struggle to accurately capture fine-scale architectural traits of saplings in forests, particularly thin branches and dense foliage, and lack the scale consistency and geo-localisation needed for repeatable, long-term ecological monitoring.
Key Innovation: Sapling-NeRF is a pipeline fusing NeRF, LiDAR SLAM, and GNSS to enable repeatable, geo-localised, and dense 3D reconstruction of individual saplings in forests, allowing for accurate measurement of traits like stem height, branching patterns, and leaf-to-wood ratios for ecological monitoring.
64. IV-tuning: Parameter-Efficient Transfer Learning for Infrared-Visible Tasks
Core Problem: Full fine-tuning of Pre-trained Visual Models (PVMs) for infrared and visible (IR-VIS) tasks leads to highly constrained and low-ranked feature spaces, which seriously impairs generalization.
Key Innovation: Proposes IV-tuning, a parameter-efficient transfer learning method that harnesses PVMs for various IR-VIS downstream tasks (e.g., salient object detection, semantic segmentation, object detection) by training only 3% of backbone parameters, preserving pretrained knowledge and achieving superior generalization, scalability, and computational efficiency.
65. Online time series prediction using feature adjustment
Core Problem: Online time series forecasting faces significant challenges due to distribution shift and delayed feedback in multi-step predictions, as current methods primarily focus on parameter selection rather than adapting to changes in underlying latent factors.
Key Innovation: ADAPT-Z (Automatic Delta Adjustment via Persistent Tracking in Z-space), a novel approach that leverages an adapter module and historical gradient information to update feature representations of latent factors, enabling robust parameter updates despite delayed feedback and outperforming state-of-the-art online learning methods.
66. G4Splat: Geometry-Guided Gaussian Splatting with Generative Prior
Core Problem: Existing methods for 3D scene reconstruction using generative priors struggle to produce high-quality reconstructions, especially in unobserved regions, due to a lack of reliable geometric supervision and ineffective mitigation of multi-view inconsistencies.
Key Innovation: Proposes G4Splat, which leverages planar structures to derive accurate metric-scale depth maps for reliable geometric supervision and integrates this guidance throughout the generative pipeline to improve visibility mask estimation, novel view selection, and multi-view consistency, resulting in accurate 3D scene completion for both observed and unobserved regions in indoor and outdoor scenarios.
67. Learning with less: label-efficient land cover classification at very high spatial resolution using self-supervised deep learning
Core Problem: Deep learning for very high-resolution land cover classification requires large volumes of representative training data, posing a significant barrier to widespread adoption for large-area mapping.
Key Innovation: Presents a novel label-efficient approach for statewide 1-m land cover classification using self-supervised deep learning (BYOL pre-training on unlabeled data) and fine-tuning with only 1,000 annotated patches, achieving high accuracy and addressing the data annotation bottleneck.
68. PCReg-Net: Progressive Contrast-Guided Registration for Cross-Domain Image Alignment
Core Problem: Deformable image registration across heterogeneous domains is challenging due to coupled appearance variation and geometric misalignment, which violates the brightness constancy assumption of conventional methods.
Key Innovation: PCReg-Net, a progressive contrast-guided registration framework that performs coarse-to-fine alignment using four lightweight modules: an initial registration U-Net, a reference feature extractor, a multi-scale contrast module to identify residual misalignment, and a refinement U-Net with feature injection, achieving high-fidelity alignment across diverse domains.
69. DDSAM-CD: Dense Feature Fusion and Difference Enhancement for the Segment Anything Model in Remote Sensing Change Detection
Core Problem: Current remote sensing change detection (RSCD) methods, even those utilizing vision foundation models (VFMs) like FastSAM, do not adequately explore and utilize the extracted multiscale features, leading to limitations in capturing temporal change information and achieving high edge-region accuracy.
Key Innovation: DDSAM-CD, a model that enhances RSCD by employing a dense feature aggregation network (DFAN) to comprehensively fuse multiscale features from FastSAM and a differential enhancement network (DEN) to capture temporal change information and significantly improve edge-region accuracy, outperforming state-of-the-art methods.
70. The Response and Attribution of Vegetation Productivity to Drought Sensitivity in Northwest China
Core Problem: Long-term patterns and mechanisms of vegetation drought sensitivity (Sdro) in arid-semiarid Northwest China remain unclear, despite drought's significant impact on vegetation productivity and expected worsening with climate change.
Key Innovation: Quantification of spatiotemporal drought trends and Sdro across major vegetation types using satellite GPP and soil-moisture records, and attribution of Sdro changes using a hierarchical segmentation mixed-effects model, clarifying where and why vegetation is becoming more drought-sensitive and identifying dominant controls like VPD and precipitation.
71. The Antarctic coastal ocean heat budget is dominated by heat loss to land ice melt
Core Problem: Ocean models often neglect the significant heat exchange required to melt ice shelves and icebergs, leading to biases in understanding the Antarctic coastal ocean heat budget and its impact on climate dynamics.
Key Innovation: Simulations reveal that heat loss to land ice melt (ice shelves and calved icebergs) constitutes the largest ocean heat sink on the Antarctic continental shelf, accounting for 60% of heat supplied across the shelf break. Omitting this sink drives enhanced heat loss to the atmosphere and nonlocal reductions in heat supply via sea ice–mediated stratification changes.
72. Spatiotemporal assessment of extreme rainfall events in the Wainganga river basin using CMIP6 climate models
Core Problem: Climate change is altering extreme rainfall patterns, leading to potential shifts in drought and flood hotspots, but a comprehensive spatiotemporal assessment of these changes using multi-model ensembles is needed for specific river basins like the Wainganga to inform sustainable planning.
Key Innovation: The study uses a multi-model ensemble of six CMIP6 Global Climate Models and fourteen rainfall-based extreme indices to conduct a spatiotemporal analysis of extreme rainfall events in the Wainganga River basin, identifying future hotspots for drought (western region) and flood (eastern ridge) and projecting a dominance of short-duration rainfall, providing crucial data for decision-makers.
73. Karst record of Holocene climate and human-induced changes in surface processes in the northern Apennines of Italy
Core Problem: Understanding the long-term interplay of climate and human activity on surface processes and their impact on sensitive Earth Critical Zone components like karst systems.
Key Innovation: Examined the Holocene sedimentary archive in Tana della Mussina Cave, revealing alternating phases of clastic deposition and speleothem precipitation linked to hydrological changes, and providing evidence of early human-induced soil erosion (deforestation, slash-and-burn) significantly increasing sediment influx into the karst system.
74. Automated detection of multiple surface defects in highway tunnel linings using lightweight deep learning model on embedded devices
Core Problem: Deploying heavy deep learning models for real-time, low-latency detection of tunnel lining defects (cracks, spalling, water leakage) on constrained embedded hardware remains challenging.
Key Innovation: Proposed an embedded deployment pipeline using a lightweight deep learning model with a feature-adaptive weighted-fusion (FAWF) mechanism and information-sharing backbone (ISB), combined with a stepwise compression strategy, achieving real-time (30+ FPS) and accurate defect detection on mobile robots.
75. A progressive failure approach for toughness evaluation of shield tunnel gasketed joints waterproofness: combined full-scale experimental and fluid-solid coupling simulation
Core Problem: Assessments of shield tunnel joint waterproofing performance have largely focused on ultimate load-bearing capacity, neglecting toughness variations and underlying physical mechanisms of the failure process.
Key Innovation: Developed a progressive failure approach combining full-scale experiments and fluid-solid coupling simulations to evaluate the toughness of shield tunnel gasketed joints, introducing multi-dimensional indicators (stress synergy coefficient, contact stress reserve ratio) and a progressive failure criterion to understand the dynamic leakage process and energy absorption capacity.
76. Formation processes and evolution model of high-fluoride groundwater in a granitic basin under complex geological conditions
Core Problem: The coupled influences of lithology, geothermal activity, and groundwater flow regimes on fluoride generation and evolution in granitic terrains are insufficiently constrained.
Key Innovation: Developed an integrated analytical framework (hydrogeochemistry, rock geochemistry, stable isotopes, leaching experiments) to elucidate fluoride enrichment mechanisms, identifying stage-dependent trajectories and the joint regulation by lithological variability, geothermal effects, and deep groundwater circulation, leading to a conceptual evolutionary model.
77. Reconstructing chlorinated solvent release histories in the vadose zone for groundwater pollution assessment using tree-ring dendrochemistry
Core Problem: Conventional monitoring networks provide concentration data but seldom reconstruct historical release events of chlorinated solvents in the vadose zone, complicating source attribution and remediation.
Key Innovation: Applied tree-ring dendrochemistry (EDXRF for Cl, δ13C) to reconstruct CCl4 release history at a legacy site, demonstrating that tree-ring Cl signals correlate with soil-gas concentrations and can identify the timing of major release events, offering a minimally invasive assessment tool.
78. Glacier Retreat Amplifies Interannual Variability in Watershed Runoff, Organic Carbon and Nutrient Yields
Core Problem: The hypothesis that future decreases in glacier runoff will reduce the stability of hydro-biogeochemical export, and how glacier decline affects interannual variability in watershed runoff and yields, needs testing.
Key Innovation: Using a decade of data from four Alaskan watersheds, this study demonstrates that glacier decline leads to lower and more variable runoff (2.6–4.3 times greater interannual variability) and increased stochasticity in biogeochemical export (up to five-fold increase in variability).
79. Spatiotemporal Dynamics of Streamflow Drought in the Larger Alpine Region
Core Problem: The difficulty in studying spatiotemporal drought evolution in detail due to many studies spatially limiting their drought analysis to specific countries or catchments.
Key Innovation: Analyzing the spatiotemporal dynamics of streamflow droughts over the larger Alpine region using spatially-explicit, high-resolution simulations from PCR-GLOBWB2.0 and a novel spatial and temporal clustering algorithm, revealing growth/recovery phases, spatially distinct sub-events, regional differences, and the interplay of multiple hydrometeorological drivers.
80. SwiftNDC: Fast Neural Depth Correction for High-Fidelity 3D Reconstruction
Core Problem: Existing depth-guided 3D reconstruction methods suffer from scale drift, multi-view inconsistencies, and require substantial refinement to achieve high-fidelity geometry.
Key Innovation: SwiftNDC, a fast and general framework built around a Neural Depth Correction field that produces cross-view consistent depth maps, generating a robust dense point cloud for accelerated and high-quality 3D Gaussian Splatting for mesh reconstruction and novel-view synthesis.
81. Quality-Aware Robust Multi-View Clustering for Heterogeneous Observation Noise
Core Problem: Deep multi-view clustering is vulnerable to complex, heterogeneous observation noise where contamination intensity varies continuously across data, a problem overlooked by existing binary noise assumptions.
Key Innovation: QARMVC, a Quality-Aware Robust Multi-View Clustering framework that uses an information bottleneck to quantify instance-level contamination intensity, integrating these quality scores into a hierarchical learning strategy for noise-robust feature learning and high-quality global consensus.
82. QuadSync: Quadrifocal Tensor Synchronization via Tucker Decomposition
Core Problem: Quadrifocal tensors, despite capturing more information than pairwise counterparts, have been considered impractical for multi-camera recovery in structure from motion (SfM).
Key Innovation: Introduces QuadSync, the first synchronization algorithm for quadrifocal tensors, using Tucker decomposition to recover 'n' cameras, and also establishes relationships and a joint synchronization algorithm for bifocal, trifocal, and quadrifocal tensors, demonstrating its effectiveness in SfM.
83. Plug, Play, and Fortify: A Low-Cost Module for Robust Multimodal Image Understanding Models
Core Problem: Missing modalities present a fundamental challenge in multimodal models, often causing catastrophic performance degradation due to an imbalanced learning process where the model develops an implicit preference for certain modalities.
Key Innovation: Proposes a Multimodal Weight Allocation Module (MWAM), a plug-and-play component that uses a Frequency Ratio Metric (FRM) to quantify modality preference and dynamically re-balances modality contributions during training, enhancing robustness and performance across diverse multimodal tasks, especially with missing data.
84. SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs
Core Problem: Traditional positional embeddings (e.g., RoPE) in 3D Large Vision-Language Models (LVLMs) are suboptimal for 3D multimodal understanding, failing to preserve 3D spatial structures and overlooking angular dependencies in point-cloud data.
Key Innovation: Introduced Spherical Coordinate-based Positional Embedding (SoPE), which maps point-cloud token indices into a 3D spherical coordinate space to unify spatial locations and directional angles, enhancing spatial awareness and geometric representation for multimodal learning.
85. Set-based v.s. Distribution-based Representations of Epistemic Uncertainty: A Comparative Study
Core Problem: The relative merits of set-based and distribution-based representations for epistemic uncertainty in neural networks are unclear due to differing semantics, assumptions, and evaluation practices, hindering principled comparison.
Key Innovation: A controlled comparative study enabling principled, like-for-like evaluation of set-based and distribution-based uncertainty representations, constructed from the same predictive distributions, providing insights into how representation choices impact practical uncertainty-aware performance.
86. Hypernetwork-based approach for grid-independent functional data clustering
Core Problem: Most existing functional data clustering methods implicitly operate on sampled grids, causing cluster assignments to depend on resolution, sampling density, or preprocessing choices rather than on the underlying functions themselves.
Key Innovation: Introduces a hypernetwork-based framework that maps discretized function observations (at arbitrary resolution and grids) into a fixed-dimensional vector space via an auto-encoding architecture with a hypernetwork encoder and an implicit neural representation (INR) decoder, enabling grid-independent and robust functional data clustering.
87. SO3UFormer: Learning Intrinsic Spherical Features for Rotation-Robust Panoramic Segmentation
Core Problem: Standard spherical Transformers for panoramic semantic segmentation overfit global latitude cues and fail catastrophically under 3D reorientations caused by unconstrained camera motions, leading to performance collapse.
Key Innovation: SO3UFormer, a rotation-robust architecture that learns intrinsic spherical features by decoupling representation from the gravity vector, using quadrature-consistent spherical attention, and a gauge-aware relative positional mechanism, demonstrating remarkable stability and superior mIoU under full SO(3) rotations compared to existing SOTAs.
88. OSDaR-AR: Enhancing Railway Perception Datasets via Multi-modal Augmented Reality
Core Problem: Railway applications suffer from a scarcity of high-quality, annotated data for safety-critical tasks like obstacle detection, and existing simulation or image-masking techniques have limitations (sim-to-real gap, lack of spatio-temporal coherence).
Key Innovation: A multi-modal augmented reality framework that integrates photorealistic virtual objects into real-world railway sequences using Unreal Engine 5, LiDAR, and INS/GNSS data, creating the OSDaR-AR dataset for railway perception.
89. SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Core Problem: Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category, and existing few-shot methods are complex, relying on memory banks or auxiliary datasets.
Key Innovation: SubspaceAD, a training-free few-shot anomaly detection method, extracts patch-level features from a frozen DINOv2 backbone and fits a Principal Component Analysis (PCA) model to estimate the low-dimensional subspace of normal variations. It achieves state-of-the-art performance on MVTec-AD and VisA datasets without training, prompt tuning, or memory banks.
90. DMAligner: Enhancing Image Alignment via Diffusion Model Based View Synthesis
Core Problem: Existing optical flow-based image alignment methods are susceptible to occlusions and illumination variations, leading to degraded visual quality and compromised accuracy in downstream tasks.
Key Innovation: DMAligner, a diffusion-based framework, enhances image alignment through alignment-oriented view synthesis. It employs a Dynamics-aware Diffusion Training approach with a Dynamics-aware Mask Producing (DMP) module to adaptively distinguish dynamic foregrounds from static backgrounds, outperforming classical methods. A new DSIA dataset was also developed.
91. Small Object Detection Model with Spatial Laplacian Pyramid Attention and Multi-Scale Features Enhancement in Aerial Images
Core Problem: Detecting small, dense, and non-uniformly distributed objects in high-resolution aerial images is challenging and inefficient.
Key Innovation: A small object detection algorithm based on a Spatial Laplacian Pyramid Attention (SLPA) module, a Multi-Scale Feature Enhancement Module (MSFEM), and deformable convolutions for feature alignment in Feature Pyramid Networks (FPN), improving detection in aerial images.
92. D-FINE-seg: Object Detection and Instance Segmentation Framework with multi-backend deployment
Core Problem: Real-time instance segmentation with transformers is less common, and existing object detectors like D-FINE need extension for this task while maintaining competitive latency and accuracy.
Key Innovation: D-FINE-seg, an instance segmentation extension of D-FINE, which adds a lightweight mask head, segmentation-aware training (including box cropped BCE and dice mask losses, auxiliary and denoising mask supervision, and adapted Hungarian matching cost), and an end-to-end pipeline for multi-backend deployment.
93. Align then Adapt: Rethinking Parameter-Efficient Transfer Learning in 4D Perception
Core Problem: Transferring 3D pre-trained models to 4D perception tasks (point cloud video understanding) is hampered by overfitting and the modality gap, due to the scarcity of 4D datasets.
Key Innovation: PointATA, an 'Align then Adapt' paradigm that decomposes transfer learning into two stages: (1) a point align embedder trained with optimal-transport theory to alleviate the modality gap, and (2) an efficient point-video adapter and spatial-context encoder to enhance temporal modeling and mitigate overfitting.
94. No Labels, No Look-Ahead: Unsupervised Online Video Stabilization with Classical Priors
Core Problem: Deep learning-based video stabilization methods require paired stable/unstable datasets, lack controllability, and are inefficient on constrained hardware, limiting their applicability to domains like UAV nighttime remote sensing.
Key Innovation: An unsupervised online video stabilization framework that instantiates a classical stabilization pipeline with a multithreaded buffering mechanism, addressing data, controllability, and efficiency issues, and introducing a new multimodal UAV aerial video dataset (UAV-Test).
95. Efficient Encoder-Free Fourier-based 3D Large Multimodal Model
Core Problem: Designing efficient 3D Large Multimodal Models (LMMs) without cumbersome visual encoders is challenging due to the unordered and large-scale nature of point clouds, limiting scalability and efficiency.
Key Innovation: Fase3D, the first efficient encoder-free Fourier-based 3D scene LMM, which uses structured superpoints, space-filling curve serialization with Fast Fourier Transform (FFT) for global context modeling, and Fourier-augmented LoRA adapters to process unordered 3D data effectively and efficiently.
96. Efficient Real-Time Adaptation of ROMs for Unsteady Flows Using Data Assimilation
Core Problem: Adapting parameterized Reduced Order Models (ROMs) to out-of-sample parameter regions for unsteady flows typically requires full retraining, which is computationally expensive, and existing methods struggle with efficient real-time adaptation using sparse observations.
Key Innovation: Proposes an efficient retraining strategy for ROMs using a VAE and transformer network, adapted to out-of-sample parameter regions with sparse data via ensemble Kalman filtering, showing that retraining can be limited to the autoencoder for lightweight, real-time adaptation.
97. Multidimensional Task Learning: A Unified Tensor Framework for Computer Vision Tasks
Core Problem: Current computer vision task formulations are constrained by matrix-based thinking, requiring structural flattening that limits the expressiveness of tasks and hinders principled configurations for spatiotemporal or cross-modal predictions.
Key Innovation: Multidimensional Task Learning (MTL), a unified mathematical framework based on Generalized Einstein MLPs (GE-MLPs) that operates directly on tensors, enabling explicit control over dimensions and unifying classification, segmentation, and detection as special cases within a strictly larger task space.
98. Bridging Geometric and Semantic Foundation Models for Generalized Monocular Depth Estimation
Core Problem: Monocular Depth Estimation (MDE) in complex scenes requires effectively fusing geometric and semantic information, and existing methods may struggle with generalization or resource demands.
Key Innovation: BriGeS, a method that fuses geometric and semantic information from pre-trained foundation models using a 'Bridging Gate' and 'Attention Temperature Scaling' to enhance MDE. It achieves state-of-the-art performance by training only the Bridging Gate, reducing resource demands and improving generalization.
99. Fast and Flexible Probabilistic Forecasting of Dynamical Systems using Flow Matching and Physical Perturbation
Core Problem: Standard physics-based ensemble forecasting often yields unphysical initial states due to Gaussian/uniform perturbations, and existing machine learning approaches (diffusion models) for learning dynamical systems are computationally expensive for probabilistic forecasting from incomplete/noisy data.
Key Innovation: A novel framework that decouples perturbation generation from propagation, using flow matching to learn physically consistent initial condition perturbations and efficient ODE integrators for ensemble propagation, achieving faster and more accurate probabilistic forecasting, validated on high-dimensional WeatherBench data.
100. Loc$^2$: Interpretable Cross-View Localization via Depth-Lifted Local Feature Matching
Core Problem: Accurate and interpretable fine-grained cross-view localization (ground-to-aerial) is challenging, with prior methods relying on global descriptors or bird's-eye-view transformations, lacking direct image-plane correspondences and interpretability.
Key Innovation: Loc$^2$, a lightweight, end-to-end trainable method that directly learns ground-aerial image-plane correspondences using weak supervision, lifts matched ground points into BEV space with monocular depth predictions, and applies scale-aware Procrustes alignment for 3 DoF pose estimation, offering state-of-the-art accuracy and strong interpretability.
101. Unsupervised Posterior Sampling for Seismic Data Recovery via Score-Based Generative Priors
Core Problem: Traditional seismic data restoration methods struggle with complex, unknown degradations, lack generalization to unseen field data, and often require retraining for each degradation type.
Key Innovation: Introduces an unsupervised Posterior Sampling Framework (PSF) built upon Score-based Generative Models (SGMs) as a seismic-aware generative prior. This enables posterior sampling across different seismic restoration tasks without retraining and incorporates an adaptive noise-level estimation for flexibility under varying conditions.
102. HLGFA: High-Low Resolution Guided Feature Alignment for Unsupervised Anomaly Detection
Core Problem: Unsupervised industrial anomaly detection (UAD) is challenging due to scarce defect samples and the need for reliable detection, with existing methods often relying on pixel-level reconstruction.
Key Innovation: HLGFA, a high-low resolution guided feature alignment framework that learns normality by modeling cross-resolution feature consistency between high-resolution and low-resolution representations, identifying anomalies where this alignment breaks down, and incorporating a noise-aware data augmentation strategy.
103. Calculation method for groundwater seepage after tunnel construction
Core Problem: Assessing the impact of tunnel construction on groundwater seepage and predicting its behavior under various influencing factors.
Key Innovation: Establishment of a quantitative formula to predict groundwater seepage under the combined influence of pressure differences, stratum permeability, and tunnel water-blocking coefficients, validated by experimental and numerical models.
104. Soil reaction to oblique relative movement of steel pipes buried in sand
Core Problem: Accurately predicting the reaction force on buried steel pipes subjected to oblique ground movement, especially considering kinematic constraints and various soil/pipe parameters.
Key Innovation: Development of a new simplified expression, based on discrete element method (DEM) simulations, for estimating the peak reaction on rigid steel pipes during oblique downwards ground movement, improving upon existing formulas.
105. SF-Net: Spatial-Frequency Feature Synthesis for Semantic Segmentation of High-Resolution Remote Sensing Imagery
Core Problem: Most conventional deep learning approaches for semantic segmentation of high-resolution remote sensing (HRRS) images primarily focus on the spatial domain, neglecting the rich textural and structural nuances found in the frequency domain, which limits comprehensive data representation.
Key Innovation: SF-Net, a spatial-frequency feature synthesis network that seamlessly integrates features across spatial and frequency domains using a multiscale convolutional grouping fusion module, Haar wavelet transform, and Mamba-enhanced global spatial feature extraction, achieving state-of-the-art performance in HRRS semantic segmentation.
106. TransGCF: A Unified Spatial–Spectral–Frequency Framework for Robust Hyperspectral Anomaly Detection
Core Problem: Prevalent reconstruction-based methods for hyperspectral anomaly detection (HAD) suffer from high false-alarm rates because they fail to adequately model complex background variations or inadvertently reconstruct high-frequency anomaly signatures, due to an inability to concurrently address distinct spatial, spectral, and frequency characteristics.
Key Innovation: TransGCF, a unified spatial–spectral–frequency framework for HAD that explicitly decouples feature extraction into specialized branches (Local Transformer for spatial context, graph convolutional network for spectral correlations, and a high-frequency elimination block for anomaly suppression), integrating these features to achieve state-of-the-art detection accuracy and reduced false alarms.
107. Frequency-Driven Visual State Space Model With Cross-Scale Spatial Aggregation for Hyperspectral and LiDAR Classification
Core Problem: Existing pixel-level land cover classification (LCC) methods using integrated hyperspectral imagery (HSI) and LiDAR data primarily operate in the spatial domain, neglecting latent correlations in the frequency domain and struggling to model complex dependencies in heterogeneous data.
Key Innovation: The Frequency-Driven State Space Model with Cross-Scale Spatial Aggregation (FSSM-CSA), which integrates frequency-guided hierarchical feature collaboration, cross-scale spatial aggregation, and frequency-spatial adaptive fusion to effectively model both spatial and frequency information for enhanced LCC accuracy.
108. PCA-Aware Attention Feature Fusion With Complex-Valued Adaptive Weighted UNet for PolSAR Classification
Core Problem: Effectively exploiting both the amplitude and phase information embedded in Polarimetric Synthetic Aperture Radar (PolSAR) imagery for robust land cover classification remains a challenge.
Key Innovation: A land classification method that integrates complex-valued modeling with feature enhancement strategies, including a PCA-based attention-aware feature selection module and an improved complex-valued adaptive weighted U-Net, to enhance semantic representation and feature discriminability for superior PolSAR classification.
109. Advanced Hyperspectral Scene Interpretation Through PCA-Enhanced Residual Learning in Complex Spectral Domains
Core Problem: Hyperspectral Imaging (HSI) classification is limited by large data dimensions, spectral redundancy, and a lack of labeled data, hindering accurate and precise land cover analysis.
Key Innovation: Hyperspectral ResNetLite, a 1-D residual convolutional neural network (CNN) that employs Principal Component Analysis (PCA) for spectral redundancy minimization and a 1-D CNN for refining feature representations, resulting in a computationally effective, accurate, and interpretable method for HSI classification.
110. Tri-CoMamba: A Tri-Complementary Mamba Framework for Multisource Remote Sensing Image Classification
Core Problem: Multisource remote sensing joint classification (e.g., HSI, LiDAR, SAR) is limited by challenges in long-range dependency modeling, cross-modal alignment, and the preservation of fine-grained spectral features.
Key Innovation: The Tri-CoMamba framework, built on the Mamba-S6 state-space model, integrates three complementary modules (CoRe-Mamba, CF-SpecMamba, and MASM) to mitigate feature mismatching, enhance spectral features, and optimize cross-modal fusion, leading to precise and computationally efficient classification of multisource data.
111. Open-access energy demand data for South and Southeast Asia
Core Problem: Fragmented, inconsistent, and difficult-to-access open-access electricity demand data in South and Southeast Asia, limiting meteorology–energy and climate–energy research, forecasting, and resilience planning, especially given the region's sensitivity to natural hazards.
Key Innovation: Presents a harmonized, open-access daily national electricity demand dataset for twelve South and Southeast Asian countries (2013–2025), compiled from diverse sources with reproducible workflows, enabling regional analysis of demand seasonality, weather–demand sensitivity, and event-based studies of disruption and recovery during extremes.
112. Addressing current and future challenges in the engineering geology community - A young engineering geologist’s perspective
Core Problem: The field of engineering geology is rapidly evolving due to global pressures (urbanization, climate change, resource demand), requiring an understanding of current challenges and future requirements, particularly from the perspective of early-career professionals, to align research and training with emerging needs and sustainable development goals.
Key Innovation: The study integrates bibliometric analysis of journal publications, international projects, and survey feedback from Young Engineering Geologists to identify a shift in engineering geology towards climate change-intensified geohazards, urban planning, and digital tools, highlighting the need for practical, interdisciplinary training and aligning YEG activities with SDGs to address future challenges.
113. Asymmetric Propagation of Hydraulic Fractures in Coal–Rock Composite Formations: Insights Based on Stress Sensitivity of Coal
Core Problem: Understanding and optimizing the asymmetric propagation of hydraulic fractures (HFs) in coal–rock composite formations, particularly the influence of coal's stress sensitivity and dynamic leak-off effects, which restrict HF growth in coal layers.
Key Innovation: Conducted true triaxial hydraulic fracturing experiments and numerical modeling, revealing that high stress sensitivity of coal significantly restricts HF propagation. Demonstrated a unique propagation pattern where overlying rock layers expand while coal layer growth is limited, and showed that higher fluid viscosity can substantially increase HF lengths by maintaining propagation in the viscosity-dominated region.
114. Simulation of treated soil structures with a bonding degradation double yield surface constitutive model and FEM implementation
Core Problem: Engineering construction in weak geological formations frequently confronts significant challenges related to foundation stability due to low shear strength, high compressibility, and high moisture content.
Key Innovation: A double yield surface constitutive model, coupling bonding degradation and fabric hardening, was introduced and implemented in ABAQUS to accurately simulate the stress-strain behavior of chemically treated soils (cement soil, polymer clays) across both brittle and ductile failure modes, demonstrating its applicability for subgrade reinforcement.
115. Dynamic climate risk modelling for agriculture: A machine learning approach integrating crop phenology into weather index insurance
Core Problem: Inaccurate estimation of planting dates and static approaches in Weather Index Insurance (WII) actuarial models increase basis risk and adverse selection, failing to adapt to temporal climate variability and integrate crop phenology.
Key Innovation: Developing a dynamic, transparent WII framework that integrates crop phenology into risk assessment and pricing using a two-step machine learning methodology: predicting planting dates with high accuracy and synchronizing remote-sensing indices with predicted phenological stages to improve county-level crop yield prediction, thereby advancing WII design for adaptive agricultural risk management.
116. Dynamic coupling reliability assessment of smart wind-photovoltaic-storage cyber-physical systems under multi-source uncertain information
Core Problem: Existing methods struggle to capture coupled power–information dynamics and their joint reliability impacts in smart Wind-PV-Storage Cyber-Physical Systems (WPS-CPSs) under multi-source uncertainties.
Key Innovation: Proposed a dynamic coupling reliability assessment framework for WPS-CPSs, constructing a heterogeneous directed network and integrating Power Flow Reliability Index (PFRI) and Information Flow Reliability Index (IFRI) via a coupling penalty mechanism to form the Dynamic Coupling Reliability Index (DCRI).
117. SCADA data-driven failure rate and reliability prediction for offshore wind turbines
Core Problem: Accurate failure rate and reliability prediction for offshore wind turbines is challenging, requiring a data-driven approach that accounts for differences from onshore turbines and leverages SCADA data.
Key Innovation: Proposed a data-driven model for offshore wind turbine failure rate prediction using SCADA data, incorporating an adaptive weighting algorithm to assess reliability-influencing factors and transform onshore device failure rates, demonstrating lower estimation errors than existing approaches.
118. Advancing understanding of parameterization effects in global hydrologic models through multi-model, multi-variable evaluation
Core Problem: Lack of comprehensive investigations into how parameter definitions in Global Hydrological Models (GHMs) impact hydrologic simulations through multi-model and multi-variable analyses, especially when relying on default parameter values.
Key Innovation: Examined parameter choices in four GHMs by optimizing them using different hydrological variables across 228 watersheds, demonstrating that optimized parameters generally outperform default values and highlighting the risk of overestimating high-flow predictions with default settings.
119. Quantifying river morphological changes using multi‐satellite observations and in situ measurements
Core Problem: Accurately estimating periodically submerged riverbed topography and quantifying morphological change using remote sensing is challenging.
Key Innovation: This study integrates optical, radar, and altimetry satellite data with river gauge measurements to estimate periodically submerged riverbed topography and quantify morphological change.
120. Improving Spatial Allocation for Energy System Coupling with Graph Neural Networks
Core Problem: Mismatched spatial resolutions in energy system analysis coupling models, where traditional aggregation methods are limited by using only a single geospatial attribute.
Key Innovation: An innovative method employing a self-supervised Heterogeneous Graph Neural Network to model high-resolution geographic units, integrating various geographical features to generate physically meaningful weights for spatial aggregation, enhancing scalability, accuracy, and physical plausibility.
121. WaveSSM: Multiscale State-Space Models for Non-stationary Signal Attention
Core Problem: Existing projection-based State-Space Models (SSMs) often rely on polynomial bases with global temporal support, which are poorly matched to signals exhibiting localized or transient structure.
Key Innovation: WaveSSM, a collection of SSMs constructed over wavelet frames, which yield localized support on the temporal dimension, making them more suitable for tasks requiring precise localization and outperforming orthogonal counterparts on real-world datasets with transient dynamics.
122. Positional-aware Spatio-Temporal Network for Large-Scale Traffic Prediction
Core Problem: Large-scale traffic flow forecasting models struggle to clearly distinguish each node and maintain a holistic view of historical data, while also facing deployment challenges due to increasing data sizes.
Key Innovation: PASTN, a lightweight Positional-aware Spatio-Temporal Network that uses positional-aware embeddings and a temporal attention module to effectively capture spatio-temporal complexities for large-scale traffic prediction, demonstrating improved effectiveness and efficiency.
123. Learning Rewards, Not Labels: Adversarial Inverse Reinforcement Learning for Machinery Fault Detection
Core Problem: Most RL-based machinery fault detection (MFD) approaches do not fully exploit RL's sequential decision-making strengths, often treating MFD as a contextual bandit problem, and require manual reward engineering or fault labels.
Key Innovation: Formulating MFD as an offline inverse reinforcement learning problem, where an agent learns reward dynamics directly from healthy operational sequences using Adversarial Inverse Reinforcement Learning, enabling early and robust fault detection by using the discriminator's learned reward as an anomaly score.
124. Vision Transformers Need More Than Registers
Core Problem: Vision Transformers (ViTs) exhibit artifacts due to 'lazy aggregation behavior,' where they use semantically irrelevant background patches as shortcuts to represent global semantics, impacting performance and interpretability.
Key Innovation: Identifies that ViT artifacts stem from lazy aggregation and proposes a solution to selectively integrate patch features into the CLS token, reducing the influence of background-dominated shortcuts and consistently improving performance across various benchmarks.
125. Revisiting Chebyshev Polynomial and Anisotropic RBF Models for Tabular Regression
Core Problem: Smooth-basis models, despite their continuously differentiable prediction surfaces suitable for surrogate optimization and sensitivity analysis, are seldom used in tabular regression where tree ensembles dominate, raising the question of their competitiveness and utility in CPU-based applied science settings.
Key Innovation: Develops and benchmarks enhanced anisotropic RBF networks, ridge-regularized Chebyshev polynomial regressors, and a smooth-tree hybrid against tree ensembles and transformers across 55 datasets, demonstrating that smooth models are statistically tied with tree ensembles on accuracy and exhibit tighter generalization gaps, recommending their inclusion in candidate pools for settings benefiting from gradually varying predictions.
126. TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series
Core Problem: Modern deep time series forecasting models typically minimize point-wise prediction loss without effectively leveraging the rich information contained in past prediction residuals from rolling forecasts, which reflect persistent biases, unmodeled patterns, or evolving dynamics.
Key Innovation: TEFL (Temporal Error Feedback Learning) is a unified learning framework that explicitly incorporates historical residuals into the forecasting pipeline during training and evaluation, using a lightweight low-rank adapter and a two-stage training procedure. It consistently improves accuracy (5-10% MAE reduction) and robustness under abrupt changes across diverse datasets and architectures.
127. Coarse-to-Fine Learning of Dynamic Causal Structures
Core Problem: Existing methods for learning dynamic causal structures in time series assume stationary or partially stationary causality, failing to capture complex, time-varying relationships efficiently and stably.
Key Innovation: DyCausal, a dynamic causal structure learning framework that uses convolutional networks for coarse-grained patterns and linear interpolation for fine-grained, time-varying causal graphs, with an acyclic constraint for efficiency and stability.
128. TabDLM: Free-Form Tabular Data Generation via Joint Numerical-Language Diffusion
Core Problem: Generating synthetic tabular data that accurately captures both structured numerical/categorical attributes and free-form text fields remains challenging, as existing diffusion models struggle with text quality and LLMs distort numerical values.
Key Innovation: TabDLM, a unified framework for free-form tabular data generation via a joint numerical-language diffusion model, which uses masked diffusion for textual/categorical features and continuous diffusion with learned specialized numeric tokens for numerical features, capturing cross-modality interactions.
129. Don't let the information slip away
Core Problem: State-of-the-art object detection models primarily focus on foreground objects, neglecting valuable contextual information provided by the background, which can significantly aid detection tasks.
Key Innovation: Association DETR, a novel object detection model that leverages background contextual information to achieve state-of-the-art results on the COCO val2017 dataset, improving detection performance.
130. Coded-E2LF: Coded Aperture Light Field Imaging from Events
Core Problem: Existing light field reconstruction methods from event cameras often require both events and intensity images, or lack theoretical support for purely event-based reconstruction.
Key Innovation: Proposes Coded-E2LF, a purely event-based computational imaging method for 4-D light field acquisition, demonstrating pixel-level accurate reconstruction from events alone and clarifying the key role of black patterns in aperture coding.
131. CGSA: Class-Guided Slot-Aware Adaptation for Source-Free Object Detection
Core Problem: Source-Free Domain Adaptive Object Detection (SF-DAOD) methods often overlook object-level structural cues, limiting their effectiveness in adapting detectors to unlabeled target domains without source data.
Key Innovation: Introduces CGSA, the first framework to integrate Object-Centric Learning (OCL) and slot-aware adaptation into SF-DAOD, using a Hierarchical Slot Awareness (HSA) module and Class-Guided Slot Contrast (CGSC) to improve domain-invariant adaptation.
132. MUG: Meta-path-aware Universal Heterogeneous Graph Pre-Training
Core Problem: Universal graph pre-training has not been explored for heterogeneous graphs, which present challenges due to diverse node/relation types and varying meta-path semantics across datasets, hindering unified representation and transferable encoding.
Key Innovation: Proposes MUG (Meta-path-aware Universal heterogeneous Graph pre-training), which introduces an input unification module for integrating diverse information into a shared representation space and trains a shared encoder to capture consistent structural patterns across meta-path views, addressing challenges in heterogeneous graph pre-training.
133. U-Net-Based Generative Joint Source-Channel Coding for Wireless Image Transmission
Core Problem: Deep learning-based joint source-channel coding (JSCC) methods for wireless image transmission either prioritize conventional distortion metrics over perceptual quality or incur high computational complexity.
Key Innovation: Proposed G-UNet-JSCC and cGAN-JSCC, two DL-based JSCC methods leveraging U-Net and adversarial training respectively, to improve both pixel-level fidelity and perceptual quality of reconstructed images in wireless transmission, with cGAN-JSCC showing greater robustness.
134. AMLRIS: Alignment-aware Masked Learning for Referring Image Segmentation
Core Problem: Referring Image Segmentation (RIS) needs improvement in explicitly estimating pixel-level vision-language alignment to enhance accuracy and robustness to diverse descriptions.
Key Innovation: Alignment-Aware Masked Learning (AML), a training strategy that explicitly estimates pixel-level vision-language alignment, filters poorly aligned regions during optimization, and focuses on trustworthy cues, leading to state-of-the-art RIS performance and enhanced robustness.
135. KMLP: A Scalable Hybrid Architecture for Web-Scale Tabular Data Modeling
Core Problem: Predictive modeling on web-scale tabular data with billions of instances and heterogeneous features faces significant scalability challenges, with features exhibiting anisotropy, heavy-tailed distributions, and non-stationarity.
Key Innovation: KMLP, a scalable hybrid deep architecture integrating a shallow Kolmogorov-Arnold Network (KAN) front-end for automatic non-linear transformations and a Gated Multilayer Perceptron (gMLP) backbone for high-order interactions, achieving state-of-the-art performance on large-scale tabular data.
136. TrajTok: Learning Trajectory Tokens enables better Video Understanding
Core Problem: Traditional video tokenization methods generate excessive and redundant tokens, limiting video efficiency and scalability, while existing trajectory-based tokenizers are complex, slow, and task-agnostic.
Key Innovation: TrajTok, an end-to-end video tokenizer module fully integrated and co-trained with video models, which performs implicit clustering over pixels in space and time to directly produce object trajectories, dynamically adapting token granularity and improving video understanding performance and efficiency.
137. SceneTransporter: Optimal Transport-Guided Compositional Latent Diffusion for Single-Image Structured 3D Scene Generation
Core Problem: Existing methods for generating part-level 3D objects from single images often fail to organize these parts into distinct instances in open-world scenes due to a lack of structural constraints within the model's internal assignment mechanism.
Key Innovation: SceneTransporter, an end-to-end framework that reframes structured 3D scene generation as a global correlation assignment problem, solving an entropic Optimal Transport (OT) objective within a compositional DiT model to enforce structural constraints and improve instance-level coherence and geometric fidelity.
138. Doubly Adaptive Channel and Spatial Attention for Semantic Image Communication by IoT Devices
Core Problem: IoT networks face significant challenges (limited bandwidth, resources, dynamic channels) for semantic image communication, and existing deep joint source-channel coding (DJSCC) methods struggle with training for various SNRs and lack comprehensive adaptive attention.
Key Innovation: Introduces Doubly Adaptive DJSCC (DA-DJSCC) which simultaneously utilizes doubly adaptive channel-wise and spatial attention modules at both transmitter and receiver, dynamically adjusting to varying channel conditions and spatial feature importance for robust and efficient semantic image communication.
139. Towards Multimodal Domain Generalization with Few Labels
Core Problem: Existing multimodal domain generalization and semi-supervised learning methods fail to effectively learn robust models from multi-source data with few labels and domain shifts, especially when combining these challenges.
Key Innovation: Introduction of the Semi-Supervised Multimodal Domain Generalization (SSMDG) problem and a unified framework (Consensus-Driven Consistency Regularization, Disagreement-Aware Regularization, Cross-Modal Prototype Alignment) to address it, along with new benchmarks.
140. MSJoE: Jointly Evolving MLLM and Sampler for Efficient Long-Form Video Understanding
Core Problem: Efficiently understanding long-form videos remains a fundamental challenge for multimodal large language models (MLLMs) due to the computational cost of processing all frames.
Key Innovation: MLLM-Sampler Joint Evolution (MSJoE), a novel framework that jointly optimizes an MLLM and a lightweight key-frame sampler through reinforcement learning to efficiently understand long-form videos by selecting a compact set of informative frames.
141. pMoE: Prompting Diverse Experts Together Wins More in Visual Adaptation
Core Problem: Existing parameter-efficient fine-tuning methods for visual adaptation typically rely on a single pre-trained model, overlooking the potential synergies and enhanced versatility that could arise from integrating diverse domain knowledge from multiple expert models.
Key Innovation: pMoE (Prompting Diverse Experts Together), a novel Mixture-of-Experts prompt tuning method that leverages multiple expert domains through expert-specialized prompt tokens and a dynamic token dispatching mechanism to enhance model versatility and performance across various visual adaptation tasks.
142. Cross-Task Benchmarking of CNN Architectures
Core Problem: There is a need for a comprehensive comparative study of dynamic convolutional neural network (CNN) architectures across various tasks (classification, segmentation, time series analysis) to understand their performance benefits over conventional CNNs.
Key Innovation: A cross-task benchmarking study comparing five variants of CNNs (vanilla, hard/soft attention, ODConv) based on ResNet-18, demonstrating that attention mechanisms and dynamic convolution methods consistently enhance accuracy, efficiency, and computational performance, especially ODConv for morphologically complex images.
143. Latent Matters: Learning Deep State-Space Models
Core Problem: Deep state-space models (DSSMs) trained by maximizing the evidence lower bound often fail to actually learn the underlying dynamics of observed sequence data for temporal predictions.
Key Innovation: A constrained optimization framework for training DSSMs and the Extended Kalman VAE (EKVAE), which combines amortized variational inference with classic Bayesian filtering/smoothing to model dynamics more accurately, improving system identification and prediction accuracy.
144. FLIGHT: Fibonacci Lattice-based Inference for Geometric Heading in real-Time
Core Problem: Estimating camera motion (heading) from monocular video is fundamental for tasks like SLAM and visual odometry, but existing methods often decrease in accuracy or become computationally expensive as noise and outlier levels increase.
Key Innovation: Proposes FLIGHT, a novel generalization of the Hough transform on the unit sphere (S(2)) to estimate camera heading. It extracts correspondences, generates great circles of compatible directions, and discretizes the unit sphere using a Fibonacci lattice for robust voting, improving accuracy and efficiency and reducing RMSE during camera pose initialization in SLAM.
145. TriLite: Efficient Weakly Supervised Object Localization with Universal Visual Features and Tri-Region Disentanglement
Core Problem: Weakly Supervised Object Localization (WSOL) methods often rely on multi-stage pipelines or expensive full fine-tuning of large backbones, leading to high training costs, and frequently suffer from partial object coverage despite recent progress.
Key Innovation: Presents TriLite, a single-stage WSOL framework that leverages a frozen, self-supervised Dinov2 Vision Transformer with minimal trainable parameters. Its core is the TriHead module, which decomposes patch features into foreground, background, and ambiguous regions, effectively improving object coverage and suppressing spurious activations, achieving state-of-the-art performance with high parameter efficiency.
146. Closing the gap on tabular data with Fourier and Implicit Categorical Features
Core Problem: Deep learning models lag behind tree-based methods for tabular data, hypothesized to be due to their bias towards uniform numerical processing and smooth solutions, making it hard to exploit non-linear interactions from categorical features.
Key Innovation: Addresses the performance gap by using statistical-based feature processing to identify target-correlated features once discretized, and mitigates the bias for overly-smooth solutions using Learned Fourier, significantly boosting deep learning models' performance on tabular data.
147. Motion-aware Event Suppression for Event Cameras
Core Problem: Event cameras generate a high volume of events from independent moving objects (IMOs) and ego-motion, which can clutter data and hinder downstream applications.
Key Innovation: The first framework for Motion-aware Event Suppression that jointly segments IMOs and predicts their future motion, enabling anticipatory suppression of dynamic events in real time, improving segmentation accuracy and benefiting downstream applications like visual odometry.
148. Sensor Generalization for Adaptive Sensing in Event-based Object Detection via Joint Distribution Training
Core Problem: The novelty of event cameras leads to a gap in data variability and a lack of extensive analysis on how intrinsic sensor parameters affect object detection model performance, hindering sensor-agnostic robustness.
Key Innovation: Provides an in-depth understanding of how intrinsic parameters affect event-data-trained object detection models and uses these findings to expand model capabilities towards sensor-agnostic robustness via joint distribution training.
149. LoBoost: Fast Model-Native Local Conformal Prediction for Gradient-Boosted Trees
Core Problem: Gradient-boosted decision trees provide strong point predictions but lack uncertainty quantification. Existing conformal prediction methods for uncertainty are often poorly adaptive under heteroscedasticity or increase cost/reduce data efficiency by requiring auxiliary models or extra data splits.
Key Innovation: LoBoost, a model-native local conformal method, reuses the fitted ensemble's leaf structure of gradient-boosted trees to define multiscale calibration groups, enabling adaptive residual quantile calibration without retraining, auxiliary models, or extra splitting, offering competitive interval quality and speedups.
150. Unveiling Deep Shadows: A Survey and Benchmark on Image and Video Shadow Detection, Removal, and Generation in the Deep Learning Era
Core Problem: The field of deep-learning-based shadow detection, removal, and generation lacks a unified understanding, standardized evaluation, and clear future directions, with inconsistencies and limitations in prior research regarding model design, resolution dependence, and cross-dataset generalization.
Key Innovation: Presents a unified survey and benchmark, introducing consistent taxonomies for architectures, supervision strategies, and learning paradigms. It reviews major datasets and evaluation protocols, re-trains representative methods for fair comparison, and outlines future directions including unified all-in-one frameworks, semantics- and geometry-aware reasoning, and the integration of physics-guided priors into multimodal foundation models.
151. From Open Vocabulary to Open World: Teaching Vision Language Models to Detect Novel Objects
Core Problem: Traditional object detection is closed-set, and Open Vocabulary Object Detection (OVD) relies on accurate prompts, struggles with misclassifying near-out-of-distribution (NOOD) objects, and ignores far-out-of-distribution (FOOD) objects, limiting its use in open-world settings.
Key Innovation: Proposes a framework enabling OVD models to operate in open-world settings by identifying and incrementally learning unseen objects. It introduces Open World Embedding Learning (OWEL) with Pseudo Unknown Embedding for FOOD objects and Multi-Scale Contrastive Anchor Learning (MSCAL) for NOOD objects, achieving state-of-the-art performance.
152. PPT: Pretraining with Pseudo-Labeled Trajectories for Motion Forecasting
Core Problem: State-of-the-art motion forecasting models rely on costly, manually annotated or post-processed trajectory datasets, which are hard to scale, lack reproducibility, and introduce domain gaps, limiting generalization across environments.
Key Innovation: Introduces PPT (Pretraining with Pseudo-labeled Trajectories), a simple and scalable pretraining framework that uses unprocessed, diverse trajectories automatically generated from off-the-shelf 3D detectors and tracking, improving generalization and performance in motion forecasting, especially in low-data and cross-domain settings.
153. Mixing It Up: Exploring Mixer Networks for Irregular Multivariate Time Series Forecasting
Core Problem: Forecasting irregularly sampled multivariate time series with missing values (IMTS) is a fundamental challenge, and the applicability of lightweight MLP-based Mixer architectures to this setting remains unexplored.
Key Innovation: IMTS-Mixer, a novel architecture that adapts Mixer principles to the IMTS setting. It introduces ISCAM (a channel-wise encoder for irregular observations) and ConTP (a continuous time decoder for forecasting at arbitrary time points), achieving state-of-the-art performance in accuracy and inference time with fewer parameters.
154. Is Exchangeability better than I.I.D to handle Data Distribution Shifts while Pooling Data for Data-scarce Medical image segmentation?
Core Problem: Data scarcity in medical imaging leads to the 'Data Addition Dilemma,' where pooling data from multiple sources can introduce distributional shifts, negatively impacting model performance, especially when the i.i.d. assumption doesn't hold.
Key Innovation: Proposes using exchangeability (instead of i.i.d.) as a more practical framework for data pooling in data-scarce medical image segmentation. Introduces a method to control foreground-background feature discrepancies across deep network layers, improving feature representations and achieving state-of-the-art performance.
155. ST-GS: Vision-Based 3D Semantic Occupancy Prediction with Spatial-Temporal Gaussian Splatting
Core Problem: Existing 3D semantic occupancy prediction methods using Gaussian Splatting for scene understanding lack sufficient multi-view spatial interaction and multi-frame temporal consistency.
Key Innovation: Proposes ST-GS, a Spatial-Temporal Gaussian Splatting framework that enhances spatial and temporal modeling through a guidance-informed spatial aggregation strategy and a geometry-aware temporal fusion scheme, achieving state-of-the-art performance and improved temporal consistency in 3D occupancy prediction.
156. PartSAM: A Scalable Promptable Part Segmentation Model Trained on Native 3D Data
Core Problem: Existing open-world 3D part segmentation methods transfer supervision from 2D foundation models, failing to capture intrinsic geometry and leading to surface-only understanding, uncontrolled decomposition, and limited generalization.
Key Innovation: Presents PartSAM, the first promptable part segmentation model trained natively on large-scale 3D data using a triplane-based dual-branch encoder and a model-in-the-loop annotation pipeline, achieving emergent open-world capabilities for accurate and comprehensive 3D part understanding.
157. Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting
Core Problem: 3D Gaussian Splatting (3DGS) methods, particularly MLP-based variants, suffer from significant redundancy and inconsistencies in occluded regions due to a lack of occlusion awareness, limiting rendering quality and speed.
Key Innovation: Proposes Proxy-GS, a novel pipeline that exploits a fast proxy system to introduce Gaussian occlusion awareness during both training (guiding densification) and inference (culling Gaussians), resulting in stronger rendering capability and faster speed, especially in heavily occluded scenarios.
158. Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction
Core Problem: Existing 3D Semantic Scene Graph Prediction methods fail to optimize the representational capacity of object and relationship features, relying excessively on Graph Neural Networks despite insufficient discriminative capability, leading to suboptimal scene graph accuracy.
Key Innovation: Designs a highly discriminative object feature encoder and employs a contrastive pretraining strategy to decouple object representation learning from scene graph prediction, enhancing object classification and relationship prediction accuracy by effectively combining geometric and semantic features.
159. Deforming Videos to Masks: Flow Matching for Referring Video Segmentation
Core Problem: Referring Video Object Segmentation (RVOS) methods often use cascaded 'locate-then-segment' pipelines, which create information bottlenecks by simplifying semantics into coarse geometric prompts and struggle to maintain temporal consistency due to decoupled segmentation.
Key Innovation: Proposes FlowRVS, a novel framework that reconceptualizes RVOS as a conditional continuous flow problem, learning a direct, language-guided deformation from a video's holistic representation to its target mask, achieving new state-of-the-art results by leveraging pretrained T2V models for fine-grained control and temporal coherence.
160. SODAs: Sparse Optimization for the Discovery of Differential and Algebraic Equations
Core Problem: Existing sparse optimization methods for discovering dynamical systems assume differential-algebraic equations (DAEs) can be reduced to ordinary differential equations (ODEs), limiting their applicability for systems with unknown constraints and timescales.
Key Innovation: Introducing Sparse Optimization for Differential-Algebraic Systems (SODAs), a data-driven method that identifies DAEs in their explicit form by sequentially discovering algebraic and dynamic components, leading to interpretable and numerically stable models without prior variable elimination.
161. Sparse Imagination for Efficient Visual World Model Planning
Core Problem: The high computational burden of world model-based planning, especially in resource-constrained environments like robotics, due to the large number of tokens processed during forward prediction.
Key Innovation: Proposing 'Sparse Imagination' for efficient visual world model planning, which leverages a sparsely trained vision-based world model with randomized grouped attention to flexibly adjust and reduce the number of tokens processed, significantly accelerating planning while maintaining control fidelity.
162. Adaptive Hybrid Caching for Efficient Text-to-Video Diffusion Model Acceleration
Core Problem: Video Diffusion Transformer (DiT) models incur high computational cost and inference latency due to their multi-step iterative denoising process, and existing single-granularity caching methods struggle to balance generation quality and inference speed.
Key Innovation: Proposing MixCache, a training-free caching-based framework that distinguishes interference between caching strategies and introduces a context-aware cache triggering strategy along with an adaptive hybrid cache decision strategy to dynamically select optimal caching granularity, significantly accelerating video generation while maintaining quality.
163. Geometric Autoencoder Priors for Bayesian Inversion: Learn First Observe Later
Core Problem: Recovering full-field information of physical systems from limited, noisy observations is a highly ill-posed problem, especially for systems with complicated variable geometries, and standard multi-system Bayesian UQ struggles with this.
Key Innovation: GABI (Geometric Autoencoders for Bayesian Inversion), a framework that learns geometry-aware generative models as highly informative, geometry-conditioned priors for Bayesian inversion, enabling robust UQ and accurate predictions without explicit PDE knowledge.
164. Throwing Vines at the Wall: Structure Learning via Random Search
Core Problem: Structure learning remains a key challenge in vine copulas, which are used for flexible multivariate dependence modeling, and existing heuristics are often suboptimal.
Key Innovation: Random search algorithms and a statistical framework based on model confidence sets to improve structure selection in vine copulas, consistently outperforming state-of-the-art approaches and providing theoretical guarantees on selection probabilities.
165. Establishing Stochastic Object Models from Noisy Data via Ambient Measurement-Integrated Diffusion
Core Problem: Conventional stochastic object models (SOMs) for medical imaging fail to capture realistic anatomy, and data-driven approaches require clean data, which is rarely available from noisy clinical measurements.
Key Innovation: Proposes AMID, an unsupervised Ambient Measurement-Integrated Diffusion with noise decoupling, which establishes clean SOMs directly from noisy measurements by aligning measurement noise with the diffusion trajectory and explicitly modeling noise coupling.
166. Denoising the Deep Sky: Physics-Based CCD Noise Formation for Astronomical Imaging
Core Problem: Astronomical imaging is noise-limited, and while calibration pipelines remove structured artifacts, stochastic noise remains unresolved, hindering the application of learning-based denoising due to scarce paired training data.
Key Innovation: Proposes a physics-based noise synthesis framework tailored to CCD noise formation in telescopes, modeling various noise sources to generate abundant paired datasets from high-SNR stacked exposures, which effectively improves photometric and scientific accuracy in denoising.
167. A validated two-phase flow model for characterizing the coupled sedimentation and self-weight consolidation of slurries
Core Problem: Accurately predicting the complex, coupled spatiotemporal evolution of concentration and stress fields during self-weight sedimentation and consolidation of slurries.
Key Innovation: A modified Eulerian two-phase flow model, derived from Navier-Stokes equations and validated by systematic experiments, that integrates sedimentation-consolidation dynamics, captures spatiotemporal evolution, and identifies phase regime transitions in multiphase particulate flows.
168. Efficient Semantic Segmentation of Transmission Line Corridor Point Clouds via Transformer With Vector Aggregation
Core Problem: Precise semantic segmentation of massive, unordered, and complex point clouds in transmission line corridors is challenging, as existing methods struggle to efficiently extract both local details and global context, often requiring high computational resources.
Key Innovation: PointVSV, a segmentation network that combines a voxelization and random sampling strategy with a shifted-patch transformer module and vector aggregation module, enabling robust global context modeling and fine-grained local feature extraction for highly accurate and efficient semantic segmentation of power transmission corridors.
169. A Novel Enhancement of the MIMIC Remote Sensing Water Vapor Product and Its Application in GNSS Positioning
Core Problem: The accuracy of remote sensing precipitable water vapor (PWV) products, such as MIMIC-TPW2, is limited, which hinders high-precision GNSS applications like precise point positioning (PPP) and climate monitoring.
Key Innovation: A novel enhancement of the MIMIC-TPW2 PWV product using generalized regression neural network (GRNN) and random forest (RF) models, which significantly reduces PWV root mean square error. The derived enhanced zenith tropospheric delays (ZTDs) then substantially improve GNSS PPP convergence and positioning accuracy.
170. Spatiotemporal Dynamics of Terrestrial Carbon Storage Driven by Land-Use and Land-Cover Change in Island Ecosystems: Insights From Long-Term Remote Sensing
Core Problem: Ambiguous ecosystem boundaries and decreased land-use and land-cover (LULC) classification accuracy in archipelago surface environments restrict the precision of carbon storage quantification and understanding the impact of LULC change.
Key Innovation: A long-term sequence analysis method employing a multifeature random forest algorithm for accurate LULC mapping and the InVEST model for quantifying carbon storage, revealing spatiotemporal dynamics of carbon storage and the significant impact of urbanization on natural ecosystems in the Zhoushan Archipelago.
171. A 30 m spatial resolution dataset of ecosystem services in China for 2000, 2010, and 2020
Core Problem: The need for high-resolution ecosystem service datasets to identify site-specific differences and inform environmental management.
Key Innovation: Produced a high spatial resolution (30m) dataset of key ecosystem services (net primary productivity, soil conservation, sandstorm prevention, water yield) for China across three decades (2000, 2010, 2020) using ecological process models, validated for high consistency.
172. GHGPSE-Net: a method towards spaceborne automated extraction of greenhouse-gas point sources using point-object-detection deep neural network
Core Problem: Existing methods for detecting and quantifying greenhouse gas (GHG) point sources from spaceborne data often require human intervention and only detect plume masks, not precise source locations, limiting their utility for regulatory applications.
Key Innovation: GHGPSE-Net, a deep learning method, simultaneously performs detection, localization, and quantification of GHG emissions from spaceborne observations, eliminating the need for traditional segmentation steps, and demonstrating high accuracy.
173. Echinoderm stereom gradient structures enable mechanoelectrical perception
Core Problem: Exploring alternative evolutionary functions of cellular solids beyond mechanical performance and leveraging natural designs for advanced sensing technologies.
Key Innovation: Discovery of unexpected mechanoelectrical perception in echinoderm stereom due to gradient cellular solids, and the creation of 3D-printed artificial structures that mimic this ability for enhanced underwater spatiotemporal sensing.
174. Effects of water and sediment variations on estuarine channel evolution: mechanisms and morphological discrimination
Core Problem: Understanding the morphological evolution of meandering tail channels, critical river-sea interaction zones, under varying flow and sediment conditions, particularly for large rivers like the Yellow River, is complex.
Key Innovation: Employs physical experiments to simulate and analyze distinct evolutionary patterns in non-estuarine and estuarine reaches during sediment-feeding and post-feeding phases. The study refines the resistance law expression and proposes a channel pattern discrimination method suitable for the lower Yellow River, validated with measured data.
175. Effects of Specimen Geometry and Bamboo Biochar Amendment on Anisotropic Shrinkage Behavior of Soils
Core Problem: Soil shrinkage characteristics, influenced by specimen shape and additives, are critical for the long-term stability of geotechnical and geoenvironmental structures like landfill covers, requiring a thorough understanding of anisotropic behavior.
Key Innovation: Both specimen shape and bamboo biochar (BB) amendment significantly influence the anisotropic shrinkage behavior of CL and SM soils, with 5% BB reducing volumetric shrinkage by 7.24-11.12% and cylindrical specimens exhibiting greater volumetric shrinkage due to better particle packing.
176. Morphology of partially vegetated desert dunes affects their response to climatic perturbations and disturbance
Core Problem: Understanding the locally variable response of desert dunefields to climatic perturbations (drought, rainfall) and disturbances (wildfire), particularly how dune morphology and antecedent conditions influence activation and stabilization.
Key Innovation: Developing new insights into desert dune activation and stabilization using remote sensing data, showing that dune response is locally variable due to interactions between wind, topography, and vegetation, and integrating these findings into a conceptual model highlighting the negative feedback (hysteresis) of active dunes on vegetation regrowth.
177. Domain-specific large language model-driven risk analysis of battery energy storage systems
Core Problem: Conventional risk analysis methods for Battery Energy Storage Systems (BESS) heavily rely on subjective expert knowledge due to data deficiency.
Key Innovation: Proposes a risk analysis method integrating a domain-specific large language model (LLM), functional resonance analysis method (FRAM), and Bayesian networks (BN). The LLM generates FRAM, which is then mapped into a BN for quantitative risk analysis, reducing reliance on subjective information and providing valuable references for BESS risk quantification and management.
178. A dual time scale degradation-tolerant control for systems with actuator performance degradation
Core Problem: Conventional fault-tolerant control struggles with gradual actuator performance degradation, and existing studies often fail to capture the intrinsic difference between slow degradation evolution and fast system dynamics.
Key Innovation: Proposed a degradation-tolerant control (DTC) strategy based on a dual time scale modeling method to adaptively compensate for performance loss by dynamically adjusting control parameters, thereby maintaining stability and extending actuator operational lifespan.
179. A pressure-chlorine driven approach to design effective district metered areas (DMA) configurations in water distribution systems
Core Problem: Conventional Graph Theory (GT) techniques have limitations in designing effective District Metered Areas (DMA) configurations in Water Distribution Systems (WDS), particularly in integrating nodal pressure and chlorine concentration values.
Key Innovation: Proposed the Graph Theory-based Pressure and Chlorine Quantities (GT-PCQ) method, which directly integrates nodal pressure and chlorine concentration into the network partitioning process, leading to improved hydraulic and quality reliability and reduced computational time.
180. Compositing high-resolution SDGSAT-1 nighttime light data by ranking of structural image features
Core Problem: The need for robust compositing methods to unlock the potential of high-resolution nighttime light (NTL) data, as raw data is often contaminated by clouds, haze, moonlight, and sensor artifacts.
Key Innovation: An automatic and scalable pixel-level compositing method for high-resolution SDGSAT-1 NTL data that ranks structural image features to select optimal observations, mitigating contamination without external masks and significantly improving data quality, spatial alignment, and geometric consistency.
181. Synergistic day–night multi-modal framework for offshore gas flare detection using SDGSAT-1 in the Southeast Asian Seas
Core Problem: Challenges in satellite monitoring of gas flaring (a significant source of greenhouse gases) due to trade-offs between low spatial resolution of nighttime observations and low signal-to-noise ratio of daytime data, leading to omissions in global flare inventories.
Key Innovation: A Day-Night Synergistic Gas Flaring Detection (DNSGFD) framework leveraging SDGSAT-1's synchronous multi-sensor capabilities, employing a "nighttime discovery and daytime confirmation" strategy with a Glimmer-enhanced Flare Disturbance Index (GFDI) to achieve high accuracy and identify previously unrecorded offshore gas flaring platforms.
182. Investigating natural biofilms on floating marine microplastics and the implications for ocean color remote sensing
Core Problem: The need for new satellites to quantify marine microplastics and the impact of natural biofilms on their spectral properties, which can affect ocean color remote sensing retrievals for marine debris detection.
Key Innovation: Characterization of the spectral properties of damp, biofilmed microplastic pieces collected from the Great Pacific Garbage Patch, identifying the significant impact of red algal biofilms on blue wavelength absorption and suggesting new narrow bands (670-680 nm) for discriminating biofilmed microplastic from phytoplankton in remote sensing.
183. Enhancing water science in Earth’s second lung: AI-generated centenary hydrological insights from two decades of satellite data in the Congo Basin
Core Problem: A lack of long-term, comprehensive terrestrial water storage anomaly (TWSA) data for the Congo Basin hinders understanding of its water challenges.
Key Innovation: Developed CM-RecNet, a climate-memory hybrid deep learning model, to reconstruct a 100-year (1923-2024) TWSA dataset for the Congo Basin from two decades of GRACE/GRACE-FO observations, providing previously unavailable data for climate change and anthropogenic impact assessment.
184. Three-dimensional mapping of key soil properties with multi-stage validation and big data
Core Problem: Many areas lack updated, high-resolution digital soil products for key properties like soil organic carbon (SOC) and pH, despite their importance for food security and carbon cycle regulation.
Key Innovation: Generated 3D digital maps of SOC and pH for the Republic of Bashkortostan across five depth intervals using a machine learning approach with extensive training data, demonstrating reliable predictions in plain regions and highlighting data gaps in mountainous areas for future improvement.
185. Transition to transpiration-dominated evapotranspiration on the Loess Plateau: spatially divergent driving mechanism and threshold effect after two decades of reforestation
Core Problem: Uncertainty regarding the hydrological implications of large-scale ecological restoration, particularly on evapotranspiration (ET), due to complex spatial heterogeneity and non-linear feedback loops between vegetation and water.
Key Innovation: Developed a novel attribution framework combining the Two-Source Energy Balance (TSEB) model with Bayesian Ridge Regression to identify spatially divergent driving mechanisms of ET and its components on the Loess Plateau, revealing a fundamental shift to transpiration-dominated ET and identifying non-linear regulation mechanisms and optimal LAI thresholds.