TerraMosaic Daily Digest: Feb 24, 2026
Daily Summary
This February 24, 2026 digest (168 selected papers from 1,058 deduplicated candidates) shows a clear split between high-impact geohazard studies and a large methodological stream. The strongest hazard-facing contributions are concentrated in landslide science: rainfall-displacement early warning design, multi-orbit InSAR monitoring of engineered slopes, and kinematic segmentation of deep-seated gravitational slope deformation. These papers are not only descriptive; they define operational variables and decision-relevant thresholds.
Beyond slope failure, the portfolio expands into flood-aware urban network recovery, permafrost and freeze-thaw mechanics, and coupled hydro-geomorphic process modeling. In parallel, the largest volume still comes from general AI/vision/time-series work. Its value to geohazards is highest when transfer is explicit, uncertainty is controlled, and physical constraints are respected.
Key Trends
The dominant signal is a shift from single-model accuracy claims toward deployable, mechanism-aware hazard intelligence.
- Landslide monitoring is moving from displacement-only tracking to process-resolved diagnostics: top studies combine kinematics, structure, and trigger information to increase warning lead time and reduce false confidence.
- Regionalization and scalability are improving: wide-area debris-flow simulation and spatially explicit susceptibility workflows indicate a transition from case studies to basin- and network-scale decision support.
- Hazard analysis is increasingly infrastructure-facing: flood-disrupted road recovery, dynamic urban risk hot-spotting, and evacuation reliability assessment link physical hazard evolution to actionable service restoration priorities.
- Cold-region and hydro-thermal mechanics remain a strong mechanistic frontier: freeze-thaw migration, ice segregation, and permafrost subgrade deformation studies provide measurable parameters for engineering design and failure prevention.
- Foundation-model volume is high, but relevance is now filter-driven: non-domain AI papers matter mainly when they demonstrate robust transfer to geohazard tasks, interpretable behavior, and uncertainty-aware deployment.
Selected Papers
This digest features 168 selected papers from 1058 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. An integrated early warning model for rainfall-induced landslides based on rainfall–displacement kinematic features
Core Problem: Large-scale rainfall-induced landslides exhibit prolonged deformation and complex mechanical processes, posing significant challenges for developing reliable early warning models.
Key Innovation: Developed a four-level dynamic early warning model (EWM) for rainfall-induced landslides by integrating 15 years of manual observations and 2 years of real-time monitoring, introducing the “one rainfall process” concept, and defining rainfall–displacement kinematic features to identify deformation stages and establish zoned EWMs.
2. InSAR-based deformation monitoring of high-fill engineered landslides: a case study at Panzhihua Airport, Southwest China
Core Problem: High-fill engineered landslides pose significant infrastructure safety challenges in mountainous regions due to progressive deformation under self-weight and external triggers.
Key Innovation: Investigated a high-fill engineered landslide using a multi-orbit Time-Series InSAR framework, innovatively combining multi-orbit LOS deformation retrieval, 2D deformation vector decomposition (vertical + slope-parallel), and spatiotemporal correlation analysis with geological structure and rainfall, providing insights for managing secondary deformation risks and early warning.
3. Kinematic segmentation of a DSGSD on the flank of the exhumed Leo Pargil Dome (northwestern Himalaya) using InSAR
Core Problem: Characterizing the internal kinematics of slow-moving deep-seated landslides (DSGSDs) in high mountain terrain is difficult from field observations alone due to their slow deformation over years to decades.
Key Innovation: Combined multi-temporal Persistent Scatterer InSAR (PS-InSAR) analysis of Sentinel-1 data with geological information to characterize the present-day kinematics and internal segmentation of the Leo Pargil DSGSD, revealing a strongly segmented displacement field influenced by inherited structures and slope geometry.
4. The UK and Ireland Geophysical Array -- Concept and Design
Core Problem: A need for a comprehensive, integrated geophysical monitoring infrastructure in the UK and Ireland to advance scientific exploration of the subsurface, understand natural phenomena, and address problems related to hazards and resources.
Key Innovation: Outlines the concept and design of the UKI Geophysical Array, a community-driven vision for an array of seismological and other geophysical sensors to maximize value from existing equipment, research Earth phenomena, solve hazard/resource problems, and inspire public engagement.
5. A comparative assessment of supervised models for landslide susceptibility mapping: a case study of Qiongzhong county, Hainan Island, China
Core Problem: Landslides cause significant economic losses and human casualties, necessitating effective landslide susceptibility assessment (LSA) for land-use planning and risk management.
Key Innovation: Investigates and provides a comparative assessment of various supervised models for landslide susceptibility mapping, using Qiongzhong county, Hainan Island, China, as a case study.
6. Exploring the spatial distribution patterns of loess landslides and identifying the key driving factors in typical regions of the upper Yellow River, China
Core Problem: Understanding the spatial distribution and key driving factors of loess landslides in the upper Yellow River region to mitigate their severe impacts on the environment and human habitation.
Key Innovation: The study explores spatial distribution patterns and identifies key driving factors for loess landslides in a specific region (Gaolan County), contributing to better understanding and potentially susceptibility mapping.
7. Quantitative evaluation and wide-area simulation of post-failure debris flows using the depth-integrated particle method
Core Problem: Numerical simulation of debris flow runout at regional scales remains challenging, especially when treating numerous slope-failure initiation points simultaneously, with existing approaches facing increasing computational demand and practical limitations.
Key Innovation: Evaluated the depth-integrated particle method (DIPM) for simulating debris flows across different spatial scales. Validated DIPM using flume experiments and two real-world case studies (calibrating Manning’s coefficient and critical deposition angle). Applied the calibrated parameters to a wide-area simulation involving 2249 landslide scarps, achieving an overall Critical Success Index (CSI) of approximately 0.60, confirming DIPM's ability to reproduce debris flow behavior at event and regional scales with moderate computational cost for scalable hazard mapping.
8. Characterizing local land-surface dynamics with spatio-temporal surface models and robust adaptive Kalman filters
Core Problem: Existing Geomorphic Change Detection (GCD) methods have limitations in characterizing changes in surface geometry, improving local uncertainty estimates, and incorporating uncertainty into derived indices, especially in complex post-landslide environments.
Key Innovation: A methodology applying a 2-D array of robust adaptive Kalman filters was developed to estimate spatial distributions of surface and surface change coefficients, enabling characterization of geometric changes and improved confidence in results for a drainage basin recovering from co-seismic landslides.
9. A novel dynamic disaster risk assessment of Urban Built Environments: an application to flood and earthquake
Core Problem: Disaster risk assessment in Urban Built Environments (UBEs) needs to consider the joint impacts of dynamic hazard conditions and spatiotemporal variations in population exposure and vulnerability, which is challenging with current static approaches.
Key Innovation: Defined a novel methodology for dynamic single-risk assessment in UBEs, combining physical vulnerability, hazard, user exposure, and vulnerability at the open-space level, applied to flood and earthquake risks, to detect recurrent and scenario-sensitive 'hot-spots' and support intervention prioritization.
10. Urban road connectivity assessment and impassable-road restoration sequencing under the impact of flood events
Core Problem: Existing methods for assessing road criticality and prioritizing restoration after flood events often fail to fully characterize global connectivity collapse or dynamic connectivity evolution under specific disaster scenarios.
Key Innovation: Analyzes urban road network connectivity loss under different rainfall recurrence periods using network efficiency indicators and proposes a connectivity-based road recovery prioritization method specifically for flood scenarios, demonstrating that flood-specific criticality differs from non-flood conditions.
11. Legacy check dams and cascading hydrogeomorphic events as drivers of large wood dynamics and storage in a mountain stream
Core Problem: The long-term influence of sequential hydrogeomorphic disturbances and legacy torrent-control structures on large wood (LW) dynamics and storage in mountain streams, and how these factors interact, is not fully understood.
Key Innovation: This study demonstrates that cascading hydrogeomorphic events (debris flow, extreme flood) and legacy check dams interact to significantly modulate large wood recruitment, storage, and fluxes in a mountain stream over decadal timescales, revealing a shift to a source-limited stage and the disproportionate influence of degraded check dams on wood retention.
12. Mechanistic investigation of hydro-thermal migration and ice segregation in permafrost subgrades using large-scale freeze–thaw test and NMR microstructural analysis
Core Problem: Hydro-thermal migration and ice aggregation are key processes driving deformation and failure of permafrost subgrade, posing serious threats to highway engineering in permafrost regions, but their underlying mechanisms require further elucidation.
Key Innovation: Conducts a large-scale freeze–thaw model test combined with internal deformation monitoring and nuclear magnetic resonance (NMR) measurements to systematically investigate hydro-thermal migration and the evolution of segregated ice in permafrost subgrades. The study reveals that embankment compaction intensifies hydro-thermal activity, water migrates bidirectionally/layeredly, and embankment deformation occurs in four distinct stages, offering theoretical recommendations for mitigating roadbed distress.
13. From Samples to Scenarios: A New Paradigm for Probabilistic Forecasting
Core Problem: State-of-the-art probabilistic time series forecasting models rely on sampling, which suffers from inherent limitations such as lacking explicit probabilities, inadequate coverage, and high computational costs.
Key Innovation: Introduces 'Probabilistic Scenarios,' a new paradigm for probabilistic forecasting that directly produces a finite set of {Scenario, Probability} pairs, avoiding Monte Carlo approximation and demonstrating superior performance with a simple linear model called TimePrism.
14. A new breaker index introducing a two-phase nearshore dissipation condition
Core Problem: Existing breaker index formulae struggle to accurately capture the complex, non-monotonic dependence of wave breaking on hydrodynamic and bathymetric conditions, leading to inaccuracies in nearshore wave models.
Key Innovation: Introducing a new physics-based, piecewise linear formula for the breaker index (γ) that explicitly models a two-phase breaking behavior governed by a dynamic relative depth threshold, significantly improving breaker wave estimation in the surf-zone.
15. Drought damage sensitivity assessment under continuous drought using a crop transpiration drought index and a three-phase disaster loss curve
Core Problem: Continuous droughts occur frequently worldwide, making effective drought damage sensitivity assessment a core link in drought risk management.
Key Innovation: Developed a method for drought damage sensitivity assessment under continuous drought conditions, utilizing crop experiments in the Huaibei Plain, a crop transpiration drought index, and a three-phase disaster loss curve.
16. An insight into the competition between the shape effects and grain crushing behavior: from single grain to aggregates
Core Problem: The competition between inherent shape effects and grain crushing behavior on the mechanical behavior of breakable coarse grains at the critical state remains insufficiently understood due to the complex nature of grain crushing characteristics.
Key Innovation: Presented a novel discrete element method (DEM) framework incorporating realistic particle shape and breakage (using bonded particle method, BPM). Conducted triaxial shear tests on granular assemblies with angular, pebble, and spherical grains. Identified two competing shape effects (irregular shapes lead to lower crushing resistance but stronger interlocking) and proposed a debonding coefficient to characterize their competition, distinguishing three influencing regimes. Provides new insight into deformation mechanisms of breakable coarse grains.
17. Dynamic forward–inverse prediction of excavations using mobilisable strength design and multi-receptive field convolutional neural network
Core Problem: Reliable and accurate prediction of retaining wall deflections in excavations is challenging due to spatial variability and sampling disturbance affecting soil parameter determination in conventional methods like MSD.
Key Innovation: Develops a dynamic approximation optimization framework integrating Mobilisable Strength Design (MSD) and a Multi-Receptive Field Convolutional Neural Network (MRFN) to progressively update soil parameters based on measured wall deflections, significantly enhancing the reliability of wall deflection prediction in multi-stage excavations.
18. Mechanisms governing wind erosion of physical crust with different cover under seasonal freeze-thaw
Core Problem: The mechanisms by which seasonal freeze-thaw cycles influence physical crust characteristics, leading to either amplification or suppression of wind erosion in cold-arid regions, remain unclear.
Key Innovation: This study reveals that seasonal freeze-thaw cycles intensify wind erosion and diminish the inhibitory effect of physical crusts, demonstrating that increased wind erosion intensity is primarily correlated with reduced crust coverage and decreased aerodynamic roughness, thereby filling a research gap in cold-arid regions.
19. Formulation and implementation of a unified hardening-softening model with strength mapping and fractional flow rule for the brittle-ductile transition in rocks
Core Problem: Accurately characterizing the complete brittle-ductile transition and post-peak behavior of rocks from initial yielding to residual flow using existing constitutive models is challenging, limiting reliable simulation of complex deformation and failure processes in geotechnical applications.
Key Innovation: Development and finite element implementation of a unified hardening-softening (UHS) constitutive model integrating a strength mapping index into the Drucker-Prager criterion and adopting fractional flow rules, validated for simulating complex deformation and failure processes in rocks, including applications to shallow tunneling and slope stability.
20. Integrating Diurnal Pulsing Signatures for AI‐Driven Tropical Cyclone Intensity Prediction
Core Problem: Existing AI models for tropical cyclone intensity prediction predominantly rely on instantaneous meteorological variables, overlooking informative temporal evolution signals like convective diurnal variation, leading to suboptimal prediction skills, especially for rapid intensification events.
Key Innovation: Integrating diurnal pulsing (DP) as a temporal evolution signal into AI models significantly improves predictions for tropical cyclone intensification rates and rapid intensification probabilities, outperforming models using only instantaneous convective indicators.
21. Climate‐Driven Changes in Wildfire Seasonality Across North America
Core Problem: The impact of climate change on the seasonal patterns of wildfires remains underexplored, making it difficult to interpret future region-specific changes and plan for adaptation.
Key Innovation: Quantifying historical changes in wildfire seasonality across North American ecoregions using satellite records, climate projections, and statistical models, revealing region-specific shifts (e.g., earlier in boreal forests, later in Mediterranean/desert regions) primarily driven by atmospheric dryness.
22. Urban heat as a growing hazard: a spatiotemporal analysis of land-use/land-cover changes and land surface temperature in Ha Noi city, Viet Nam
Core Problem: Understanding the spatiotemporal relationship between land-use/land-cover changes and land surface temperature to assess the growing urban heat island (UHI) hazard and its implications in rapidly growing tropical cities like Ha Noi.
Key Innovation: Using Landsat imagery, biophysical indices, Contribution Index, and GIS to analyze LULC changes and LST over 24 years, quantifying UHI intensity, applying the Urban Thermal Field Variance Index (UTFVI) to assess ecological thermal stress, and identifying LULC transitions as primary drivers of urban thermal risk.
23. Unraveling Node, Place, and Resilience During Disasters: Evident from a Typhoon in Hong Kong
Core Problem: Public transport-land use integration rarely considers performance during natural disasters, leading to mismatches in demand and functionality, and existing models lack a comprehensive resilience aspect.
Key Innovation: Introduces 'resilience' (Node Resilience and Place-Resilience) into the node-place model to examine transport-land use dynamics during disasters, demonstrating its importance in maintaining trip dynamics during a typhoon.
24. Reliability assessment of evacuation systems at the urban spatial scale using a two-stage evacuation simulation under temporary and fixed refuge scenarios
Core Problem: The dynamic performance of large-scale, staged evacuation systems during severe disasters is not sufficiently explored, leading to challenges in assessing urban evacuation reliability.
Key Innovation: Proposes a two-stage evacuation simulation model at the urban spatial level to analyze demand points and shelters under temporary and fixed refuge scenarios, identifying coverage gaps and assessing overall system performance, emphasizing spatial equity in shelter distribution.
25. A novel approach for soil moisture retrieval from Sentinel-1 SAR via temporal stability-based backscatter analysis
Core Problem: Difficulty in retrieving high-resolution soil moisture (SM) from Sentinel-1 SAR due to noise from surface heterogeneities, and limitations of existing masking techniques (like dynamic masking) that can remove valid SM signals or retain noisy pixels.
Key Innovation: Introduction of a novel masking approach based on temporal stability analysis (TSA) that identifies and retains pixels reliably capturing SM dynamics, improving correlations and reducing ubRMSE against both LSM-derived and in situ SM compared to no masking and dynamic masking across diverse landscapes and spatial scales.
26. Applicability of the Richards equation in infiltration simulation: A comparative study with the two-phase flow model
Core Problem: The Richards equation (RE), commonly used for simulating water infiltration in the vadose zone, has limitations due to its assumption of infinite air-phase mobility, and a systematic assessment of its applicability across diverse hydrogeological conditions is incomplete.
Key Innovation: Evaluated the RE model's performance against a two-phase flow model, revealing its reliability degradation under saturated/ponding conditions, high initial soil moisture, and fine-textured soils due to its failure to capture elevated pore-air pressures. Proposed quantitative guidelines for assessing RE model suitability, especially for scenarios with potential extensive confining layers.
27. Fill-spill process-guided hydrologic modeling: enhanced identification of hydrologically sensitive zones and simulations in semi-arid basins
Core Problem: Inaccurate flood simulations in semi-arid regions due to limitations in representing depression-storage and threshold-activated connectivity in traditional hydrological models.
Key Innovation: Proposed the Hydrologically Sensitive Fill-Spill Zone (HSFSZ) concept and developed the CASC2D-HSFSZ model, which dynamically delineates ponding-prone areas and significantly improves flood simulation accuracy in semi-arid regions by explicitly accounting for spatial heterogeneity.
28. An orthotropic OSB-PD model for anisotropic fracture behaviors of geomaterials: An orientation-angle-aware energy density approach
Core Problem: Conventional ordinary state-based peridynamic (OSB-PD) models for orthotropic materials are spatially discretization-dependent and struggle to accurately capture anisotropic fracture behaviors in geomaterials with inherent discontinuities and heterogeneity, particularly in applications like tunnel engineering.
Key Innovation: Development of a novel orthotropic OSB-PD constitutive model using an orientation-angle-aware energy density approach and short-range repulsive forces, validated for robust convergence and accurate crack initiation and propagation in orthotropic materials, offering insights for anisotropic damage analysis in tunnel-surrounding rock masses.
29. Modeling dynamic responses of three-dimensional high-speed train-ballasted track-subgrade coupled system subjected to spatial differential subgrade settlement
Core Problem: The relationship between subgrade settlement and rail deformation is unclear, and the impact of spatial differential subgrade settlement on the dynamic responses of ballasted tracks and the running safety/comfort of high-speed trains has not been adequately quantified.
Key Innovation: Establishes a 3D numerical model of a train-ballasted track-subgrade coupled system considering spatial differential subgrade settlements, proposing an innovative iterative algorithm to determine real-time track-subgrade contact. The model quantifies how settlement amplitude and wavelength affect unsupported sleepers, wheel-rail interaction, and identifies areas prone to damage and higher dynamic subgrade stress, providing a theoretical basis for railway operation and maintenance.
30. Deep Learning‐Based Tracking of Subduction Zones in Mantle Convection Models
Core Problem: Traditional methods for identifying and tracking subduction zones in mantle convection models rely on arbitrary thresholds and lack sufficient spatial context, making robust and consistent detection challenging.
Key Innovation: Adapts a deep-learning workflow using Fully Convolutional Networks (FCNs) for semantic segmentation of subduction zones from geodynamic simulations, providing improved spatial coherence, robustness, and consistent tracking of subduction zone evolution over time, without arbitrary thresholds.
31. FedAvg-Based CTMC Hazard Model for Federated Bridge Deterioration Assessment
Core Problem: Assessing bridge deterioration collaboratively across municipalities is impractical due to data governance constraints on sensitive inspection records, preventing shared model training.
Key Innovation: A federated framework using Federated Averaging (FedAvg) to estimate a Continuous-Time Markov Chain (CTMC) hazard model for bridge deterioration, enabling collaborative training of a shared benchmark model without transferring raw data, thereby facilitating evidence-based life-cycle planning while preserving data sovereignty.
32. CaDrift: A Time-dependent Causal Generator of Drifting Data Streams
Core Problem: Evaluating the robustness of machine learning methods under evolving data streams with concept drift is challenging due to the lack of controlled, time-dependent synthetic data generators.
Key Innovation: Introduces Causal Drift Generator (CaDrift), a time-dependent synthetic data generator based on Structural Causal Models (SCMs), capable of producing data streams with controlled distributional and covariate shifts and occasional perturbations, enabling robust evaluation of methods under evolving data.
33. GeoPT: Scaling Physics Simulation via Lifted Geometric Pre-Training
Core Problem: Scaling neural simulators for physics simulation is bottlenecked by the prohibitive cost of generating high-fidelity training data, and pre-training on static geometry alone ignores dynamics, leading to negative transfer on physics tasks.
Key Innovation: GeoPT, a unified pre-trained model for general physics simulation based on 'lifted geometric pre-training.' It augments geometry with synthetic dynamics to enable dynamics-aware self-supervision without physics labels, significantly reducing labeled data requirements and accelerating convergence across various fluid and solid mechanics benchmarks.
34. CGSTA: Cross-Scale Graph Contrast with Stability-Aware Alignment for Multivariate Time-Series Anomaly Detection
Core Problem: Multivariate time-series anomaly detection is challenging due to evolving inter-variable dependencies and inevitable noise. Existing methods often use single-scale graphs or instance-level contrast, and learned dynamic graphs can overfit noise without a stable anchor, leading to false alarms or misses.
Key Innovation: Proposes CGSTA, a framework with Dynamic Layered Graph Construction (DLGC) for multi-scale variable relations and Contrastive Discrimination across Scales (CDS) for structure-aware learning. It also introduces Stability-Aware Alignment (SAA) to maintain a per-scale stable reference from normal data, guiding dynamic graphs to suppress noise, achieving optimal performance on PSM and WADI benchmarks.
35. Long-Term Multi-Session 3D Reconstruction Under Substantial Appearance Change
Core Problem: Existing Structure-from-Motion (SfM) pipelines fail in long-term environmental monitoring scenarios (e.g., coral reefs) due to their reliance on post-hoc alignment of independently reconstructed sessions and inability to handle substantial visual and structural appearance changes over months or years.
Key Innovation: Enforced cross-session correspondences directly within a joint SfM reconstruction by combining complementary handcrafted and learned visual features, enabling robust reconstruction of a single coherent 3D model from imagery captured years apart, and ensured scalability by restricting expensive learned feature matching to likely cross-session image pairs.
36. Hybrid Fusion: One-Minute Efficient Training for Zero-Shot Cross-Domain Image Fusion
Core Problem: Traditional image fusion methods lack adaptability and performance, while deep learning approaches, despite achieving state-of-the-art results, suffer from critical inefficiencies due to slow, resource-intensive, patch-based training, creating a significant gap with full-resolution inference.
Key Innovation: Proposes a novel hybrid framework that resolves this trade-off by utilizing a learnable U-Net to generate a dynamic guidance map that directs a classic, fixed Laplacian pyramid fusion kernel. This decoupling enables remarkably efficient full-resolution training (approx. one minute), eliminates the train-inference gap, achieves SOTA-comparable performance, and demonstrates powerful zero-shot generalization across diverse tasks.
37. T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation
Core Problem: Imputing missing values in multivariate time series is challenging, especially under diverse and heavy missing patterns, as corrupted temporal features hinder effective cross-variable information transfer, amplifying reconstruction errors.
Key Innovation: T1 (Time series imputation with 1-to-1 channel-head binding), a CNN-Transformer hybrid architecture that achieves robust imputation through Channel-Head Binding, enabling selective information transfer by adaptively down-weighting corrupted temporal patterns while preserving reliable cross-variable connections through unaffected channels.
38. Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones
Core Problem: Fast-flying drones suffer from severe motion blur in images and significant drift/noise in pose estimates, making high-fidelity 3D reconstruction with Neural Radiance Fields (NeRFs) challenging for rapid inspection tasks.
Key Innovation: A unified framework that leverages asynchronous event streams alongside motion-blurred frames to reconstruct high-fidelity radiance fields from agile drone flights, by embedding event-image fusion into NeRF optimization and jointly refining event-based visual-inertial odometry priors, recovering sharp radiance fields and accurate camera trajectories without ground-truth supervision.
39. Spa3R: Predictive Spatial Field Modeling for 3D Visual Reasoning
Core Problem: Vision-Language Models (VLMs) have superficial 3D spatial understanding, and current methods for bridging this gap either rely on explicit 3D modalities or burden the language model with implicit 3D reconstruction from sparse cues.
Key Innovation: Spa3R, a self-supervised framework that learns a unified, view-invariant spatial representation directly from unposed multi-view images using Predictive Spatial Field Modeling (PSFM), enabling VLMs to ground language reasoning in a global spatial context and achieve state-of-the-art 3D VQA.
40. LST-SLAM: A Stereo Thermal SLAM System for Kilometer-Scale Dynamic Environments
Core Problem: Thermal Simultaneous Localization and Mapping (SLAM) remains difficult in dynamic large-scale outdoor environments due to unreliable feature extraction, unstable motion tracking, and inconsistent global pose and map construction.
Key Innovation: LST-SLAM, a novel large-scale stereo thermal SLAM system combining self-supervised thermal feature learning, stereo dual-level motion tracking, geometric pose optimization, and a semantic-geometric hybrid constraint to suppress dynamic features. It also uses an online incremental bag-of-words model for loop closure.
41. Not Just How Much, But Where: Decomposing Epistemic Uncertainty into Per-Class Contributions
Core Problem: Bayesian deep learning summarizes epistemic uncertainty with a single scalar (mutual information), which cannot distinguish whether a model's ignorance involves a benign or safety-critical class, making it insufficient for safety-critical classification with asymmetric failure costs.
Key Innovation: Decomposes mutual information into a per-class vector C_k(x), which quantifies epistemic uncertainty for each class, improving selective prediction (e.g., reducing selective risk by 34.7% for diabetic retinopathy) and out-of-distribution detection, and revealing asymmetric shifts invisible to standard MI.
42. Towards Robust Real-World Multivariate Time Series Forecasting: A Unified Framework for Dependency, Asynchrony, and Missingness
Core Problem: Real-world multivariate time series forecasting faces fundamental challenges of complex inter-channel dependencies, asynchronous channel sampling, and pervasive missing values, which existing architectures typically address only partially or with simplifying assumptions.
Key Innovation: Proposes ChannelTokenFormer, a Transformer-based forecasting framework with a flexible architecture designed to explicitly capture cross-channel interactions, accommodate channel-wise asynchronous sampling, and effectively handle missing values. It demonstrates superior robustness and accuracy on public benchmark datasets and a private real-world industrial dataset under challenging conditions.
43. Peering into the Unknown: Active View Selection with Neural Uncertainty Maps for 3D Reconstruction
Core Problem: Active view selection (AVS) for 3D reconstruction remains challenging, as it requires identifying the minimal set of views for accurate and efficient 3D reconstruction without the high computational overhead of learning radiance fields or 3D Gaussian Splatting for uncertainty estimation.
Key Innovation: Introduces a novel AVS approach guided by neural uncertainty maps predicted by UPNet, a lightweight feedforward deep neural network. UPNet takes a single input image and outputs a predicted uncertainty map, which is then aggregated to suppress redundant candidate viewpoints and select the most informative ones. This achieves comparable reconstruction accuracy with half the viewpoints and significantly reduces computational overhead (up to 400x speedup) compared to baseline methods, generalizing to novel object categories.
44. FREQuency ATTribution: benchmarking frequency-based occlusion for time series data
Core Problem: Deep neural networks, despite their performance, are black boxes, and existing interpretability methods do not adequately address the analysis of time-series-based networks, particularly regarding robustness to signal fluctuations.
Key Innovation: Presents FreqAtt, a framework for post-hoc interpretation of time-series analysis that uses frequency-based attribution. It evaluates relevant frequencies and either filters the signal or marks relevant input data, demonstrating that frequency-based attribution, especially when combined with traditional attribution on frequency-optimized signals, provides strong performance and robustness across different metrics.
45. Chlorophyll-a Mapping and Prediction in the Mar Menor Lagoon Using C2RCC-Processed Sentinel 2 Imagery
Core Problem: Traditional in situ measurements for monitoring chlorophyll-a (Chl-a) in coastal lagoons like Mar Menor are spatially and temporally limited, hindering the anticipation of harmful algal blooms (eutrophication) and effective mitigation.
Key Innovation: An end-to-end, validated methodology for high-resolution, depth-specific Chl-a mapping and prediction using C2RCC-processed Sentinel 2 imagery integrated with buoy-based ground truth and various ML/DL algorithms, enhancing early-warning capabilities for eutrophication.
46. SpecAware: A Spectral-Content Aware Foundation Model for Unifying Multi-Sensor Learning in Hyperspectral Remote Sensing Mapping
Core Problem: The inherent heterogeneity of Hyperspectral Imaging (HSI) data, particularly spectral channel variation across sensors, limits model generalization and adaptability for cross-sensor joint learning in land-use and land-cover (LULC) mapping. Existing HSI foundation models underutilize sensor meta-attributes and image semantic features.
Key Innovation: Introduces SpecAware, a novel spectral-content aware foundation model for unifying multi-sensor HSI learning. It uses a meta-content aware module and a HyperEmbedding module with a sample-conditioned hypernetwork to dynamically generate matrix factors for channel-wise encoding, enabling adaptive processing of variable spectral channels within a unified multi-sensor joint pre-training framework. Also introduces the Hyper-400K dataset.
47. Changes in Real Time: Online Scene Change Detection with Multi-View Fusion
Core Problem: Online Scene Change Detection (SCD) from unconstrained viewpoints is an extremely challenging problem, with existing online methods being significantly less accurate than offline approaches.
Key Innovation: Introduces the first online SCD approach that is pose-agnostic, label-free, and multi-view consistent, operating at over 10 FPS and surpassing state-of-the-art offline methods. It uses a self-supervised fusion loss, PnP-based fast pose estimation, and a fast change-guided update strategy for 3D Gaussian Splatting.
48. Ecological mapping with geospatial foundation models
Core Problem: The value, limitations, and practical considerations of Earth observation foundation models for high-impact ecological applications, such as land use and land cover mapping, remain insufficiently characterized, and traditional baselines may lack generalization and transferability.
Key Innovation: Systematically evaluates and demonstrates that fine-tuned geospatial foundation models (Prithvi-EO-2.0 and TerraMind) consistently outperform traditional baselines (ResNet-101) in ecological mapping tasks, showing improved generalization and transfer across domains, while highlighting the importance of dataset alignment and higher-resolution inputs.
49. Closure to “Instrumentation and Monitoring of a Steel-Reinforced MSE Wall”
Core Problem: Engaging with the findings and methodologies presented in the original paper on instrumentation and monitoring of steel-reinforced MSE walls.
Key Innovation: Providing additional insights, clarifications, or critical perspectives on the instrumentation and monitoring techniques for steel-reinforced MSE walls.
50. Discussion of “Instrumentation and Monitoring of a Steel-Reinforced MSE Wall”
Core Problem: Engaging with the findings and methodologies presented in the original paper on instrumentation and monitoring of steel-reinforced MSE walls.
Key Innovation: Providing additional insights, clarifications, or critical perspectives on the instrumentation and monitoring techniques for steel-reinforced MSE walls.
51. Community vulnerability to dust-related and industrial air pollution in Kuwait
Core Problem: Assessing the vulnerability of communities in Kuwait to the combined challenges of frequent dust storms and intensive industrial air pollution.
Key Innovation: The paper likely provides an assessment of community vulnerability to dust storms and industrial air pollution in Kuwait, though specific methods or findings are not detailed in the abstract.
52. Physics-informed neural networks for groundwater: evidence, limits, and a roadmap
Core Problem: The need for a systematic review of Physics-informed neural networks (PINNs) in groundwater science to understand current research hotspots, strengths, and methodological bottlenecks.
Key Innovation: Presents a dual-track framework (bibliometric analysis + methodological synthesis) to review PINNs in groundwater, identifying research hotspots, major strengths, key bottlenecks (training stability, efficiency, boundary conditions), and proposing a development roadmap for future advancements.
53. Meso- and macroscale investigation on impact penetration mechanisms in weathered sand
Core Problem: The mechanisms governing impact penetration in highly weathered sand are poorly understood, limiting advances in rapid, invasive impact penetration technologies for assessing in situ soil mechanical properties in remote or hazardous environments.
Key Innovation: A meso- and macroscale investigation integrating laboratory experiments, high-speed imaging (DIC), and DEM simulations revealed a unique 'twin peaks' phenomenon in deceleration and cone tip resistance for highly weathered sand, elucidating the underlying mechanisms and advancing theoretical support for CIPT interpretation.
54. Interface behaviour of fibre optic strain sensors in cement-improved soil
Core Problem: Understanding the interface behavior of distributed fibre optic strain sensors (DFOSS) in cement-improved and generally layered/heterogeneous ground conditions is crucial for accurate strain measurements and effective monitoring.
Key Innovation: Experimental pull-out and column compression tests, interpreted with analytical and finite element models, demonstrated that sensor type significantly influences interface bond strength and shear stiffness, revealing localized strain decoupling and fully coupled zones in layered soils, applicable to DFOSS in heterogeneous ground.
55. Spectral reflectance time series of four soils while drying from over-saturated to air-dry
Core Problem: Understanding the relationship between spectral changes and underlying soil moisture status as soil dries, particularly the slow and subtle changes during much of the drying time and the rapid changes when pore water is depleted.
Key Innovation: Documentation of time series laboratory measurements of spectral reflectance for four distinct soils drying from oversaturated to air-dry, describing continuous changes, documenting a spectral Fresnel component, and presenting a rationale for a metric to detect the end of constant evaporation, providing insights for remote sensing of soil moisture and characterization.
56. Three-dimensional numerical investigation of blast-induced rock failure under in-situ stress using the material point method
Core Problem: Understanding the complex interaction between explosive stress waves and pre-existing geological stresses in blast-induced rock failure, especially under deep in-situ stress conditions, is challenging due to the difficulty in analyzing stress wave propagation, crack formation, and failure mechanisms.
Key Innovation: Development of a 3D computational framework based on coupled GIMP and CPDI material point methods, incorporating novel algorithms for in-situ stress initialization and non-reflective boundary conditions, which revealed distinct stages of blast-induced rock damage and the significant influence of in-situ stress on tensile failure and overall failure morphology.
57. Global Imaging of the Mantle Transition Zone Using SS Precursors Shows Lower Mantle Convection Patterns that are Broadly Linked to Tectonics
Core Problem: Understanding mantle convection and material transport, particularly the role of mantle transition zone thickness variations, is central to knowledge of geochemical reservoirs, hydration cycles, and Earth's evolution, but global imaging with excellent coverage and robust models is needed.
Key Innovation: Uses SS precursors and a newly developed method with extensive seismic data (1990–2021) to globally image the mantle transition zone, revealing thickened zones beneath subduction zones and thinned zones beneath oceans, broadly linking lower mantle convection patterns to tectonics and thermal perturbations.
58. Numerical Modeling of Purely Active (Plume‐Produced) Continental Rifting and Break‐Up
Core Problem: The processes of purely active continental rifting and break-up, driven solely by rising mantle plumes without far-field tectonic forces, are still poorly understood despite their potential link to supercontinent fragmentation.
Key Innovation: Presents systematic numerical modeling demonstrating that purely active continental rifting and break-up can be achieved within 10 Myr, but only under specific conditions: a continuously fed, buoyant plume (Δρ ≤ 30 kgm⁻³) and a relatively warm, weak continental plate (Moho temperature ≥ 750°C), suggesting external tectonic stresses are usually required for successful break-up in the Phanerozoic.
59. In-context Pre-trained Time-Series Foundation Models adapt to Unseen Tasks
Core Problem: Existing Time-series Foundation Models (TSFMs) typically require fine-tuning to adapt to new, unseen tasks, limiting their real-world applicability and generalization capabilities.
Key Innovation: Proposes In-Context Time-series Pre-training (ICTP) to augment TSFMs with In-Context Learning (ICL) capabilities, enabling them to dynamically adapt to unseen time-series tasks during test-time inference without requiring fine-tuning.
60. GSNR: Graph Smooth Null-Space Representation for Inverse Problems
Core Problem: Ill-posed inverse problems in imaging lead to infinitely many solutions due to the unconstrained null-space of the sensing matrix, causing common image priors to bias reconstructions.
Key Innovation: Proposes Graph-Smooth Null-Space Representation (GSNR), a mechanism that imposes structure only into the invisible (null-space) component of the signal using a null-restricted graph Laplacian and low-dimensional spectral graph modes, significantly improving reconstruction quality in various inverse problems.
61. Large-scale Photorealistic Outdoor 3D Scene Reconstruction from UAV Imagery Using Gaussian Splatting Techniques
Core Problem: Existing 3D reconstruction methods from UAVs, particularly NeRF-based approaches, suffer from high latency and lower rendering performance, hindering real-time, high-fidelity applications for aerial perception.
Key Innovation: An end-to-end pipeline combining live UAV video acquisition, synchronized sensor fusion, camera pose estimation, and 3D Gaussian Splatting optimization to achieve continuous, low-latency, high-fidelity 3D scene reconstruction suitable for real-time augmented perception applications.
62. CLIPoint3D: Language-Grounded Few-Shot Unsupervised 3D Point Cloud Domain Adaptation
Core Problem: Vision-language models (VLMs) like CLIP are fragile under domain shifts, particularly when adapting from synthetic to real-world 3D point clouds, and conventional 3D domain adaptation methods are often inefficient.
Key Innovation: Introduces CLIPoint3D, the first framework for few-shot unsupervised 3D point cloud domain adaptation built upon CLIP, which leverages a frozen CLIP backbone, knowledge-driven prompt tuning, entropy-guided view sampling, and optimal transport-based alignment losses to bridge source-target distribution gaps efficiently and effectively.
63. VINA: Variational Invertible Neural Architectures
Core Problem: Normalizing Flows (NFs) and Invertible Neural Networks (INNs) lack theoretical guarantees on approximation quality under realistic assumptions, whether for posterior inference in INNs or for generative modeling with NFs.
Key Innovation: Introduces a unified framework for INNs and NFs based on variational unsupervised loss functions, deriving theoretical performance guarantees for posterior and distributional accuracy under weaker, more practically realistic assumptions. It provides general design principles and practical guidelines, demonstrating effectiveness on a realistic ocean-acoustic inversion problem.
64. Pip-Stereo: Progressive Iterations Pruner for Iterative Optimization based Stereo Matching
Core Problem: Iterative stereo matching, while accurate, is computationally intensive and difficult to deploy on edge hardware due to its dependence on Recurrent Neural Networks (RNN) and redundant computations.
Key Innovation: Introduces a progressive iteration pruning strategy, a collaborative monocular prior transfer framework, and FlashGRU (a hardware-aware RNN operator) to achieve real-time, high-fidelity stereo matching on edge devices with significant speedup and memory reduction.
65. PFGNet: A Fully Convolutional Frequency-Guided Peripheral Gating Network for Efficient Spatiotemporal Predictive Learning
Core Problem: Pure convolutional models for spatiotemporal predictive learning, while efficient, struggle to adaptively capture spatially varying motion patterns due to their fixed receptive fields.
Key Innovation: Proposes PFGNet, a fully convolutional framework that dynamically modulates receptive fields through pixel-wise frequency-guided gating. Its core Peripheral Frequency Gating (PFG) block extracts localized spectral cues and adaptively fuses multi-scale large-kernel peripheral responses with learnable center suppression, enabling structure-aware spatiotemporal modeling efficiently without recurrence or attention, achieving SOTA or near-SOTA performance with fewer parameters and FLOPs.
66. Real-time Motion Segmentation with Event-based Normal Flow
Core Problem: Directly processing sparse raw event data from event-based cameras for real-time vision tasks like motion segmentation is highly inefficient, severely limiting the applicability of state-of-the-art methods for dynamic scene understanding.
Key Innovation: Proposed a normal flow-based motion segmentation framework for event-based vision. This framework leverages dense normal flow learned from event neighborhoods, formulates motion segmentation as an energy minimization problem solved via graph cuts, and achieves real-time performance (nearly 800x speedup) and high accuracy by efficiently estimating motion models with a limited number of candidates.
67. RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction
Core Problem: Existing 3D Gaussian splatting SLAM methods struggle in dynamic environments with moving objects, hindering reliable tracking and 3D reconstruction. The potential of 4D reconstruction for 4D-aware SLAM remains largely underexplored.
Key Innovation: Proposes RU4D-SLAM, a robust and efficient framework for 4D scene reconstruction. It introduces temporal factors into 3D representation, incorporates uncertainty-aware perception of scene changes, motion blur rendering, and extends per-pixel uncertainty modeling to handle blurred images. A semantic-guided reweighting mechanism and a learnable opacity weight support adaptive 4D mapping.
68. DA-Cal: Towards Cross-Domain Calibration in Semantic Segmentation
Core Problem: Existing unsupervised domain adaptation (UDA) methods for semantic segmentation often neglect network calibration quality, leading to misalignment between prediction confidence and actual accuracy, which poses a significant risk in safety-critical applications. Performance degrades when soft pseudo-labels replace hard pseudo-labels due to poor calibration.
Key Innovation: Proposes DA-Cal, a dedicated cross-domain calibration framework that transforms target domain calibration into soft pseudo-label optimization. It introduces a Meta Temperature Network to generate pixel-level calibration parameters, employs bi-level optimization to establish the relationship between soft pseudo-labels and UDA supervision, and utilizes complementary domain-mixing strategies to prevent overfitting and reduce domain discrepancies.
69. Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting
Core Problem: Existing 3D Gaussian Splatting (3DGS) Dropout methods for sparse-view conditions suffer from a neighbor compensation effect, weakening regularization, and overlook the contribution of high-degree spherical harmonic coefficients (SH) to overfitting.
Key Innovation: DropAnSH-GS, a novel anchor-based Dropout strategy that simultaneously removes selected Gaussians and their spatial neighbors to disrupt local redundancies, and extends Dropout to color attributes by randomly dropping higher-degree SH, mitigating overfitting and enabling flexible post-training model compression.
70. BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting
Core Problem: Recovering accurate 3D boundary representation (B-rep) models from unstructured 2D image data, which is challenging for existing deep learning methods, especially regarding generalization and dependence on dense point clouds.
Key Innovation: BrepGaussian, a novel framework that learns 3D parametric representations from 2D images using a Gaussian Splatting renderer and a two-stage learning framework for disentangling geometry and feature learning, demonstrating superior performance in CAD reconstruction.
71. Region of Interest Segmentation and Morphological Analysis for Membranes in Cryo-Electron Tomography
Core Problem: Identifying and quantitatively analyzing regions of interest (ROIs) and complex, continuous structures like membranes in cryo-ET is challenging, often requiring indirect derivation from full structure segmentation.
Key Innovation: TomoROIS-SurfORA, a two-step framework for direct, shape-agnostic ROI segmentation (using deep learning trainable with small datasets) and quantitative morphological surface analysis (extracting features like inter-membrane distances, curvature, and roughness from point clouds and surface meshes), applicable to complex geometries.
72. Inspectorch: Efficient rare event exploration in solar observations
Core Problem: Large volumes of solar observation data cannot be fully analyzed with conventional methods, and popular machine learning techniques often overlook rare, unusual events due to their low frequency.
Key Innovation: Inspectorch, an open-source framework, utilizes flow-based models (flexible density estimators) to efficiently identify rare events in multidimensional solar observations by assigning probabilities to samples, allowing computational resources to be focused on physically relevant anomalies.
73. Comparing Implicit Neural Representations and B-Splines for Continuous Function Fitting from Sparse Samples
Core Problem: Limited direct comparisons of intrinsic representation capabilities between classical methods (B-splines) and emerging Implicit Neural Representations (INRs) for continuous function fitting from sparse samples.
Key Innovation: Empirical demonstration that INRs (with positional encoding) achieve lower normalized root-mean-squared error, sharper edge transitions, and fewer oscillatory artifacts than oracle-tuned B-splines for continuous function fitting from sparse data, supporting their superior representation capacity.
74. DANCE: Doubly Adaptive Neighborhood Conformal Estimation
Core Problem: Conformal prediction methods for classification often use logit scores, leading to inefficient or overly conservative prediction sets, especially for pre-trained models not calibrated to the target task.
Key Innovation: Introduction of DANCE, a doubly locally adaptive nearest-neighbor based conformal algorithm that combines two novel nonconformity scores directly from data embeddings, fitting a task-adaptive kernel regression model to produce efficient and robust prediction sets for uncertainty quantification.
75. Two Models for Surface Segmentation using the Total Variation of the Normal Vector
Core Problem: Efficiently and accurately segmenting surfaces represented by triangular meshes based on the similarity of their normal vector field to a given set of label vectors, especially in the presence of noise.
Key Innovation: Proposes two variational models for surface segmentation using different total variation measures of the normal vector as regularizers, optimized with split Bregman (ADMM) and a manifold Newton scheme, demonstrating improved noise removal and computational efficiency for the second model, with potential for terrain analysis.
76. Light of Normals: Unified Feature Representation for Universal Photometric Stereo
Core Problem: Existing universal photometric stereo methods struggle to decouple illumination and normal information, leading to loss of high-frequency geometric details and poor generalization under arbitrary, unknown lighting conditions.
Key Innovation: LINO UniPS introduces Light Register Tokens with light alignment supervision and an Interleaved Attention Block for better feature decoupling. It also uses a Wavelet-based Dual-branch Architecture and Normal-gradient Perception Loss to preserve finer geometric details, along with a new synthetic dataset (PS-Verse) and curriculum training.
77. Seeing Through the Noise: Improving Infrared Small Target Detection and Segmentation from Noise Suppression Perspective
Core Problem: Existing CNN-based methods for Infrared Small Target Detection and Segmentation (IRSTDS) primarily focus on feature enhancement, leading to increased false alarms due to noise, and fail to effectively suppress noise from a frequency domain perspective.
Key Innovation: Proposes NS-FPN (noise-suppression feature pyramid network) which integrates a low-frequency guided feature purification (LFP) module to suppress noise features and a spiral-aware feature sampling (SFS) module to fuse target-relevant features, significantly reducing false alarms and improving performance.
78. Uncertainty Propagation Networks for Neural Ordinary Differential Equations
Core Problem: Existing neural Ordinary Differential Equations (ODEs) predict only state trajectories, lacking explicit uncertainty quantification in continuous-time modeling, which is crucial for robust predictions.
Key Innovation: Introduces Uncertainty Propagation Networks (UPN), a novel family of neural differential equations that simultaneously model both state evolution and its associated uncertainty by parameterizing coupled differential equations for mean and covariance dynamics, enabling principled uncertainty quantification and handling irregular data.
79. Unbiased Object Detection Beyond Frequency with Visually Prompted Image Synthesis
Core Problem: Prior debiasing methods for object detection are limited by sample representation diversity, and naive generative augmentation often preserves biases. Simply generating more data for rare classes is suboptimal due to incomplete proxies for data needs and lack of fidelity in layout-to-image synthesis.
Key Innovation: A generation-based debiasing framework for object detection that introduces a representation score (RS) to diagnose representational gaps beyond frequency, and employs visually prompted image synthesis with generative alignment for high-quality, unbiased layout generation, significantly narrowing performance gaps for underrepresented object groups.
80. Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding
Core Problem: Lack of strong open-weight and open-data video-language models (VLMs), especially those with grounding capabilities (pointing/tracking in pixels), hinders progress in the open-source community.
Key Innovation: Molmo2, a new family of open-source VLMs that are state-of-the-art among open-source models and demonstrate exceptional new capabilities in point-driven grounding across image and video tasks, supported by a collection of 7 new video datasets and 2 multi-image datasets collected without proprietary VLMs.
81. All-Optical Segmentation via Diffractive Neural Networks for Autonomous Driving
Core Problem: Conventional deep neural networks for semantic segmentation and lane detection in autonomous driving incur high energy costs due to extensive analog-to-digital conversions and large-scale image computations, hindering low-latency, real-time responses.
Key Innovation: Proposes an all-optical computing framework for RGB image segmentation and lane detection using Diffractive Optical Neural Networks (DONNs), which performs image processing via light diffraction at the speed of light, significantly reducing computation energy costs and analog-to-digital conversion overhead.
82. Universal 3D Shape Matching via Coarse-to-Fine Language Guidance
Core Problem: Establishing dense semantic correspondences between strongly non-isometric 3D shapes across different object categories remains challenging, with prior approaches limited to near-isometric assumptions and homogeneous subject types.
Key Innovation: Proposing UniMatch, a coarse-to-fine framework that leverages class-agnostic 3D segmentation, multimodal large language models (MLLMs) for part identification, and vision language models (VLMs) for text embeddings to guide the learning of dense correspondences via a rank-based contrastive scheme, enabling universal matching for inter-class and non-isometric shapes.
83. An Efficient LiDAR-Camera Fusion Network for Multi-Class 3D Dynamic Object Detection and Trajectory Prediction
Core Problem: Service mobile robots require efficient and accurate 3D dynamic object detection and trajectory prediction in complex environments, often with limited computational resources.
Key Innovation: An efficient multi-modal framework integrating LiDAR and camera inputs, featuring a Unified modality detector with Mamba and Transformer (UniMT) for high-accuracy, fast 3D object detection, and a Reference Trajectory-based Multi-Class Transformer (RTMCT) for efficient and diverse trajectory prediction, achieving real-time performance on a wheelchair robot.
84. Experimental Observations of Plate Anchor Keying Process in Transparent Soil: Performance of Different Keying Flap Designs
Core Problem: Difficulty in directly observing the keying process of plate anchors in opaque natural soils, hindering the understanding of anchor capacity mobilization and optimal design for offshore engineering.
Key Innovation: Utilization of a transparent soil technique for direct visual observation of plate anchor keying, comparing different flap designs, and providing insights into anchor-soil interaction mechanisms for optimized offshore anchor designs.
85. WD-DCCAN: A wavelet-decomposed dual-channel cross attention network with transfer learning for intelligent prediction of short-term irregular motion of containership
Core Problem: Precise and efficient prediction of ship motions is essential for safe operations, but existing methods struggle with generalizability across diverse hull geometries and high training costs.
Key Innovation: Proposing WD-DCCAN, a Wavelet-Decomposed Dual-Channel Cross Attention Network with transfer learning, which integrates wavelet transform and a dual-channel cross-attention mechanism for robust and efficient short-term multi-step motion prediction of containerships.
86. Extremely short-term ship motion prediction using real-sea motion datasets based on the VMD-LSTM-FDR network
Core Problem: Ship motion in real sea is inherently random and uncertain, significantly impacting operational safety, necessitating accurate extremely short-term prediction.
Key Innovation: Proposing an extremely short-term ship motion prediction model integrating Variational Mode Decomposition (VMD), LSTM networks, and a Frequency Domain Rectification (FDR) algorithm, achieving high accuracy across different ship types and sea conditions.
87. Spatial–Frequency Domain Joint Learning With Shape Constraints for Fine-Grained Aircraft Detection in SAR Imagery
Core Problem: Fine-grained aircraft detection in SAR images is challenging due to high azimuth sensitivity, discrete scattering, significant intraclass variance, and topological fragmentation, with existing methods not fully utilizing frequency-domain features.
Key Innovation: Proposes SAR-SFNet, a dual-domain feature learning architecture with shape constraints. It integrates fractional Gabor transform and Fourier's global cues for spatial-frequency domain joint learning and uses class-aware shape constraints to improve fine-grained aircraft detection performance in SAR imagery.
88. An Efficient Time Reversal Technique Based on CSF-MUSIC for Unexploded Ordinances Localization
Core Problem: Conventional time reversal (TR) techniques for UXO detection require large antenna arrays and have high computational complexity, limiting practical applicability and real-time performance.
Key Innovation: Proposes a common-offset-based space–frequency multiple signal classification (CSF-MUSIC) algorithm. It reconstructs the SF-MDM into a CSF-MDM, operates in a dual-antenna mode, and uses optimized iterative QR decomposition to reduce computational time by 89% while maintaining precise UXO localization and imaging.
89. Democratizing planetary-scale analysis: An ultra-lightweight Earth embedding database for accurate and flexible global land monitoring
Core Problem: The immense computational resources and storage volumes required for global-scale Earth Observation (EO) analysis, which hinder widespread scientific adoption and the execution of planetary-scale studies.
Key Innovation: Introduction of the Embedded Seamless Data (ESD), an ultra-lightweight, 30-m global Earth embedding database spanning 25 years, which achieves a ~340-fold data volume reduction while maintaining high reconstructive fidelity and outperforming raw reflectance data in land-cover classification tasks.
90. Analyzing the spatiotemporal distribution of climate extremes under the CMIP6 climate model in the upper Blue Nile Basin, Ethiopia
Core Problem: Assessing the current and future spatiotemporal distribution of climate extremes (temperature and rainfall) in the Upper Blue Nile Basin under CMIP6 climate models to understand their impacts and inform adaptation strategies.
Key Innovation: Applying bias correction methods and statistical tests (Sen’s slope, Mann-Kendall) to CMIP6 models to analyze trends in 15 standard climate extreme indices, identifying significant increasing trends in warm days/nights and heavy/very heavy precipitation in the study area.
91. An XGBoost-SHAP framework for interpretable and probabilistic flood susceptibility mapping
Core Problem: Flood susceptibility assessments require rigorous spatial analysis, but complex machine learning models often lack interpretability, hindering stakeholder confidence and broader adoption.
Key Innovation: Presented an interpretable flood susceptibility mapping framework by coupling Extreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP), providing precise insights into individual feature influence on flood probabilities and enhancing model transparency for disaster management planning.
92. Comparison of electrical resistivity tomography and frequency domain electromagnetic methods for mapping seawater intrusion in a shallow coastal aquifer (Northern Italy)
Core Problem: The limitations of ERT for large-area seawater intrusion mapping and the ongoing debate regarding the robustness and calibration of electromagnetic induction (EMI) data.
Key Innovation: Compares two FDEM instruments with ERT surveys, demonstrating FDEM's effectiveness for mapping shallow salinity when calibrated with ERT data, offering a faster and cost-effective alternative for long-distance profiles and 3-D spatial analysis of seawater intrusion.
93. Rate dependent evolution of fracture process zone in sandstone: Insights from large specimen double torsion tests
Core Problem: The rate dependence of fracture process zone evolution in quasi-brittle rocks, which is key to understanding fracture behavior, remains insufficiently understood.
Key Innovation: Meter-scale double torsion tests with a combined monitoring system quantified the rate-dependent evolution of the fracture process zone in sandstone, revealing a negative logarithmic relationship between its length and crack propagation rate, providing mechanical insight into rock fracture processes.
94. Robust reconstruction of seamless daily VIIRS nighttime light imagery with cloud mask refinement and multi-strategy spatiotemporal gap-filling
Core Problem: Pervasive data gaps in daily VIIRS nighttime light (NTL) data due to cloud contamination and poor-quality observations, which limit its potential for characterizing human activity changes and disaster impact assessment.
Key Innovation: A novel spatiotemporal gap-filling framework that refines the original VNP46 cloud mask and employs a multi-strategy approach leveraging spatiotemporal information from valid observations and existing data gaps, demonstrating robust performance in reconstructing NTL intensity and providing insights into post-hurricane recovery.
95. Scenario-adaptive cross-modal multistep temporal prediction of heat release rate in tunnel fires
Core Problem: Traditional temperature sensors for tunnel fire situation awareness are prone to thermal damage and failure in extreme fire conditions, limiting accurate heat release rate (HRR) prediction.
Key Innovation: A hybrid deep learning framework integrating computer vision (EfficientNet-B3) and temporal models (e.g., Seq2Seq) for scenario-adaptive, cross-modal, multi-step HRR prediction using flame image sequences, demonstrating high accuracy and robustness in diverse tunnel fire scenarios.
96. Groundwater drought in the United States: spatial and temporal variability
Core Problem: Understanding the spatial and temporal variability of groundwater drought across the contiguous United States (CONUS) and the limitations of current monitoring data (wells vs. remote sensing) for accurate representation.
Key Innovation: Quantified the spatial and temporal variability of groundwater drought (duration, severity, mean levels) across CONUS from 1981-2020 using well data and GRACE-DADM, highlighting disparities between data sources and the need for more long-term well monitoring.
97. A prototype hyper-resolution groundwater digital twin for the contiguous United States: integrating physics-based modeling, machine learning, and observations
Core Problem: Advancing large-scale, hyper-resolution groundwater modeling for operational decision-making, integrating diverse data sources and improving prediction accuracy of water table depth (WTD).
Key Innovation: Developed a prototype hyper-resolution groundwater digital twin for CONUS by training an adjusted random forest model to downscale physics-based simulations and bias-correct to observations, producing daily 1 arcsec WTD maps with improved accuracy and operational capability.
98. Porous grain-based modeling of coral reef limestone: revealing heterogeneity and strain rate effects
Core Problem: The strong heterogeneity and complex pore/crystal morphology of deep coral reef limestone (CRL) make it challenging to comprehensively understand its mechanical behavior, damage mechanisms, and the effects of strain rate, especially given the scarcity of deep CRL samples.
Key Innovation: Development of a refined porous grain-based modeling strategy using the finite-discrete element method, incorporating detailed pore characteristics from CT scans, to reveal the influence of pore structure, crystal morphology, and strain rate on the mechanical parameters and macro–micro damage mechanisms of deep CRL, addressing sample-dependent issues in numerical models.
99. Investigation of acid dissolution patterns along rough-walled fractures with effects of mineral compositions and fracture geometries
Core Problem: The mechanism of acid–rock fracture interaction and the resulting acid dissolution patterns in heterogeneous carbonate rocks with complex fracture geometries and mineral compositions are poorly understood, making it difficult to quantify acidizing efficiency for reservoir stimulation.
Key Innovation: Development of a novel hydrochemical coupled thin-layer acidizing model embedding a virtual rough-fracture component, which accurately captures acid dissolution patterns by incorporating nonlinear flow and multireaction processes, clarifying competitive interactions among carbonate heterogeneities and providing a quantitative basis for optimizing acid-treatment parameters.
100. Self‐Consistent Models of Earth's Mantle and Core From Long‐Period Seismic and Tidal Constraints
Core Problem: Fundamental questions about Earth's mantle and core structure, particularly radial anelastic seismic structure and associated uncertainties, remain, and existing radial seismic reference models often lack uncertainty estimates.
Key Innovation: Inverts a large set of normal-mode frequencies and quality factors, along with astronomic-geodetic data, using parameterized models and a stochastic sampling approach to quantify uncertainties, revealing considerable deviations from PREM (e.g., denser outer core, less dense inner core) and providing uncertainty measures on all inverted properties.
101. Quantifying Accretion of Intra‐Oceanic Arcs to Continent: Numerical Modeling of Their Crustal Composition and Rheological Property
Core Problem: The diversity in crustal composition and rheological properties of intra-oceanic arcs (IOAs) and how this diversity affects their accretion processes and efficiencies to continental margins, a key process in crustal growth, remains unclear.
Key Innovation: Conducted 2-D geodynamic modeling to explore how compositional differentiation and rheological stratification influence IOA accretion, revealing that nascent IOAs with partially molten zones at the Moho can achieve high accretion efficiencies (up to 76.4%), while mature arcs with cold thermal gradients show complete subduction and lowest efficiencies, aligning with geophysical observations and ancient orogenic belts.
102. Global Desert Variations During 1985–2024 Associated With Effective Water Availability
Core Problem: Uncertainty in estimates of global desert area variations under accelerating climate change, with current estimates differing significantly, and a lack of high-resolution, long-term global mapping.
Key Innovation: Presenting the first 40-year global mapping of desert area changes at 30m resolution using Landsat imagery, revealing a declining trend (2.27×10^4 km^2/yr) driven by episodic increases in water inputs and vegetation buffering, with climatic factors playing a primary role.
103. Glacier Equilibrium‐Line Altitude Change Across Alaska and Adjacent Canada Indicates a Cold, Dry Little Ice Age and Weaker Aleutian Low
Core Problem: Understanding hydroclimate shifts during past periods like the Little Ice Age (LIA) is key to projecting future glacier melt and sea-level rise, but the climatic signature of the LIA in Alaska and adjacent Canada remains to be fully quantified.
Key Innovation: Quantifying changes in equilibrium-line altitude (ΔELA) for 215 Alaskan glaciers from the LIA maximum to present using remote sensing and GIS, indicating a rise of 170±8m and suggesting the LIA was characterized by colder, drier conditions and a weak, westward-displaced Aleutian Low.
104. IMOVNO+: A Regional Partitioning and Meta-Heuristic Ensemble Framework for Imbalanced Multi-Class Learning
Core Problem: Class imbalance, overlap, and noise degrade data quality and model reliability in multi-class learning, especially where complex inter-class relationships make minority-majority structures unclear and traditional methods fail.
Key Innovation: IMOVNO+, a two-level framework that jointly enhances data quality (via conditional probability, regional partitioning, overlapping-cleaning, and smart oversampling) and algorithmic robustness (via meta-heuristic ensemble pruning) for imbalanced multi-class tasks, showing consistent superiority over state-of-the-art methods.
105. MultiModalPFN: Extending Prior-Data Fitted Networks for Multimodal Tabular Learning
Core Problem: Existing foundation models for tabular data, like TabPFN, struggle to integrate heterogeneous modalities such as images and text, limiting their applicability in real-world multimodal datasets.
Key Innovation: Presents Multi-Modal Prior-data Fitted Network (MMPFN), which extends TabPFN to handle tabular and non-tabular modalities in a unified manner. It uses per-modality encoders and modality projectors (multi-head gated MLP and cross-attention pooler) to transform non-tabular embeddings into tabular-compatible tokens, addressing attention imbalance and extracting richer context.
106. N4MC: Neural 4D Mesh Compression
Core Problem: Efficiently compressing time-varying 3D mesh sequences is challenging due to their irregular nature and high temporal redundancy, limiting storage, transmission, and real-time processing.
Key Innovation: Presents N4MC, a neural 4D compression framework that converts irregular mesh frames into regular 4D tensors, uses an auto-decoder for spatial and temporal correlation, and a transformer-based interpolation model for motion compensation, achieving superior rate-distortion performance and real-time decoding.
107. gQIR: Generative Quanta Image Reconstruction
Core Problem: Capturing high-quality images from only a few detected photons using SPAD sensors is challenging due to sparse, noisy, binary photon detections and noise statistics that differ from standard restoration pipelines.
Key Innovation: Presents gQIR, an approach that adapts large text-to-image latent diffusion models to the photon-limited domain of quanta burst imaging, leveraging structural and semantic priors while handling Bernoulli photon statistics and integrating latent-space restoration with burst-level spatio-temporal reasoning to produce photometrically faithful and perceptually pleasing reconstructions.
108. MedCLIPSeg: Probabilistic Vision-Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation
Core Problem: Medical image segmentation is challenging due to limited annotations, ambiguous anatomical features, and domain shifts, and the potential of vision-language models like CLIP for dense, text-guided medical image segmentation is underexplored.
Key Innovation: Presents MedCLIPSeg, a novel framework that adapts CLIP for robust, data-efficient, and uncertainty-aware medical image segmentation by leveraging patch-level CLIP embeddings through probabilistic cross-modal attention and a soft patch-level contrastive loss, demonstrating improved accuracy, efficiency, and generalizability across diverse medical datasets.
109. Elimination-compensation pruning for fully-connected neural networks
Core Problem: Traditional pruning techniques for neural networks assume that expendable weights have small impact, but this idea can be generalized. Simply removing weights might not be optimal; compensating for their removal with adjacent bias perturbations could improve efficiency and information preservation.
Key Innovation: Introduces a novel pruning method where the importance of each weight is computed considering the output behavior after an optimal perturbation of its adjacent bias. These perturbations are applied directly after weight removal, efficiently computable by automatic differentiation, demonstrating intrinsic efficiency in diverse machine learning scenarios.
110. SD4R: Sparse-to-Dense Learning for 3D Object Detection with 4D Radar
Core Problem: The inherent sparsity and noise of 4D radar point clouds present significant challenges for accurate 3D object detection, especially in scenes with a small number of points, and existing densification methods often lack robustness.
Key Innovation: Proposed SD4R, a novel framework that transforms sparse radar point clouds into dense representations using a foreground point generator (FPG) to mitigate noise and a logit-query encoder (LQE) to enhance pillarization for robust feature representations, achieving state-of-the-art 3D object detection.
111. MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision
Core Problem: Generating crisp (one-pixel-wide) edge maps remains a fundamental challenge in edge detection, as existing methods rely on non-differentiable, hand-crafted post-processing algorithms (NMS, thinning) that hinder end-to-end optimization.
Key Innovation: Proposed MatchED, a lightweight and plug-and-play matching-based supervision module that enables joint end-to-end learning of crisp edges by performing one-to-one matching between predicted and ground-truth edges, substantially improving performance and achieving state-of-the-art results without traditional post-processing.
112. Deep unfolding of MCMC kernels: scalable, modular & explainable GANs for high-dimensional posterior sampling
Core Problem: Markov chain Monte Carlo (MCMC) methods are computationally intensive, especially in high-dimensional Bayesian computation, while existing push-forward generative models like GANs lack modularity and generalizability to changes in the likelihood function.
Key Innovation: Introduced a novel GAN architecture design by applying deep unfolding to Langevin MCMC algorithms, creating scalable, modular, and explainable models for high-dimensional posterior sampling. This design allows key model parameters to be specified at inference time, offering robustness to likelihood changes, and achieves high sampling accuracy and computational efficiency while retaining physics consistency and interpretability.
113. Computing a Characteristic Orientation for Rotation-Independent Image Analysis
Core Problem: Standard neural networks lack inherent rotation invariance, requiring computationally demanding data augmentation or specialized architectural modifications to achieve robustness to geometric transformations in computer vision.
Key Innovation: General Intensity Direction (GID), a preprocessing method that estimates a global orientation for each image and aligns it to a canonical reference frame, allowing standard models to process inputs consistently across rotations without modifying network architecture while preserving spatial structure.
114. MAST: A Multi-fidelity Augmented Surrogate model via Spatial Trust-weighting
Core Problem: Existing multi-fidelity surrogate modeling methods suffer from expensive training or rely on global correlation assumptions, leading to poor performance when fidelity relationships vary spatially or under tight budget constraints.
Key Innovation: Introduces MAST, a method that blends corrected low-fidelity observations with high-fidelity predictions using explicit discrepancy modeling and distance-based weighting, producing a single heteroscedastic Gaussian process that shows robust performance across varying budgets and fidelity gaps.
115. Le-DETR: Revisiting Real-Time Detection Transformer with Efficient Encoder Design
Core Problem: Current real-time Detection Transformer (DETR) models are challenging to reproduce due to excessive pre-training overheads on backbones, hindering research into novel architectures for high-accuracy, low-latency object detection.
Key Innovation: Le-DETR, a real-time Detection Transformer that achieves state-of-the-art performance with significantly reduced pre-training costs (80% fewer images) by proposing EfficientNAT backbones and redesigning the hybrid encoder with local attention, enhancing both performance and inference speed.
116. Optimizing Occupancy Sensor Placement in Smart Environments
Core Problem: Achieving accurate real-time zone occupancy recognition for energy savings in smart environments requires careful sensor placement, but existing methods lack an automatic optimization approach that considers geometric constraints and predicts accuracy.
Key Innovation: An automatic sensor placement method that determines optimal sensor layouts for a given number of sensors and predicts counting accuracy, formulated as an integer linear programming (ILP) problem and solved with the branch and bound method, demonstrated on various office environments.
117. Seeing Through Words: Controlling Visual Retrieval Quality with Language Models
Core Problem: Short and underspecified user queries in text-to-image retrieval lead to semantic ambiguity, visual collisions, and lack explicit control over the quality of retrieved images.
Key Innovation: A quality-controllable retrieval paradigm that leverages a generative language model to enrich short queries with contextual details and explicit image quality notions, enabling flexible, transparent, and controllable retrieval results compatible with existing VLMs.
118. Mask-HybridGNet: Graph-based segmentation with emergent anatomical correspondence from pixel-level supervision
Core Problem: Graph-based medical image segmentation requires manually annotated landmarks with point-to-point correspondences, which are rarely available, hindering clinical adoption despite its benefits for anatomical structure representation.
Key Innovation: Mask-HybridGNet, a framework that trains graph-based models directly using standard pixel-wise masks, eliminating the need for manual landmark annotations, and implicitly learns stable anatomical correspondences across patients.
119. Standard Transformers Achieve the Minimax Rate in Nonparametric Regression with $C^{s,\lambda}$ Targets
Core Problem: Lack of rigorous theoretical investigation into the capabilities of standard Transformer models, specifically their ability to approximate functions and achieve optimal rates in nonparametric regression.
Key Innovation: First work proving that standard Transformers can approximate H"older functions with arbitrary precision and achieve the minimax optimal rate in nonparametric regression for these target functions, providing theoretical justification for their powerful capabilities.
120. Amortized Bayesian inference for actigraph time sheet data from mobile devices
Core Problem: Need for statistical methods for high-resolution actigraph data that are congruent with AI frameworks for transfer learning and amortization, ensuring full propagation and quantification of uncertainty.
Key Innovation: Devising amortized Bayesian inference using a hierarchical dynamic linear model for actigraph time sheets, enabling probabilistic imputation and statistical learning of time-varying impacts of explanatory variables on acceleration magnitude for a cohort of subjects.
121. F10.7 Index Prediction: A Multiscale Decomposition Strategy with Wavelet Transform for Performance Optimization
Core Problem: Improving the accuracy and generalization capability of F10.7 index forecasting, as existing methods may lack optimal performance and generalization across different solar activity conditions.
Key Innovation: Proposing a novel F10.7 index forecasting method using wavelet decomposition to feed multiscale signals (F10.7, approximate, and detail signals) into an iTransformer model, demonstrating significant performance improvements over baselines and superior generalization, marking the first application of wavelet decomposition in F10.7 prediction.
122. Functional Continuous Decomposition
Core Problem: Traditional smoothing algorithms for non-stationary time-series data lack parametric optimization with guaranteed continuity and physical interpretability for analyzing local and global patterns.
Key Innovation: Functional Continuous Decomposition (FCD), a JAX-accelerated framework that performs parametric, continuous optimization on mathematical functions, transforming raw time-series data into M modes with up to C1 continuous fitting. FCD features also enhance CNN performance.
123. Empirically Calibrated Conditional Independence Tests
Core Problem: Conditional independence tests (CIT) often fail to provide frequentist guarantees in practice due to inaccurate asymptotic guarantees in small samples or skewed test behavior from unaccounted dependencies in large samples with misspecified models.
Key Innovation: Empirically Calibrated Conditional Independence Tests (ECCIT), a method that optimizes an adversary to select features and response functions to maximize a miscalibration metric, then fits a monotone calibration map to adjust base-test p-values, achieving valid FDR with higher power.
124. A Benchmark for Deep Information Synthesis
Core Problem: Current LLM evaluation benchmarks do not adequately assess agents' ability to solve real-world tasks requiring deep information synthesis and inference beyond simple fact retrieval from multiple sources.
Key Innovation: Introduces DEEPSYNTH, a novel benchmark with 120 realistic, time-consuming tasks across 7 domains and 67 countries, requiring information gathering, synthesis, and structured reasoning, revealing that current agents struggle with hallucinations and reasoning over large information spaces.
125. Revisiting the Generalization Problem of Low-level Vision Models Through the Lens of Image Deraining
Core Problem: Low-level vision models struggle to generalize to unseen image degradations, and the underlying mechanism for this failure is not primarily limited network capacity but rather a 'shortcut learning' phenomenon driven by the relative complexity between image content and degradation patterns.
Key Innovation: Reveals that generalization issues are caused by 'shortcut learning' and proposes two principled strategies: balancing the complexity of training data to redirect network focus toward content reconstruction, and leveraging strong content priors from pre-trained generative models to physically constrain the network onto a high-quality image manifold, improving robustness and generalization.
126. HoloLLM: Multisensory Foundation Model for Language-Grounded Human Sensing and Reasoning
Core Problem: Embodied agents in smart homes need to understand human behavior through diverse sensory inputs and natural language, but existing Vision-Language Models (VLMs) are limited by visual data, and there is a scarcity of aligned modality-text data for rare sensors (LiDAR, infrared, mmWave radar, WiFi) and heterogeneity in their physical signal representations.
Key Innovation: Introduces HoloLLM, a Multimodal Large Language Model (MLLM) that integrates uncommon sensing modalities (LiDAR, infrared, mmWave radar, WiFi) for language-grounded human perception and reasoning. It uses a Universal Modality-Injection Projector (UMIP) to enhance modality embeddings with fine-grained, text-aligned features and a human-VLM collaborative data curation pipeline to generate paired textual annotations, significantly improving sensing accuracy.
127. Trajectory-aware Shifted State Space Models for Online Video Super-Resolution
Core Problem: Existing online video super-resolution (VSR) methods are limited in long-range temporal modeling, often using only one previous frame, and struggle with efficient spatio-temporal information aggregation.
Key Innovation: TS-Mamba (Trajectory-aware Shifted State Space Models) leverages long-term trajectory modeling and low-complexity Mamba to efficiently aggregate spatio-temporal information. It constructs video trajectories to select similar tokens and uses a Trajectory-aware Shifted Mamba Aggregation (TSMA) module with shifted SSMs blocks to enhance spatial continuity and supervise trajectory generation.
128. Learning Unified Representations from Heterogeneous Data for Robust Heart Rate Modeling
Core Problem: Heart rate prediction faces significant challenges from data heterogeneity, stemming from diverse device sources (varying feature sets) and distinct user physiological patterns, limiting real-world performance.
Key Innovation: Proposes a framework that learns unified latent representations robust to both source and user heterogeneity, employing a random feature dropout strategy for device variations and a history-aware attention module with contrastive learning for individual physiological differences.
129. Spatial-DISE: A Unified Benchmark for Evaluating Spatial Reasoning in Vision-Language Models
Core Problem: Existing benchmarks are inadequate for assessing the spatial reasoning ability of Vision Language Models (VLMs), particularly the intrinsic-dynamic spatial reasoning crucial for human-like spatial cognition and real-world applications.
Key Innovation: Spatial-DISE, a unified benchmark and dataset based on a cognitively grounded taxonomy (Intrinsic-Static, Intrinsic-Dynamic, Extrinsic-Static, and Extrinsic-Dynamic spatial reasoning), along with a scalable and automated pipeline to generate diverse and verifiable spatial reasoning questions.
130. CuriGS: Curriculum-Guided Gaussian Splatting for Sparse View Synthesis
Core Problem: Extending 3D Gaussian Splatting (3DGS) to sparse-view settings remains challenging due to supervision scarcity and overfitting caused by limited viewpoint coverage.
Key Innovation: Presents CuriGS, a curriculum-guided framework for sparse-view 3D reconstruction using 3DGS. It introduces student views (pseudo-views sampled around ground-truth poses) with varying perturbation levels, regularized via depth-correlation and co-regularization, and periodically promotes best-performing students to the training set to augment sparse training views.
131. Pluggable Pruning with Contiguous Layer Distillation for Diffusion Transformers
Core Problem: Diffusion Transformers (DiTs) have high computational costs due to large parameter counts, hindering deployment in resource-constrained settings for image generation.
Key Innovation: A flexible structured pruning framework (PPCL) for DiT architectures that identifies redundant layer intervals and uses a plug-and-play teacher-student alternating distillation scheme to achieve a 50% parameter reduction with less than 3% degradation in key objective metrics.
132. Less is More: Data-Efficient Adaptation for Controllable Text-to-Video Generation
Core Problem: Fine-tuning large-scale text-to-video diffusion models for new generative controls (e.g., camera parameters) typically requires vast, high-fidelity datasets that are difficult to acquire.
Key Innovation: A data-efficient fine-tuning strategy that learns new generative controls from sparse, low-quality synthetic data, yielding superior results compared to models fine-tuned on photorealistic 'real' data.
133. Seeing What Matters: Visual Preference Policy Optimization for Visual Generation
Core Problem: Existing reinforcement learning pipelines for visual generative models rely on a single scalar reward per sample, ignoring the rich spatial and temporal structure of visual content, which hinders correction of localized artifacts and modeling of fine-grained perceptual cues.
Key Innovation: Visual Preference Policy Optimization (ViPO), a GRPO variant that lifts scalar feedback into structured, pixel-level advantages using a Perceptual Structuring Module, redistributing optimization pressure toward perceptually important regions and improving alignment with human preferences.
134. Classification and reconstruction for single-pixel imaging with classical and quantum neural networks
Core Problem: Traditional CMOS/CCD cameras face challenges in imaging outside the visible spectrum, and accelerating high-dimensional single-pixel visualization for practical applications is needed, potentially via quantum machine learning.
Key Innovation: Simulation of single-pixel imaging using Hadamard basis patterns combined with classical and parameterized quantum neural networks for image classification and reconstruction, demonstrating competitive classification accuracies and exploring quantum machine learning's potential.
135. Characterizing State Space Model and Hybrid Language Model Performance with Long Context
Core Problem: Transformer architectures suffer from quadratic computational and memory overhead, hindering their use for processing continuous and/or long-context inputs on local devices, necessitating new architectures like State Space Models (SSMs) and hybrid models.
Key Innovation: A comprehensive comparative benchmarking revealing that SSMs and hybrid models offer near-linear scaling, becoming up to 4x faster than Transformers at very long contexts (~57K tokens) with ~64% reduced memory footprint on consumer and embedded GPUs, making them well-suited for on-device AI for long-context inferences.
136. Distribution-informed Online Conformal Prediction
Core Problem: Many online conformal prediction methods, designed for fully adversarial environments, produce overly conservative prediction sets, even when predictable data patterns exist.
Key Innovation: Conformal Optimistic Prediction (COP), an online conformal prediction algorithm that incorporates underlying data patterns (via estimated cumulative distribution function of non-conformity scores) to produce tighter prediction sets while retaining valid coverage guarantees.
137. Double QMIX: A global optimization framework for multi-agent cooperative path planning in maritime SAR
Core Problem: Traditional multi-agent cooperative path planning for maritime SAR operations overlooks global search efficiency and practical asynchronous agent entry, leading to suboptimal rescue outcomes.
Key Innovation: Proposing Double QMIX, an adaptive multi-agent deep reinforcement learning framework for cooperative path planning in maritime SAR, which achieves global optimization by fusing local value functions and incorporating a Double DQN-inspired mechanism, demonstrating improved search success and practical applicability.
138. Flow control over a horizontal cylinder by symmetrically distributed flexible body
Core Problem: Reducing hydrodynamic loads on marine structures, such as horizontal cylinders, under various flow conditions (uniform currents and wave-currents) is crucial for marine engineering applications.
Key Innovation: Experimentally demonstrated that symmetrically distributed flexible bodies significantly reduce lift on a horizontal cylinder under both uniform currents and wave-current conditions, revealing the control mechanism through pressure distribution and unsteady flow characteristics, and identifying optimal parameter combinations.
139. Investigated effects of cylinder proximity on the hydrodynamic performance of axisymmetric AUV by using CFD
Core Problem: The hydrodynamic performance of Autonomous Underwater Vehicles (AUVs) operating near offshore structures or in confined environments is strongly influenced by wall proximity effects, impacting energy efficiency and operational safety.
Key Innovation: Conducted CFD simulations to investigate the hydrodynamic performance of an axisymmetric AUV near a cylindrical offshore structure, revealing that resistance coefficient increases significantly (up to 64.3%) when h/LA < 0.9 due to pronounced pressure asymmetry and enhanced flow separation.
140. A Field Parcel Scale Algorithm for Mapping Potato Distribution Using Multitemporal Sentinel-2 Images
Core Problem: Lack of accurate remote sensing techniques for mapping potato distribution at the field-parcel scale, with most attention focused on major crops.
Key Innovation: A cropland field-parcel-scale methodology integrating Canny edge detection, watershed segmentation, and a random forest classifier with multitemporal Sentinel-2 imagery to accurately map potato distribution, achieving high overall accuracy.
141. Regulatory Effects of Urban Morphology on Near-Surface Wind Speed in China Revealed by Observed and Remote Sensing Data
Core Problem: Previous studies overlooked multidimensional aspects of urban morphology and coupled effects on surface wind speed (SWS) at large scales due to technological limitations.
Key Innovation: Integrates SWS data from 183 stations with satellite-derived 2D–3D urban morphological indicators across China (1975–2020) to quantitatively assess impacts. It reveals that building surface density, near-surface air temperature, and building-topographic height difference jointly dominate SWS variation, with coupled effects being more significant than single factors.
142. Annual 10-m high-resolution cropland maps for Southeast Asia since 2019 using AlphaEarth embeddings
Core Problem: Persistent challenges in generating high-precision cropland data for Southeast Asia due to cloud cover, fragmented farming, shifting cultivation, and complex phenology, particularly for sloping croplands.
Key Innovation: Development of SEA_Cropland10, a 10-m annual cropland dataset for SEA using a random forest model integrating Sentinel-1/2 and AlphaEarth embeddings, significantly improving accuracy and detection of sloping cropland compared to existing global products.
143. Regional drought assessment using multi-site probabilistically integrated precipitation by Bayesian network
Core Problem: Drought monitoring is challenging due to uneven meteorological station distribution and strong climatic variability, hindering effective regional assessment.
Key Innovation: Proposed a Regional Standardized Precipitation Drought Index (RSPDI) for regional drought assessment by integrating multi-station precipitation dynamics using Bayesian Network (BN) theory and standardizing the combined series with a K-Component Gaussian Mixture Model (K-CGMM), demonstrating enhanced spatial coherence and robust probabilistic consistency.
144. Urban waterlogging in coastal cities: a composite multicriteria index approach from Bangladesh
Core Problem: Urban areas, particularly rapidly urbanizing coastal cities with inadequate drainage, face critical challenges from waterlogging, requiring effective spatial vulnerability assessment.
Key Innovation: Developed a Composite Waterlogging Vulnerability Assessment Index (CWLI) for urban areas by selecting six key factors (surface runoff, aspect, distance from drainage, tidal influence, solid waste, surface geology) and assigning weights using Fuzzy Analytic Hierarchy Process, producing a spatial vulnerability map to identify high-risk zones.
145. Assessing flood hazards and evacuation safety in metro stations: insights from Paral·lel station (Barcelona)
Core Problem: Underground transport systems, particularly metro stations, are highly vulnerable to increasing extreme rainfall events, posing severe threats to passenger safety during floods.
Key Innovation: Assessed flood hazards and evacuation safety in a metro station using a high-resolution two-dimensional hydraulic model (Iber software) with site-specific topographic data, simulating four inflow scenarios to evaluate water depth, flow velocity, and evacuation feasibility based on human-centered hazard indicators, identifying critical zones for emergency response.
146. Multi-Index evaluation of meteorological drought across Türkiye: a temporal and seasonal perspective
Core Problem: Türkiye, a drought-prone region, requires a comprehensive evaluation of meteorological drought dynamics across temporal and seasonal scales, especially considering the increasing influence of temperature-driven evapotranspiration.
Key Innovation: Evaluated the temporal and seasonal dynamics of meteorological drought across Türkiye using three distinct drought indices (SPI, SPEI, PNI) at multiple timescales with long-term precipitation and temperature data, revealing increasing drought severity after 2000, temperature-driven summer drying, and the added value of temperature-based indices for early warning and water resource management.
147. Characterizing extreme climate events at different time scales and their contributions to agricultural drought and flooding areas
Core Problem: A systematic understanding of multi-scale impacts of regional hydroclimatic extremes on agricultural water disasters, particularly during critical crop growth stages, is limited.
Key Innovation: Investigated climate extremes in Hunan Province (1961–2020) at seasonal, hydrological, and crop-growing scales using Modified Mann–Kendall trend analysis, correlation analysis, and random forest (RF) models, characterizing trends in precipitation and temperature extremes and quantifying their contributions to agricultural drought and flooding areas.
148. The devastating impact of the category 5 hurricane Otis on Mexico’s pacific coast
Core Problem: The severe and widespread damage to infrastructure and housing in Acapulco caused by Hurricane Otis, exacerbated by building code shortcomings and rapid intensification, leading to wind speeds far exceeding design thresholds.
Key Innovation: Presents a detailed post-event analysis of the impacts of Hurricane Otis, attributing damage primarily to extreme wind speeds and deficiencies in construction quality and building codes, emphasizing the urgent need for code revision.
149. High-resolution mapping of air pollution in the Yangtze River Delta: monthly 1 km resolution of PM2.5 data using spatiotemporal random forest algorithm
Core Problem: The challenge of accurately estimating PM2.5 concentrations at high resolution over large regions due to reliance on coarse datasets or high computational costs of ultra-fine mapping.
Key Innovation: Developed a spatiotemporal random forest (ST-RF) model to reconstruct monthly PM2.5 concentrations at 1 km resolution across the Yangtze River Delta, integrating various environmental and spatiotemporal features, demonstrating robust predictive ability and outperforming conventional models.
150. Comparison of physicochemical properties of reclaimed and undisturbed land and their associations with biomass in semi-arid mining areas of China
Core Problem: Ecological restoration in semi-arid grassland mining areas is constrained by limited topsoil and arid climate, making the recovery of soil physicochemical properties critical for ecosystem reconstruction.
Key Innovation: Conducted a comparative study of soil properties and their associations with vegetation biomass in reclaimed (3, 7, 11 years) and undisturbed land in a semi-arid mining area, revealing non-linear recovery trajectories and identifying key regulators (pH, STN, SOM) for biomass, suggesting integrated anthropogenic and biological restoration strategies.
151. Spatiotemporal variations and fluxes of dissolved nitrogen and carbon in a karst resurgence river
Core Problem: Insufficient understanding of coupled carbon-nitrogen dynamics in Karst Resurgence Rivers, which are critical linkages between surface and subsurface hydrological systems.
Key Innovation: Investigated spatiotemporal variations and transport fluxes of dissolved carbon and nitrogen in the Panyang River (a karst resurgence river) through monthly and rainfall-event sampling, highlighting distinct migration mechanisms and the differing effects of dilution and accumulation on C/N concentrations between surface and groundwater.
152. Network metrics for bridge prioritization considering bridge condition, accessibility, equity and uncertainty
Core Problem: Traditional bridge prioritization methods overlook localized accessibility impacts, broader network disruptions, and require extensive historical data, failing to adequately address equity and service continuity under probabilistic failure scenarios.
Key Innovation: Introduced a diagnostic framework integrating Population-Weighted Betweenness Centrality (PWBC), Average Annual Daily Traffic (AADT), and an Efficiency Sensitivity (ES) metric to evaluate bridge importance under probabilistic failure scenarios, identifying critical infrastructure and revealing disproportionate impacts on vulnerable communities.
153. Fusion of heterogeneous data for robust degradation prognostics
Core Problem: Predicting the remaining useful life (RUL) of industrial assets robustly, especially with respect to uncertainties from individual model-based and data-driven approaches, is challenging.
Key Innovation: Introduces a modular methodology for fusion of heterogeneous data in degradation prognostics, combining kernel-based sensitivity analysis, a Bayesian framework, and Kalman-based smoothing, along with an aggregate surrogate modeling strategy, to reduce output uncertainty.
154. Reliability analysis of dependent multistate phased mission systems with Vine-copulas
Core Problem: Existing reliability models for phased mission systems (PMSs) often assume known component dependencies, which is unrealistic, and struggle with unknown/heterogeneous dependencies in multistate PMSs (MS-PMSs).
Key Innovation: Proposes a Vine-copula-based modeling framework integrating Vine-copula dependency modeling with modular decomposition and PMS-BDD analysis for accurate and scalable reliability evaluation of MS-PMSs with unknown and heterogeneous component dependencies.
155. Reliability modeling for dependent competing failure processes with degrading self-healing mechanism
Core Problem: Most existing reliability models for systems with self-healing mechanisms assume a constant healing rate, neglecting its dependence on the system’s degradation state.
Key Innovation: Introduces a novel reliability modeling framework for dependent competing failure processes with degrading self-healing, where the self-healing rate explicitly varies as system degradation progresses, using piecewise-deterministic Markov processes and stochastic semigroup theory.
156. Quantifying potential cyber-attack risks in CNC systems under zero-subjectivity closed-loop Dempster–Shafer theory FMECA and rule-based Bayesian network modelling
Core Problem: Dedicated cybersecurity risk assessment methods for CNC systems are scarce, and domain-specific threat probability databases are lacking in smart manufacturing.
Key Innovation: Develops a risk quantification framework integrating Failure Mode, Effects, and Criticality Analysis (FMECA) with a Rule-Based Bayesian Network (RBN), based on Zero-Subjectivity Closed-Loop Dempster–Shafer (ZDS) theory, to fuse expert opinions and achieve hierarchical probabilistic modeling of attack risks.
157. Enhancing power grid cybersecurity against FDI attacks via deep Q-network-based moving target defense
Core Problem: Cybersecurity threats like False Data Injection (FDI) attacks pose significant risks to modern power systems, and existing defense methods are less intelligent or cost-effective in maximizing attack detectability with minimal operational cost.
Key Innovation: Proposes an Intelligent Moving Target Defense (iMTD) framework using a Deep Q-Network (DQN) to dynamically modify transmission line reactances, obscuring system parameters from attackers while ensuring minimal disruption to power flow and cost.
158. Optimal Bayesian maintenance policy for gear shafts under variable operating conditions with partially observable information
Core Problem: Gear shaft failures lead to substantial costs, and sensor-based condition monitoring provides only partially observable information, complicating maintenance decisions.
Key Innovation: Presents a novel optimal Bayesian maintenance policy under partially observable information, employing a hidden semi-Markov model (HSMM) and updating conditional reliability via Bayes' theorem, integrated into a cost-minimizing semi-Markov decision process (SMDP) to identify optimal downtime scheduling.
159. Automatic tree-level based forest inventories retrieval via ultra-high resolution UAV images and deep learning
Core Problem: The challenge of efficiently and accurately monitoring forest dynamics and quantifying carbon stocks at the individual tree level, especially in dense natural forests, using cost-effective RGB UAV imagery which has limited spectral signals compared to multispectral sensors.
Key Innovation: An Individual Tree Crown (ITC)-based framework leveraging ultra-high resolution UAV RGB data and a novel Multi-Task Convolutional Neural Network (ITCMNet) to simultaneously and accurately identify individual tree crowns, discriminate tree species, and assess tree vitality, enabling precise forest investigations and improved carbon stock estimation.
160. TLNet: A deep learning framework for tree detection in forest point clouds using multi-layered forest structure
Core Problem: Consistent and accurate individual-tree detection from LiDAR point clouds across diverse sites and sensors, hindered by severe occlusion, irregular point density, and platform-dependent sampling geometry.
Key Innovation: Tree-Layer Network (TLNet), a deep learning framework that slices forest point clouds into horizontal slabs, embeds 3D points via sparse-voxel convolutions, fuses grid feature maps in bottom-up and top-down passes with learnable scalar gates, and predicts tree-location heatmaps, achieving high F1 scores across multiple LiDAR platforms and forest types.
161. MHATCN: Integrating local patterns and long-range dependencies for snow albedo forecasting
Core Problem: Existing snow albedo prediction approaches struggle to simultaneously capture short-term variability and long-range temporal dependencies, and often lack emphasis on physical interpretability and systematic evaluation of model robustness in complex mountainous environments.
Key Innovation: Developed MHATCN, an enhanced Temporal Convolutional Network integrating multi-head attention, for snow albedo forecasting in mountainous regions. This model effectively captures both local temporal patterns and long-range dependencies, achieving improved predictive accuracy (RMSE 0.029, R2 0.873) and robustness compared to baseline models.
162. Groundwater controls on legacy antibiotics and pesticides in an intensive agricultural headwater catchment
Core Problem: Investigating the occurrence and long-term persistence of legacy antibiotics and pesticides in groundwater and understanding the role of groundwater in their storage and release in agricultural systems.
Key Innovation: Integrated field measurements (CFCs as age tracers) with a 3D physically-based hydrogeological model and particle tracking to simulate groundwater transit time distributions, demonstrating groundwater's critical role in storing and releasing legacy pollutants (e.g., atrazine and metolachlor detected decades after ban).
163. Microbial N2O production and the functional communities regulated by groundwater flow regime
Core Problem: The mechanism regulating N2O production in response to groundwater flow regimes remains poorly understood, despite groundwater being a potentially important source of N2O emissions.
Key Innovation: Integrated 15N isotope tracing, metagenomic analysis, qPCR, and microbial cultivation to investigate N2O production pathways and functional microbial communities in a groundwater system, demonstrating that microbial N2O production and communities are regulated by biogeochemical gradients driven by groundwater flow, identifying discharge zones as N2O hotspots.
164. Perched groundwater recharge and subsurface flow dynamics in check-dam systems of the Loess Plateau
Core Problem: The mechanistic role of check dams in perched groundwater recharge and subsurface flow dynamics, particularly in semi-arid environments, remained unresolved.
Key Innovation: Developed a new conceptual model of "precipitation-driven-perched groundwater" dynamics in check-dam systems, revealing the formation, migration, and precipitation response mechanisms of perched groundwater as "dynamic hydrological regulators" through integrated monitoring and tracer tests.
165. Integrating NMR logging with three-dimensional steady-state hydraulic tomography for improved mapping of hydraulic conductivity distributions
Core Problem: Mapping high-resolution hydraulic conductivity (K) distributions in heterogeneous aquifer systems, especially with limited calibration data for hydraulic tomography (HT), often results in smooth tomograms.
Key Innovation: Demonstrated that integrating NMR logging data as initial guesses significantly improves the resolution of K tomograms and drawdown prediction performance in 3D steady-state HT analysis, even with sparse calibration data, for complex aquifer systems.
166. Developing the fusion of MODFLOW simulation and data-driven approaches for river-aquifer recharge modeling
Core Problem: Efficiently and accurately modeling complex river-aquifer interactions and recharge for aquifer management often requires significant computational time and resources with traditional numerical models.
Key Innovation: Developed a fusion approach combining MODFLOW simulation with data-driven machine learning (ML) algorithms (e.g., GPR) to accurately predict river-aquifer recharge with reduced computational time, demonstrating its utility for aquifer management programs.
167. Physics-informed cross-learning for seismic acoustic impedance inversion and wavelet extraction
Core Problem: Seismic acoustic impedance inversion is challenging due to factors like seismic wavelets and low-frequency initial models, and existing deep learning methods often lack direct and effective physical constraints, leading to instability in inversion results.
Key Innovation: Development of a deep learning framework for simultaneous acoustic impedance inversion and seismic wavelet extraction, coupled with a physics-informed cross-learning strategy, to impose effective physical constraints and achieve significant improvements in accuracy on both synthetic and field datasets.
168. Intelligent assessment of subgrade compaction quality under variable moisture conditions using roller acceleration response
Core Problem: Spatiotemporal fluctuations in moisture content during subgrade construction significantly compromise compaction quality assessment, as existing studies often assume optimal moisture conditions or analyze within limited moisture ranges, underestimating vibration response drift.
Key Innovation: Proposes an adaptive assessment framework integrating multi-domain features (time, frequency, time–frequency, and entropy) with a self-attention mechanism (1D-ResCNN-SA) for intelligent, moisture-aware subgrade compaction quality assessment. This method achieves adaptive feature extraction and dynamic weighting, outperforming mainstream approaches in both compaction state identification and degree prediction across complex moisture conditions (LMC, OMC, HMC).