Initiated by Dr. Xin Wei, University of Michigan
Ongoing development by the community

TerraMosaic Daily Digest: August 9, 2026

August 9, 2026 TerraMosaic Daily Digest

Daily Summary

Direct geohazard papers on August 9 concentrate on observation units and state variables that better match physical processes. Wildfire studies show that short-term prediction improves when forecast zones are learned from historical ignition patterns rather than imposed grids, that sub-pixel fire temperatures can be retrieved in real time through full-physics VSWIR inversion, and that compact UAV detectors benefit more from real than synthetic training imagery. Flood and landslide studies make a parallel shift away from uniform image-only mapping: one regional framework compares cross-zone with zone-constrained learning before combining susceptibility with exposure and vulnerability, another attains centimeter-level flood-depth estimation by selectively tuning causally relevant vision-language cross-attention, and landslide papers fuse deformation with image features, reconstruct 3D slope motion with InSAR plus surface-parallel-flow constraints, and adapt frozen foundation models with terrain, material, and rainfall guidance.

Mechanistic geotechnical and seismic studies sharpen the diagnosis of how instability develops. High-fidelity discrete-element modelling resolves a basal-failure to retrogressive-collapse sequence in rainfall-triggered waste-dump landslides, controlled experiments separate surface-load and pile-load pathways in cover-collapse sinkholes, and shaking-table studies show that water level governs whether railway embankments liquefy, spread laterally, erode, or fail as immersed landslides, while weak-interlayer rock slopes exhibit clear prestress and shaking thresholds. Seismic risk contributions add a Croatia-specific ground-motion model and a multivariate fragility framework for masonry buildings under earthquake sequences. Hydroclimatic papers likewise become more state-resolving: the 2022 European mega-drought is attributed chiefly to circulation and subsidence rather than moisture supply, dust-layer structure is linked to near-surface heat stress during an Iberian dust-heatwave event, and downscaled climate projections, GPU urban-heat modelling, HeatCast benchmarking, and finer Local Climate Zone mapping all move heat and drought assessment toward operational spatial detail.

Separate from those direct hazard findings, the larger transferable cohort is defined by calibration, robustness, and physically structured generalization rather than by demonstrated geohazard validation. The selected papers audit segmentation labels without reference masks, adapt frozen backbones for change detection and causal spatial disaggregation, refine precipitation nowcasts, strengthen cross-view localization and 2D-3D matching, and improve restoration or reconstruction under blur, low light, event streams, infrared degradation, and mixed-quality imagery. Physics-oriented contributions advance neural-operator splitting, geometry-agnostic Fourier operators, improved PINN training, sparse-dynamics recovery, invertible forward-inverse modelling, and solver-level adjoints, while conformal prediction with missing modalities, digital-twin falsification, uncertainty attribution, long soil-moisture and forcing datasets, and root-zone moisture fusion provide a more rigorous evidential base for future hazard-monitoring systems.

Key Trends

The clearest movement across the August 9 selections is from generic prediction pipelines toward process-aware sensing, threshold-resolving mechanics, and transferable geospatial AI that is explicitly calibrated for uncertainty and distribution shift.

  • Process-aware hazard sensing is displacing generic image interpretation: Across direct hazard papers, observation pipelines are being tied to physically meaningful structure: wildfire forecast zones are learned from ignition geography, landslide mapping absorbs terrain, material, rainfall, and deformation cues, 3D slope motion is constrained to plausible flow directions, and cavity detection benefits from explicit multi-view GPR fusion.
  • Forecast systems are being redesigned around spatial heterogeneity and distributional structure: Regional flood-landslide mapping compares global and zone-constrained learning, tropical-cyclone forecasting jointly predicts imagery and atmospheric fields, missing Rmax values are reconstructed with temporal deep learning, precipitation nowcasts preserve phase while refining amplitudes, and coastal sea-level ensemble assessment exposes joint-distribution errors that aggregate scores can miss.
  • Instability mechanics are being resolved through state transitions and thresholds: The direct geotechnical papers repeatedly identify discrete failure regimes rather than smooth risk gradients: waste-dump landslides progress from basal failure to retrogressive collapse, sinkholes follow distinct collapse pathways under different load configurations, railway embankments switch among liquefaction, spreading, erosion, and sliding with water level, and reinforced rock slopes show seismic and prestress thresholds in load redistribution.
  • Heat and drought hazard analysis is moving toward finer attribution and urban-scale diagnostics: Drought and thermal studies combine synoptic attribution with local detail: circulation and subsidence are isolated as dominant drivers of the 2022 European mega-drought, dust-layer stratification is linked to surface thermal stress, Odisha projections are statistically downscaled to hazard metrics, and GPU SOLWEIG workflows, HeatCast, and refined Local Climate Zone mapping push urban heat assessment to neighborhood scales.
  • Transferable geospatial AI is prioritizing calibration, uncertainty, and physical coherence: Non-hazard-specific papers concentrate less on headline accuracy than on whether outputs can be trusted under sparse labels, degraded imagery, missing modalities, or shifted domains. Label-auditing metrics, conformal regression, digital-twin falsification, uncertainty decomposition, operator-based generalization, improved PINN propagation, and geometry-preserving 3D reconstruction collectively define a more disciplined technical substrate for later geohazard applications without constituting those applications themselves.

Selected Papers

The selected papers separate into direct geohazard contributions and transferable sensing or AI methods. The first group reports hazard-specific findings across landslides, floods, wildfire, earthquakes, drought, and heat, while the second contributes reconstruction, forecasting, calibration, and physics-based tools that may support later geohazard workflows without themselves constituting hazard validation.

1. Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: wildfire Relevance: 8/10

Core Problem: Improve short-term wildfire prediction by replacing arbitrary grid cells with better spatial forecasting units.

Key Innovation: Builds fire-zone prediction units from historical ignitions using watershed detection plus K-means, outperforming grid discretization across regions and models.

2. Mechanistic Interpretability-Guided Selective Fine-Tuning of Vision-Language Models for Centimeter-Level Flood Depth Estimation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: flood Relevance: 8/10

Core Problem: Estimate real-time centimeter-level urban flood depth from street imagery.

Key Innovation: Uses mechanistic interpretability to fine-tune only causally relevant VLM cross-attention layers with near-dense performance.

3. Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: wildfire Relevance: 8/10

Core Problem: Retrieve sub-pixel wildfire temperatures from VSWIR imaging spectroscopy in real time.

Key Innovation: Uses fast full-physics GPU inversion on AVIRIS-3 and shows transferable temperature posteriors to coarser spaceborne settings.

4. Spatial Heterogeneity-Aware Multi-Hazard Susceptibility and Risk Mapping at Regional Scale

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: flood and landslide Relevance: 8/10

Core Problem: Flood and landslide controls vary spatially, making uniform regional susceptibility models unreliable.

Key Innovation: Compares cross-zone versus zone-constrained spatial learning and integrates susceptibility with exposure-vulnerability risk mapping.

5. Tropical Cyclone Forecasting via Latent Rectified Flow using Satellite Imagery and Atmospheric Fields

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: tropical cyclone Relevance: 8/10

Core Problem: Existing deep cyclone forecasters either omit atmospheric fields, sample slowly, or decouple track prediction from generated physics.

Key Innovation: Jointly forecasts satellite and atmospheric fields in one pass and fine-tunes tracks with differentiable steering-flow rewards.

6. GeoPhysAdapter: Scale-Matched Geophysical Adaptation for Cross-Domain Landslide Mapping with Vision Foundation Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: landslide Relevance: 8/10

Core Problem: Cross-domain landslide segmentation suffers from high-confidence false positives when models face new regions and events.

Key Innovation: Adds scale-matched terrain, material, and rainfall guidance with body-level adaptation on top of a frozen vision foundation model.

7. Remote Sensing, Vol. 18, Pages 2692: DIGSFNet: Deformation-Integrity-Guided Symmetric Fusion Network for High-Risk Landslide Extraction from Multi-Source Remote Sensing Images

Source: Remote Sensing (MDPI) Type: landslide remote-sensing extraction study Geohazard Type: landslide Relevance: 8/10

Core Problem: Extract high-risk landslides from combined InSAR and optical imagery without fragmented masks or inefficient inference.

Key Innovation: Deformation-integrity-guided symmetric fusion plus a distilled INT8 student for accurate, fast operational deployment.

8. Unveiling the retrogressive traction-collapse mechanism of rainfall-induced landslides in high waste rock dumps via high-fidelity 3D discrete element modeling

Source: Bull. Eng. Geol. & Env. Type: rainfall-induced landslide mechanism modeling Geohazard Type: waste-dump landslide Relevance: 8/10

Core Problem: Resolve the progressive failure mechanism of a high waste-rock dump slope under rainfall infiltration and basal softening.

Key Innovation: A high-fidelity 3D DEM framework with coupled EPPB-JKR constitutive behavior reproduced a two-stage basal-failure to retrogressive-collapse process validated by field observations.

9. Formation mechanism of cover collapse sinkholes induced by urban static loads in water-rich strata: an experimental investigation

Source: Bull. Eng. Geol. & Env. Type: sinkhole formation mechanism study Geohazard Type: sinkhole Relevance: 8/10

Core Problem: Explain how static urban loads trigger cover-collapse sinkholes in water-rich strata.

Key Innovation: Reduced-scale experiments distinguished surface-load and pile-load collapse pathways and formalized a three-stage cavity-growth to collapse evolution model.

10. An integrated InSAR and SPF-constrained workflow for 3D landslide displacement reconstruction

Source: Bull. Eng. Geol. & Env. Type: InSAR landslide displacement reconstruction study Geohazard Type: landslide Relevance: 8/10

Core Problem: Reconstruct full 3D landslide displacement histories from line-of-sight InSAR observations.

Key Innovation: An SBAS-InSAR plus surface-parallel-flow constrained inversion workflow recovered 3D motion fields and linked deformation timing to precipitation.

11. Towards a ground motion model for Croatia

Source: Bull. Earthquake Eng. Type: regional seismic hazard model Geohazard Type: earthquake Relevance: 8/10

Core Problem: Build a region-specific ground motion model for Croatia despite sparse strong-motion data.

Key Innovation: A stochastic rock-site GMM calibrated with Croatian seismological parameters and nonlinear site amplification captured local tectonic behavior across 69,120 simulated scenarios.

12. Impact of Dataset Composition on Embedded Real-Time UAV Wildfire Detection Using Compact YOLO Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: wildfire Relevance: 7/10

Core Problem: Test whether augmentation and synthetic images improve embedded UAV wildfire detection.

Key Innovation: Shows real non-augmented training data outperform synthetic and augmented alternatives for compact YOLO deployment.

13. Contrastive Mask Fidelity: Reference-Free Auditing of Ground-Truth Masks in Remote Sensing Semantic Segmentation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing segmentation label auditing Relevance: 7/10

Core Problem: Remote-sensing segmentation masks can be coarse or wrong, making standard overlap metrics misleading.

Key Innovation: Scores masks directly against image evidence with a reference-free vision-language auditing metric and uses it to improve supervision quality.

14. TriView-YOLO: Early Multi-View Fusion for Ground Penetrating Radar Cavity Detection in Soft, High-Water-Content Soils

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: subsurface cavity and road-collapse hazard Relevance: 7/10

Core Problem: Cavity detection in soft water-saturated soils is especially difficult because GPR signals attenuate where cavities form most readily.

Key Innovation: Fuses three co-registered GPR views in a 9-channel YOLO detector trained on expert-verified field surveys from high-risk soft-ground roads.

15. Deep Learning Imputation of Missing Radius of Maximum Winds (Rmax) Values in Tropical Cyclone Best-Track Data

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: coastal cyclone hazard Relevance: 7/10

Core Problem: Missing radius-of-maximum-winds values in best-track archives weaken probabilistic tropical cyclone hazard assessments.

Key Innovation: Benchmarks temporal deep learning, physics-informed inputs, and transfer learning for Rmax reconstruction and shows temporal models best preserve storm variability.

16. Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: dynamic hazard forecasting transfer Relevance: 7/10

Core Problem: Video diffusion models generate plausible frames but fail to learn transferable physical dynamics for out-of-distribution motion.

Key Innovation: Imposes explicit latent kinematic integration so the model extrapolates physical dynamics far beyond its training distribution.

17. Do AI Forecast Ensembles Sample the Correct Conditional Distribution?

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: coastal_flooding_and_sea_level Relevance: 7/10

Core Problem: AI sea-level forecast ensembles can appear skillful marginally while failing to sample the correct joint conditional distribution across stations.

Key Innovation: A shuffle-based permutation decomposition showing that joint-structure failures are missed by the energy score but exposed by the variogram score.

18. Test-time Generalization for Physics through Neural Operator Splitting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: physics-based hazard modeling Relevance: 7/10

Core Problem: Pretrained neural operators generalize poorly to unseen PDE coefficients, initial conditions, or novel physics.

Key Innovation: The method composes a dictionary of pretrained operators at test time to approximate unseen dynamics without weight updates.

19. Brief communication: Atmospheric moisture and near-surface temperature anomalies: key drivers in the 2022 European mega-drought

Source: NHESS Type: drought hazard mechanism study Geohazard Type: drought Relevance: 7/10

Core Problem: Explain why the 2022 European mega-drought intensified despite regional moisture uptake.

Key Innovation: Lagrangian moisture-temperature framework showing that circulation and subsidence, not moisture supply, controlled drought severity.

20. GLIDE-SOL: a GPU-accelerated global lightweight infrastructure for diagnostic environmental modeling with SOLWEIG

Source: GMD Type: heat-stress environmental modeling system Geohazard Type: urban heat Relevance: 7/10

Core Problem: Make urban outdoor thermal-stress modeling fast enough for large-scale practical deployment.

Key Innovation: GPU-accelerated global SOLWEIG workflow with automated data ingestion and bias-reducing corrections.

21. High-resolution machine-learning downscaled climate projections and extreme hazard assessment over monsoon-dominated Odisha, India

Source: Geomatics, Nat. Haz. & Risk Type: climate hazard assessment study Geohazard Type: extreme heat and drought Relevance: 7/10

Core Problem: Produce robust high-resolution climate projections and assess future extremes over monsoon-dominated Odisha.

Key Innovation: Grid-wise XGBoost bias-correction/downscaling with transferability testing and ETCCDI hazard metrics.

22. Remote Sensing, Vol. 18, Pages 2693: Vertical Variability of Temperature and Moisture in a Compound Dust-Heatwave Scenario at South-Western Iberian Peninsula: Implications for Surface Thermal Stress and CCN Predictions

Source: Remote Sensing (MDPI) Type: compound atmospheric hazard study Geohazard Type: dust storm and heatwave Relevance: 7/10

Core Problem: Characterize thermodynamic structure and stress implications of a Saharan dust-heatwave episode over Iberia.

Key Innovation: Lidar-plus-ERA5 analysis linking dust-layer structure to near-surface heat stress and CCN-favorable conditions.

23. Seismic failure mechanisms of railway embankments under variable water level conditions: Insights from transparent soil shaking table tests

Source: Bull. Eng. Geol. & Env. Type: seismic embankment failure experiment Geohazard Type: seismic embankment failure and landslide Relevance: 7/10

Core Problem: Determine how water-level conditions alter seismic failure modes of railway embankments and foundations.

Key Innovation: Transparent-soil shaking-table tests revealed water-dependent transitions among liquefaction, lateral spreading, erosion, and catastrophic immersed embankment landsliding.

24. Centrifugal shaking-table test study on the dynamic responses of rock slope containing weak interlayer reinforced with prestressed composite structure

Source: Geoenvironmental Disasters Type: seismic rock-slope reinforcement experiment Geohazard Type: seismic rock-slope instability Relevance: 7/10

Core Problem: Quantify how seismic intensity and anchor prestress control the dynamic response of reinforced weak-interlayer rock slopes.

Key Innovation: Centrifugal shaking-table tests identified dual thresholds in seismic intensity and prestress and clarified load redistribution within pile-anchor composite support systems.

25. Multivariate seismic intensity measures and vulnerability assessment framework for multistory masonry structures under near-field and far-field earthquake sequences

Source: Bull. Earthquake Eng. Type: earthquake vulnerability assessment framework Geohazard Type: earthquake Relevance: 7/10

Core Problem: Estimate failure probabilities and cumulative damage of multistory masonry structures under near-field and far-field earthquake sequences.

Key Innovation: A multivariate intensity-measure framework combined field damage matrices with new fragility and cumulative-failure models tuned to regional seismic design acceleration.

26. Hierarchical graph neural networks for stochastic slope stability analysis: A coarse-to-fine surrogate modeling approach

Source: Computers and Geotechnics Type: AI slope stability analysis method Geohazard Type: landslide Relevance: 7/10

Core Problem: Accelerate stochastic slope stability analysis without losing spatial detail in failure prediction.

Key Innovation: A hierarchical coarse-to-fine graph neural network surrogate for probabilistic slope-stability calculations.

27. HeatCast: A Benchmark for Neighborhood-Scale LST Forecasting across 124 U.S. Cities

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: urban heat Relevance: 6/10

Core Problem: Establish a shared 30 m benchmark for neighborhood-scale urban land surface temperature forecasting.

Key Innovation: Releases a large multi-city Landsat LST benchmark with fixed splits, LCZ-stratified metrics, and reference baselines.

28. AdaDINO: Pair-Aware In-Backbone Adaptation of Frozen DINO for Efficient Remote Sensing Change Detection

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: surface change Relevance: 6/10

Core Problem: Single-image vision foundation models miss cross-temporal interactions needed for bi-temporal change detection.

Key Innovation: Adds pair-aware in-backbone temporal interaction to frozen DINO plus efficient chunk pruning for change detection.

29. CLAM: Causal Spatial Disaggregation to Infer Local Effects From Coarse Data

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: multi-hazard Relevance: 6/10

Core Problem: Inferring local causal effects from coarse aggregated spatial observations is difficult.

Key Innovation: Jointly learns causal mechanisms and spatial disaggregation mappings using high-resolution contextual covariates.

30. FreCast: Refining Radar Echo Intensity via Phase-Preserving Amplitude Residual Diffusion for Precipitation Nowcasting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: rainfall and flood precursor Relevance: 6/10

Core Problem: Radar nowcasts capture precipitation layout but misestimate local echo intensity.

Key Innovation: Two-stage nowcasting that refines pixelwise echo amplitudes with phase-preserving residual diffusion under structural constraints.

31. GeoAI-based post-segmentation quality validation of building footprints via spatial feature engineering

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: geospatial feature extraction qa Relevance: 6/10

Core Problem: Building-footprint extraction from imagery often yields topologically poor vectors unsuitable for GIS ingestion.

Key Innovation: Detects vector errors with engineered geometric, spatial-context, and raster predictors in a post-segmentation GeoAI pipeline.

32. OGG-FR: Orthogonal Gradient Gaming and Frequency Rectification for Unmanned Aerial Vehicle Infrared Image Super-Resolution

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: uav thermal remote sensing Relevance: 6/10

Core Problem: Lightweight UAV infrared super-resolution suffers from conflicting pixel- and frequency-domain training objectives.

Key Innovation: Separates redundant and orthogonal gradient components and rectifies frequency updates based on conflict and confidence.

33. VeinCast: Physics-Guided Dynamic Field Graphs with Graph-Conditioned Fusion for Global Medium-Range Weather Forecasting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: meteorological hazard drivers Relevance: 6/10

Core Problem: Global weather models need better state-dependent interactions across many atmospheric fields.

Key Innovation: Builds a dynamic field graph and graph-conditioned latent fusion scheme to inject relational physical guidance into 69-field forecasts.

34. Warp-free Cross-view Geo-localization via Feature-space Consensus Mining

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: geospatial localization Relevance: 6/10

Core Problem: Geometric warping creates distorted supervision for street-to-satellite matching under severe viewpoint change.

Key Innovation: Mines joint-view semantic consensus directly in feature space and distills it into single-view encoders.

35. MoNo: Multiscale Optimal Transport Neural Operator for Solving PDEs on General Geometries

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: physical hazard simulation transfer Relevance: 6/10

Core Problem: Neural operators on general geometries suffer unstable latent-token assignments that degrade multiscale PDE learning.

Key Innovation: Formulates cross-scale latent assignment as entropy-regularized optimal transport to build balanced multiscale neural operators.

36. ADEx-FNO: A Unified Ambient-Domain Framework for Fourier Neural Operators on Varying Geometries

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geofluid_modeling Relevance: 6/10

Core Problem: Fourier neural operators struggle with varying geometries and mismatched discretizations across physical domains.

Key Innovation: A deterministic ambient-domain extension framework that embeds varying geometries in a common cube without changing FNO layers.

37. Unsupervised Point Cloud Registration with Self-Distillation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: LiDAR and radar change-detection transfer Relevance: 6/10

Core Problem: Deep point-cloud registration usually needs expensive ground-truth poses and struggles to scale.

Key Innovation: Learns registration by self-distillation with a teacher plus robust solver, eliminating pose labels.

38. ProPINN: Demystifying Propagation Failures in Physics-Informed Neural Networks

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: PDE hazard-modeling transfer Relevance: 6/10

Core Problem: PINNs suffer propagation failure because boundary supervision does not reliably reach the interior domain.

Key Innovation: Links failure to low nearby-point gradient correlation and introduces ProPINN to improve propagation.

39. NeuralDMD: Interpretable Neural Representation of Dynamics from Sparse and Noisy Measurements

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: weather and geophysical data assimilation transfer Relevance: 6/10

Core Problem: Need interpretable reconstruction of spatiotemporal dynamics from sparse noisy measurements without ground truth simulators.

Key Innovation: Combines neural implicit fields with dynamic mode decomposition to recover continuous dynamics and forecast from sparse data.

40. Multi-Year Geospatial Reasoning using Interannually-Consistent Historical Predictions as a Free Input Modality

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: cross-hazard mapping Relevance: 6/10

Core Problem: Operational EO classifiers ignore their own historical prediction archives and partner masks when mapping each new year.

Key Innovation: The method feeds confidence-scaled historical class tokens and consistent ancillary masking into the model to reason across years.

41. Causal Falsification of Digital Twins

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: digital twin validation Relevance: 6/10

Core Problem: Observational data generally cannot certify that a digital twin is causally correct without strong assumptions.

Key Innovation: The authors reframe assessment as causal falsification and provide a sound statistical test for finding when a twin is wrong under confounding.

42. An invertible generative model for forward and inverse problems

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: geophysical inversion and forecasting Relevance: 6/10

Core Problem: It is difficult to build one generative model that supports both forward simulation and posterior inference for inverse problems.

Key Innovation: The reversible simulator merges conditional normalizing flows into a single invertible model trainable directly from paired samples.

43. Remote Sensing, Vol. 18, Pages 2699: Superpixel-Level Joint-Sparse and Graph-Regularized Framework for Hyperspectral Image Classification

Source: Remote Sensing (MDPI) Type: hyperspectral classification method Geohazard Type: landslide-mapping transfer Relevance: 6/10

Core Problem: Improve hyperspectral image classification when labels are sparse and spatial coherence matters.

Key Innovation: Superpixel-level joint sparse representation with graph Laplacian regularization.

44. Rethinking 3D Segmentation from Individual LiDAR Scans: Incidence-Aware Sampling on the SIP Benchmark

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain and infrastructure mapping Relevance: 5/10

Core Problem: Improve 3D segmentation from sparse individual LiDAR scans with limited coverage and density bias.

Key Innovation: Introduces incidence-aware manifold-space sampling that stabilizes segmentation across resolutions without backbone changes.

45. SeqLoc: Beyond the Single Frame for Cross-View Geo-Localization in Feature-Sparse Scenes

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: geospatial localization Relevance: 5/10

Core Problem: Fix single-frame cross-view geo-localization failures in rural, feature-sparse scenes.

Key Innovation: Adds online sequence belief aggregation with entropy tempering, map-guided relocalization, and peak-anchored smoothing.

46. IRPol-Fuse: Energy-structure coordination for infrared polarization fusion under low visibility

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: low-visibility hazard sensing Relevance: 5/10

Core Problem: Fuse infrared and polarization imagery without losing weak structural cues in low-visibility scenes.

Key Innovation: Coordinates thermal energy and polarization structure with dedicated injectors and a new low-visibility evaluation dataset.

47. FlexSplat: Flexible Feed-Forward 3D Gaussian Splatting without Point Cloud Correspondence

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain reconstruction Relevance: 5/10

Core Problem: Feed-forward Gaussian-splatting methods usually require known camera poses for multi-view reconstruction.

Key Innovation: Jointly predicts camera parameters and depth so compact 3D Gaussian reconstruction works without pose input.

48. SUMI: Scalable Unified Model for 3D Point Cloud Inference

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain reconstruction Relevance: 5/10

Core Problem: Coarse-to-fine point-cloud completion often loses local geometric detail in the refinement stage.

Key Innovation: Uses diffusion-enhanced cross-attentive refinement to recover local geometry while preserving global shape.

49. When Does An Extra View Help? Adapting Single-View 3D Reconstruction with Extra Imagery

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain reconstruction Relevance: 5/10

Core Problem: Single-view 3D reconstruction lacks a mechanism to incorporate an auxiliary test-time image.

Key Innovation: Adapts single-view reconstructions using one extra image through zero-shot and optimized contrastive schemes.

50. VTO: Visual Tool Orchestration for Video Anomaly Detection

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: operational hazard monitoring Relevance: 5/10

Core Problem: Video anomaly agents lack robust multi-step tool orchestration and fine-grained supervision.

Key Innovation: Uses process-supervised RL with a tool hierarchy and cognitive evaluator for anomaly reasoning.

51. Test-Time Prototype Adaptation for Open-Vocabulary Semantic Segmentation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard mapping Relevance: 5/10

Core Problem: Open-vocabulary segmentation methods usually need host-specific architectural changes for adaptation.

Key Innovation: Builds training-free class prototypes from unlabeled deployment images and fuses them with host logits at test time.

52. Open-World Semantic Segmentation with Sensitivity Modeling

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard mapping Relevance: 5/10

Core Problem: Closed-world segmentation overconfidently misclassifies unseen or anomalous content.

Key Innovation: Adds a sensitivity decoder that captures local activation instability alongside known-class and unknown-region signals.

53. ERF-GS: Reconstructing Fast Motion from Disjoint Event-RGB Viewpoints

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: rapid scene mapping transfer Relevance: 5/10

Core Problem: Fast motion and disjoint event/RGB viewpoints break dynamic scene reconstruction.

Key Innovation: Event-RGB Gaussian splatting that injects event cues into optimization and densification.

54. SC-Diff: Semantically Calibrated Diffusion for Visible-to-Infrared Image Translation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: multispectral thermal imaging transfer Relevance: 5/10

Core Problem: Visible-to-IR diffusion translation loses object layout and semantic consistency.

Key Innovation: Uses semantic masks to condition inputs and calibrate denoising self-attention at token level.

55. EvTrajGS: Accurate and Efficient 3D Gaussian Splatting from Unposed Event Streams

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: rapid terrain and structure mapping transfer Relevance: 5/10

Core Problem: Event-stream 3DGS needs accurate poses without expensive SLAM-style optimization.

Key Innovation: Continuous-time trajectory refinement with temporally coupled pose updates and loss-reweighted event sampling.

56. JSGS: JPEG State-Guided Supervision for 3D Gaussian Splatting from Mixed-Quality Views

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: heterogeneous photo-to-3D mapping transfer Relevance: 5/10

Core Problem: Mixed-quality JPEG views inject compression artifacts into multi-view 3DGS supervision.

Key Innovation: Uses view-specific JPEG observation operators and block-disagreement regularization during splat optimization.

57. Semi-Dense Matching Uncertainty Is Not Just Local Confidence

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: photogrammetric uncertainty transfer Relevance: 5/10

Core Problem: Semi-dense matchers underestimate uncertainty by ignoring catastrophic coarse assignment failures.

Key Innovation: Calibrated two-component Laplace mixture with posterior-weighted geometric refitting.

58. OccAnyScene: Towards Unified Indoor-Outdoor 3D Occupancy Predictio

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain and built-environment completion transfer Relevance: 5/10

Core Problem: One occupancy model must generalize across indoor and outdoor camera setups and voxel regimes.

Key Innovation: Frustum-parameterized Gaussian lifting built on pretrained depth priors for unified cross-scene occupancy.

59. AeroReformer2: Spoken-Query Referring Segmentation for Aerial Images

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing scene parsing Relevance: 5/10

Core Problem: Remote-sensing referring segmentation lacks spoken-query benchmarks and efficient speech-image grounding models.

Key Innovation: Introduces a speech-enabled aerial benchmark and a bilateral network with token-preserving speech encoding and efficient cross-modal attention.

60. Damage Classification for 3D Point Cloud Data via 3D Data Analysis and Vision Foundation Model-based 2D Projections

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: damage mapping from point clouds Relevance: 5/10

Core Problem: Fine-grained damage classification from 3D point clouds is expensive and data-limited.

Key Innovation: Compares topological 3D descriptors with low-cost 2D projection pipelines that leverage vision foundation models.

61. RMR-Net: Degradation-Evidence-Guided Road-Image Restoration for Defect Detection

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: infrastructure damage monitoring Relevance: 5/10

Core Problem: Image degradation hides cracks and pothole boundaries needed by road defect detectors.

Key Innovation: Uses degradation-evidence-guided restoration conditioned on corruption cues and optimized for a frozen downstream detector.

62. Right Answer, Wrong Heat: Explanation-Aware Evaluation and Thermal-Grounded Feedback for MLLMs on Infrared Images

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: infrared remote sensing explainability Relevance: 5/10

Core Problem: Infrared image question answering can be correct while relying on weak or non-thermal evidence.

Key Innovation: Separates answer accuracy from thermal grounding and improves explanation fidelity with a training-free feedback loop.

63. GRASP: Granularity-Aware Region Alignment and Semantic Prototype Learning for Fine-Grained Cross-Modal Understanding in Drone Views

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: aerial scene understanding Relevance: 5/10

Core Problem: Background clutter and visually similar objects weaken fine-grained drone image-text alignment.

Key Innovation: Combines region-focused alignment with semantically perturbed hard negatives built from a foreground prototype codebook.

64. Degraded Infrared Small Object Detection via Degradation-Adapted Physics-Guided Restoration

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: thermal hazard monitoring Relevance: 5/10

Core Problem: Fog and sensor nonuniformity suppress infrared target contrast and degrade small-object detection.

Key Innovation: Detects degradation type and severity, performs branch-specific physics-guided restoration, and then fuses results for robust detection.

65. Beyond Uniform Restoration: Empowering All-in-One Restoration with Pixel-Level Multimodal Guidance

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing and monitoring imagery restoration Relevance: 5/10

Core Problem: Unified restoration models treat all pixels uniformly even when degradation type and severity vary locally.

Key Innovation: Estimates pixel-level visual prompts and combines them with text and visual guidance for localized restoration decisions.

66. XFeat Revisited: Reproducibility and Evaluation of a Lightweight Image Matcher

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: image registration and co-registration Relevance: 5/10

Core Problem: The claimed accuracy-efficiency tradeoffs of a lightweight matcher are hard to trust because paper, code, and evaluation details diverge.

Key Innovation: Reproduces XFeat, ablates its design, and extends evaluation to thermal-visible and multimodal remote-sensing matching.

67. MSP-Net: Manifold-Guided Spectral Prompt Network for Hyperspectral Object Tracking

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hyperspectral remote sensing Relevance: 5/10

Core Problem: Fixed band-order fusion makes hyperspectral trackers brittle across sensors and changing scene conditions.

Key Innovation: Reconstructs band manifolds for adaptive spectral grouping and forms target-conditioned prompts that update during tracking.

68. TeaMatch: Teachable Cross-Modal Representation Learning for 2D-3D Matching

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: 3D mapping and registration Relevance: 5/10

Core Problem: Existing 2D-3D matchers fail under low overlap, noise, and ambiguous geometry because they optimize only within a single model.

Key Innovation: Optimizes teacher representations for recoverability by weak degraded-input students using geometry-aware meta supervision.

69. FedOrbit: Adaptive Personalized Federated Learning for Non-IID LEO Satellite Constellations

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote sensing systems for hazard monitoring Relevance: 5/10

Core Problem: Non-IID orbital data and irregular visibility make standard federated learning unreliable across LEO satellite constellations.

Key Innovation: Combines continuous inter-satellite training, class-aware hierarchical aggregation, and similarity-adaptive personalization for better per-orbit remote-sensing accuracy.

70. Real-Time Climate Risk Assessment for Supply Chain Resilience: A Data-Driven Nowcasting Framework for Colombian Agriculture

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: climate and agricultural risk Relevance: 5/10

Core Problem: Agricultural supply chains need operational short-term climate risk signals, but weather data and logistics decisions are poorly integrated.

Key Innovation: Links precipitation nowcasting with threshold-based risk mapping to create an actionable early-warning framework for Colombian agriculture.

71. Leveraging generative models to assist Monte Carlo sampling

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: probabilistic hazard simulation transfer Relevance: 5/10

Core Problem: High-dimensional multimodal sampling remains a bottleneck in scientific inference and simulation.

Key Innovation: Synthesizes how flows and diffusion models can assist or exactly accelerate Monte Carlo sampling without requiring conventional training data.

72. Conformal Calibration for Multi-Modal Regression with Missing Modalities

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard Relevance: 5/10

Core Problem: Prediction intervals in multimodal regression become miscalibrated when modalities disagree or are missing.

Key Innovation: A modality-aware conformal layer that uses disagreement-based continuous scaling and stratified recalibration while preserving coverage guarantees.

73. Eikonal Regularisation in Physics-Informed Neural Networks for Three-Dimensional Level-Set Advection: Transferability of Two-Dimensional Design Principles

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: mass_flow_simulation Relevance: 5/10

Core Problem: The best eikonal-regularization strategy for PINN level-set advection in 2D was unclear in 3D and across random seeds.

Key Innovation: A preregistered multi-seed 3D benchmark showing how optimal eikonal weights track signed-distance departure and improve reproducibility.

74. Physics-Informed Learning for Robust Acoustic Localization with Calibrated Uncertainty

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: acoustic_hazard_monitoring Relevance: 5/10

Core Problem: Classical outdoor acoustic localization breaks under multipath, near-field effects, and complex propagation.

Key Innovation: A physics-informed learned correction to a hyperbolic solver plus geometry-aware calibrated uncertainty estimates.

75. Coordinate-Residual Physics-Driven Neural Network for Electromagnetic Inverse Scattering

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: geophysical imaging transfer Relevance: 5/10

Core Problem: 3D electromagnetic inverse scattering is ill-posed, noisy, and computationally expensive.

Key Innovation: Reconstructs contrast fields directly from coordinates with a residual physics-driven network, avoiding unstable preliminary reconstructions.

76. AirFlow: Context Preserving and Multi-Rate State Modeling for Air Quality Forecasting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: air quality hazard Relevance: 5/10

Core Problem: Pollutant channels have distinct periodicity and drift that shared temporal models fail to capture.

Key Innovation: Routes pollutants through statistic-guided normalization and a dual-stream multi-rate state model.

77. OATS: Online Data Augmentation for Time Series Foundation Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: cross-hazard forecasting Relevance: 5/10

Core Problem: Static augmentation strategies do not adapt synthetic time-series generation to different training stages.

Key Innovation: OATS generates step-specific synthetic series with a diffusion model guided by valuable samples and an explore-exploit controller.

78. Scalable extensions to given-data Sobol' index estimators

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: cross-hazard model interpretation Relevance: 5/10

Core Problem: Existing given-data Sobol estimators become memory-heavy and biased for models with very large input dimension.

Key Innovation: The paper generalizes the estimator, adds a streaming algorithm, and proposes asymptotic screening heuristics for large models.

79. A dataset of soil water measurements at multiple scales and depths on China’s Loess Plateau

Source: ESSD Type: hydroclimate dataset paper Geohazard Type: soil-moisture and loess-slope transfer Relevance: 5/10

Core Problem: Document multi-scale, multi-depth soil-water dynamics across the Loess Plateau over 20 years.

Key Innovation: Rare long-duration dataset extending to five meters depth from plot to transect scales.

80. A 30-Year 1-km Daily Precipitation and Air Temperature Dataset for the Po River District (Italy)

Source: ESSD Type: high-resolution hydroclimate dataset paper Geohazard Type: flood and drought forcing data Relevance: 5/10

Core Problem: Build a 30-year 1 km daily precipitation and air-temperature dataset for the Po River District.

Key Innovation: Multi-source interpolation workflow delivering high-resolution forcing data for regional hydrologic applications.

81. Retrieving root-zone soil moisture from land surface modelling and GRACE/-FO and validating its dynamics with in-situ data over West Africa

Source: HESS Type: soil-moisture retrieval study Geohazard Type: drought and hydrologic hazard transfer Relevance: 5/10

Core Problem: Estimate and validate root-zone soil moisture over West Africa using models, GRACE/FO, and in situ data.

Key Innovation: Cross-comparison showing complementary value of CCI soil moisture and GRACE-based water-storage products.

82. Remote Sensing, Vol. 18, Pages 2698: KSR-Huber: A Robust Method for Wind Vector Retrieval from Doppler Wind Lidar Observations

Source: Remote Sensing (MDPI) Type: atmospheric lidar retrieval method Geohazard Type: severe-weather monitoring transfer Relevance: 5/10

Core Problem: Recover accurate 3D wind vectors from noisy Doppler wind-lidar observations.

Key Innovation: KNN local priors plus Huber IRLS for outlier-robust wind retrieval.

83. Remote Sensing, Vol. 18, Pages 2690: Improving Local Climate Zone Mapping at Fine Spatial Scales Using Urban Morphology, Spectral Information, and Machine Learning

Source: Remote Sensing (MDPI) Type: urban climate remote-sensing study Geohazard Type: urban heat Relevance: 5/10

Core Problem: Improve fine-scale local climate zone mapping using morphology, spectral data, and machine learning.

Key Innovation: Fusion of Sentinel-2, PRISMA, and urban canopy parameters with training-sample refinement for higher-resolution LCZ mapping.

84. An approach based on geometric and spectral features for automatic identification of rock mass discontinuities from multi-view UAV photogrammetry

Source: Engineering Geology Type: UAV rock-discontinuity mapping method Geohazard Type: rock-slope hazard mapping transfer Relevance: 5/10

Core Problem: Automatically identify rock mass discontinuities from multi-view UAV photogrammetry.

Key Innovation: Fuses geometric and spectral cues for automated discontinuity extraction from UAV-derived rock outcrops.

85. Evolving perceptions of drought risk and management: Insights from a repeated survey of Swedish municipalities

Source: IJDRR Type: drought risk governance study Geohazard Type: drought Relevance: 5/10

Core Problem: Track how Swedish municipalities perceive and manage drought risk over time.

Key Innovation: A repeated-survey design reveals evolving local drought-risk awareness and management practices.

86. LAPAC: A self-adaptive atmospheric correction method for fine-resolution satellite imagery using pseudo-invariant features from Landsat archive

Source: Remote Sensing of Env. Type: satellite atmospheric-correction method Geohazard Type: remote-sensing preprocessing transfer Relevance: 5/10

Core Problem: Retrieve accurate surface reflectance from fine-resolution satellite imagery without concurrent external atmospheric inputs.

Key Innovation: LAPAC uses Landsat-archive pseudo-invariant features plus precomputed radiative-transfer constraints for self-adaptive pixel-wise atmospheric correction.

87. GF7S: A large-scale GaoFen-7 stereo dataset and self-supervised adaptation of FoundationStereo for high-resolution DSM generation

Source: ISPRS J. Photogrammetry Type: stereo remote-sensing dataset and DSM method Geohazard Type: terrain-modeling transfer Relevance: 5/10

Core Problem: Improve high-resolution DSM generation from GaoFen-7 stereo imagery.

Key Innovation: A large stereo dataset and self-supervised adaptation of FoundationStereo for high-resolution DSM production.

88. HSMLA: Hierarchical Softmax Multi-scale Linear Attention for Efficient Vision Transformers

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: image segmentation enabling Relevance: 4/10

Core Problem: Reduce the quadratic cost of self-attention in high-resolution dense prediction.

Key Innovation: Combines linear attention, selective softmax refinement, and multi-scale tokens for faster dense prediction.

89. Tracing sources of epistemic uncertainty in deep learning predictions: homo- and hetero-scedastic linearized estimators

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard model calibration enabling Relevance: 4/10

Core Problem: Disentangle aleatoric scarcity effects from epistemic uncertainty in deep learning predictions.

Key Innovation: Adapts classical homoscedastic and heteroscedastic estimators to scalable linearized deep models using approximate Fisher information.

90. BRUCE: Benchmarking Robustness Under Corruption Escalation for Scientific Vision-Language Reasoning

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard VLM robustness enabling Relevance: 4/10

Core Problem: Measure how scientific VLM reasoning degrades under escalating visual corruption.

Key Innovation: Defines corruption-escalation metrics and failure taxonomies for robustness analysis in scientific visual reasoning.

91. Vision-Language Grounding as Bidirectional Concept Correspondence

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: multimodal mapping enabling Relevance: 4/10

Core Problem: Recover all visually referential text spans and their matching image segments without pre-specified phrases.

Key Innovation: Reframes grounding as bidirectional concept correspondence using bridge tokens that jointly predict text masks, image masks, and matches.

92. SegDem: Segmentation helps Demosaicing

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing imaging enabling Relevance: 4/10

Core Problem: Improve demosaicing by injecting structural understanding rather than relying only on local pixel cues.

Key Innovation: Transfers instance-segmentation decoder representations into RAW reconstruction and anchors both tasks to shared structure.

93. LAD-COD: Language-Aligned Dense Perception for Camouflaged Object Detection

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard imagery Relevance: 4/10

Core Problem: Camouflaged object segmentation lacks top-down guidance in dense visual features.

Key Innovation: Aligns language-conditioned target semantics with hierarchical visual features through dual visual fusion.

94. From Uncertainty to Failure Attribution: Self-Diagnosing Models for Failure Attribution under Distribution Shift

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general geospatial Relevance: 4/10

Core Problem: OOD methods detect shift but usually do not explain why predictions fail under shift.

Key Innovation: Learns joint uncertainty and structured failure-attribution signals for different shift mechanisms.

95. Distilling Physical Priors into Streaming World Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: physical process simulation Relevance: 4/10

Core Problem: Streaming world models lack stable physical priors and produce implausible long-horizon dynamics.

Key Innovation: Builds a physics-video dataset and distills a physics-aware teacher into a causal streaming generator with temporal credit routing.

96. Advantage-Guided Gate: Reshaping Open-Ended Reasoning for Vision-Based Spatial Intelligence

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: spatial scene understanding Relevance: 4/10

Core Problem: Open-ended visual reasoning drifts and accumulates decision errors over multi-step reasoning.

Key Innovation: Uses advantage-guided gating over reasoning trees to select higher-value intermediate steps and trajectories.

97. Evidence-RL: Towards Evidence-intensive Visual Reasoning

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard image interpretation Relevance: 4/10

Core Problem: Vision-language models answer from shortcuts instead of causal local visual evidence.

Key Innovation: Uses counterfactual evidence disentanglement inside RL post-training to reward evidence-dependent reasoning.

98. EvBS: Event-guided Blur Synthesis for Domain-adaptive Motion Deblurring

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard imagery Relevance: 4/10

Core Problem: Deblurring models fail under domain shift because motion and scene content are entangled.

Key Innovation: Uses event-guided intrinsic and extrinsic blur synthesis to adapt deblurring models to target domains.

99. Learning Structural Illumination for Unsupervised Low-light Enhancement

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard imagery Relevance: 4/10

Core Problem: Unsupervised low-light enhancement confounds spatial illumination structure with exposure level.

Key Innovation: Learns relative illumination structure from reliable bright regions and uses scene-adaptive exposure references.

100. Wiener Representation Filtering for VLM Hallucination Suppression

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard image interpretation Relevance: 4/10

Core Problem: Vision-language models hallucinate nonexistent objects and relations in visual descriptions.

Key Innovation: Applies a training-free Wiener-style representation filter that suppresses hallucination-associated modes.

101. Retrieval-Augmented Generation-Based Color Restoration for Low-Light Image Enhancement

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard imagery Relevance: 4/10

Core Problem: Low-light enhancement methods recover brightness but leave persistent color casts.

Key Innovation: Decouples color correction as retrieval-augmented post-processing using reference-guided color modulation.

102. Your VLM Already Knows When: Training-Free Temporal Grounding by Asking Yes or No

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard event timing Relevance: 4/10

Core Problem: VLMs recognize events but fail badly at identifying when they occur in time.

Key Innovation: Recasts temporal grounding as a coarse-to-fine sequence of binary questions scored from first-token probabilities.

103. RayLift: Lifting Complementary Ray-Wise Evidence with 3D Geometry Priors for Semantic Scene Completion

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain and scene reconstruction transfer Relevance: 4/10

Core Problem: Stereo-depth errors propagate into camera-based semantic scene completion.

Key Innovation: Lifts uncertainty-aware complementary ray evidence with frozen 3D foundation priors into voxel features.

104. Towards Adaptive Super-Resolution and Quality Assessment via Test-Time Adaptation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: monitoring imagery enhancement transfer Relevance: 4/10

Core Problem: Video SR and no-reference quality models fail under unseen degradations.

Key Innovation: Test-time adaptation pipeline that uses perceptual quality guidance and region-aware refinement.

105. eBIRD: Event-based Intensity Image Reconstruction Using Controllable Diffusion Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: high-speed monitoring imaging transfer Relevance: 4/10

Core Problem: Sparse asynchronous events make intensity-image reconstruction difficult.

Key Innovation: ControlNet-conditioned diffusion reconstruction tailored to event streams.

106. MotionCraft: Latent World Modeling with Sparse Attention for Visual Upscaling

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: monitoring imagery enhancement transfer Relevance: 4/10

Core Problem: VSR must balance motion fidelity, long-range temporal modeling, realism, and efficiency.

Key Innovation: Latent world-model-based motion-aware VSR with adaptive sparse attention and controllable fidelity-smoothness tradeoffs.

107. Trajectory Design and Budgeted Querying for Digital Twin Calibration

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: physical-system calibration transfer Relevance: 4/10

Core Problem: Calibration data and privileged parameter queries are expensive in digital twins.

Key Innovation: Jointly optimizes excitation trajectories, uncertainty-aware parameter estimation, and budgeted oracle querying.

108. Degradation-Guided Underwater Image Restoration with Task-Oriented Latent Control

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: subsea hazard imaging transfer Relevance: 4/10

Core Problem: Underwater restoration underuses degradation cues and propagates degraded skip features.

Key Innovation: Combines degradation-guided dynamic feature modulation with latent control of skip features.

109. Efficient Test-Time Scaling for LLM-based Time Series Forecasting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hazard-signal forecasting transfer Relevance: 4/10

Core Problem: Scaled LLM forecasters are costly and drift on long-horizon global shape.

Key Innovation: Coarse-to-fine shape prediction guides fixed-step residual refinement with fewer tokens.

110. High-Quality Exposure Correction with Diffusion-Based Image Generation Priors

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: field imagery restoration transfer Relevance: 4/10

Core Problem: Exposure correction methods miss extreme exposure regions and lose perceptual quality.

Key Innovation: Single-step exposure correction that injects pretrained diffusion priors via joint cross-attention.

111. LASA: Language-and-Source-Anchored Alignment for Domain Generalized Semantic Segmentation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: cross-domain mapping transfer Relevance: 4/10

Core Problem: Style randomization and normalization weaken domain-generalized segmentation on unseen domains.

Key Innovation: Anchors style transfer and decoder alignment with source features and language priors.

112. Dynamic Distribution-Aware Uncertainty Tracking in Vision-Language Representation Learning

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: trustworthy hazard ai Relevance: 4/10

Core Problem: Post-hoc uncertainty methods for VLMs fail when test distributions shift away from the source domain.

Key Innovation: Models distributional evidence with a Gaussian mixture and updates uncertainty dynamically at inference time.

113. View-Adaptive Renderer for View-Consistent 2D-to-3D Generation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: terrain and scene reconstruction Relevance: 4/10

Core Problem: Inconsistent synthesized views degrade monocular 3D reconstruction quality.

Key Innovation: Adds view-adaptive neural renderers and attention-based multi-view fusion to correct viewpoint-specific errors.

114. When Latents Forget Pixels: Restoring Fidelity in Diffusion Transformer Super-Resolution

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing image restoration Relevance: 4/10

Core Problem: Latent diffusion super-resolution hallucinates details because compressed latents forget low-resolution pixel evidence.

Key Innovation: Grounds super-resolution with pre-VAE pixel features on both the conditioning path and decoder side while freezing the main backbone.

115. Multi-Submap Implicit Neural SLAM with Local-to-Global Loop Closure for Large-Scale Scene Reconstruction

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: field mapping and reconstruction Relevance: 4/10

Core Problem: NeRF-based SLAM struggles to scale because of forgetting, drift, and weak loop closure in large scenes.

Key Innovation: Uses progressive neural submaps, dual-tier loop closure, and online inter-submap distillation for scalable reconstruction.

116. Multimodal Federated Learning under Dual-Axis Modality Missingness

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: distributed multimodal sensing Relevance: 4/10

Core Problem: Federated multimodal learning breaks when clients and samples both have incomplete modality sets.

Key Innovation: Separates confidence-tempered private adaptation from shared federated optimization to handle dual-axis missingness.

117. Revisiting the Current Frame: Physical-Trace-Guided Network Output Correction for Video Restoration

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: field imagery restoration Relevance: 4/10

Core Problem: Temporally aggregated restoration outputs can be unreliable where reference frames are inconsistent or occluded.

Key Innovation: Estimates a spatial trust field from physical-trace evidence to blend restoration outputs with the current observation.

118. Preserve More Details: Mitigating Content Drift in Real-World Image Super-Resolution

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing image enhancement Relevance: 4/10

Core Problem: Diffusion-based real-image super-resolution drifts from low-quality evidence and alters fine details and semantics.

Key Innovation: Uses a dual-pathway one-step diffusion model with detail-conditioned structure injection and detail-modulated semantic guidance.

119. MeanSR: Restoration Trajectory Learning for One-Step Perceptual Super-Resolution

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote-sensing image enhancement Relevance: 4/10

Core Problem: Distillation-free one-step super-resolution still does not explicitly learn the restoration trajectory from degraded to sharp images.

Key Innovation: Learns a low-resolution-conditioned average velocity field with stage-aware temporal sampling for one-step perceptual super-resolution.

120. RecoverFly: A Failure-Aware Reinforcement Learning Post-Training Framework for Aerial Vision-Language Navigation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: UAV hazard-response robotics Relevance: 4/10

Core Problem: Behavior-cloned UAV vision-language-action policies lack corrective learning for failure cases and long-tail scenes.

Key Innovation: Revisits unresolved failures, uses token-level RL for grammar-constrained actions, and applies a long-tail scene curriculum with policy regularization.

121. Hierarchical rank-evolving representation for physics-informed neural networks

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: PDE modeling methodology Relevance: 4/10

Core Problem: Tensor PINNs depend on manually chosen low ranks that often miss the true structure of multivariate solutions.

Key Innovation: Introduces hierarchical rank-evolving representations that adaptively reveal ranks and improve high-dimensional PDE solving.

122. From Semantic Grounding to Decision Optimization: A Unified Framework for Long-Horizon UAV Vision-Language Navigation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: aerial field robotics Relevance: 4/10

Core Problem: Long-horizon UAV navigation suffers from weak landmark grounding, poor history use, and unstable decisions in local traps.

Key Innovation: Unifies instruction-grounded semantic enhancement, relevance-aware temporal aggregation, and topology-aware decision optimization.

123. Marrying Optimal Transport and ODEs for Unified Continuous-Time 4D Reconstruction and Tracking

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: dynamic scene reconstruction Relevance: 4/10

Core Problem: Existing 4D reconstruction and tracking models cannot produce kinematically coherent states at arbitrary timestamps.

Key Innovation: Learns continuous velocity fields with optimal-transport-guided flow matching and ODE-based integral-consistency training.

124. Confusion-Geometry Rebalancing for Long-Tailed Adversarial Training

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general hazard classification transfer Relevance: 4/10

Core Problem: Long-tailed adversarial training overfits head classes and misses which class boundaries drive robustness collapse.

Key Innovation: Builds a directed confusion-geometry graph from robust errors to reweight losses and sharpen vulnerable decision boundaries.

125. Evaluating Generative Time-Series Models on Data with Point Masses

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: hydrometeorological forecasting transfer Relevance: 4/10

Core Problem: Generative time-series benchmarks can badly mis-evaluate models when sparse event occurrence in evaluation windows differs from the dataset distribution.

Key Innovation: Defines matched evaluation for point-mass series and shows metric choice and occurrence structure can reverse model rankings.

126. NTIRE 2026 Low-light Enhancement: Twilight Cowboy Challenge

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: remote imaging transfer for hazard monitoring Relevance: 4/10

Core Problem: Burst low-light imaging must jointly handle severe noise, mixed illumination, and frame misalignment from hand motion.

Key Innovation: Provides a challenging real-world benchmark and establishes strong burst-restoration baselines for nighttime image enhancement.

127. The Knowing-Saying Gap: When Probes See Errors that Confidence Misses

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: reliable hazard-AI monitoring transfer Relevance: 4/10

Core Problem: Internal probes can detect corrupted reasoning context, but it is unclear whether that helps predict final answer failure better than stated confidence.

Key Innovation: Shows a robust knowing-saying gap and maps which probe-based interventions rescue errors without breaking correct traces.

128. Crowd-Sourced Geographies of Income: Using Google Maps Points of Interest as High-Frequency Proxies for Sub-Municipal Income Estimation in Sao Paulo, Brazil

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: exposure_and_vulnerability_mapping Relevance: 4/10

Core Problem: Sub-municipal income data are sparse and outdated between censuses in large cities.

Key Innovation: Using Google Maps POI composition with decomposition plus regression to estimate sector-level income at high spatial detail.

129. Search over the Visual World: Persistent Visual Memory, Layered Indexes, and Source-Grounded Evidence

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard_monitoring Relevance: 4/10

Core Problem: Video retrieval over continuous cameras and streams needs persistent memory, task-specific context selection, and inspectable source evidence.

Key Innovation: A visual-world search framework with persistent visual memory, layered indexes, and source-grounded evidence over live and archived video.

130. ARC: Augmented-Rank Conformalization for Changepoint Localization --- Finite-Sample Validity and Distribution-Robust Efficiency

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard_monitoring Relevance: 4/10

Core Problem: Conformal changepoint localization guarantees coverage, but practical confidence sets can become inefficient under skew, heavy tails, and shift.

Key Innovation: Rank-based augmented conformal scores whose confidence sets are invariant to monotone marginal transforms.

131. Differentiate the Solver, Not the Equation: Reverse-Sweep Adjoints for Block Implicit Simulation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard_modeling Relevance: 4/10

Core Problem: Unrolled autodiff is memory-heavy and equation-level implicit differentiation ignores the executed local solver structure.

Key Innovation: Reverse-sweep adjoints that differentiate block implicit solvers directly through local reverse-order updates.

132. End-to-End Neural Decomposition with Koopman Operators for Time-Series Forecasting

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard_forecasting Relevance: 4/10

Core Problem: Standard Koopman-based forecasters struggle when nonlinear signals contain both time-varying periodic and trend components.

Key Innovation: An end-to-end model that jointly decomposes signals and learns separate frequency-dependent and frequency-independent Koopman dynamics.

133. CPDA: Class-Conditional Path Distribution Alignment for Unsupervised Time-Series Domain Adaptation

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: general_geohazard_monitoring Relevance: 4/10

Core Problem: Time-series domain adaptation methods often align only global marginals and miss class-conditional temporal structure.

Key Innovation: A class-conditional signature-spectral kernel discrepancy that aligns latent path distributions rather than only pooled features.

134. Walk-on-Spheres Monte Carlo and deep neural network approximations of elliptic PDEs with drift and killing

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: numerical-method transfer Relevance: 4/10

Core Problem: Need scalable Monte Carlo and neural approximations for elliptic PDEs with drift and killing.

Key Innovation: Extends Walk-on-Spheres estimators and proves polynomial-complexity neural approximations for the stochastic representations.

135. Analogical Learning for Cross-Scenario Generalization: Framework and Application to Intelligent Localization

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: sensor-network localization transfer Relevance: 4/10

Core Problem: Models trained across diverse scenarios overfit scenario-specific absolute representations.

Key Innovation: Uses reference-frame style relative metrics and a paired-transformer architecture for analogical learning across scenarios.

136. An Information-Theoretic Framework for Feature Construction in Out-of-Distribution Detection

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: deployment-robustness transfer Relevance: 4/10

Core Problem: OOD features for neural detectors lack a principled construction framework.

Key Innovation: Combines ID-OOD separation and information bottleneck objectives to derive better OOD shaping features.

137. Physics-Informed Policy Iteration for High-Dimensional Hamilton--Jacobi--Bellman Equations: Interior Error Bounds without Boundary Data

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: PDE and control-method transfer Relevance: 4/10

Core Problem: High-dimensional HJB control problems are hard to solve reliably without boundary data.

Key Innovation: Alternates neural residual policy evaluation and pointwise policy improvement with interior error bounds and rollout safeguards.

138. SplitGaussian: Reconstructing Dynamic Scenes via Visual Geometry Decomposition

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: photogrammetry transfer Relevance: 4/10

Core Problem: Dynamic Gaussian-splatting reconstructions entangle moving and static content, causing geometry and temporal artifacts.

Key Innovation: Separates static and dynamic branches so only the dynamic component deforms over time.

139. Adversarially Robust Few-Shot Anomaly Detection with Vision Foundation Models

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: infrastructure anomaly transfer Relevance: 4/10

Core Problem: Few-shot anomaly detectors using frozen foundation features are vulnerable to adversarial perturbations.

Key Innovation: Adds a differentiable distance proxy plus input and feature purification defenses without backbone retraining.

140. Progressive Learned Image Compression for Machine Perception

Source: ArXiv (Geo/RS/AI) Type: Preprint Geohazard Type: cross-hazard monitoring Relevance: 4/10

Core Problem: Machine-oriented image codecs lack progressive bitstreams that can stop at task-sufficient quality levels.

Key Innovation: PICM-Net combines trit-plane progressive coding, task-adaptation modules, and a confidence-driven adaptive decoder for efficient progressive transmission.

141. Satellite-to-simulation reconstruction of digital sea ice fields from optical imagery

Source: Ocean Engineering Type: cryosphere remote sensing reconstruction Geohazard Type: sea ice Relevance: 4/10

Core Problem: Reconstruct digital sea-ice fields from optical satellite imagery for simulation use.

Key Innovation: Satellite-to-simulation pipeline that converts optical imagery into digital sea-ice fields.

142. Spatio-temporal prediction zone for phase-resolved wave forecasting in directional sea states

Source: Ocean Engineering Type: wave forecasting method Geohazard Type: coastal wave hazard transfer Relevance: 4/10

Core Problem: Forecast phase-resolved directional wave fields in real time from remote measurements.

Key Innovation: Simplified assimilation and a spatio-temporal prediction zone for multidirectional sea states.

143. CSRFormer: An Instantaneous Global-ocean Clear-sky Radiative Flux Dataset Derived From CERES

Source: ESSD Type: climate remote-sensing dataset paper Geohazard Type: climate forcing transfer Relevance: 4/10

Core Problem: Create an instantaneous global-ocean clear-sky radiative-flux dataset matched to satellite overpass times.

Key Innovation: CERES-derived machine-learning dataset for clear-sky top-of-atmosphere fluxes over oceans.

144. High resolution modelling of temporal and spatial dynamics in soil conditions in subarctic Finland using HydroBlocks model

Source: HESS Type: land-surface modeling study Geohazard Type: cold-region hydrogeomorphic transfer Relevance: 4/10

Core Problem: Model spatiotemporal soil, snow, and ice conditions across subarctic Finland at high resolution.

Key Innovation: HydroBlocks configuration and evaluation for Finnish snow, soil moisture, temperature, and ice dynamics.

145. Remote Sensing, Vol. 18, Pages 2689: Airborne Streak Tube Imaging LiDAR-Based Effective Reconstruction of Urban Water Areas

Source: Remote Sensing (MDPI) Type: LiDAR signal-enhancement and water-surface reconstruction method Geohazard Type: bathymetry and flood-mapping transfer Relevance: 4/10

Core Problem: Recover usable underwater LiDAR echoes for 3D reconstruction of urban water areas under severe attenuation.

Key Innovation: MSAGAN combines morphology-aware attention, spectral-temporal feature fusion, and adaptive loss design to enhance weak streak-tube echoes and densify point clouds.

146. A revised physically-based shadow kernel for improved hotspot simulation in urban thermal anisotropy models

Source: Remote Sensing of Env. Type: urban thermal remote-sensing model improvement Geohazard Type: urban heat transfer Relevance: 4/10

Core Problem: Improve hotspot simulation in urban thermal anisotropy models.

Key Innovation: A revised physically based shadow kernel for more realistic urban thermal-directionality modeling.

147. A hyperparameter tuning strategy for an LSTM model to simulate reservoir outflows: large-scale evaluation across 441 dams in the CONUS

Source: Journal of Hydrology Type: hydrologic AI forecasting study Geohazard Type: reservoir and flood-management transfer Relevance: 4/10

Core Problem: Find robust hyperparameter-tuning strategies for LSTM simulation of reservoir outflows across many dams.

Key Innovation: A large-sample evaluation across 441 CONUS dams to identify scalable LSTM tuning behavior for reservoir-outflow prediction.

148. A meshfree (SPH)-grid (FDM) coupled framework for hydro-mechanical fracturing in deformable porous medium

Source: Computers and Geotechnics Type: hydro-mechanical fracture simulation framework Geohazard Type: fracture-process modeling transfer Relevance: 4/10

Core Problem: Simulate coupled fluid-driven fracture initiation and propagation in deformable porous media without predefined cracks.

Key Innovation: A coupled SPH-FDM framework with dynamic Darcy-to-cubic-law flow transition and damage-driven fracture detection.

149. Progressive Failure Mechanism and Prediction Model of Geogrid‑Reinforced Embankments under Long‑Term Service Conditions in Xinjiang, China

Source: Transportation Geotechnics Type: embankment failure prediction study Geohazard Type: embankment instability transfer Relevance: 4/10

Core Problem: Explain and predict long-term progressive failure of geogrid-reinforced embankments.

Key Innovation: A title-indicated mechanism-plus-prediction framework for service-life embankment instability assessment.