TerraMosaic Daily Digest: July 29, 2026
Daily Summary
Landslide intelligence advances from regional susceptibility maps toward decision-scale evidence. The first building-level national assessment for the United States integrates susceptibility, 128 million structures and socioeconomic conditions, locating 6.5 million people in high-susceptibility terrain and exposing a marked rural vulnerability gap. Complementary studies resolve post-earthquake vegetation recovery, optical landslide boundaries, active periods in Sentinel-1 coherence, monitoring geometry for open-pit slopes and reservoir-wave overtopping.
Terrain and hydroclimate are treated as causal controls rather than map background. Cloud-resolving experiments attribute more than 300 mm of Hurricane Helene rainfall to Appalachian uplift, while a 796-basin analysis shows that atmospheric-riverine compound heatwaves have tripled since the 1980s. DEM resolution shifts reconstructed flood hydrographs, coastal GRACE products inherit ocean leakage, and long climate, cyclone and vulnerability records expose where scale choices alter hazard conclusions.
Reliability is increasingly tested against the evidence a model actually uses. Visual-credit audits, intermediate-state tests, local-cue search and long-horizon visual retrieval distinguish grounded reasoning from language-prior success. Coverage-aware token pruning and class-conditional conformal prediction expose overconfidence and minority undercoverage, while point-cloud distillation, registration and unsupervised segmentation address scale mismatch and long-tailed three-dimensional observations.
Key Trends
Across slope hazards, climate extremes and multimodal models, useful prediction depends on resolving exposure, activity, scale and evidence rather than reporting a final score alone.
- Exposure Is Moving to the Building Scale: A national United States database links susceptibility to individual structures and social vulnerability, separating where landslides can occur from who and what is exposed.
- Landslide Monitoring Is Becoming Activity-Sensitive: Coherence loss, geometric sensitivity, optical validation and optimized GB-SAR placement shift monitoring from static inventories toward the timing and observability of slope movement.
- Terrain and Compound Climate Are Coupled: Orographic rainfall, atmospheric-riverine heat, flood hydrographs and coastal water-storage estimates show that relief, transport and observation scale must be resolved together.
- Visual Reliability Depends on Evidence Coverage: Visual-credit tests, token-pruning experiments, local-cue search and tool-use benchmarks measure whether a model finds, retains and follows the image regions needed for its answer.
Selected Papers
The July 29 literature is led by a national landslide-exposure database and a cluster of monitoring, mapping, susceptibility and recovery studies that connect slope processes to operational decisions. Hydroclimate, earthquake and infrastructure papers extend the same logic across interacting hazards, while multimodal AI studies test whether predictions remain calibrated and visibly grounded.
1. Landslide Exposure in the United States
Core Problem: No national building-level assessment has linked landslide susceptibility, structures and socioeconomic vulnerability across the United States.
Key Innovation: The Landslide Exposure Database integrates national susceptibility, 128 million structures and socioeconomic data, identifying 6.5 million people in high-susceptibility terrain and distinct rural and urban policy contexts.
2. Threefold increase in atmospheric–riverine compound heatwaves under climate change
Core Problem: Atmospheric and riverine heatwaves are usually assessed separately even though their coincidence compounds pressure on water, energy and food systems.
Key Innovation: Observations and deep learning across 796 basins show a threefold rise in compound events and quantify the faster growth of riverine heatwave frequency, duration and intensity.
3. RABTracks: Reanalysis-Augmented Best Tracks for Tropical Cyclones
Core Problem: Historical tropical-cyclone records omit or shorten weak early and late storm stages, limiting consistent global life-cycle analysis.
Key Innovation: RABTracks augments IBTrACS with five analyses and two trackers from 1940 onward, extending tracks and adding consistent intensity, size and cyclone-phase information.
4. Unraveling Spatiotemporal Heterogeneity and Driving Mechanism of Vegetation Recovery in Co‐Seismic Landslide Areas via Nonlinear Trajectories and Bayesian Networks
Core Problem: Post-seismic vegetation recovery varies sharply across disturbed mountain terrain, but its timing and environmental constraints remain difficult to separate.
Key Innovation: Landsat trajectories from 2008-2023 combine generalized additive models and a Bayesian network to locate five-to-seven-year recovery inflections and terrain-specific climate constraints.
5. AS-UNet: A Lightweight U-Net with Asymmetric Strip Attention and Joint Gating for RGB Optical Landslide Segmentation
Core Problem: Small, irregular landslides resemble bare soil and acquire blurred boundaries after image resizing, challenging rapid RGB-only mapping.
Key Innovation: AS-UNet combines asymmetric strip attention, gated skip connections and a composite imbalance-aware loss, reaching 77.82 percent test IoU with 8.634 million parameters.
6. Enhancing Landslide Susceptibility Assessment With a Novel IV ‐ LR ‐ RF Coupled Model: A Comparative Study in Lvhe Town, Qinling‐Daba Mountains, China
Core Problem: Township-scale susceptibility assessment with few historical landslides needs both nonlinear skill and interpretable controls that transfer to new events.
Key Innovation: The slope-unit IV-LR-RF pipeline reaches 0.967 test accuracy and 0.937 AUC, while placing all four independently observed 2025 landslides in moderate- or high-susceptibility zones.
7. Back-analysis of the rock landslide at the northeastern part of Kef Essenoun open-pit mine (NE Algeria): Post-disaster observations and numerical assessment
Core Problem: The failure surface and controlling weak layer of the 2007 Kef Essenoun mine landslide required reconstruction from post-disaster evidence.
Key Innovation: Field observations, limit-equilibrium analysis and finite elements converge on planar sliding through a phosphate-marl layer, with factors of safety of 0.944 and 0.83.
8. LADI: Landslide Activity Detection Index based on InSAR Coherence and Sensitivity
Core Problem: Fast landslide movement can destroy InSAR coherence, but conventional activity detection often requires phase unwrapping and time-series inversion.
Key Innovation: LADI combines geometric sensitivity and coherence loss through a harmonic mean, detecting optically validated active periods from 90 Sentinel-1 interferograms without unwrapping or inversion.
9. Exploratory Q-Network: A DRL Framework for MOP-based GB-SAR Deployment in Open-Pit Coal Mine Slope Monitoring
Core Problem: Selecting a GB-SAR deployment for open-pit slope monitoring is a high-dimensional multi-objective problem in which standard reinforcement learning can stall in local optima.
Key Innovation: Multi-head actions, uncertainty-aware rewards and perturbation-based exploration improve GB-SAR coverage and coherence in open-pit deployment experiments.
10. Three‐Dimensional Numerical Investigation of Landslide Generated Impulse Waves: Generation, Propagation and Overtopping
Core Problem: Reservoir geometry, wave scattering and overtopping interact in three dimensions, while empirical formulas may not identify which waves control total overflow.
Key Innovation: A numerical model validated against a 1:200 physical experiment shows that multiply scattered subsequent waves, not the leading crest, dominate dam overtopping volume.
11. Orographic Effects on Precipitation From Hurricane Helene
Core Problem: Hurricane rainfall over steep terrain is difficult to attribute because storm dynamics and orographic lifting operate simultaneously.
Key Innovation: Paired global cloud-resolving simulations with and without the Appalachians isolate more than 300 mm of upslope enhancement and diagnose the microphysical pathway.
12. Spatiotemporal evolution and economic effects of regional vulnerability under compound disasters: evidence from the Pearl River Delta, China
Core Problem: Coastal vulnerability assessments rarely represent the reinforcing economic and ecological effects of typhoons, extreme rainfall and floods over time.
Key Innovation: A compound-disaster vulnerability index couples typhoon, rainfall and flood exposure with economic evolution to identify persistent and emerging Pearl River Delta hotspots.
13. Mass Loss From Thwaites Glacier Continues Even Without Ocean Melting
Core Problem: It is uncertain whether continued mass loss from Pine Island and Thwaites glaciers can be reversed by removing present-day ocean melt.
Key Innovation: Three ice-sheet models are subjected to zero ocean melt, revealing Pine Island readvance but continued Thwaites retreat for roughly 150 years.
14. SAGINGeo: A Space–Aerial–Ground Integrated Framework for VGI Geolocalization in Multidisaster Scenarios
Core Problem: Disaster imagery submitted from the ground is hard to geolocate when methods ignore the complementary geometry of aerial and satellite views.
Key Innovation: SAGINGeo learns a shared space-aerial-ground representation and releases SAGINDisaster to test volunteered-image localization across multiple disaster settings.
15. Highs and lows: exploring digital elevation model resolution effects on hydrograph model precision
Core Problem: Flood hydrographs derived from watershed geometry can change substantially with DEM resolution, but the resulting model error is rarely measured against events.
Key Innovation: Five DEM resolutions are propagated through two geomorphological hydrograph models and tested against 64 floods, exposing scale-specific performance reversals.
16. Comparative population exposure to compound hot–dry vs. hot–wet extremes in the rapidly urbanizing Pearl River Basin: drivers and implications
Core Problem: Population exposure to compound hot-dry and hot-wet extremes has not been compared consistently within rapidly urbanizing river basins.
Key Innovation: One-kilometre climate and socioeconomic grids separate hot-dry from hot-wet exposure and attribute their divergence to hazard, population and urban growth.
17. Source parameters of shallow earthquakes in the Indian-Arabian-Eurasian plate collision zone
Core Problem: Source properties across the Indian-Arabian-Eurasian collision system remain difficult to compare because regional earthquake solutions are methodologically uneven.
Key Innovation: A single teleseismic waveform-inversion workflow estimates 98 source mechanisms and reveals contrasting thrust geometry across the two collision zones and Makran transition.
18. Seafloor Displacements Across the Mid‐Ocean Ridge Eruption Cycle Modulated by Mush Zone Viscosity
Core Problem: Seafloor deformation through a mid-ocean-ridge eruption cycle is difficult to interpret when magma-lens inflation and mush relaxation overlap.
Key Innovation: Finite-element eruption cycles couple magma-lens inflation, dike drainage and Maxwell mush relaxation to identify viscosity-dependent uplift and subsidence.
19. Marine Heatwave Imprints on Salinity in the Northwest Atlantic
Core Problem: Marine heatwaves are usually defined from temperature alone, leaving the roles of water-mass advection, salinity and stratification in their vertical structure unresolved.
Key Innovation: Satellite salinity and temperature with ocean reanalysis distinguish high-salinity, vertically extensive Gulf Stream events from fresher, surface-confined shelf heatwaves.
20. Metis: Memory Foundation Model
Core Problem: Agent memory is usually bolted on as an external store, leaving persistence and retrieval outside end-to-end foundation-model learning.
Key Innovation: Metis formalizes a persistent internal memory state and native store-use procedures, then trains both as part of the model backbone.
21. Ouroboros-Spatial: Closing the Data-Model Loop for Spatial Reasoning
Core Problem: Static spatial-reasoning datasets waste training on examples that remain too easy or too hard as model capability changes.
Key Innovation: A proposer generates spatial questions plus executable labels, while a solver's confidence selects the next training frontier in an iterative data-model loop.
22. CMIP6-MedPlus dataset: climate projections for the Mediterranean region using statistical downscaling
Core Problem: Global climate projections are too coarse and biased for many Mediterranean assessments that require daily regional precipitation and temperature.
Key Innovation: CMIP6-MedPlus combines CNN spatial refinement with bias correction to provide open 0.25-degree daily projections for 1985-2100 under two emissions pathways.
23. Spectral nudging impacts on precipitation downscaling in the Conformal Cubic Atmospheric Model, version CCAM-2504: insights from summer 2011
Core Problem: Regional precipitation downscaling is sensitive to spectral-nudging wavelength, variables and update frequency, especially during extreme climate modes.
Key Innovation: Sixteen controlled CCAM runs vary wavelength, height, frequency and nudged variables, identifying short-wavelength hourly pressure-wind-temperature nudging as most robust.
24. A distribution-aware Riemannian transformer for intelligent simulation and early warning of risks in geological energy and underground spaces
Core Problem: Geological-energy and underground-space observations are noisy, heterogeneous and often non-Euclidean, weakening conventional image-based warning models.
Key Innovation: RGSTNet embeds feature distributions on a Riemannian geometry before transformer learning, targeting non-Euclidean dependence and noisy risk classes.
25. Seismic Response Analysis and Energy Identification of Corrugated Steel Utility Tunnel Based on Shaking Table Test
Core Problem: The frequency-dependent transfer of earthquake energy between soil and corrugated-steel utility tunnels is not captured by peak response alone.
Key Innovation: Shaking-table measurements and Hilbert-Huang analysis identify frequency-localized energy transfer and the transition from coherent to incoherent tunnel-soil motion.
26. Seismic Response and Failure Mechanisms of Tunnels Constructed Using the Pipe Roof Preconstruction Method
Core Problem: The earthquake response of pipe-roof preconstruction tunnels is poorly constrained once soil interaction, composite separation and local damage evolve together.
Key Innovation: A refined 3D soil-structure model tracks cracking, compression, racking and steel-concrete separation under nonlinear earthquake histories.
27. Interface-slip localization and vulnerability amplification in historic masonry pagoda walls under seismic loading
Core Problem: Weak mortar interfaces in historic masonry can localize earthquake damage, but continuum models blur the transition from cracking to sliding.
Key Innovation: A cohesive-interface 3D micro-model reproduces experimental strength within 3 percent and shows earthquake loading shifting dissipation toward basal interface slip.
28. Simplified discontinuum out-of-plane seismic analysis of archetypal Italian masonry church gabled façades
Core Problem: Historic church facades can overturn as discrete macroelements, while simplified seismic checks struggle to represent geometry and block interaction.
Key Innovation: Evidence-based archetypal geometries are tested in a simplified 3DEC discontinuum framework to map facade overturning across structural parameters.
29. Directional peak factors of strong-motion response spectra: a stochastic field representation on the circle
Core Problem: Orientation-independent response spectra discard directional peak fluctuations that can matter for strong-motion demand.
Key Innovation: More than 2,000 strong-motion pairs are represented as a pi-periodic stochastic field, preserving period- and orientation-dependent peak-factor structure.
30. Optimization of coal mine gas extraction: Hydraulic fracturing and machine learning
Core Problem: Coal-seam gas extraction depends nonlinearly on hydraulic-fracture geometry, making both parameter choice and production forecasting difficult.
Key Innovation: Staged Abaqus fracture simulations generate physical predictors for RF, MLP, LSTM, GRU and hybrid forecasts of gas-extraction volume.
31. A signal-processing enhanced CPTu framework for evaluating small-strain shear modulus of marine soils
Core Problem: Direct measurement of small-strain marine-soil stiffness is costly, while common CPTu correlations do not transfer reliably across sediment types.
Key Innovation: Field CPTu signals and laboratory resonant-column tests are combined in a soil-type-aware framework for estimating small-strain shear modulus at three marine sites.
32. Sealing performance evaluation of salt cavern hydrogen storage under multi-interlayer salt deposit geological conditions
Core Problem: Non-salt interlayers may create preferential hydrogen leakage paths in bedded salt caverns despite the low permeability of rock salt.
Key Innovation: Measured hydrogen and methane permeability is embedded in coupled THM simulations to quantify horizontal seepage through multiple cavern interlayers.
33. Drone-mounted VLF-EM survey for electrical resistivity imaging and its application in complex coastal saline permafrost environments
Core Problem: Saline coastal permafrost varies sharply over engineering distances, yet dense resistivity imaging is difficult to acquire safely and efficiently.
Key Innovation: A drone-mounted VLF-EM survey with constrained 2D tipper inversion is tested synthetically and over 36,000 square metres of Alaskan polygonal tundra.
34. Long-term monitoring of active layer, ice wedges, and frost cracking conditions in northern Victoria Land, Antarctica (2004–2023)
Core Problem: Antarctic active-layer thickness, ice-wedge temperature and frost cracking do not necessarily respond together over decadal climate variability.
Key Innovation: A 2004-2023 altitudinal transect jointly tracks air, surface, permafrost, active-layer and cracking indicators, revealing distinct site-specific trajectories.
35. Risk-informed predictive emergency control for safety and resilience enhancement of canal-tunnel water conveyance systems
Core Problem: Fixed emergency controls cannot balance tunnel overpressure, canal overtopping and water loss as a conveyance disturbance evolves.
Key Innovation: State-dependent adaptive model predictive control changes safety and efficiency priorities as mixed-flow transients move through a 40 km canal-tunnel system.
36. Morphological optimization of urban small watershed blue–green infrastructure to enhance stormwater management benefits: a case study of the Xuanwu Lake Watershed, Nanjing
Core Problem: Urban blue-green infrastructure is usually optimized by area or type rather than the joint morphology of networks, corridors and distributed storage.
Key Innovation: Twenty-six SWMM scenarios jointly vary river networks, bioretention, green roofs, configuration and connectivity to identify morphology-driven stormwater gains.
37. Joint optimization of spatiotemporal convolution and time–frequency domain: A new method for spatial interpolation of broadband ground motions
Core Problem: Rapid post-earthquake assessment needs broadband ground-motion histories at uninstrumented sites without running a full source-to-site physical simulation.
Key Innovation: A ConvLSTM3D-LSTM model combines station geometry with time- and frequency-domain losses to synthesize 0.1-25 Hz histories at unobserved sites.
38. Nonlinear macro-modeling and machine-learning surrogates for seismic assessment of precast column–foundation connections
Core Problem: Detailed cyclic models of precast column-foundation connections are too expensive for broad seismic assessment and design exploration.
Key Innovation: A validated nonlinear macro-model generates 100,000 cyclic-response samples for machine-learning surrogates spanning four precast connection types.
39. Seismic performance analysis of a double-layer wheel spoke cable-net roof stadium under spatially varying seismic motions including rotational components
Core Problem: Large cable-net roofs face spatially variable and rotational earthquake input that conventional uniform translational analyses omit.
Key Innovation: A refined stadium model is driven by synthesized spatially variable motions that include both translational and rotational components over heterogeneous soil.
40. Multivariate return period theory of three-dimensional mainshock and aftershock ground motions and its application in seismic analysis of a tunnel
Core Problem: Seismic design commonly separates horizontal, vertical, mainshock and aftershock demand, obscuring their joint return period and accumulated damage.
Key Innovation: A multivariate Gaussian-mixture formulation defines joint return periods for three-dimensional mainshock-aftershock pairs and carries them into tunnel analysis.
41. An inerter-enhanced bistable nonlinear energy sink to suppress seismic ground excitations
Core Problem: Passive seismic absorbers often trade broadband effectiveness against stroke demand and sensitivity to earthquake frequency content.
Key Innovation: A ball-screw inerter is integrated with a bistable nonlinear energy sink and validated on a shaking table under varied starting states and earthquake inputs.
42. Hybrid test for ground-tunnel interaction subjected to seismic loadings
Core Problem: Hybrid tests of buried structures lack a general interface that transmits global soil-domain response to a physical tunnel specimen.
Key Innovation: A displacement-based actuator interface couples numerical half-space propagation to a physical soil-tunnel subdomain and is demonstrated in hybrid tests.
43. Shaking table tests of self-centering railway bridge piers with replaceable sacrificial components
Core Problem: Self-centering railway piers reduce residual drift but can lack the stiffness and replaceable damage mechanisms needed for serviceability.
Key Innovation: A scaled shaking-table programme compares free-rocking and sacrificial-component piers, quantifying stiffness, isolation and self-centering trade-offs.
44. Response of a circular tunnel in saturated potentially liquefiable ground with an annular rubber soil mixture cushion under moving train loading. A three-dimensional finite element study
Core Problem: A tunnel in saturated, potentially liquefiable soil must reduce lining demand without creating unacceptable movement under repeated train loading.
Key Innovation: Coupled solid-fluid finite elements show a rubber-soil cushion cutting peak lining force by about 62 percent while explicitly tracking added movement and settlement.
45. Hybrid AI and finite element framework for seismic assessment of 3D-Printed recycled concrete building wall systems under soil-structure interaction
Core Problem: Seismic response of 3D-printed recycled-concrete walls couples print anisotropy, material variability and soil-structure interaction.
Key Innovation: High-fidelity finite-element simulations encode directional material properties and SSI, then train AI surrogates for fast cyclic-response prediction.
46. Teleseismic S‐Wave Coda Autocorrelation and vp/vs Distribution Over the Antarctic Ice Sheet
Core Problem: Sparse Antarctic shear-wave observations limit estimates of ice-sheet elastic structure, mechanical state and internal layering.
Key Innovation: Autocorrelation of both P- and S-wave teleseismic coda yields 28 new vp/vs estimates and expands Antarctic coverage to 61 sites.
47. Stable Water Isotopes Improve Calibration and Flow Path Identification in Integrated Hydrological Models
Core Problem: Integrated hydrological models can match discharge and groundwater levels while representing the catchment's internal flow paths incorrectly.
Key Innovation: Joint calibration to discharge, groundwater levels and stream isotopes narrows parameter equifinality and discriminates plausible surface-subsurface flow paths.
48. Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models
Core Problem: When image evidence conflicts with a language prior, it is unclear whether multimodal models fail to encode the visual attribute or fail to use information they already represent.
Key Innovation: WhatIfVis combines reconstruction, image-prior conflicts, activation patching and steering across six models to localize and control failures in visual-context use.
49. Simultaneous Coverage and Efficiency Guarantee in Online Conformal Prediction
Core Problem: Online conformal prediction can satisfy signed average coverage while remaining miscalibrated for long intervals or producing unhelpfully wide sets.
Key Innovation: The method controls unsigned interval miscoverage and regret in set size against a shifting comparator, joining validity and efficiency in one online guarantee.
50. See2Think: Do Multimodal Models Really Use Intermediate Visual States?
Core Problem: Final-answer benchmarks do not reveal whether multimodal models generate faithful intermediate images or actually use those visual states in later reasoning.
Key Innovation: See2ThinkBench contributes 1,200 visually dependent tasks, while controlled corruption of intermediate states tests generation, rendering, feedback uptake and behavioral dependence.
51. What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations
Core Problem: A predictive latent may appear physical without revealing which hidden parameters are actually identifiable from its inputs and objectives.
Key Innovation: Certificate-gated interventions first prove each physical parameter observable, then map how inputs and prediction targets determine its retention in the latent.
52. Visual Credit Audit for Multimodal Spatial Reasoning
Core Problem: Closed spatial benchmarks can award correct answers even when the image contributes no decision-relevant evidence beyond textual priors.
Key Innovation: Visual Credit Audit compares image-conditioned decisions with text-only, blank and permuted-image controls to compute dependence-credited correctness.
53. Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark
Core Problem: Marginal conformal coverage can conceal near-total undercoverage of rare, costly classes in imbalanced high-stakes decisions.
Key Innovation: Across 3,150 runs on 15 datasets, Mondrian conformal prediction restores minority-class coverage and cost-aware abstention routes uncertain cases to human review.
54. Explainable and Resource-Efficient Spatial Reasoning in Multimodal LLMs for Decision-Critical Applications
Core Problem: Multimodal models often hallucinate fine-grained spatial relations, while depth-only prompting cannot explain same-plane projective and topological judgments.
Key Innovation: ByDeWay-V2 injects detector-derived depth and human-readable spatial predicates without training, improving spatial F1 while operating on CPU within a 40-token budget.
55. Chaos Is a LADDER: Domain Generalization Beyond Invariance via Reweighting
Core Problem: Domain-invariant predictors fail when the mapping from causal content to outcome genuinely changes across environments.
Key Innovation: LADDER disentangles causal content from style, then uses target style to reweight source-specific predictors without treating style as a direct cause.
56. LiDARDraft: Generating LiDAR Point Cloud from Versatile Inputs
Core Problem: Controllable LiDAR generation struggles to reconcile simple text or image inputs with the complex geometry and sampling distribution of point clouds.
Key Innovation: Text, images and point clouds are converted into a common 3D layout that supplies semantic and depth controls to a range-map diffusion generator.
57. MapTab: A Diagnostic Benchmark for Long-Horizon Multi-Criteria Multimodal Reasoning on Heterogeneous Topological Graphs
Core Problem: Multimodal models lack a controlled test of long-horizon route reasoning that combines map topology with several tabular constraints.
Key Innovation: MapTab contributes 196,800 route queries spanning 160 metro systems and tests joint reasoning over visual networks, time, price and other criteria.
58. When Does Visual Token Pruning Improve Calibration? The Role of Evidence Coverage in MLLMs
Core Problem: Visual-token pruning is usually judged by task accuracy, obscuring how removed evidence changes calibration and overconfidence.
Key Innovation: Controlled pruning experiments measure retained evidence coverage alongside accuracy and calibration, exposing overconfidence when decisive visual tokens are removed.
59. VistaHop: Benchmarking Long-Horizon Visual DeepSearch
Core Problem: Existing visual benchmarks rarely require repeated inspection and long-horizon traversal of fine-grained evidence across an image collection.
Key Innovation: VistaHop pairs a long-horizon visual-search benchmark with a tool-use arena that requires iterative inspection rather than one-shot image answering.
60. LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models
Core Problem: High-resolution multimodal models may contain the decisive local detail yet fail to select and use it amid redundant visual context.
Key Innovation: LOCUS trains local-cue selection as a search process so models learn to isolate decisive high-resolution regions without requiring extra inference-time inputs.
61. Frequency-Aware Prompt Learning for Remote Sensing Image Super-Resolution
Core Problem: Remote-sensing super-resolution models with static features often blur scene-specific edges and textures under compound degradation.
Key Innovation: A Fourier-aware transformer and learned prompts adapt high-frequency restoration to the degradation and content of each remote-sensing scene.
62. CAM-CD: Change-Aware Mamba for Remote Sensing Change Detection
Core Problem: Uniform state-space scans spend equal capacity on unchanged background and sparse, geometrically diverse changes in bi-temporal imagery.
Key Innovation: Spatial change priors steer Mamba state transitions toward likely change regions, suppressing static background with little parameter overhead.
63. Exploring recovery complexity in the context of consecutive disasters
Core Problem: A second disaster can interrupt recovery from the first and alter vulnerability across social domains, yet recovery models usually isolate events.
Key Innovation: A four-domain framework traces how incomplete recovery changes settlements, health, economies and governance before the next event arrives.
64. A community-driven metadata model for describing and integrating Controlled Source Seismic data in multidisciplinary data-driven Earth Sciences
Core Problem: Controlled-source seismic datasets are difficult to find, compare and combine because acquisition and access metadata are inconsistent.
Key Innovation: An EPOS community model standardizes acquisition, access and reuse metadata so controlled-source seismic surveys can participate in FAIR multidisciplinary workflows.
65. GEE-DisALEXI: cloud-based implementation of the DisALEXI model for evapotranspiration monitoring using Google Earth Engine
Core Problem: Field-scale evapotranspiration is valuable for drought and water management, but established retrieval models remain difficult to run consistently over large regions and climate zones.
Key Innovation: GEE-DisALEXI ports a two-source energy-balance workflow to Google Earth Engine, validates it from flux sites to basins, and compares evaporative anomalies with the U.S. Drought Monitor.
66. Limits to GRACE-based groundwater storage monitoring in a 23 000 km 2 coastal basin: evidence from the Lower Kutai Basin, Indonesia
Core Problem: Coastal GRACE groundwater estimates can be dominated by ocean leakage and unresolved surface-water terms in basins near the product's spatial limit.
Key Innovation: Paired land-ocean residual tests expose leakage correlations up to 0.68 and show when inferred groundwater changes become physically implausible.
67. Unveiling cascading lag effects of wetland methane emissions: Evidence from Lake Chad in Africa
Core Problem: Tropical wetland methane seasonality remains uncertain because precipitation, inundation, vegetation succession and emissions operate with different and poorly resolved lags.
Key Innovation: Satellite methane inversions and high-resolution open-water and inundated-vegetation maps resolve a three-month precipitation-to-inundation lag followed by a four-month emissions lag.
68. Are low-stress zones possible with ultradeep and intense extrusion?
Core Problem: Direct stress measurements below six kilometres are scarce, leaving the origin of apparent low-stress zones in intensely extruded crust unresolved.
Key Innovation: Three-dimensional seismic data, deep-well logs and reservoir geomechanics are integrated to test how extrusion and structure create localized stress relief.
69. Wave climates and cay-reef dimensions are strong predictors of cay geomorphic activity within the Great Barrier Reef, Australia over the past eight years
Core Problem: Low reef islands can shift rapidly, but the relative control of wave climate and reef-cay geometry is poorly constrained across sites.
Key Innovation: Forty-seven hydrodynamic and geometric variables across 16 Great Barrier Reef cays are reduced to the wave and reef dimensions that best predict mobility.
70. GeoGS: Geometric Prior-Guided Gaussian Splatting for robust urban reconstruction from sparse views
Core Problem: Sparse-view urban Gaussian splatting becomes incomplete or geometrically unstable when camera overlap, texture and visibility are limited.
Key Innovation: GeoGS initializes from CityGML LoD2 surfaces and adds foundation-model depth cues, using complementary global and local priors where views are sparse.
71. MapGlue: Multimodal remote sensing image matching
Core Problem: Cross-modal remote-sensing matching lacks both a globally varied benchmark and a representation robust to radiometric, geometric and viewpoint differences.
Key Innovation: MapData supplies 121,781 aligned map-visible pairs, while MapGlue combines saliency-guided keypoints with semantic information for cross-modal matching.
72. An interpretable Earth observation framework for estimating lake dissolved organic carbon in Sweden using Sentinel-2 and ERA5-Land data
Core Problem: Lake dissolved organic carbon is ecologically important but difficult to estimate consistently from Earth observation across seasons and regions.
Key Innovation: Sentinel-2, ERA5-Land and geography are combined with interpretable XGBoost and spatially separated validation to produce quarterly 10 m DOC maps.
73. Earth Observation–Based Machine learning for operational forecasting of harmful algal blooms in the Arabian Gulf and Sea of Oman
Core Problem: Coastal harmful-algal-bloom forecasts often fit historical variability without demonstrating skill in strictly future, operational periods.
Key Innovation: Region-specific bloom thresholds and unchanged operational decision thresholds are evaluated on independent and strictly forward 2021-2024 periods.
74. Quantifying Water Resource Changes across Natural and Managed Landscapes in the Indus Basin Drainage Catchments: A Geographically Weighted Machine Learning Approach for GRACE Downscaling
Core Problem: GRACE water-storage observations are too coarse to distinguish management and climate signals among individual Indus Basin catchments.
Key Innovation: Hybrid geographically weighted random forest downscales three GRACE products to 1 km and separates terrestrial and groundwater-storage change across managed and natural catchments.
75. Riverine land cover mapping through semantic segmentation of multispectral point clouds
Core Problem: Riverine land-cover mapping must separate spectrally and geometrically similar gravel, sand, vegetation and water in irregular point clouds.
Key Innovation: Point Transformer v2 fuses spectral and geometric LiDAR channels, and sparse annotations from a second river test cross-site generalization.
76. Evolution of water and sediment inflows to the largest reservoir system in Southwest China: Historical patterns and future predictions
Core Problem: After cascade-reservoir development shifted sediment supply toward tributaries, the historical controls and future trajectory of those inflows remained poorly constrained.
Key Innovation: Long-term observations, basin descriptors and interpretable monthly models with KGE of at least 0.80 reconstruct tributary water-sediment controls and project 2026-2056 inflows.
77. Mapping temperate groundwater-dependent terrestrial ecosystems in Denmark
Core Problem: Groundwater-dependent ecosystems are hard to map in temperate regions because their vegetation can resemble surrounding landscapes outside drought periods.
Key Innovation: Positive-unlabelled learning combines drought-season NDVI, delayed spring green-up and height above drainage to find seven times more candidate ecosystems than field maps alone.
78. Multi-modal dual-target spatio-temporal graph learning for regional evapotranspiration estimation
Core Problem: Regional evapotranspiration estimates inherit different biases from meteorological formulas and satellite retrievals, and their spatial dependence is underused.
Key Innovation: Modality-specific graph encoders, recurrent temporal learning and a learned fusion gate are trained jointly against satellite ET and Penman-Monteith reference ET.
79. Future climate and land-use change reshape watershed best management practices portfolio for nutrient control
Core Problem: Watershed nutrient controls chosen for the present climate may fail as climate and land use alter runoff, loading and management trade-offs.
Key Innovation: CMIP6 and land-use scenarios are crossed in a fixing-changing design, then management portfolios are ranked by nutrient reduction, cost and yield loss.
80. R-SLPR: Region-based Small-to-Large Point-cloud Registration with Contrastive Learning
Core Problem: Registering a small, incomplete point cloud against a much larger reference is unstable when overlap is low and the matching region is unknown.
Key Innovation: R-SLPR proposes candidate regions with a Fibonacci grid, learns local matches contrastively and refines anchors iteratively under severe scale mismatch.
81. Long-Tailed 3D Point Cloud Dataset Distillation
Core Problem: Compact synthetic point-cloud training sets can preserve dominant classes while erasing rare objects when the source distribution is long-tailed.
Key Innovation: The method co-optimizes synthetic point sets, labels and an augmentation network so a compact distilled set better preserves minority-class performance under long-tailed training.
82. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization without Ray Matching
Core Problem: Indoor localization against repetitive floorplans is intrinsically multimodal, so committing early to one pose can discard valid alternatives.
Key Innovation: A diffusion-based coarse stage preserves multiple pose hypotheses before visual-floorplan matching refines them to sub-metre locations.
83. A Closer Look at Dynamic Scene Graph Generation In the Era of Multimodal Large Language Models
Core Problem: Recall-heavy scene-graph metrics reward redundant relations and poorly represent the precision, relevance and temporal continuity needed in practice.
Key Innovation: The benchmark supplements recall with practical precision, temporal relation and relevance measures, then tests architectures against those deployment-oriented criteria.
84. Distill, Diffuse, Segment: Unsupervised 3D Semantic Segmentation for Autonomous Driving Based on Multi-Level Distillation and Graph Diffusion
Core Problem: Dense LiDAR annotation is costly, while small and rare safety-relevant objects are easily lost during unsupervised representation transfer.
Key Innovation: Multi-level distillation and graph diffusion transfer semantics without point-wise labels while explicitly preserving small and long-tailed LiDAR classes.
85. Frequency-Enhanced and Multimodal Fusion Visual State Space Model for Tree Crown Segmentation Using UAV Remote Sensing Images
Core Problem: Tree-crown boundaries in steep UAV imagery are obscured by shadows, terrain and overlapping spectral signatures.
Key Innovation: FMF-VMUNet combines frequency attention, a canopy-height model and transformer fusion to recover crown edges under terrain and shadow interference.
86. A Slope-Stratified Modeling Strategy for Canopy Height Estimation: Fusing GEDI LiDAR and Multisatellite Data With XGBoost
Core Problem: GEDI-based canopy-height models lose accuracy in complex terrain when topographic effects are treated as a single global relationship.
Key Innovation: Separate slope strata are learned from GEDI and multisatellite predictors, reducing relief-induced bias instead of forcing one canopy-height model across all terrain.
87. Fourier-Calibrated Dynamic Prompts for Universal Hyperspectral Image Restoration
Core Problem: A single hyperspectral restoration model must recognize unknown degradation type and severity rather than rely on fixed textual prompts.
Key Innovation: Spatial statistics and Fourier energy identify degradation prototypes, then condition continuous prompts on both degradation class and intensity.
88. A Plug-and-Play Data-Centric Training Framework for Efficient Hyperspectral Image Classification
Core Problem: Hyperspectral classifiers incur high training cost when spatial correlation and spectral redundancy are ignored in data selection.
Key Innovation: PACE creates physically consistent spatial-spectral pseudo-labels and progressively constrains training samples without changing the underlying classifier.
89. A Hybrid Visual Mamba Network With Global-Local Perception for Land Use and Land Cover Semantic Segmentation
Core Problem: Land-cover segmentation needs long-range context without sacrificing the local edges and small objects that define mapped classes.
Key Innovation: Hybrid visual Mamba blocks capture long-range dependencies while cross-scale attention reconciles local detail across LoveDA and UAVid scenes.
90. Narrowband Suppressive Interference Suppression for UAV-Borne SAR Based on TSA-ISN
Core Problem: UAV-borne SAR imagery is vulnerable to narrowband interference, while onboard processing must preserve phase under tight compute limits.
Key Innovation: TSA-ISN works directly on intermediate-frequency time-series data with temporal and channel attention, preserving phase for resource-limited UAV platforms.
91. An Integrated Deep Learning Approach to Estimate Canopy Height and Uncertainty by Combining Seasonal Optical, SAR and Limited GEDI LiDAR Data over Ontario’s Managed Forests, Canada
Core Problem: Canopy-height products need uncertainty estimates when optical, SAR and sparse GEDI observations vary seasonally and spatially.
Key Innovation: Seasonal optical and SAR signals are fused with limited GEDI samples in a deep model that predicts both canopy height and spatial uncertainty.
92. A Geographic Consistency-Constrained Cross-Modal Super-Resolution Matching Method for UAV Geo-Localization
Core Problem: UAV geolocation across ground and overhead imagery fails when super-resolution improves appearance without preserving geographic correspondence.
Key Innovation: Cross-modal super-resolution matching is regularized by geographic consistency, preserving location identity while reconciling UAV and overhead scale differences.
93. Automatic Extraction of Oxbow Lakes from Landsat Imagery Using Deep Learning Across the Amazon Basin
Core Problem: Oxbow-lake inventories across the Amazon are incomplete because manual mapping cannot follow thousands of evolving floodplain features.
Key Innovation: A basin-scale deep-learning workflow extracts oxbow lakes from the Landsat archive and replaces local manual inventories with consistent automated mapping.
94. Quantifying Tidal Asymmetry of Suspended Sediment Concentration in Macro-Tidal Embayments: A Sentinel-2 Based Framework
Core Problem: Tidal asymmetry in suspended sediment is difficult to map across large embayments from sparse in-situ measurements.
Key Innovation: A Sentinel-2 retrieval framework maps suspended-sediment concentration through tidal phases and derives spatially distributed flood-ebb asymmetry.
95. Tracking forest carbon and growth in large-scale passive restoration using multi-temporal airborne lidar
Core Problem: Regional restoration monitoring cannot infer subtle forest-carbon gains reliably when airborne LiDAR surveys differ in terrain offset and pulse density.
Key Innovation: Inter-survey terrain and pulse-density corrections permit 30 m forest-height and carbon-change mapping, then support 30-year restoration projections.
96. Influence of extensive gold mining in the southwestern Amazon on the spatial and temporal dynamics of suspended sediment transport in the Madre de Dios River
Core Problem: Gold-mining impacts on Amazonian suspended sediment are documented at river endpoints, but along-channel, cross-sectional and grain-size pathways remain weakly quantified.
Key Innovation: Harmonized Sentinel-2 and Landsat records plus field sampling separate two mining phases, reveal paleochannel reactivation and trace cohesive sediment relevant to mercury transport.
97. Extracting Linear Structures from LiDAR Data by Msplit Estimation and Different Segmentation Approaches
Core Problem: Linear features in LiDAR point clouds are sensitive to segmentation width, computational cost and outliers in the observation set.
Key Innovation: Msplit estimation is paired with a hybrid disjoint-window segmentation scheme that roughly halves runtime while retaining sliding-window accuracy.
98. Monitoring Coastal Surface Roughness Suppression Associated with Algal Blooms and Pollutants Using NASA SWOT Observations
Core Problem: Algal blooms and surface pollutants can damp short ocean waves, but the detectability of that roughness signal in SWOT observations is uncertain.
Key Innovation: SWOT observations are compared with coastal conditions to test whether wave-damping roughness anomalies can reveal blooms and pollutant films.
99. ParkingTwin: Training-free streaming 3D reconstruction for parking-lot digital twins
Core Problem: Parking-lot digital twins must be reconstructed from sparse moving views despite dynamic occlusion, lighting change and edge-compute limits.
Key Innovation: Structural priors support a training-free streaming pipeline that reconstructs geometry despite forward-view degeneration, occlusion and lighting change.
100. A transferable framework for roof vegetation mapping across tropical cities based on open-source multimodal data and morphology-aware UDASS
Core Problem: Roof-vegetation mapping does not transfer well across tropical cities when high-resolution imagery and labeled target data are unavailable.
Key Innovation: Open satellite, Google and building-footprint data feed a multimodal encoder, while morphology-aware pseudo-label correction enables unsupervised city-to-city transfer.