TerraMosaic Daily Digest: July 26, 2026
Daily Summary
Earthquake analysis is moving beyond stationary empirical summaries toward structure-aware and region-specific representations of rupture and shaking. A diffusion model recasts aftershock forecasting as conditional generation of spatiotemporal rate and magnitude fields and recovers anisotropic, fault-controlled patterns that fixed kernels miss. At the fault scale, critical slip distance is reinterpreted as a scale-dependent structural property of self-affine roughness rather than a constant material parameter, while tectonic tremor beneath the Antarctic-South American margin reveals active deep slow deformation where ordinary seismicity offers little guidance. At the regional scale, new probabilistic hazard models for the East Anatolian Fault Zone explicitly include cascading multi-segment rupture, and strong-motion simulations of the Luding earthquake show how wave-facing slopes, dip-aligned peak ground velocity, and vertical shaking concentrate landslide-prone amplification.
Water-driven instability is another dominant thread, with water resolved not only as a trigger but also as an observable state variable that can be reconstructed, forecast, and mechanistically partitioned. Physical experiments on reservoir landslides show that buoyancy and drawdown seepage destabilize slopes at different stages of water-level fluctuation, and flume experiments on landslide dams show that seepage can either accelerate failure or stabilize coarse-grained dams depending on grain-size sorting and inflow. Field monitoring of a cold-region levee demonstrates that freeze-thaw driven densification can lower hydraulic conductivity by orders of magnitude and slow internal water response through time. In parallel, flood studies extract river bathymetry from SWOT water-surface observations, formalize forecaster expertise into explicit operational workflows, and distill black-box rainfall-runoff models into interpretable hydrological concepts that remain faithful to basin-scale behavior.
Remote-sensing research increasingly emphasizes dynamic monitoring, cross-event transfer, and uncertainty propagation rather than static map production alone. High-resolution SAR time series detect post-earthquake reconstruction without labels, while the BRIGHT challenge shows that instance-level building damage mapping still suffers a steep generalization gap even when public baselines are greatly surpassed. Several studies strengthen the sensor-method interface itself: SAR pretraining targets are redesigned around structural stability under speckle, large curated SAR and hydrologic datasets are released for Alpine and continental-scale learning, and an H-A-V-Q framework links inland-water height, area, storage, and water quality into a single uncertainty-aware system. Environmental hazard monitoring follows the same pattern, from TROPOMI-based methane quantification over coal mines to test-time-adaptive chlorophyll retrieval and phylum-resolved cyanobacterial bloom severity mapping, producing operational geoscience products explicitly evaluated against domain shift.
Key Trends
Five methodological shifts connect structure-aware hazard prediction, coupled hydrologic controls, time-series SAR, realistic transfer tests and explicit uncertainty propagation.
- Generative and hybrid models are replacing fixed hazard kernels: Several papers replace rigid empirical forms with models that learn structured physical variability: QuakeGen generates aftershock fields instead of fitting isotropic point-process kernels, hybrid rainfall-runoff models retain accuracy under warmer climates better than pure LSTMs, and post-FWI flow matching injects geological priors without rerunning inversion.
- Hydrologic controls are being resolved as coupled process systems: The strongest landslide and flood papers do not treat water as a single forcing term. They separate buoyancy from seepage during reservoir drawdown, track failure-mode transitions in seepage-prone landslide dams, document time-evolving levee conductivity under freeze-thaw, and quantify reversed surface-water and groundwater exchange under reservoir regulation.
- Time-series SAR is becoming a full life-cycle hazard sensor: Across recovery mapping, ground-motion services, Alpine benchmarks, and structural SAR pretraining, SAR is used not only for deformation snapshots but for long-duration monitoring, cross-scene transfer, and early reconstruction detection. The emphasis is shifting from isolated images to temporally stable, physics-aware representations.
- Transferability and realistic validation are now central scientific questions: A recurring message is that performance collapses under spatial, climatic, or event transfer. The BRIGHT challenge exposes cross-event fragility in all-weather damage mapping, spatial cross-validation changes land-cover model rankings, and controlled experiments on Burgers dynamics show that missing physical regimes in training data sharply limit extrapolation.
- Operational geoscience products are being built with explicit uncertainty budgets: Multiple papers now carry uncertainty through the product rather than adding it after the fact: inland-water monitoring propagates errors from height and area into storage and quality, rockfill-dam inversion tolerates modeled uncertainty inside the objective function, covariance-boosted Gaussian processes prevent overconfident irregular-field prediction, and satellite methane inversions attribute much of their spread to wind forcing.
Selected Papers
Across earthquakes, floods, landslides, water quality, and deformation monitoring, static averages are giving way to dynamic, physically anchored representations evaluated across regions, events, and sensing conditions.
1. Earthquake Aftershock Forecasting using Conditional Generative Models
Core Problem: Operational aftershock forecasting still relies on point-process models with fixed temporal decay and isotropic spatial kernels that miss fault-controlled anisotropy and sequence-specific productivity.
Key Innovation: QuakeGen reframes the task as conditional generation of spatiotemporal fields, using a diffusion model to forecast aftershock rate and maximum magnitude from recent seismicity and forecast horizon, outperforming the USGS Reasenberg-Jones baseline on global sequences.
2. Flood Inundation Modeling Using the Surface Water Ocean Topography Mission
Core Problem: Flood inundation models are often limited by the absence of reliable river bathymetry, especially on rivers too narrow for standard satellite altimetry assumptions.
Key Innovation: The study derives bathymetry from multi-temporal SWOT water-surface heights, introduces a Tulip-shaped misfit to suppress problematic observations, and shows that SWOT-derived bathymetry can drive flood simulations with accuracy comparable to sonar-based models.
3. Critical slip distance on rough faults
Core Problem: The physical origin and scale dependence of critical slip distance remain unresolved, leaving a gap between laboratory friction parameters and meter-scale natural ruptures.
Key Innovation: The paper derives critical slip distance from self-affine fault roughness, shows that rate-and-state friction emerges as a mean-field description of fractal asperity dynamics, and introduces fault retentivity as a multiscale measure of rupture arrest capacity.
4. Monitoring Post-Disaster Urban Recovery Using High-Resolution SAR Time Series and Unsupervised Learning: Evidence from the 2023 T\"urkiye-Syria Earthquake
Core Problem: Post-disaster reconstruction is difficult to monitor when labeled recovery data are sparse and recovery signals evolve over time across heterogeneous urban settings.
Key Innovation: Using COSMO-SkyMed time series and unsupervised anomaly detection, the framework maps persistent reconstruction signatures and distinguishes rebuilt districts, cleared areas, and container settlements while complementing nighttime-light indicators.
5. Failure process and characteristics of landslide dams under different seepage effects: Experimental study
Core Problem: Failure in landslide dams is usually studied under isolated seepage factors, even though real dams experience coupled effects of inflow, geometry, grain sorting, and seepage intensity.
Key Innovation: Fifty-four flume experiments reveal shifts from overtopping to mixed overtopping-seepage and pure seepage failure, identify coarsening and seepage-stable modes in coarse dams, and yield a validated predictor for failure-mode transition.
6. Characteristics and triggering mechanism of buoyancy-driven and seepage-driven landslides: insights from physical model tests
Core Problem: Reservoir-landslide instability under fluctuating water levels is not well partitioned into the respective roles of buoyancy, seepage, and permeability-dependent deformation timing.
Key Innovation: Physical model tests separate buoyancy-driven and seepage-driven mechanisms, show instability peaks at high water level and during drawdown, and identify soil pressure as an earlier warning indicator than displacement.
7. Tectonic Tremor Reveals Slow Deformation Along the Antarctic–South American Subduction Zone
Core Problem: The Antarctic-South American subduction zone remains so weakly characterized that even the existence of active deep plate-boundary deformation was uncertain.
Key Innovation: Continuous seismic observations reveal tectonic tremor at 30-40 km depth, with migration and tidal modulation that directly indicate slow deformation and help constrain the geometry of the deep plate interface.
8. Advancing All-Weather Building Damage Mapping to the Instance Level: Outcomes and Insights from the 2026 Bright Challenge
Core Problem: Building-level all-weather damage mapping from pre-event optical and post-event SAR data remains weakly generalized across disaster events and hazard types.
Key Innovation: The BRIGHT challenge extends the dataset to roughly 291,000 instance-labeled buildings across 16 events and shows that staged or late optical-SAR fusion and optical-dominant localization help, even though performance still collapses sharply on unseen events.
9. Context-Aware Concept Distillation for Trustworthy Flood Prediction
Core Problem: Flood forecasting models can be accurate yet remain unusable for high-stakes decisions because their explanations do not resolve into hydrologically meaningful causal concepts.
Key Innovation: Context-Aware Concept Distillation discovers an unsupervised hydrological concept vocabulary and modulates it with basin characteristics through a residual hypernetwork, producing interpretable surrogates with strong fidelity across 5,203 basins.
10. SARATR-X-v2: Scale-Aware Structural Pre-Training for SAR Foundation Models
Core Problem: Masked-reconstruction pretraining for SAR lacks a principled target that is simultaneously stable under speckle and compatible with the multiple spatial scales demanded by downstream tasks.
Key Innovation: SARATR-X-v2 builds a unified supervision target from fixed structural extractors across six receptive fields, yielding state-of-the-art transfer on 12 benchmarks and far lower representation drift under synthetic speckle perturbation.
11. HydroAgent: Formalizing Forecaster Expertise into Skill-Orchestrated Flood Forecasting Workflows
Core Problem: Tacit flood-forecaster expertise is effective in practice but difficult to formalize, audit, and reuse inside model-driven warning workflows.
Key Innovation: HydroAgent encodes explicit expert rules as bounded skills around LLM reasoning, improving scheme selection on a strong forecasting baseline and reproducing observed peak flow and flood volume within prior judgment ranges for most tested events.
12. Haicheng and Tangshan Earthquakes as potential Dragon-Kings
Core Problem: It remains unclear whether some great earthquakes are simply the tail of Gutenberg-Richter statistics or arise from distinct maturation dynamics that make them true outliers.
Key Innovation: A combined KDE clustering and sequential outlier-testing framework identifies dragon-king signatures for the Haicheng and Tangshan earthquakes, with especially strong pre-mainshock evidence for Haicheng.
13. FLAME: Physics-Guided Neural Operators for Onboard Satellite Methane Detection in Hyperspectral Imagery
Core Problem: Methane plume detection from orbital hyperspectral data must operate under tight onboard compute budgets without sacrificing physical fidelity or false-positive control.
Key Innovation: FLAME embeds methane absorption physics into a neural-operator architecture, achieving the best benchmark accuracy, nearly tripling false-positive reduction over the strongest learned baseline, and meeting onboard latency constraints.
14. Comprehensive probabilistic seismic hazard assessment for the East Anatolian Fault Zone: a multi-segment rupture framework with physics-based seismo-tectonic assessment using the 2023 Kahramanmaraş earthquake sequence
Core Problem: Seismic hazard along the East Anatolian Fault Zone is underestimated when source models ignore cascading rupture across multiple fault segments.
Key Innovation: The study builds a 162-branch PSHA logic tree with individual, limited multi-segment, and extensive multi-segment rupture models and cross-checks the tectonic framework with SPECFEM3D simulations of the 2023 sequence.
15. The role of slope orientation and source mechanism in topographic amplification: insights from the 2022 Ms 6.8 Luding earthquake
Core Problem: Topographic amplification is often treated generically, despite the fact that slope orientation relative to rupture-controlled wave propagation can strongly alter landslide-triggering motion.
Key Innovation: Three-dimensional spectral-element simulations of the Luding earthquake show that wave-facing slopes, PGV aligned with slope dip, and strong vertical motion jointly focus energy into the most failure-prone sectors.
16. Building-specific seismic damage prediction of Italian historical churches: A gradient boosting approach on a large empirical database
Core Problem: Population-level seismic fragility tools cannot estimate damage for individual historical churches with distinct geometry, maintenance status, and prior deterioration.
Key Innovation: A LightGBM model trained on multi-event Italian survey records predicts both continuous and categorical damage, while SHAP analysis identifies PGA, maintenance state, apse presence, nave height, and pre-existing damage as physically coherent controls.
17. A Plate‐Based Model for Reconciling Geodetic and Geologic Vertical Deformation Across Taiwan
Core Problem: Short-term geodetic deformation and millennial geological uplift in Taiwan have remained difficult to reconcile within a single mechanically consistent framework.
Key Innovation: A three-dimensional elastic plate model jointly fits geological uplift and geodetic data, showing that basal detachment slip drives most long-term uplift while interseismic locking concentrates on upper-crustal faults.
18. Influence of the Polar‐Eurasia Pattern on Interannual Variations in Extratropical Transition of Western North Pacific Tropical Cyclones
Core Problem: Interannual variability in the extratropical transition of western North Pacific tropical cyclones lacks a clear large-scale climate control.
Key Innovation: The paper links lower ET fractions in positive Polar-Eurasia years to weaker baroclinic development at 30-40 N, making recurving cyclones travel farther before reaching a transition-favorable environment.
19. Embeddings based Anomaly Detection for Cleaning Global Crop Type Reference Datasets
Core Problem: Global crop-type reference datasets contain spatially variable label noise that simple rule-based cleaning cannot identify without damaging valid regional samples.
Key Innovation: Locality-aware anomaly detection on geospatial foundation-model embeddings flags mislabeled samples at rates well above chance and improves WorldCereal training when used conservatively for removal or down-weighting.
20. From Machine Learning to Large-Scale EO Products: Best Practices for Making Maps
Core Problem: Large-scale EO maps can inherit hidden errors from early pipeline decisions, yet the field still lacks a concise end-to-end set of product-building best practices.
Key Innovation: The paper organizes operational map making around six coupled themes - infrastructure, preprocessing, dataset design, uncertainty quantification, production, and validation - to expose where product credibility is gained or lost.
21. Hybrid AI-Physical Modeling for Penetration Bias Correction in X-band InSAR DEMs: A Greenland Case Study
Core Problem: X-band InSAR DEMs over snow and ice suffer from penetration bias that physics-only corrections do not fully remove and pure ML models do not robustly generalize.
Key Innovation: The proposed hybrid framework combines parametric physical modeling with machine learning, cutting DEM error statistics over Greenland while generalizing better than purely learned corrections across acquisition scenarios.
22. CPAZMaL: A New Dataset of X-Band SAR Time Series in the French Alps for Machine Learning
Core Problem: Machine-learning development for Alpine SAR time series is constrained by the scarcity of high-resolution, temporally consistent datasets in steep, glacier-rich terrain.
Key Innovation: CPAZMaL releases three years of PAZ X-band dual-polarization time series over Mont-Blanc with glacier and nonglacier classes plus baselines for classification, forecasting, domain adaptation, and transfer learning.
23. A Hybrid Approach Using Statistics and Unsupervised Learning to Investigate the Impact of Ionospheric Storm Disturbances on Pre-Earthquake Total Electron Content Anomalies
Core Problem: Potential TEC earthquake precursors are easily confounded by geomagnetic forcing and by spatially diffuse anomalies that fail physical consistency tests.
Key Innovation: A hybrid framework combining z-tests, k-means clustering, and spatial verification isolates a negative TEC anomaly during storm recovery before the Afghanistan event and finds a similarly validated feature before Tohoku.
24. Hybrid models generalize better to warmer climate conditions than process-based and purely data-driven models
Core Problem: High-performing rainfall-runoff models may fail under future warmer conditions even when they excel on historical or spatial generalization benchmarks.
Key Innovation: Differential split-sample tests show that hybrid models preserve observed change signals under warming better than pure LSTMs, particularly by limiting exaggerated shifts in mean and high flows.
25. Ground Motion Monitoring System of InSAR.Hungary: Results and Validation Findings
Core Problem: Nationwide ground-motion monitoring requires deformation products that are not only dense and automated but also validated and interpretable for practical infrastructure and hazard use.
Key Innovation: The InSAR.Hungary system extends country-scale Sentinel-1 persistent-scatterer monitoring with GNSS-referenced deformation products and an explicit results-and-validation focus aimed at operational ground-motion interpretation.
26. Peak and residual shear strength of coral sand under progressively accelerated ring shear
Core Problem: The rate dependence of peak and residual shear strength in coral sand is poorly constrained, even though it directly affects progressive shoreline failure on reclaimed reef islands.
Key Innovation: Ring-shear experiments across 0.01 to 100 mm/min reveal a nonmonotonic rate effect, persistent dry-sample strengthening, and micromechanical shifts that refine constitutive descriptions of sliding coral sand.
27. Fast Fourier Convolutional GAN for 30 m Clear-Sky Land Surface Temperature Gap-Free Reconstruction
Core Problem: Cloud contamination leaves fine-resolution land-surface temperature products with large missing regions that are difficult to reconstruct when gaps dominate a scene.
Key Innovation: A multimodal Fast Fourier Convolution GAN leverages global receptive fields and SAR-guided auxiliary inputs to recover 30 m clear-sky LST, including scenes with more than 70% cloud gaps.
28. Contrastive Parameter Disentanglement for Multi-modal Remote Sensing Image Generation
Core Problem: Remote-sensing image generation usually treats modalities separately, making it hard to synthesize optical, infrared, and SAR imagery that stays semantically aligned across modalities.
Key Innovation: Contrastive parameter disentanglement separates shared semantics from modality-specific attributes inside LoRA adapters, while a structure-transfer mechanism preserves cross-modal geometric alignment during generation.
29. Post-FWI Injection of Learned Priors Using a Flow Matching Model
Core Problem: Generative prior injection in full-waveform inversion often improves geology at the cost of another expensive inversion cycle.
Key Innovation: The paper applies flow matching as a post-FWI refinement step, using the FWI model and optional well logs as guidance to sharpen structure and even correct depth misties without rerunning inversion.
30. Accuracy potential of visual localization exploiting high-end street-level imagery
Core Problem: Visual localization is often proposed as a GNSS complement, but its true accuracy potential is uncertain because sub-centimeter outdoor ground truth is rarely available.
Key Innovation: A scalable street-level reconstruction and PnP pipeline is paired with the FHNW Muttenz dataset, demonstrating centimeter-level translation and sub-tenth-degree rotation accuracy over a 10 km network.
31. Which Workloads Belong in Orbit? A Workload-First Framework for Orbital Data Centers Using Semantic Abstraction
Core Problem: The rapid growth of data-intensive EO and AI workloads raises the question of which computations should move into orbit rather than remain on the ground.
Key Innovation: The study proposes a workload-first orbital-compute framework and demonstrates that semantic abstraction can reduce EO payloads by more than 99%, making in-orbit processing attractive for selective tasks.
32. Machine Learning for Cloud Detection in IASI Measurements: A Data-Driven SVM Approach with Physical Constraints
Core Problem: Reliable cloud detection from hyperspectral infrared measurements alone remains uncertain across seasons, surfaces, and climate zones.
Key Innovation: CISVM uses IASI radiances with supervised SVM classification and physically informed spectral reductions to reproduce large-scale operational cloud behavior while exposing how performance varies with geography and season.
33. Dynamic and thermal analysis of dust storm processes based on vertical observation data
Core Problem: Dust-storm initiation in the Taklimakan Desert is poorly observed in the vertical, leaving the relative roles of dynamic forcing and thermal preconditioning unresolved.
Key Innovation: Dual-gradient tower data combined with ERA5 and back trajectories show that spring storms retain stronger pressure-gradient control, whereas summer storms exhibit stronger pre-onset thermal buildup and boundary-layer growth.
34. EARLS: a runoff reconstruction dataset for Europe
Core Problem: Europe lacks a harmonized, uncertainty-aware runoff reconstruction dataset at the scale needed for continental hydrologic analysis and ungauged-basin studies.
Key Innovation: EARLS reconstructs daily streamflow from 1953 to 2023 for more than 10,000 basins using a single LSTM trained on over 5,000 basins, with uncertainty included in the released product.
35. A Global Paired Argo Dataset for Ocean Thermal and Salinity Responses to Tropical Cyclones (2000–2024)
Core Problem: Studies of tropical-cyclone ocean response are slowed by the effort needed to pair storm tracks with quality-controlled subsurface observations in a consistent reference frame.
Key Innovation: This paired Argo dataset standardizes more than three million storm-related profile pairs on common depth levels and storm-relative coordinates, reproducing canonical thermal and salinity responses at global scale.
36. Holocene ocean-atmosphere coupling and Mediterranean sensitivity to Atlantic circulation: Lessons from the Late Bronze Age collapse
Core Problem: The Late Bronze Age collapse is often linked to drought, but the mechanism connecting long-term aridification and abrupt hydroclimate extremes remains debated.
Key Innovation: Transient Holocene simulations show that Atlantic internal variability can align across centennial to millennial scales and intensify drought on top of an orbitally driven drying background in the Eastern Mediterranean.
37. Field observations of time-dependent seepage behavior in a full-scale river levee under cold-region conditions
Core Problem: Levee seepage analyses usually assume fixed hydraulic properties even though cold-region embankments evolve through freeze-thaw and post-construction densification.
Key Innovation: Three years of field observations on a full-scale levee show declining water content, increasing dry density, and a two- to three-order reduction in saturated conductivity, which together slow internal water-level response.
38. Evolution of Source Mechanism and Local Stress Inversion During Grouting in a High-Steep Slope Based on Microseismic Monitoring
Core Problem: Rock-mass damage and stress redistribution during curtain grouting in steep slopes are difficult to resolve from conventional monitoring alone.
Key Innovation: Microseismic monitoring coupled with hybrid moment-tensor and stress inversion reveals that grouting mainly triggers shear and mixed-mode slip on pre-existing fractures under a horizontally oriented maximum principal stress.
39. Hybrid and symbolic regression-based modeling of pore water pressure generation in sands under cyclic loading
Core Problem: Predicting excess pore-water pressure under cyclic loading remains difficult when loading conditions and soil descriptors vary across realistic liquefaction scenarios.
Key Innovation: Symbolic regression and a hybrid logarithmic model trained on 125 triaxial tests and validated on new irregular-wave experiments achieve high predictive accuracy while preserving interpretable control by cycle state, CSR, density, and fines.
40. Unraveling valley-channel-bed morphology of the Yarlung Tsangpo Grand Canyon in the Tibetan Plateau: The role of lateral sediment supply
Core Problem: The geomorphic effect of lateral sediment delivery from mass wasting on the lower Yarlung Tsangpo has not been quantified over long canyon reaches.
Key Innovation: Remote sensing and drone-supported field mapping across 270 km show that massive hillslope-derived boulders create step-pool nodes that dissipate flow energy and inhibit both bedrock incision and lateral widening.
41. An uncertainty-aware multi-source remote sensing framework for integrated inland water monitoring: linking water level, surface extent, storage, and water quality
Core Problem: Height, area, storage, and water quality are usually monitored as separate inland-water products, obscuring how uncertainty in one variable propagates into the others.
Key Innovation: The review proposes an integrated H-A-V-Q framework that couples altimetry, water mapping, bathymetry, and quality retrievals, with explicit uncertainty propagation and SWOT-enabled reduction of temporal mismatch.
42. NRNN: A physics-informed neural framework for non-linear rock modelling and deep tunnel analysis
Core Problem: Purely data-driven neural rock models can fit laboratory curves yet still produce physically inadmissible stresses and unstable boundary-value predictions.
Key Innovation: NRNN embeds analytical constitutive relations directly into the network architecture, sharply reducing stress error and reconstructing tunnel-scale full-field mechanical response without spatial error accumulation.
43. Physics-informed machine learning of hyporheic exchange driven by channel blockage and riverbed morphodynamics
Core Problem: Hyporheic exchange under channel blockage is strongly shaped by scour and deposition, but most predictive models do not include evolving riverbed morphology.
Key Innovation: A hierarchical physics-informed ML framework parameterizes blockage-driven morphodynamics with synthetic cases and identifies flow-morphology interactions and effective submergence as the dominant controls on exchange flux.
44. Gravity effects on two-phase displacement in scaled geotechnical centrifuge models: insights from two-dimensional CFD-VOF simulations and microfluidic experiments
Core Problem: Geotechnical centrifuge models of two-phase flow need pore-scale similarity rules to determine when in-situ particle sizes remain valid and when scaling becomes mandatory.
Key Innovation: Theory, CFD-VOF simulations, and microfluidic experiments show that gravity is negligible in tight pores but requires particle-size downscaling in gravity-dominated media according to a 1/n geometric to n^2 gravity relation.
45. Variations in Subducted Plate Crustal Thickness Along the Perú–Chile Margin and Implications for Slab Dip
Core Problem: The thickness of the already subducted Nazca plate crust is poorly known, limiting tests of how crustal buoyancy influences slab dip and flat-slab geometry.
Key Innovation: Receiver-function analysis maps 6-22 km crustal thickness down to roughly 100 km depth and shows that broad thickened crust beneath flat-slab segments may be as important as ridge subduction itself.
46. AI Model Performances Depend on Adequate Training Data Representing Diverse Physical Mechanisms
Core Problem: The relationship between mechanism diversity in training data and the generalization of AI models for physical systems is still weakly understood.
Key Innovation: Experiments on a controllable Burgers system show that missing or weakly represented physical regimes sharply degrade extrapolation, implying that physically consistent synthetic simulations are needed for extremes and climate shifts.
47. PriSAR: 3D Geometric-Prior-Guided Diffusion for Parameter-Controlled SAR Image Generation
Core Problem: Controllable SAR image generation remains difficult when only sparse azimuth views are available and geometry-aware conditioning is weak.
Key Innovation: PriSAR injects lightweight multi-bounce ray-tracing priors from 3D models into a diffusion framework, improving both structural similarity and azimuth consistency for intermediate-view completion.
48. Direction-adaptive Mamba: Spatial-Frequency Dual-Domain Collaborative Learning for PolSAR Image Classification
Core Problem: Standard Mamba backbones for PolSAR classification neglect directional anisotropy and fine boundary information that are physically important in scattering analysis.
Key Innovation: Direction-adaptive scanning, contourlet-based frequency decomposition, and dual-domain collaborative learning jointly preserve long-range context and anisotropic structural cues, improving classification on three PolSAR datasets.
49. Covariance-Boosted Gaussian Processes for Spatiotemporal Irregularities
Core Problem: Highly parameterized nonstationary Gaussian processes tend to overfit sparse irregular data and to produce overconfident uncertainty estimates in integrity-critical settings.
Key Innovation: Covariance-Boosted Gaussian Processes iteratively update weak covariance priors using partially whitened observations, then restrict priors and inflate posterior uncertainty to maintain robust three-nines integrity.
50. Autoregressive One-Step Generative Modeling for Dynamical System Forecasting
Core Problem: Autoregressive surrogates for physical dynamics must roll out quickly without losing the long-horizon statistical structure of turbulence and other chaotic systems.
Key Innovation: MeLISA uses a one-step MeanFlow transition kernel together with window-consistency and time-increment consistency losses to preserve spectra, kinetic energy, and mixing behavior while remaining as fast as neural operators.
51. FUSE-Flow: A Decoupled Framework for Calibration and Stateless Real-Time Multi-View Point Cloud Fusion
Core Problem: Real-time multi-view 3D reconstruction suffers when camera calibration, fusion, and optimization are tightly entangled and errors accumulate over time.
Key Innovation: FUSE-Flow decouples geometry-aligned extrinsic calibration from confidence-guided stateless point-cloud fusion, improving dynamic stability and scalability without relying on global bundle adjustment.
52. SwinEADFormer: An Edge-Aware Dynamic Swin Transformer for Building Change Detection in High-Resolution Remote Sensing Images
Core Problem: Uniform cross-temporal interaction in building change detection lets unchanged areas generate spurious responses when true changes are sparse.
Key Innovation: SwinEADFormer conditions temporal interaction on estimated change evidence through a difference-aware router and discrepancy fusion, improving boundary-sensitive building change mapping.
53. SpecGateNet: Spectral-Guided Fusion Network for Cloud and Cloud Shadow Segmentation in Optical Remote Sensing Imagery
Core Problem: Cloud and cloud-shadow segmentation still struggles to balance local boundaries and global context when decoder fusion treats all spatial locations identically.
Key Innovation: SpecGateNet derives spatial fusion gates from Fourier-amplitude energy, combining spectral guidance with a CNN-Swin dual-branch encoder to sharpen thin-cloud and shadow boundaries using RGB-only inputs.
54. Spatially Robust Land Cover Classification with Multi-Seasonal Sentinel-2 Imagery: A Comparison of CNN, UNet++, ConvNeXt and ViT
Core Problem: Land-cover benchmarks can overstate model quality when spatial autocorrelation leaks information between training and testing regions.
Key Innovation: Using strict four-fold spatial cross-validation on multi-seasonal Sentinel-2 composites, the study shows that convolutional models transfer more reliably than a scratch-trained ViT and isolates winter Blue and summer SWIR channels as the most influential features.
55. Analysis and evaluation of hydraulic conductivity using various field data: an example from Southwest China
Core Problem: Hydraulic conductivity in fractured rock masses depends on interacting depth, rock quality, weathering, and unloading effects that simple empirical formulas cannot resolve.
Key Innovation: A MARS-based model trained on packer tests maps conductivity from depth, RQD, and a weathering-unloading index, while preserving interpretable parameter effects for seepage and water-curtain design.
56. Typical surface water–groundwater interaction mechanisms in the three gorges reservoir area: coupled evidence from hydrochemistry and hydrogen–oxygen isotopes
Core Problem: The way reservoir-stage regulation reorganizes shallow surface-water and groundwater exchange in the Three Gorges area has not been quantified with coupled tracers.
Key Innovation: Hydrochemistry, hydrogen-oxygen isotopes, and a two-end-member mixing model show strong riparian infiltration from surface water but groundwater discharge from both banks, driven by regulation-induced reversal of hydraulic gradients.
57. New models for arias intensity, cumulative absolute velocity, and significant durations for the Alborz region of northern Iran
Core Problem: Northern Iran lacks region-specific models for secondary ground-motion measures such as Arias intensity, cumulative absolute velocity, and significant duration.
Key Innovation: Mixed-effects residual corrections adapt existing reference GMMs to Alborz using magnitude, Joyner-Boore distance, and Vs30, producing locally calibrated models with unbiased residual behavior.
58. Quantifying methane emissions from large coal mine using Sentinel-5 Precursor satellite data and Gaussian plume model
Core Problem: Bottom-up coal-mine methane inventories in China remain highly uncertain because source-specific emission rates are poorly constrained.
Key Innovation: TROPOMI methane enhancements are inverted with a Gaussian plume model that accounts for mixed plumes from multiple sources, yielding mine-scale emission estimates that diverge from major global inventories.
59. Pioneering Test-Time training for robust Fine-Grained chlorophyll-a retrieval from Sentinel-2: Unveiling China’s lake eutrophication dynamics (2018–2025)
Core Problem: Large-scale chlorophyll-a retrieval degrades under atmospheric-correction inconsistencies and shifting optical properties across lakes and seasons.
Key Innovation: A dual-task test-time training architecture adapts during deployment through spectral reconstruction, improving fine-grained chlorophyll retrieval and enabling a national analysis of 2,621 Chinese lakes from 2018 to 2025.
60. From pixels to Algae: Assessing cyanobacterial bloom severity in Northeastern lakes through Multi-Class phytoplankton density retrieval
Core Problem: Bloom monitoring that ignores algal phyla cannot resolve which taxa are driving water-quality deterioration or how severity evolves across lakes.
Key Innovation: The study retrieves densities of major algal phyla from satellite data, establishes a cyanobacterial severity classification, and shows that Cyanophyta dominates bloom intensification in Northeast China.
61. Modeling preferential flow and deep drainage in a long-term furrow-irrigated cotton–wheat rotation on a shrink–swell clay soil in semi-arid Australia
Core Problem: Single-porosity flow models often miss preferential deep drainage in shrink-swell agricultural soils, obscuring when and where high-loss flow paths activate.
Key Innovation: HYDRUS-2D simulations demonstrate that a dual-porosity formulation reproduces event drainage and captures seasonal downward shifts in preferential flow from upper to deeper soil layers.
62. Shrub encroachment suppresses soil preferential flow in desert steppes and increases the risk of soil aridity
Core Problem: How shrub encroachment alters preferential flow and aridity risk in desert steppes is poorly constrained at the pore and flow-path scale.
Key Innovation: Computed tomography and dye tracing show that shrub encroachment suppresses herbaceous roots, enriches mineral particles, weakens macropore-driven preferential flow, and thereby intensifies soil aridity risk.
63. Identification of environmental drivers controlling pyrogenic carbon accumulation and mobility in the Da Xing'an Mountains, China
Core Problem: Pyrogenic carbon in burned soils is treated as persistent, but the environmental controls on its production, accumulation, and retention remain poorly resolved.
Key Innovation: Field sampling after the Black Dragon Fire combined with piecewise structural equation modeling shows that pre-fire fuel availability dominates charcoal production while gentle slopes favor in situ preservation.
64. Applications of artificial intelligence models in seawater intrusion: progress, challenges, and future directions
Core Problem: Applications of AI to seawater intrusion are expanding, but the field remains fragmented and hampered by weak physical consistency and limited interpretability.
Key Innovation: The review organizes AI advances across prediction, monitoring, vulnerability mapping, and management, and proposes a grey-box roadmap that embeds hydrogeologic constraints and adaptive multi-source learning.
65. An uncertainty-guided and knowledge-informed inversion framework for spatial-temporal numerical modeling of rockfill dams
Core Problem: Inverse calibration for rockfill dams often optimizes fit alone, ignoring uncertainty sources and whether the recovered material parameters remain physically plausible.
Key Innovation: An uncertainty-guided, knowledge-informed inversion framework combines surrogate ensembles, empirical parameter constraints, a modified TPE optimizer, and weighted TOPSIS to produce rational calibrations.
66. Identification of rock failure modes based on waveform characteristics of acoustic emission
Core Problem: Acoustic-emission monitoring needs waveform-level criteria that can distinguish tensile, mixed-mode, and shear rock failure before macroscopic instability is obvious.
Key Innovation: Moment-tensor-labeled AE signals are analyzed in time and frequency domains to map failure-mode signatures and to show a common progression from tensile cracking toward shear-dominated failure.