TerraMosaic Daily Digest: July 20, 2026
Daily Summary
Landslide research spans image segmentation, changing susceptibility and failure mechanics. DSFF-Net and TM-Net both target fragmented, elongated failures in complex mountain imagery, but their cross-benchmark results also expose the cost of geographic heterogeneity. In the Belluno Alps, multi-temporal inventories and cumulative rainfall improve a spatiotemporal susceptibility model; at Po Shan Road, regional seismic waves are transferred into a large-deformation runout simulation. Field geophysics at Durung Drung glacier further links discontinuous frozen moraine, accelerating retreat and lake expansion to weakening high-mountain flood barriers.
Warning and surveillance systems make spatial structure operational. TREAD asks a regional earthquake-warning Transformer to learn station-to-target distance while predicting ground motion. At Kusatsu-Shirane, a remotely operated droneport sustained near-daily crater-lake observations over 184 days and delivered live imagery to the Japan Meteorological Agency; snow-related outages also define its present operating limit. The distinction between exposure and vulnerability is equally consequential at global scale: GDP exposure dominates tropical-cyclone economic loss, whereas population vulnerability is the stronger determinant of mortality.
Long-horizon modelling changes the scale of the risk question. National flood simulations indicate at least 0.5 million additional UK residents exposed to an undefended 1-in-200-year coastal flood by 2100 and as many as 13 million under the highest 2300 sea-level storyline. At the observational end, the Garrotxa inventory reconciles 1,049 local hazard events, the Glacial Lake Observatory provides annual lake outlines from 2017-2024, and a six-class satellite benchmark supports single-image multi-hazard detection. Controlled Earth-observation experiments meanwhile show that patch size and decoder design can explain more performance variance than foundation-model identity.
Key Trends
Five shifts connect terrain-aware mapping, unattended monitoring, spatially explicit warning, loss attribution and longer hazard records.
- Landslide models are tested against terrain heterogeneity: Cross-benchmark segmentation, optimized slope units and multi-temporal inventories reveal where image architecture, terrain partitioning and changing rainfall evidence each matter.
- Monitoring moves toward unattended operation: A semi-autonomous volcano droneport and beam-aware deformation radar reduce the need for continuous on-site observation while making weather and platform limits measurable.
- Geometry is built into warning and process models: Station-to-target distance guides earthquake alerts, regional seismic tractions drive local runout, and moraine structure is resolved jointly with glacier motion and retreat.
- Exposure and vulnerability separate across loss types: Long-horizon coastal simulations quantify population exposure, while global cyclone attribution distinguishes the economic role of GDP exposure from the mortality role of population vulnerability.
- Hazard records become longer and more reproducible: Century-scale multi-hazard inventories, annual glacial-lake outlines and open satellite benchmarks support temporal comparison beyond isolated events or one-off maps.
Selected Papers
The leading papers connect direct landslide observation with unattended hazard surveillance, distance-aware warning and long-horizon exposure modelling.
1. Quantifying UK coastal flood exposure under future sea-level rise to 2300
Core Problem: Coastal flood assessments rarely extend far enough to test how multi-century sea-level commitment changes the scale of population exposure.
Key Innovation: National 25 m flood modelling is coupled to physically consistent UK sea-level storylines through 2300, showing at least 0.5 million additional people exposed by 2100 and up to 13 million under the highest 2300 storyline.
2. Remote surveillance of volcanic unrest at Kusatsu-Shirane volcano, Japan, using a semi-autonomous drone docking station
Core Problem: Volcano drone surveys are constrained by short flight range and the need to place personnel near an unrest zone for each launch.
Key Innovation: A remotely operated docking station sustained near-daily Kusatsu-Shirane flights over 184 days, quantified crater-lake discoloration related to magmatic-fluid supply and shared live video with the national monitoring agency.
3. TREAD: A Distance-Aware Transformer for Regional Earthquake Early Warning
Core Problem: Regional earthquake-warning models must infer how shaking propagates from an evolving station network to target locations within seconds.
Key Innovation: TREAD adds station-to-target distance regression as a physically interpretable auxiliary task, embedding spatial geometry in a Transformer and improving warning performance across CWA and KiK-net thresholds.
4. Exploring the impact mechanisms of tropical cyclone losses: a global analysis integrating hazard, exposure, and vulnerability (1990-2019)
Core Problem: Tropical-cyclone mortality and economic loss are often pooled even though hazard, exposure and vulnerability can control the two outcomes differently.
Key Innovation: A 1990-2019 global attribution separates those controls: GDP exposure dominates direct economic loss, population vulnerability dominates mortality, and the burden shifts sharply across national income groups.
5. Dual-stream feature fusion network for landslide detection
Core Problem: Landslide segmentation must retain small boundaries and high-level context without making large-area inventory mapping prohibitively expensive.
Key Innovation: DSFF-Net combines dual feature streams, hierarchical fusion and enhanced spatial attention, reaching 97.08% F1 on Bijie and 70.20% on Landslide4Sense with 24.51 million parameters.
6. Mapping Landslide-Affected Land Surfaces in Complex Mountainous Landscapes Using a Twin-Path Multi-Scale Deep Learning Network
Core Problem: Bare soil, terraces, roads and shadows produce false landslide detections, while fragmented and elongated failures are easily lost during decoding.
Key Innovation: TM-Net concentrates refinement in a twin-path multi-scale decoder, leading the reported fused benchmark and XBLD comparisons while exposing a large performance gap between the two test domains.
7. Debris variability and velocity estimation using Electrical Resistivity Tomography and DInSAR techniques of Durung Drung glacier, northwestern Himalaya
Core Problem: Surface retreat alone cannot show whether a moraine-dammed lake is expanding against ice-rich support or a weakening, hydrologically connected debris barrier.
Key Innovation: Electrical resistivity tomography, DInSAR and DGPS jointly resolve moraine structure, 120-138 m per year glacier motion, accelerating snout retreat and lake growth at Durung Drung.
8. Analysis of Causes and Prevention Measures for a Landslide Geological Disaster in a Mining Area of Longsheng County, Guangxi
Core Problem: Mining-area slope failure in Longsheng reflects interacting geological, rainfall and engineering controls that cannot be addressed by one stabilization measure.
Key Innovation: The case study links terrain, lithology and structure with infiltration and disturbance, then organizes slope cutting, drainage, retaining works, ecological restoration and monitoring as one prevention system.
9. Prototyping the Seismogenic Fault Injection Test (SFIT)
Core Problem: Injection-induced earthquake forecasts cannot be stress-tested efficiently before a full subsurface operation begins.
Key Innovation: The proposed Seismogenic Fault Injection Test optimizes a two-stage injection protocol to maximize disagreement among statistical, physics-based and machine-learning forecasts, turning a short controlled test into a diagnostic of model assumptions.
10. Natural hazard events dataset for the Garrotxa Region (Catalonia, Spain): a foundational step toward multi-hazard risk assessment
Core Problem: Global disaster repositories underrepresent local, recurrent events and provide little evidence for interactions among hazards.
Key Innovation: The Garrotxa dataset reconciles official inventories, historical press, scientific literature, oral records and citizen science into 1,049 cross-validated earthquake, landslide, rockfall, flood, wildfire and subsidence events spanning 1900-2023.
11. Glacial Lake Observatory (GLO): annual dataset of glacial lakes in Nepal and transboundary catchments (2017-2024)
Core Problem: Glacial-lake inventories are rarely updated consistently enough to identify anomalous expansion and emerging outburst-flood exposure.
Key Innovation: The Glacial Lake Observatory provides annual, deep-learning-derived lake outlines for Nepal and transboundary catchments from 2017-2024, establishing a repeatable basis for lake-change and lake-glacier monitoring.
12. A Framework Integrating Slope-Unit Parameter Optimization and Ensemble Machine Learning for Landslide Susceptibility Mapping
Core Problem: Landslide susceptibility maps are sensitive to the scale and geometry of slope units, yet these choices are often fixed before model evaluation.
Key Innovation: The study jointly optimizes slope-unit parameters and ensemble-learning performance across multiple DEM resolutions, tying terrain partition quality to predictive skill instead of selecting the mapping unit heuristically.
13. Enhancing landslide susceptibility assessment through ensemble learning and PCoA
Core Problem: High-dimensional conditioning factors can obscure unstable or redundant structure in ensemble landslide-susceptibility models.
Key Innovation: Principal coordinate analysis is used with multiple machine-learning learners and ensemble strategies to reduce factor redundancy and test whether the transformed representation improves susceptibility discrimination.
14. Permafrost slope failure and mitigation techniques in cold regions
Core Problem: Warming permafrost destabilizes slopes through coupled thermal, hydraulic and mechanical change, but mitigation guidance remains dispersed across disciplines.
Key Innovation: The review links thaw processes and failure modes to monitoring and stabilization options, organizing cold-region slope mitigation around the mechanisms that control ice-rich ground degradation.
15. From static to dynamic landslide susceptibility: a multi-temporal inventory-based approach in the Belluno Alps (NE Italy)
Core Problem: Static susceptibility models cannot represent how cumulative rainfall and successive failures alter the probability of instability through time.
Key Innovation: A spatiotemporal generalized additive model combines multi-temporal inventories, terrain variables and rainfall windows in the Belluno Alps and outperforms a conventional cross-sectional formulation.
16. Regional contrasts in environmental conditions associated with retrogressive thaw slumps across the Arctic and the Qinghai-Tibetan Plateau
Core Problem: Retrogressive thaw slumps are often modelled with common controls even though Arctic and high-altitude permafrost occupy different climatic and soil regimes.
Key Innovation: Multivariate comparison and maximum-entropy models show that thawing degree days dominate Arctic occurrence, freezing conditions matter more on the Qinghai-Tibetan Plateau, and both regions require locally calibrated controls.
17. Fractal-wavelet diagnostics of liquefaction-induced transport corridor deformation after the 6 February 2023 Kahramanmaraş earthquakes: Evidence from the Hatay airport access road
Core Problem: Liquefaction damage along transport corridors is spatially irregular, making conventional profile statistics insufficient for locating the most consequential deformation.
Key Innovation: GNSS profiles are analysed with detrended fluctuation and continuous wavelet methods, then checked with Google Earth Engine imagery to localize deformation hotspots along the Hatay airport access road after the 2023 earthquakes.
18. Towards efficient multi-hazard detection from single satellite imagery via spectral-semantic graph reasoning
Core Problem: Most satellite disaster detectors target one hazard and require before-after imagery, limiting rapid multi-class recognition from an on-orbit observation.
Key Innovation: SSF-GNN combines predefined spectral indices with learned semantic features and graph reasoning; together with the new six-class S2MHD dataset, it reaches 88.37% overall accuracy from a single multispectral image.
19. A traction-based SEM-MPM framework for multiscale large deformation analyses of coseismic landslides with case study in Hong Kong
Core Problem: Coseismic-landslide simulation must resolve regional wave propagation and local runout without making the full domain computationally intractable.
Key Innovation: A traction interface transfers spectral-element seismic forces into a material-point large-deformation model and reproduces the Po Shan Road setting with lower computational demand than one uniformly resolved simulation.
20. Now We Know? A Systematic Comparison of TerraMind and THOR
Core Problem: Aggregate geospatial-foundation-model rankings do not reveal whether performance differences arise from pretraining, patch scale, decoder capacity or fine-tuning choices.
Key Innovation: A controlled ten-task comparison of TerraMind and THOR finds that patch size and decoder design explain more variance than model identity and replaces a single leaderboard with task- and architecture-specific evidence.
21. AGENTS4GEOS: agentic platform for open-source multi-physics simulation
Core Problem: Training surrogates for subsurface multiphysics requires physically consistent simulations whose setup, execution and post-processing remain labor intensive.
Key Innovation: Agents4GEOS connects 52 domain-aware tools to the open GEOS solver through the Model Context Protocol, allowing an orchestrated agent to create inputs, inspect meshes, compute properties, run analyses and diagnose outputs against executed simulations.
22. Delineate Anything v2: A Global Foundation Model for Field Delineation
Core Problem: General vision foundation models often merge adjacent fields or miss weak boundaries because cropland topology, texture and physical scale differ from ordinary images.
Key Innovation: Delineate Anything v2 combines the 73-million-instance FBIS-73M corpus with resolution-specific parcel curation and a 100-country benchmark, doubling mAP at 0.5 over its predecessor while mapping Ukraine in 5.4 hours on a consumer workstation.
23. Vibration-Induced Phase Error Compensation in Ground-Based Phased Array Radar for Deformation Monitoring
Core Problem: Platform vibration produces beam-dependent interferometric phase errors that limit rapid ground-based radar deformation monitoring.
Key Innovation: A grouped phase-error model with adjacent-beam weighting compensates the error according to beam position, targeting the measurement stability needed for geological-disaster monitoring.
24. Enhancing Detection of Inundated Areas Using Novel Hybrid PolSAR-Metaheuristic-Deep Learning Models
Core Problem: Cloud cover and polarimetric misclassification reduce the reliability of rapid inundation maps during floods.
Key Innovation: Metaheuristic feature selection is coupled to deep semantic segmentation over polarimetric SAR channels, decompositions and textures to identify a compact feature set with improved flood delineation.
25. Severe convective weather in Italy: Current understanding and Research Priorities in the TIM campaign
Core Problem: Italian severe-storm observations are concentrated in the north, leaving central and southern storm environments and mountain-to-plain intensification less resolved.
Key Innovation: The TIM campaign coordinates pan-European observations and two Italian airborne platforms to measure storm structure, aerosols, complex-orography effects and sea-land interactions for improved warnings and climate-impact estimates.
26. RTA-FireNet: a lightweight RGB-thermal asymmetric detector for early forest fire detection from UAV imagery
Core Problem: Symmetric RGB-thermal fire detectors are often too computationally expensive for real-time inference on small UAV platforms.
Key Innovation: RTA-FireNet uses a visible-light primary branch with selective thermal assistance in a 2.2-million-parameter YOLO11 design, reaching 66.0% average precision and 94.0% F1 on RGBT-3M at 9.9 GFLOPs.
27. Evaluating Waterlogging Risk Inequality in a Megacity
Core Problem: Urban flood-risk maps rarely show whether a small share of residents bears a disproportionate share of waterlogging risk within the same city.
Key Innovation: The Beijing framework combines machine-learned susceptibility, entropy-weighted resilience and population exposure with Gini indices and Lorenz curves, revealing high absolute risk in the urban core but much greater inequality in outer suburbs and mountain districts.
28. A UAV–TLS Point-Cloud Fusion Framework for Three-Dimensional Characterization of Mining-Induced Surface Cracks
Core Problem: UAV crack maps usually describe mining-induced ground damage in two dimensions and cannot recover crack depth or three-dimensional geometry.
Key Innovation: A residual-attention U-Net extracts crack centerlines from UAV orthomosaics and spatially links them to terrestrial-laser-scanning points, producing measurable semantic 3D crack clouds in real topography.
29. A novel forest fire scenario deduction method based on belief rule base
Core Problem: Forest-fire response depends on anticipating how a scenario will evolve, but expert rules and interacting scenario attributes make that deduction uncertain.
Key Innovation: A belief-rule-base model uses rank correlation to select attributes and information gain with differential evolution to optimize its rules; it is tested on the 2022 Fairview fire and five-fold cross-validation of 347 additional records.
30. The contribution of fine-grained exposure data to flood risk assessment along the Niger River
Core Problem: Rural flood assessments in low- and middle-income regions often lack asset-level exposure and damage data needed to target practical mitigation.
Key Innovation: Along 58 km of the Niger River, hydraulic scenarios validated with Sentinel-2 are combined with mapped buildings, crops and wells from open, local and field data, quantifying where losses concentrate and supporting nine site-specific structural measures.
31. Spatiotemporal variability of precipitation and meteorological drought in Pakistan: trends, patterns, and machine learning prediction
Core Problem: National averages obscure strong regional shifts in Pakistan's precipitation and meteorological drought, limiting identification of the variables that precede drought conditions.
Key Innovation: Kriged precipitation statistics, standardized precipitation indices and machine-learning models map changing drought frequency; random forests identify lagged three-month SPI, precipitation and surface pressure as the leading predictors.
32. Rule-based damage assessment of the 2023 Kahramanmaraş earthquakes: a decision tree and association analysis approach
Core Problem: Sparse post-earthquake observations require damage models that remain interpretable and can separate the influence of site conditions from conventional shaking metrics.
Key Innovation: Station-neighbourhood damage labels are analysed with nested leave-one-out decision trees and association rules, identifying engineering-bedrock depth, Vs30 and predominant frequency as stronger discriminators than the selected ground-motion variables.
33. Spatiotemporal patterns and predictive modeling of precipitation-induced railway disasters in China
Core Problem: Railway rainfall risk is difficult to generalize when flood, geological and subgrade failures are recorded in separate and incomplete sources.
Key Innovation: A long-term Chinese inventory shows that 83.8% of incidents occur from June to August, identifies flooding and geological failures as the dominant classes, and derives precipitation-based recurrence thresholds for representative rail sections.
34. Linking Coupled Reservoir Modeling to Dynamic Rupture Modeling: An Automated Workflow to Simulate Induced Seismicity in Faulted Geothermal Systems
Core Problem: Reservoir and rupture simulators are commonly coupled in one direction, so stress redistribution from an induced earthquake does not alter the subsequent injection model.
Key Innovation: The workflow closes that loop by returning dynamically computed slip and stress changes to a three-dimensional thermo-hydro-mechanical reservoir simulation after each triggered event.
35. Rotational Ground Motion Model and Design Spectrum for the Himalayan Region
Core Problem: Seismic codes omit rotational shaking largely because direct observations are scarce, leaving rocking and torsional demand unrepresented in Himalayan design spectra.
Key Innovation: Rotational motions simulated from NGA-West2 records are validated against Kefalonia observations, learned with a neural ground-motion model, transferred to Himalayan conditions and propagated through probabilistic hazard analysis to a regional design spectrum.
36. Post-glacial landforms dictate fluvial sediment dynamics during extreme floods: Morphological adjustments and flood hazard implications
Core Problem: Post-glacial terraces and fans are often treated as sediment buffers even though extreme floods can remobilize their unconsolidated deposits.
Key Innovation: Pre- and post-flood LiDAR from the Nicola River shows that adjacent post-glacial landforms supplied much of the 2021 event's sediment, demonstrating that the same landform can buffer frequent flows and feed rare extremes.
37. Strategic placement of emergency check dams to reduce connectivity and mitigate post-fire soil erosion in South-Central Chile
Core Problem: Emergency check dams are frequently installed after wildfire without testing whether their locations intersect the most connected, erosive parts of a catchment.
Key Innovation: Connectivity, RUSLE and maintenance mapping show that many structures occupied low-erosion areas and that 47-67% remained functional after two years, supporting placement based on erosion-connectivity maps rather than convenience.
38. Integrating hydrodynamic flood scenarios and shophouse-specific depth-damage functions for risk assessment in Ho Chi Minh City, Vietnam
Core Problem: Citywide flood-loss models can misallocate risk when they ignore both the generating flood process and the damage response of locally common building forms.
Key Innovation: Six return periods for fluvial, pluvial and coastal flooding are paired with depth-damage functions for irregular and peri-urban shophouses, showing that irregular shophouses dominate expected annual damage across all three flood types.
39. A hydraulic-geometry informed discharge retrieval algorithm for ungauged basins using ICESat-2 and Sentinel-1/2: Application to Qinghai-Tibet Plateau
Core Problem: River discharge remains poorly observed in ungauged mountain basins where width-only satellite rating curves are weak.
Key Innovation: ICESat-2 cross sections and water-surface slope are fused with Sentinel-derived width in a hydraulic-geometry formulation, reaching mean R-squared of 0.87 and Nash-Sutcliffe efficiency of 0.84 across twelve Qinghai-Tibet Plateau reaches.
40. Wetting-induced collapse of loess: Multistage microstructural evolution of particles and pores
Core Problem: Loess-collapse studies often compare only the states before and after wetting, leaving the sequence of particle and pore reorganization unresolved.
Key Innovation: Oedometer testing, microscopy and U-Net++ segmentation divide collapse into four stages and show coordinated aggregate breakup, particle realignment and macropore conversion as water content and loading change.
41. The Fate of Snowmelt: Global Partitioning Into Runoff and Evaporation
Core Problem: Global estimates do not resolve how much snowmelt becomes streamflow and how much returns to the atmosphere across contrasting climates and elevations.
Key Innovation: FLEX-Global, an empirical partitioning equation and inverse estimates from three other hydrological models converge on 53-71% of snowmelt becoming runoff and 29-47% becoming evaporation, with sharply different arid, humid and mountain regimes.
42. Physics-Informed Super-Resolution of Atmospheric Data
Core Problem: Machine-learning downscaling can sharpen atmospheric fields while violating the hydrostatic, multivariable relationships required for credible extreme-event analysis.
Key Innovation: PISR adds multi-scale losses derived from hydrostatic primitive equations and introduces a Normalized Physical Consistency metric; tests on ERA5, CERRA and COSMO improve physical consistency, reconstruction and heatwave or extreme-wind detection.
43. Learning Semantic-Robust Change Detection via Semantic-Invariant Self-Distillation
Core Problem: Illumination, shadow and atmospheric differences can resemble land-cover change, producing false alarms when remote-sensing models learn appearance rather than semantics.
Key Innovation: SCDistill combines semantic-invariant self-distillation with diffusion-generated appearance perturbations, improving disturbance resistance across semantic and binary change detection as well as change captioning.
44. Incomplete Observations Boost Evolutionary Performance in Ocean Modeling
Core Problem: Data-driven ocean models usually require complete reanalyses, preventing them from learning directly from the sparse and noisy observations available in practice.
Key Innovation: A continuous-state hidden Markov model couples neural initial and transition operators to a masked-Gaussian observation model, using expectation-maximization and Langevin reconstruction to learn ocean evolution from incomplete measurements.
45. Disentangling Forced and Internal Climate Variability in Single Realizations using Dynamic Mode Decomposition with Control
Core Problem: A single observed climate record mixes externally forced change with internal variability, while common single-realization methods omit either forcing predictors or system dynamics.
Key Innovation: PullbackDMDc treats forcing as a dynamical control and estimates pullback-attractor modes, matching or exceeding established forced-response baselines while retaining interpretable internal variability for model evaluation.
46. NGPS: GPS-Denied Aerial Geo-Localization and 2.5D Reconstruction via Deep Satellite Image Matching and Multi-Rate Sensor Fusion
Core Problem: High-altitude UAVs lose absolute position when GPS is unavailable, whereas visual-inertial odometry drifts and onboard sensors report at incompatible rates.
Key Innovation: NGPS combines deep matching to georeferenced satellite imagery, confidence-weighted UKF fusion, velocity-guided search and time-ordered multi-rate updates, obtaining 2.94 m position RMSE and real-time georeferenced 2.5D reconstruction.
47. A Geometry-Aware AI Emulator for the Coupled Whole Atmosphere from Earth Surface to the Ionosphere and Thermosphere
Core Problem: Whole-atmosphere simulations resolve surface-to-ionosphere coupling at a computational cost that limits ensembles, sensitivity tests and uncertainty analysis.
Key Innovation: CAM-NET combines a spherical Fourier neural operator with a lightweight plasma-variable adapter, reproducing dominant ionosphere-thermosphere morphology in stable multi-day rollouts while documenting loss of fine mesospheric structure.
48. Mechanical response of surrounding rock supported by NPR long and short anchor cables under seismic action
Core Problem: Conventional anchor systems adapt poorly to the large deformation of deep soft-rock tunnels subjected to seismic loading.
Key Innovation: A coordinated negative-Poisson-ratio long- and short-cable system is represented with strain-softening elastoplastic and improved quasi-static solutions; numerical validation shows more uniform cable force and over 20% lower peak displacement.
49. Two-Stage Training Single-Source Domain Generalization Network for Cross-Scene Hyperspectral Image Classification
Core Problem: Single-stage domain generalization can move class centers while aligning scenes, leaving unstable prototypes and blurred boundaries in hyperspectral classification.
Key Innovation: TSTNet first learns prototype-anchored source features, then freezes the prototype classifier while aligning feature and output distributions, improving transfer to unseen hyperspectral scenes.
50. IDACP: Information-Driven Adaptive Cooperative Planning for Ground-Sensor-Conditioned UAV Remote Sensing
Core Problem: UAV survey planners commonly ignore uncertainty already reduced by ground sensors, wasting flight time over well-constrained locations.
Key Innovation: IDACP fuses heterogeneous ground and aerial observations in one Gaussian process, then uses uncertainty-guided waypoint selection, adaptive stopping and Voronoi fleet coordination to reduce error by 12-32% while using 35-55% less flight energy.
51. FSSTNet: Frequency Enhanced Spectral-Spatial Transformer for Methane Plumes Segmentation
Core Problem: Weak methane plumes are difficult to separate from heterogeneous surfaces and atmospheric noise when spatial and spectral resolution trade against one another.
Key Innovation: FSSTNet fuses a SWIR transformer stream with a methane-product convolutional stream, then separates amplitude and phase in a frequency-aware decoder to improve weak-plume segmentation across AVIRIS-NG, WorldView-3 and EMIT.
52. Toward Meter-Scale Soil Moisture Mapping: UAV Microwave Radiometry at L-Band With Machine Learning Models and Physical Models
Core Problem: Satellite L-band soil-moisture products are too coarse for field decisions, but transferring their physical retrievals directly to UAV scale is unreliable in complex terrain.
Key Innovation: A UAV-mounted L-band radiometer is evaluated with three physical and three machine-learning retrievals; localized UL-SCA parameters improve physical transfer, while random forests reach about 0.02 m3/m3 RMSE across most land-cover classes.
53. LGRA-Net: Progressive Local-Global Representation Alignment Network for ORSI Salient Object Detection
Core Problem: Clutter, weak boundaries and scale variation expose a mismatch between frozen SAM2 representations and salient-object detection in optical remote-sensing imagery.
Key Innovation: LGRA-Net progressively aligns shallow structure and high-level semantics inside a unified local-global refinement path, reaching a mean absolute error of 0.0226 on the large ORSI-4199 benchmark.
54. Dynamic Capture and Evaluation of Coastal Dune Evolution by Integrating InSAR and ORS
Core Problem: Slow coastal-dune change is difficult to quantify when shoreline movement and localized surface deformation are observed by separate sensors and over different time scales.
Key Innovation: Optical imagery tracks the 2008-2022 coastline while InSAR resolves 2021-2023 dune motion, separating a slight mean erosion trend from localized uplift and subsidence near the Qilihai River and southern coast.
55. Multimodal Fusion ConvLSTM Model for Long-Term Snow Cover Image Prediction
Core Problem: Long-range snow-cover forecasts must reconcile optical and SAR observations while preserving the interacting spatial and temporal controls on snow distribution.
Key Innovation: Freq-CDDFuse integrates the two sensor modalities before a residual ConvLSTM models temporal evolution, reducing baseline error metrics by 11-28% in the reported experiments.
56. Comparative Multi-Indicator Satellite Analysis of Vegetation Recovery in Postfire Sandy Coastal Areas (Case Study of the Curonian Spit and Kinburn Peninsula)
Core Problem: Single vegetation indices are unreliable over bright, sparsely vegetated coastal sands, and comparable long-term evidence from different climate regimes is scarce.
Key Innovation: A five-index Landsat record from 1984-2024 separates recovery trajectories at two Eastern European coasts: normalized recovery reaches 1.46 at managed, humid Smiltyne but remains 0.38 on the drier, repeatedly disturbed Kinburn Peninsula.
57. Deep Learning-Based 4D Reconstruction of Arctic Ocean Hydrography and derived Geostrophic Currents from Satellite and In Situ Observations
Core Problem: Sparse profiles leave Arctic subsurface temperature, salinity and circulation poorly resolved between observing campaigns.
Key Innovation: An observation-constrained LSTM combines satellite surface variables with in situ profiles to reconstruct three-day, four-dimensional hydrography for 2011-2021 on a pan-Arctic grid and higher-resolution gateway grids, from which geostrophic currents are derived.
58. A combination of time-variable gravity field solutions from multi-satellite datasets (1993-2024) via constrained collocation model
Core Problem: GRACE and GRACE-FO gravity records contain mission gaps and begin too late for many multi-decadal mass-change analyses.
Key Innovation: A Tikhonov-regularized constrained collocation model combines and denoises several satellite solutions into gapless monthly gravity fields to degree and order 60 for 1993-2024.
59. Continental-scale fern savannah wildfires during end-Triassic greenhouse warming
Core Problem: The widespread darkening of end-Triassic pollen and spores could reflect burial heating rather than wildfire, leaving its ecological meaning unresolved.
Key Innovation: Palynomorph Darkness Index measurements, controlled heating, microcharcoal and pyrolytic hydrocarbon records jointly attribute the signal to continental-scale fires in fern savannahs during greenhouse warming and forest collapse.
60. True-Triaxial Investigation of Hydraulic Fracturing in Granite and Sandstone Using Active-Passive Seismic Monitoring
Core Problem: Hydraulic-fracture geometry and seismic expression vary with host-rock structure, complicating monitoring across crystalline and sedimentary geothermal reservoirs.
Key Innovation: True-triaxial tests combine active transmission with passive microseismic monitoring, distinguishing tortuous, mixed-mode fracture growth in granite from more planar sandstone fractures and capturing aseismic deformation before failure.
61. Seismic retrofitting strategies for unreinforced masonry buildings in Europe: a review of techniques for improving seismic performance
Core Problem: Evidence for retrofitting unreinforced masonry is scattered across experiments, simulations and case-specific applications, obscuring trade-offs with durability and heritage conservation.
Key Innovation: A structured review of 192 studies organizes interventions into connectivity, diaphragm, masonry-integrity and added-element strategies, comparing their governing mechanisms, feasibility, durability and conservation compatibility.
62. Seismic performances of geogrid-wrapped tire-faced retaining walls: influence of geogrid wrapping structures and binding properties
Core Problem: The seismic response of tire-faced reinforced-soil walls depends on geogrid wrapping and tire binding, but displacement-based design guidance is limited.
Key Innovation: Shaking-table tests vary wrapping geometry, binding mode, thickness and strength, link geogrid strain to local wall deformation and derive preferred connections for different seismic intensities.
63. Terrestrial laser scanning point cloud-based simulation of microtopographic loess slope in water erosion
Core Problem: Rainfall rapidly modifies loess-slope microtopography, yet dense point clouds are costly and interpolation-based DEMs can smooth erosion features.
Key Innovation: Terrestrial laser scanning and NURBS surface reconstruction retain higher DEM accuracy than conventional interpolation across three slope gradients, while a 5 mm simplification threshold sharply reduces modelling time.
64. Frequency-domain and uncertainty-aware framework for the detection of decadal topographic changes in steep mountain basins: A case study of the Lanyang-Yilan region
Core Problem: Mixed radar, optical-stereo and LiDAR DEM stacks retain long-wavelength bias and spatially variable error that can create false erosion or deposition and inflate sediment budgets.
Key Innovation: The workflow combines phase-correlation registration, FFT bias correction, slope-roughness-dependent detection limits and 2,000-run Monte Carlo budgets, reducing apparent mobilized volume by about 75-92% in a Taiwanese mountain basin.
65. Label-efficient 3D forest mapping: Self-supervised and transfer learning for instance segmentation, semantic segmentation, and species classification
Core Problem: Dense annotation of forest point clouds is too expensive to scale instance, semantic and species mapping across complex stands.
Key Innovation: Self-supervision, domain adaptation and hierarchical transfer learning improve all three tasks in one open framework, including a 16.98-point AP50 gain for instance segmentation and 21% lower training energy use.
66. Optical remote sensing of sea surface wind streaks: Retrievals of wind directions and wind waves
Core Problem: Optical wind streaks are rarely exploited because their direction can be ambiguous and their texture can be confused with swell.
Key Innovation: Cross-spectral analysis with inter-band timing retrieves unambiguous wind direction and separates wind waves from swell; 157 streak scenes among 965 Sentinel-2 buoy matchups yield a 7.6-degree directional RMSE.
67. Arctic sea ice motion retrieval via vision foundation model and cross-sensor SAR
Core Problem: Single-sensor SAR revisit times limit sea-ice-motion coverage, while radiometric and geometric differences impede cross-sensor image matching.
Key Innovation: DINOv2 features and physics-constrained multiscale optical flow align Sentinel-1, ALOS-2, RCM and C-SAR observations, keeping velocity error within 0.73 cm/s while increasing effective Arctic coverage days by 53% and area by 75%.
68. Generation of high-precision precipitable water vapor products from morphed integrated microwave imagery
Core Problem: Near-real-time MIMIC precipitable-water products contain systematic bias, striping, boundary discontinuities and excessive smoothing.
Key Innovation: Spherical-cap harmonic analysis followed by XGBoost calibration produces 2016-2024 North American water-vapour fields at two resolutions, reducing GNSS-validated RMSE from 4.1 mm to 1.4-1.6 mm.
69. A data-driven runoff-sediment prediction system with explainable hybrid deep learning models
Core Problem: Runoff and sediment loads respond nonlinearly to weather, climate variability and human intervention, yet many forecasts provide neither calibrated intervals nor an explanation of their drivers.
Key Innovation: Wavelet-coherence screening, variational decomposition and a Transformer-LSTM are coupled to interval prediction and hierarchical SHAP analysis, reaching Nash-Sutcliffe efficiencies of 0.920 for runoff and 0.873 for sediment.
70. Divergent responses of runoff and sediment partitioning to bedrock exposure rate on karst slopes under extreme rainfall
Core Problem: How exposed bedrock redistributes surface runoff, subsurface flow and sediment on karst slopes during extreme rain remains poorly constrained.
Key Innovation: Factorial rainfall experiments separate rainfall-intensity control of runoff from bedrock-exposure control of sediment and identify a response transition near 30% exposure, where surface and underground partitioning changes markedly.
71. Silica‑calcium composite amendments increase the erosion resistance of soda saline-alkali soils under consecutive extreme rainfall events
Core Problem: The erosion response of soda saline-alkali slopes to repeated extreme rain and the mechanisms by which amendments suppress that erosion remain uncertain.
Key Innovation: Rainfall experiments show that calcium-chloride plus nano-silica cuts soil loss by 38.00% and runoff by 17.77%, linking the improvement to higher soil resistance, altered runoff power and increased cohesion.
72. Unmeasurable parameter-informed surface settlement prediction model based on CVAE
Core Problem: Shield-tunnel settlement models omit influential ground and construction variables that cannot be measured directly, limiting transfer between adjacent sections.
Key Innovation: UP-SSP uses a conditional variational autoencoder to represent those latent factors and a section-aware training and inference strategy to carry their influence into surface-settlement predictions.
73. A comprehensive assessment and comparison of streamflow simulations performance between the national water model versions 3.0 and 2.1 across the continental United States
Core Problem: National Water Model version 3.0 had not been evaluated across the continental United States for both ordinary daily flow and annual flood peaks.
Key Innovation: A 1979-2020 comparison at more than 4,300 USGS gauges shows broad gains over version 2.1 in daily and peak-flow metrics, while retaining pronounced regional, seasonal and human-impact-dependent errors.
74. Hazard or Anomaly? Evaluating VLMs for Understanding Dangers and Discrepancies
Core Problem: Binary safe-or-unsafe tests cannot reveal whether a vision-language model recognizes a physical hazard or merely reacts to an unusual scene.
Key Innovation: Separate hazard and anomaly labels across two datasets and multiple prompts show that current models often treat contextual irregularity as danger, exposing failures hidden by binary safety accuracy.
75. SkyEV: RGB-Event UAV detection and tracking dataset and baseline
Core Problem: Existing counter-UAV datasets underrepresent ego-motion, tiny targets, lens variation and compression-free sensing, weakening evaluation at operational detection ranges.
Key Innovation: SkyEV provides synchronized uncompressed RGB and event data under realistic camera motion and optics, plus a unified loader and multimodal detection baseline for small airborne targets.
76. Cross-Modal UAV Object Tracking: State-Aware Representation Learning and A Unified Benchmark
Core Problem: Switching between visible and thermal UAV sensors creates abrupt changes in both target appearance and apparent position that conventional trackers do not model explicitly.
Key Innovation: SARLA learns separate modality-state and spatial-state representations with a shift-prediction loss; the accompanying CM-UOT benchmark contains 1,079 cross-modal sequences and more than 671,000 frames.
77. CGMap: A Geospatially Aware Deep Learning Framework for Crop Gap Mapping Using UAV
Core Problem: Manual sugarcane emergence counts are error-prone and do not provide georeferenced maps of gaps that can guide field intervention.
Key Innovation: CGMap detects seedlings in UAV imagery with YOLOv8, normalizes row orientation with a minimum-spanning tree, infers missing plants from point spacing and exports the resulting gap map in GIS-ready WKT format.
78. Point Ladder Tuning: Parameter-Efficient Hierarchical Adaptation for 3D Point Cloud Understanding
Core Problem: Parameter-efficient tuning of heavily downsampled point-cloud tokens preserves global semantics but cannot recover local geometry discarded by the frozen backbone.
Key Innovation: Point Ladder Tuning builds a raw-point feature pyramid, fuses it with intermediate backbone features and generates instance-conditioned prompts, reaching leading classification and dense-prediction results with as little as 0.36% trainable parameters on PointGPT-L.
79. Predictive model for saturated hydraulic conductivity of thawed fine-grained permafrost
Core Problem: Freeze-thaw restructuring changes pore geometry and flow paths, so conductivity relations calibrated for unfrozen soils can underestimate thawed permafrost permeability.
Key Innovation: A thaw-consolidation model links conductivity to cryostructure degradation, effective stress and compression using three index properties, a soil-type term and a pore-ice proxy, then validates the estimates against historical laboratory data.
80. Efficient Multiscale Hierarchical Feature Extraction and Fusion Network for Hyperspectral Image Classification
Core Problem: Hyperspectral classifiers must extract multiscale spatial-spectral structure from few labels without the parameter cost of repeatedly optimized deep kernels.
Key Innovation: MSHNet-P uses differently sized image patches as fixed three-dimensional convolution kernels, removes redundant responses with layerwise PCA and classifies the compact fused representation with an SVM.
81. AS2G-Transformer: An Adaptive Spatial-Spectral Guided Transformer Network for Hyperspectral and Multispectral Image Fusion
Core Problem: Processing all spatial and spectral channels together can blur local detail and global context when hyperspectral and multispectral images are fused.
Key Innovation: AS2G-Transformer reconstructs spatial and spectral information in separate adaptive branches, then uses guided cross-attention to preserve both spectral fidelity and spatial detail across three benchmark datasets.
82. Decade-long isotope dataset of rainfall and non-rainfall waters in the central Namib Desert
Core Problem: Hyper-arid water-cycle studies lack long records that distinguish rainfall from fog and dew using the same isotope measurements.
Key Innovation: The Gobabeb dataset combines 731 isotope samples from rainfall, fog, dew and groundwater collected during 2014-2023 with spatial fog surveys and six decades of monthly water-input and Kuiseb River flood records.
83. The impact of calibration strategies on future evapotranspiration projections: a SWAT-T comparison of three hydrological modeling approaches in West Africa
Core Problem: Future evapotranspiration estimates can change with calibration targets even when the climate forcing and hydrological model are unchanged.
Key Innovation: Three SWAT-T strategies calibrated against discharge, leaf-area index and evapotranspiration show that adding leaf-area information improves seasonal water-loss dynamics and materially shifts CMIP6-based projections for the Beterou catchment.
84. Hydrological Evolution of Siling Co over the Past 38 Years: Lake Area, Water Level Monitoring, and Water Storage Estimation Based on Multi-Source Remote Sensing
Core Problem: Long-term water-storage change in Siling Co is difficult to reconstruct because consistent lake levels are sparse and the relative climatic drivers remain uncertain.
Key Innovation: MNDWI lake areas are tied to HYDROWEB levels and checked against GRACE and meteorology, reconstructing 1988-2025 expansion of 805.83 km2, a 14.33 m level rise and 30.42 km3 of added storage.
85. Monitoring Spatiotemporal Forest Fragmentation in Urban Landscapes: An Improved Urban Fringe Mapping Approach Using Time-Series Remote Sensing Data
Core Problem: Urban-fringe forest fragmentation is difficult to compare through time because the fringe boundary itself moves and differs among land, population, economic and thermal indicators.
Key Innovation: A fused 2004-2024 boundary and landscape-metric analysis shows fringe forest cover falling from 24.72% to 18.37% and fragmentation intensifying earlier and more strongly than in urban-core forests.
86. Recharge landforms in a semi-confined karst aquifer revealed by geophysical and high-resolution UAV-based LiDAR surveys
Core Problem: Caprock can conceal recharge pathways in semi-confined karst, making surface morphology alone insufficient to identify focused infiltration and hydraulic connection.
Key Innovation: High-resolution UAV-LiDAR morphometry and three-dimensional doline volumes are combined with electrical resistivity tomography to distinguish doline types, trace infiltration and link aquifer connectivity to faults and fractures.
87. DFELD-YOLO: A Decoupled Lightweight Detection Model for UAV-Based Wind Turbine Blade Damage Inspection
Core Problem: Lightweight UAV detectors lose tiny defects during downsampling and struggle to represent long cracks against cluttered blade backgrounds.
Key Innovation: DFELD-YOLO combines pixel-preserving downsampling, a compact interference-suppression block and cross-shaped stripe attention, improving mAP at 0.5 by 3.4 points with 2.05 million parameters and 5.9 GFLOPs.
88. Remote Sensing Reveals How Agricultural Restructuring Spatially Organizes Non-Cropland Ecological Restoration in the Mu Us Sandy Land
Core Problem: Greening beside expanding cropland can reflect either genuine ecological recovery or a shift in land-use pressure, which conventional regional trends do not distinguish.
Key Innovation: Fractional vegetation cover is decomposed by land class and related to an Agricultural Opportunity Space Index, showing that non-cropland improvement supplied 84.30% of greening and strengthened along the restructuring gradient.
89. Gated Spiking Attention Transformer for Hyperspectral Image Classification
Core Problem: Hyperspectral classifiers are difficult to deploy on edge or on-orbit hardware because conventional attention is computationally and energetically expensive.
Key Innovation: GS-SpikeFormer replaces matrix attention with binary gated spiking operations and groupwise convolutions, using 0.35 million parameters and four time steps while retaining high accuracy on four benchmark datasets.
90. Heterogeneous Dual-Path Wide Residual Network for Accurate Land Use and Land Cover Image Classification
Core Problem: Land-cover classifiers must reconcile local texture, long-range context and large appearance differences across sensors and spatial resolutions.
Key Innovation: HDP-WRN fuses an ECA-equipped local convolution path with a Ghost-convolution and CBAM global path, reporting 95.11-98.67% accuracy across four benchmarks while noting protocol differences in external comparisons.
91. WaveUAV-YOLO: A Lightweight Architecture for UAV Object Detection with Frequency-Domain Edge Preservation and Stable Gradient Fusion
Core Problem: Edge-deployed UAV detectors often erase high-frequency boundaries of sub-pixel and elongated targets during downsampling and feature fusion.
Key Innovation: WaveUAV-YOLO combines Haar-wavelet downsampling, asymmetric convolutions and mean-normalized aggregation, retaining weak edge signals while running at about 50 frames per second in the reported desktop-GPU test.
92. Assessment of Feature Selection Methods for Machine Learning-Based Chlorophyll-a Retrieval Across Optical Water Types
Core Problem: Chlorophyll retrieval is unstable across optical water types because redundant bands, nonlinear interactions and regional sampling can select different predictors.
Key Innovation: Seven feature selectors and five machine-learning models are tested with spatially blocked stratified splits and robustness filters, yielding compact predictors that transfer from MERIS bands to independent MODIS and GlobColour matchups with explicit uncertainty diagnostics.
93. Fine-Scale Spartina alterniflora Mapping Using Advanced Deep Learning and High-Resolution UAV Imagery
Core Problem: Moderate-resolution imagery misses small, early Spartina alterniflora invasions where water ripples and native vegetation create similar textures.
Key Innovation: UAV-SaSeg supplies high-resolution training data and DINOsegNext couples DINOv3 features with multi-frequency filtering, reaching 0.7420 IoU at 8.5 ms per image for edge-ready coastal mapping.
94. Land Use/Land Cover Classification of the Qinghai Lake Basin Using Multitemporal Sentinel-1/2 Imagery
Core Problem: Single-date imagery and class imbalance limit 10 m land-cover maps in the topographically and ecologically heterogeneous Qinghai Lake Basin.
Key Innovation: QLBLC-10 fuses multitemporal Sentinel-1 and Sentinel-2 data with DEM features in an SVM; ablation raises apparent accuracy from 77.07% to 91.24%, while area-weighted validation reports the lower, more realistic 81.50% accuracy.