Acoustic remote-reflection logging is widely used for characterizing geological structures surrounding boreholes, but in anisotropic formations, reflected-wave propagation exhibits strong directional dependence and mode-coupling effects, and conventional numerical simulations become computationally expensive for large-scale remote-detection problems. Existing asymptotic studies mainly focus on isotropic or simplified anisotropic models, leaving a gap in fast analytical methods for reflections from tilted transversely isotropic (TTI) interfaces received in vertically transversely isotropic (VTI) boreholes. To address this, a high-frequency asymptotic method based on generalized ray theory and the method of steepest descent (MSD) is developed for remote reflections in VTI boreholes. Under far-field conditions, borehole radiation, interface reflection, and reflected-wave reentry are asymptotically treated via successive MSD approximations. Generalized ray theory with Snell’s law determines the reflection path, yielding a compact frequency-domain formulation that requires only a single inverse Fourier transform to obtain transient responses. Finite-difference simulations validate the asymptotic solution at the waveform level for TTI formations, demonstrating good agreement in reflected-wave arrivals and waveform characteristics under representative anisotropic conditions. The method is applied to the Mesaverde Group and Green River Formation to analyze remote reflection responses in realistic sedimentary environments.
Remote sensing change detection underpins applications such as urban monitoring and disaster assessment. However, performance is often constrained by scarce pixel-level annotations and by severe class imbalance, where unchanged pixels vastly outnumber changed ones. To address these challenges, we introduce ChangeMix, a change-aware cut-and-paste augmentation for semi-supervised change detection that dynamically constructs an instance bank of change regions from both scarce ground truth and high-confidence pseudo-labels. ChangeMix stores paired pre- and post-change image crops including their corresponding change and confidence masks, and pastes multiple instances onto strongly augmented unlabeled image pairs for class balanced self-training. Furthermore, by randomly swapping the temporal order of the pre- and post-change patches before pasting, we vary the imaging conditions under which the pasted changes appear and keep appearance and disappearance events balanced. To demonstrate its applicability to building change detection, we validated ChangeMix as a framework- and model-agnostic module. We integrated it into three distinct semi-supervised learning frameworks (FixMatch, UniMatch V1, and UniMatch V2) and tested it across multiple network architectures (DPT with DINOv2 backbones and DeepLabV3+). Extensive experiments show consistent improvements across the benchmarks. At the 20% label budget with the DPT/DINOv2-Base architecture, averaged over three random seeds, ChangeMix increases changed-class Intersection-over-Union by 0.58–0.85 points on LEVIR-CD and 0.91–1.29 points on WHU-CD across the three frameworks. ChangeMix provides a practical and effective tool that can complement existing semi-supervised pipelines to address critical data-related challenges in remote sensing. The code will be made publicly available at https://github.com/IonutMotoi/ChangeMix.
For mapping high-precision seafloor topography, inverting bathymetry from satellite altimetry-derived gravity data offers a highly cost-effective approach. To assess the quality and reliability of a seafloor topographic map, it is essential to quantify the uncertainty in inverted bathymetry. Existing inversion methods for bathymetry implicitly assume that errors follow an idealized white noise distribution. However, this assumption may be violated due to the inherent complexity and variability of deep-sea topography, with inversion residuals exhibiting spatial heterogeneity that correlates with topographic undulations. Building on advances in interval prediction, i.e., conformal prediction (CP), this study suggests a new inference framework for seafloor topography that incorporates locally weighted inductive conformal prediction (LWICP) to quantify uncertainty in seafloor topography mapping. The most outstanding advantage of LWICP is that it is data-driven and introduces a local weighting strategy to address spatial heterogeneity in inversion residuals, enabling spatially adaptive uncertainty quantification for bathymetric point predictions from different inversion models. LWICP is applicable to various bathymetry inversion methods, including three primary methods demonstrated in this paper. Experiments across three deep-sea study areas with distinct geomorphological characteristics in the Western Pacific show that the proposed LWICP method achieves empirical coverage rates of 90.03%–92.01% and 95.04%–95.58% at nominal confidence levels of 0.90 and 0.95, respectively. Compared with the three baseline uncertainty quantification methods, LWICP reduces the average uncertainty by 24.01%–38.43%, while adaptively capturing the dependence of inversion uncertainty on topographic variability. The source code for the proposed LWICP method is publicly available at https://doi.org/10.5281/zenodo.22139793.
Hyperspectral image (HSI) and LiDAR fusion improves land-cover classification by combining spectral discrimination with geometric structure. However, many existing methods still treat LiDAR as an auxiliary feature stream and perform fusion through fixed operators such as concatenation, summation, or attention. Such designs do not explicitly adapt hyperspectral spatial aggregation according to local geometric context. Moreover, conventional linear or MLP-based embeddings may be limited in modelling nonlinear spectral-spatial-geometric interactions. To address these issues, this paper proposes a Geometry-Conditioned Implicit Spatial–Frequency Fusion Network (GFFNet) for HSI-LiDAR classification. GFFNet formulates multimodal fusion as a geometry-conditioned implicit aggregation problem, where LiDAR-derived structural cues guide how neighbouring hyperspectral features are aggregated. Specifically, HSI and LiDAR inputs are first projected into a shared latent space using Kolmogorov-Arnold Network (KAN)-based nonlinear embeddings. A Spatial Implicit Aggregation Unit then uses a KAN-parameterized implicit function to map neighbouring HSI features, query-level LiDAR geometry, and relative spatial offsets into adaptive candidate features and aggregation weights. In this way, the fusion operator varies with local geometric structure rather than remaining fixed across the image. To complement geometry-guided local aggregation, GFFNet further introduces a complex-valued frequency-domain branch for global structural modelling and a local residual pathway for preserving fine-grained spatial details. Experiments on Houston 2013, MUUFL, and Trento demonstrate that GFFNet achieves 96.98%, 92.90%, and 99.65% overall accuracy, respectively, outperforming representative HSI-LiDAR fusion methods. Ablation, KAN-to-MLP replacement, visualization, and complexity analyses verify the effectiveness of the proposed geometry-conditioned implicit fusion framework.
Improvements to the Time-Series Ratio (TSR) algorithm for radar-based soil moisture retrievals are presented and analyzed using data collected in the SMEX02 field campaign. This study advances the baseline TSR algorithm by incorporating a SMAP climatology-based dataset of soil moisture bounds disaggregated to 200 m using a thermal-hydraulic algorithm, a modified solution process that includes an initial guess for soil moisture, and an additional retrospective averaged data product. Dual-frequency, dual-polarization, and combined dual-frequency/dual-polarization methods are also presented as well as an innovative “re-scaling method” to systematically address vegetation impacts. Each modified approach is shown to improve SMEX02 campaign performance in terms of soil moisture root mean square error (RMSE), unbiased rms error (ubRMSE), and correlation relative to previously reported baseline results. The refined algorithms are also shown to meet or outperform a target ubRMSE of 0.06 cm3/cm3. The findings demonstrate the advantages of integrating multiple frequencies and/or polarizations in the TSR algorithm to provide high-resolution soil moisture retrievals and provided an updated synthetic aperture radar soil moisture sensing approach.
