In magnetotelluric (MT) forward modeling, the double-curl operator is associated with a large null space. In addition, the strong electrical contrast between air and subsurface media leads to a large condition number of the equations system, causing traditional Krylov subspace methods to suffer from slow convergence or even divergence, especially at low frequencies. The existing geometric multigrid (GMG) methods based on structured hexahedral meshes offer high computational efficiency, yet their discretization with regular grids is difficult to adapt to complex geological structures, limiting their ability in modeling practical earth structures. In contrast, the unstructured grids exhibit strong adaptability and allow for more accurate representation of complex geological boundaries. By combining unstructured meshes with geometric multigrid (GMG), it is possible to improve the efficiency of iterative solutions while maintaining good adaptability to complex structures. To this end, we propose a finite-element forward modeling method based on an unstructured geometric multigrid framework. The method constructs nested unstructured meshes progressively from coarse to fine, establishes restriction and prolongation operators between grids through basis function interpolation and integration, employs a V-cycle strategy to solve the linear system, and finally achieves three-dimensional (3D) MT forward modeling efficiently. The accuracy of the proposed algorithm is checked against the analytical solution for a homogeneous half-space model. Subsequently, a double-block model, a Dublin Test Model 1 (DTM-1), and a subduction-zone mantle plume model are constructed to validate the effectiveness of the proposed algorithm. The results demonstrate that the proposed geometric multigrid method outperforms traditional Krylov subspace methods (BICGSTAB-ILU, BICGSTAB-SOR, GMRES-ILU, and GMRES-SOR) in terms of iterations, computational time, and stability. It is more suitable for large-scale 3D MT forward modeling under complex geological conditions.
In the field of cross-domain hyperspectral image (HSI) classification, single-source domain generalization frameworks are currently being explored. Most existing methods generate out-of-domain data from the source domain using techniques like style transfer, photometric, and geometric transformations. However, the domain expansion strategies designed by these methods often ignore the characteristics of mixed pixels in hyperspectral images. Additionally, current HSI cross-domain classification methods mainly focus on domain-invariant representations, neglecting the use of domain-specific knowledge for cooperative inference in the target domain. To address these limitations, we propose the Spectral Unmixing-guided Prompt Network (SUPnet), which integrates unmixing and classification through a coordinated mechanism. This approach generates out-of-domain data based on spectral unmixing and employs a prompt learning strategy to facilitate multi-domain-specific representation collaborative inference. Specifically, in order to estimate abundance values and endmembers in an unsupervised manner, a spectral unmixing generator (SUG) is designed, which constructs extended domains by perturbing endmembers weighted by abundance values. Furthermore, a multi-level prompt learning (MPL) is developed to perform prompt learning at both the domain and endmember levels, thereby extracting specific knowledge and correlation between different domains to improve target domain inference. Experimental results on five cross-domain HSI datasets covering multiple platforms (including UVA, airborne, and satellite) demonstrate that the proposed method significantly outperforms several current state-of-the-art approaches. Notably, SUG proves to be more efficient than many of advanced unmixing methods. The codes will be available from the website: https://github.com/YuxiangZhang-BIT/IEEE_TGRS_SUPnet.
Diffusion models have recently achieved remarkable performance in image super-resolution (SR), but their high computational cost limits practical deployment in remote sensing applications. To address this issue, we propose SlimDiffSR, an efficient diffusion-based framework for real-world remote sensing image SR. Unlike existing single-step diffusion methods that rely on fixed timesteps, we first introduce an uncertainty-guided timestep assignment strategy to construct a stronger single-step teacher model, where reconstruction difficulty is explicitly linked to diffusion timesteps, enabling adaptive generative strength. Building upon this teacher, we further present a structured pruning strategy tailored to remote sensing imagery, which systematically removes redundant semantic modules and replaces standard operations with lightweight designs, including frequency-separable convolution, direction-separable convolution, and a query-driven global aggregation module. These components explicitly exploit the unique characteristics of remote sensing data, such as sparse high-frequency details, strong directional patterns, and long-range spatial dependencies. To strengthen the distillation process, we introduce MMD-based distribution alignment alongside MSE feature supervision, improving PSNR by 0.23 dB. Extensive experiments on multiple remote sensing benchmarks demonstrate that SlimDiffSR achieves a favorable balance between efficiency and reconstruction quality. In particular, it attains up to 200× inference acceleration and a 20× reduction in model parameters compared with multi-step diffusion models, while achieving competitive perceptual quality and clearly outperforming existing efficient diffusion baselines in efficiency. The code is available at https://github.com/wwangcece/SlimDiffSR.
Characterizing the spatial distribution of subsurface geological structures is essential in engineering geology, geophysics, and related fields. Directional borehole radar (DBR), an array-based extension of conventional borehole radar, enables three-dimensional (3D) imaging within a single borehole for subsurface characterization. However, in complex geological environments, subsurface heterogeneity, multi-target coexistence, and strongly coherent scattering often result in severe temporal overlap of echoes, limiting the separability of target responses and degrading the stability and accuracy of conventional 3D imaging methods. To address this limitation, we propose a short-time direction-of-arrival (ST-DOA)-based 3D imaging method for DBR. ST-DOA analysis is applied to migrated multi-channel radar data to construct a joint time–azimuth representation, which transforms temporally overlapping echoes into separable features in the time–azimuth domain. This enables effective separation of multi-target responses, followed by echo localization and spatial reconstruction to achieve 3D imaging. We validate the proposed method using synthetic data, laboratory experiments, simulated data, and field measurements. The results demonstrate stable and reliable imaging performance under strong echo overlap and complex scattering conditions, with improved imaging accuracy compared with existing methods. The proposed method improves the imaging capability of DBR, providing new possibilities for accurate mapping of complex subsurface geological structures.
Joint classification of multispectral (MS) and panchromatic (PAN) imagery has achieved remarkable success, aiming to provide more detailed and accurate interpretations of ground objects. However, when labeled samples are insufficient, the generalization performance of deep learning-based methods can be significantly affected. Therefore, we explore a novel cross-modal contrastive learning paradigm, the Dual-Key MoCo, to achieve few-shot joint representation learning using unlabeled samples. We also utilize an early-stage adversarial fusion approach to mitigate the loss of unique features from a single source. Specifically, 1) we propose an early-stage adversarial fusion mechanism that leverages correlation to construct adversarial fusion constraints, thereby highlighting unique frequency information and reducing similar redundant information; 2) we propose a contrastive learning paradigm, Dual-Key MoCo, that enhances intra-class invariance and inter-class discriminability by maximizing similarity between anchors and positives while minimizing similarity with negatives; 3) we construct a cross-modal self-supervised pre-training framework that leverages sufficient unlabeled samples to develop a robust joint representation prior. This provides a robust pre-training model for few-shot PAN and MS joint classification tasks. The theoretical analyses and experimental results on multiple popular datasets comprehensively illustrate the robustness and effectiveness of our proposed method under a few-shot learning. Our source code is accessible at: https://github.com/Xidian-AIGroup190726/DKMoCo.
