The emergence of large vision-language models (VLMs) has significantly enhanced the efficiency and flexibility of geospatial interpretation. However, general-purpose VLMs remain suboptimal for remote sensing (RS) tasks. Existing geospatial VLMs typically adopt a unified modeling strategy and struggle to differentiate between task types and interpretation granularities, limiting their ability to balance local detail perception and global contextual understanding. In this paper, we present SkyMoE, a Mixture-of-Experts (MoE) vision-language model tailored for multimodal, multi-task RS interpretation. SkyMoE employs an adaptive router that generates task- and granularity-aware routing instructions, enabling specialized large language model experts to handle diverse sub-tasks. To further promote expert decoupling and granularity sensitivity, we introduce a context-disentangled augmentation strategy that creates contrastive pairs between local and global features, guiding experts toward level-specific representation learning. We also construct MGRS-Bench, a comprehensive benchmark covering multiple RS interpretation tasks and granularity levels, to evaluate generalization in complex scenarios. Extensive experiments on 21 public datasets demonstrate that SkyMoE achieves state-of-the-art performance across tasks, validating its adaptability, scalability, and superior multi-granularity understanding in remote sensing.

The distortion of magnetotelluric (MT) data due to various noise has emerged as a universal yet challenging issue in geophysical exploration. When MT surveys conducted in mining areas and urban areas, there are always full of spatially and temporally random artificial electromagnetic sources, violating the conventional plane-wave assumption, resulting in near-field effects. While, most MT data processing methods have challenges of difficulty in suppressing near-field effect, even with losing useful signals. Thus, we propose a novel method based on multiscale decomposition combined with coherence clustering. First filtering out small-scale signals from the time series, then classifying MT data into useful signals, incoherent noisy data, and coherent noisy data. This approach generates a mixed MSD-Coclustering structure with enhanced robustness in restoring weak natural signals while attenuating highly coherent anthropogenic interference, particularly in strong interference environments. The performance of MSD-Coclustering was tested using synthetic data and MT measured data, the signal-to-noise ratio (SNR) of noisy data is improved by nearly ten times, with no limitation to noise properties and data quality. Notably, the proposed method effectively suppresses the near-field effect without loss of useful signals and economic costs generated by setting up remote reference stations. This advancement significantly improves the interpretation accuracy and reliability of MT data in complex environments, with MT application fields extending.

Based on the spatial structure correlation in different geophysical parameters, we propose a new 3-D joint inversion method for frequency-domain airborne electromagnetic (AEM) and airborne magnetic (AirMag) data by incorporating a local Pearson correlation constraint (LPCC). For each iteration, the entire model is separated into multiple subdomains and the Pearson correlation coefficients of resistivity and magnetization in the subdomain are employed as the additional regularization term to do the joint constraint. This new regularization term is continuously updated in the inversion process to ensure that the resistivity and magnetization models in two separated inversions converge to a similar spatial structure. As a statistics technology, the LPCC-based joint inversion scheme not only has the advantages of the conventional joint inversions, but also can implement the structural constraints in different scales by selecting different sizes of the subdomain. This provides the flexibility for solving multiscale problems. Synthetic examples show that the joint inversion can improve the overall inversion resolution by combining the high vertical resolution of the EM method and large exploration depth and high horizontal resolution of the magnetic method. In the application to field survey datasets, the joint inversion delivers better results than those of separate inversions, which further verifies the effectiveness of our method.