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.

Clustering is a fundamental task in graph data mining, including both node-level and graph-level clustering. While the former has been extensively explored to capture local structures and features, the latter has gained attention for its ability to capture global relationships and high-level abstractions. However, existing methods often address these two tasks in isolation, which not only wastes computational resources but also fails to fully leverage the knowledge from both levels to improve each other, hindering consistent performance improvement. To this end, we propose a novel Unified Graph Clustering Network called UGCN, which employs both local and global graph information to address node- and graph-level clustering collaboratively. In detail, we design a dual-branch projector that performs joint learning at both node and graph levels. The first branch extracts node-level features and projects them into distinct cluster layers, where the derived prototypes are used to refine graph attributes and highlight clustering-friendly substructures. In parallel, the second branch captures subgraph embeddings and aggregates them into discriminative graph-level representations. we align the two branches through joint contrastive objectives to establish a bidirectional interaction: refined prototypes guide subgraph and graph-level clustering, while graph-level pseudo-labels provide feedback to enhance node-level clustering. Extensive experimental results across seven datasets demonstrate that our method significantly outperforms existing state-of-the-art approaches.

The marine vibrators (MVibs) have become indispensable in marine industrial exploration due to its controllable energy output, high repeatability, and environmental compatibility. MVib-based blended acquisition improves exploration efficiency by reducing sampling duration and operational costs, but it introduces significant challenges in source separation due to blending noise. Since the subsequent processing method of MVibs require precorrelation data, this article presents a deblending framework for precorrelation MVib data by integrating a prior network into the inversion framework. To address the scarcity of labeled MVib data for deep learning, a data augmentation strategy is proposed in which natural images are converted into seismic-like data using band-pass filtering to train the prior network. This approach enhances model generalization and alleviates the limitations of MVib datasets. The proposed framework is validated through multi-MVib simulations and large scale open sea trials. Experimental results demonstrate that the method effectively suppresses blending noise while preserving key signal features, outperforming conventional deblending techniques and demonstrating strong practical value in real-world marine exploration.

Seed germination rate is a critical factor in agricultural productivity. Traditional approaches to germination assessment necessitate human scrutiny, introducing subjectivity and diminishing operational efficiency. While fully supervised deep learning approaches offer objectivity, reproducibility and efficiency, they require large-scale and high-quality labeled datasets, which are often challenging to obtain. To address this limitation, this study introduce a Semi-Supervised Germination Detection (SSGD) method built upon the Soft Teacher framework. SSGD employs a Faster R-CNN detector with ResNet50-FPN feature extraction network in both the teacher model and the student model. To ensure higher learning stability, the parameters of the teacher model are dynamically improved by an exponential moving average (EMA) updating mechanism. This study conducted a comprehensive evaluation of SSGD on the publicly available Pennisetum glaucum (PG), Secale cereale (SC), and Zea mays (ZM) datasets. Remarkably, with only 10% of labeled data, SSGD achieved mAP50 scores of 0.954, 0.928, and 0.961 on PG, SC, and ZM, respectively, surpassing fully supervised methods trained on 100% labeled data, including YOLOv3, FCOS, Cascade R-CNN and Faster R-CNN. Moreover, SSGD consistently outperformed the Faster R-CNN baseline across various annotation ratios (1%, 10%, 20% and 30%). Notably, even when the PG dataset’s labeling rate dropped to only 1%, SSGD maintained a high mAP50 of 0.94, exceeding the baseline by 5 percentage points. These findings underscore SSGD’s strong adaptability to limited data and further emphasize the effectiveness of semi-supervised learning in seed germination detection.

In the field of Earth observation and computer vision, oriented object detection for remotesensing images is a crucial task that aims to locate objects more accurately in complex scenes containing a large number of densely arranged, large aspect ratio, and arbitrarily oriented objects. Although recently proposed methods have achieved remarkable performance, there are still several challenges to address: 1) interference from complex backgrounds, 2) imbalanced and mismatched label assignments caused by tiny objects and objects with large aspect ratios, and 3) misalignment between the tasks of classification and localization. In this article, we propose a frequency-assisted dynamic learning network (FADL-Net) to overcome the crucial challenges. Concretely, we introduce a spatial-spectral feature pyramid network to adaptively capture global long-range dependency feature representations containing various frequency domains. Meanwhile, to produce more reliable training samples for objects with extreme shapes, we design a geometric aware dynamic label assignment to dynamically mitigate the imbalance and mismatch in label assignment in a coarse-to-fine manner, thereby achieving more stable optimization during the training process. Moreover, we propose a joint-learning rotated quality loss that addresses the inconsistency between classification and localization by dynamically adapting the weights of different samples in the training stage. Extensive experiments on several public remote sensing datasets demonstrate that our method performs favorably against state-of-the-art detection approaches.

Semisupervised object detection (SSOD) has garnered significant interest for its capability to enhance the detection performance by leveraging large amounts of unlabeled data. However, current SSOD methods primarily focus on detecting horizontal objects, with little research devoted to the detection of arbitrary-oriented objects in remote sensing images. Drawing inspiration from this limitation, this article proposes a semisupervised oriented object detection framework (S$^{2}$O-Det) to reduce annotation costs while improving detection performance in a semisupervised manner. Initially, the proposed task-consistent learning aims to alleviate the inconsistencies between classification and localization, which provides consistent confidence for the pseudolabels. Subsequently, the introduced coarse-to-fine sample mining employs dense prediction for pseudolabel assignment, adopting a divide-and-conquer approach to independently identify consistent and reliable labels for both classification and localization tasks. Finally, a probabilistic distillation loss ensures the harmonization of the probability distributions across the teacher and student feature domains, thereby reciprocally enhancing the learning competencies. Experimental results on the DOTA-v1.0 and DOTA-v1.5 datasets demonstrate that S$^{2}$O-Det achieves promising performance across different labeling ratios.

Suppressing random noise in seismic data is a significant problem in seismic data processing. Often, there is serious aliasing between the effective signal and random noise, affecting the identification of weak signals, and even resulting in great difficulties in the suppression of conventional seismic signals. We propose an improved attention-guided convolutional neural network (ADNet) to eliminate seismic interference noise. After a sufficient amount of training, the network removes noise by transferring seismic data features learned from a synthetic dataset to tests with complex field data. Our workflow consists of four parts. First, in the model, we improve the feature enhancement module (FEM) and attention module (AM), increase the convergence speed, and enhance the expressive ability. Second, we use 2-D synthetic data to verify the ability of the model to suppress noise in seismic records. Third, we use 2-D real seismic data to further verify the denoising effect of the improved ADNet. Fourth, we convert the 3-D simulated seismic data and field data into 2-D data for processing and reorganize the 2-D denoising results into 3-D data. By comparing the noise suppression outcomes of several classic denoising methods, simulations and actual experiments show that the improved ADNet effectively maintains the signal amplitude, reduces the network depth, and better suppresses seismic noise. Hence, we believe that our model can be widely applied in the field of seismic data processing.