Jue Wang

Object detection in remote sensing imagery is a fundamental and crucial task within the remote sensing domain. However, small object detection in remote sensing images remains a challenge for SOTA methods. This is because small objects often have large aspect ratios, exhibit characteristic variations across different orientations, and are distributed within complex backgrounds. Current methods, which employ square convolutions, often retain excessive background information during feature extraction, leading to suboptimal detection performance. In this paper, an orthogonal-nesting orientation-aware small object detection network based on YOLOv10 is proposed to achieve small object detection in complex backgrounds. ONOA-YOLO employs an orthogonal-nesting module to enhance the features of small objects from multiple orientations, improving the sensitivity of the network to their intrinsic features. Additionally, it incorporates a background-suppression module based on large convolutional kernels. This module integrates the contextual information of small objects as guiding cues with other features, enabling the suppression of background interference while preserving the extracted small object features.

Unsupervised domain adaptation (UDA) has become a research hotspot in remote sensing scene classification (RSSC) to reduce dependence on labeled samples and mitigate performance degradation caused by domain shift. Most existing UDA approaches rely on convolutional neural networks (CNNs), which are constrained by limited receptive fields and struggle to capture global contextual dependencies in complex remote sensing scenes. Vision Transformers (ViTs) provide strong global modeling capability; however, their sequence-based representation lacks explicit constraints on spatial neighborhood structures, resulting in insufficient perception of fine-grained local structures, thereby causing attention misdirection under domain shift. To address this critical problem, a Neighbor Consistency Modulation (NCM) module is first designed to extract local structural consistency representations and fuse them with global tokens, explicitly enhancing fine-grained land object perception. Furthermore, to achieve reliable category-level feature alignment, an Adaptive Probability Pairwise Alignment (APPA) method is introduced. Specifically, an adaptive threshold is designed to retain reliable target pseudo-labels, and discriminator-free pairwise constraints are proposed to align these retained samples with source samples via probability matching, thereby stabilizing cross-domain representations. Ultimately, these two complementary components are seamlessly integrated to form a novel transformer-based method, termed Neighbor-Consistent Transformer (NCT). Extensive experiments on three public remote sensing datasets demonstrate the superiority of the proposed NCT, achieving state-of-the-art performance in cross-domain adaptation. The code will be made publicly available at https://github.com/CccLYChen/NCT-UDA.

Utilizing joint observations from the Magnetospheric Multiscale (MMS) and OMNI spanning from 2017 to 2021, we investigate the response of the Earth's plasma sheet (PS) to solar wind (SW) forcing. The PS is divided into three key regions in the downtail ( X < −10 R _E , Geocentric Solar Magnetospheric), that is, the current sheet (CS), central plasma sheet (CPS), and plasma sheet boundary layer, using a hybrid filter–decision tree model (HFDTM). The PSs of different regions were analyzed under four upstream SW speed and interplanetary magnetic field (IMF) conditions: low‐speed northward (LSNW), low‐speed southward (LSSW), high‐speed northward (HSNW), and high‐speed southward (HSSW). For each regime, we perform a comprehensive statistical analysis of key physical parameters, including the convective electric field (E _CY ), kinetic electric field (E _K ), and ion density (n _i ) and temperature (T _i ). Our main finding are as follows: (a) Ec _Y in the plasma sheet shows a monotonic increase with rising SW–E _CY0 across all 3 PS regions under southward IMF, but ambiguous correlation with SW under northward IMF; (b) n _i shows a strong correlation with the solar wind under both IMF orientations, particularly during low‐speed conditions; and (c) E _K and T _i consistently display weak correlations across all regimes, highlighting the dominant role of internal processes in plasma sheet evolution. These results support a dual‐path coupling mechanism between the solar wind and the magnetosphere, characterized by persistent mass loading across all regimes, and an energy transfer pathway that is strong under the southward IMF but weak and primarily governed by internal energy relaxation processes under northward IMF.

In recent years, remote sensing domain generalization semantic segmentation has attracted increasing attention due to significant domain shifts caused by variations in sensors, imaging conditions, and geographic regions. Vision foundation models (VFMs) possess strong general-purpose feature extraction capabilities and can be effectively transferred across diverse data, showing great potential for remote sensing domain generalization semantic segmentation. However, existing VFM-based spatial domain parameter-efficient fine-tuning methods struggle to handle pronounced cross-domain intra-class variations in remote sensing. To address this issue, the Adaptive Frequency Aware Adapter (AFA-Adapter) is proposed, which adaptively selects frequency components to improve cross-domain intra-class feature consistency. Building upon this, the Spatial Multi-Prototype Adapter (SMP-Adapter) is proposed, which clusters multiple prototypes for features of each land-cover category to model complex intra-class diversity. Land-cover features are then weighted by their nearest intra-class prototype, thereby enhancing the discriminability of easily confused features at feature cluster boundaries. By integrating these two modules, we propose a Frequency–Spatial Dual Domain Fine-Tuning Network (D²FT-Net), which effectively alleviates cross-domain intra-class variations and improves the generalization capability of VFMs for remote sensing domain generalization semantic segmentation. Extensive experiments under four cross-domain settings demonstrate the effectiveness of the proposed D²FT-Net, which achieves an average mIoU improvement of 1.09% over state-of-the-art methods, with the best gain reaching 1.64%. The source code will be released at https://github.com/ssshen0315/D2FT-Net.

Robust building extraction is a critical component of Earth observation (EO) applications, facilitating urban planning and disaster response. While atmospheric conditions or imaging mechanisms often hinder single-sensor interpretation, the joint exploitation of optical and Synthetic Aperture Radar (SAR) imagery offers a viable solution for all-weather monitoring. However, effective multimodal fusion in EO tasks faces unique challenges: noise amplification in low-level multimodal features and semantic misalignment at deep-level descriptions. In this work, we present a frequency-guided detail and semantic fusion network (FSFNet) for high-precision building extraction from co-registered optical and SAR imagery. FSFNet couples two complementary modules: a frequency-guided detail enhancement module (FDEM) that operates on shallow representations and a semantic feature fusion module (SFFM) that operates on deep representations. FDEM uses successive wavelet decompositions and cross-frequency fusion to decouple and asymmetrically combine low- and high-frequency information, while a spatial–frequency fusion component reinforces fine boundary cues from optical high-frequency signals. SFFM enforces modality-consistent semantic representations by applying a similarity-driven warping alignment that projects heterogeneous deep maps into a shared space for dense alignment and fusion. The integrated design produces boundary-aware and semantically consistent joint representations. Extensive experiments on public datasets demonstrate that our proposed FSFNet outperforms single-sensor baselines and other state-of-the-art multimodal methods. Qualitatively, our method achieves more complete building-body extraction while markedly improving boundary fidelity across diverse scenes.

Remote sensing object detection is critical for earth observation and analysis. However, accurately detecting small objects remains a significant challenge due to complex and heterogeneous backgrounds as well as the inherently limited information these objects provide. While numerous specialized algorithms have been proposed to enhance small object detection, the weak feature representation of objects within cluttered scenes continues to severely limit performance, making this a persistent and open research problem. Therefore, in this article, to enhance the discriminative representation of small objects in complex background, a novel two-phase rotation normalized prompting oriented small object detection (RNP-OSOD) framework is proposed. In the first mask generation phase, a rotation normalized prompt for segment anything model (RNP-SAM) is proposed to obtain high-quality instance and semantic mask annotations. In this method, image patches containing objects are cropped, enlarged and rotation normalized to acquire RNP before being fed into SAM, which then produces accurate mask supervision for training. In the second mask supervising object detection phase, to amplify the feature responses of small objects and balance instances across different scales, a scale-adaptive instance focusing (SAIF) loss is proposed by employing the generated masks as supervisory signals. This loss comprises an area-guided semantic reconstruction loss, called scale-adaptive semantic (SAS) loss and a detection-decoupled instance segmentation loss. To rigorously evaluate the proposed framework, we conduct extensive experiments on the SODA dataset and create a new dataset, Small-DOTA, constructed by filtering and downsampling DOTA to match SODA’s characteristics. Experimental results demonstrate that our method achieves state-of-the-art performance across both benchmarks. The source code will be available at https://github.com/cbq233333/RNP-OSOD.

Semantic segmentation and cloud removal via optical-Synthetic Aperture Radar (SAR) fusion are crucial tasks in Earth observation. Existing works have proposed multitask architectures to simultaneously address segmentation and cloud removal, but they lack targeted modeling of under-cloud features, leading to blurred ground-object reconstruction and poor fine-grained recognition performance in thick-cloud regions. To address these challenges, this paper proposes a Cloud-Guided Mamba-based Multitask Network (CGM²Net) for end-to-end multimodal segmentation and cloud removal. Specifically, CGM²Net utilizes a dual-stream encoder and integrates three key modules at each stage: cloud-guided feature restoration (CGFR) module, which, under cloud-mask-guided gating, exploits spatial- and channel- wise cross-modal correlations to reconstruct sub-cloud optical features and suppress SAR speckle; cloud-guided semantic interaction (CGSI) module, which performs mask-aware state-space modulation to enable selective, region-aware cross-modal semantic exchange; and Mamba-Based Fusion (MF) module, which adaptively fuses enhanced multimodal features to fully exploit modal complementary information. Two task-specific decoders synergistically optimize segmentation and cloud removal, thereby promoting mutual enhancement between the semantic prior and visual restoration. Experiments on M3M-CR and LuojiaSET-OSFCR demonstrate that CGM²Net achieves state-of-the-art performance on both tasks. Ablation studies and feature visualizations further validate the complementary roles and effectiveness of the proposed modules. The source code will be available at https://github.com/1lee1338/CGM2Net.

Accurately detecting tiny objects in remote sensing imagery remains challenging, primarily due to weak detail representation and complex background interference, which stem from inadequate modeling of fine-grained features and limited ability to distinguish targets from visually similar contexts. Existing advanced DETR-based detectors often struggle with these issues owing to insufficient sensitivity to subtle details and limited capability for contextual modeling. To overcome these limitations, we propose a Hierarchical Detail and Context Enhancement Detection Transformer (HDC-DETR), a hierarchical query-based detection framework specifically designed for tiny object detection in remote sensing imagery. HDC-DETR explicitly enhances frequency-domain detail perception and spatial contextual modeling through two dedicated modules. The Frequency-Statistics Adaptive Attention Module (FSAAM) employs discrete wavelet transform to decompose multi-scale features and introduces a Local–Global Probability-Density Attention (LG-PDA) to selectively enhance statistically distinctive high-frequency edges and textures, thereby improving fine-grained detail discriminability. Meanwhile, the Hierarchical Context Modulation Block (HCMB) adaptively aggregates multi-scale spatial contexts and employs spatial gating mechanisms to selectively enhance foreground features and suppress ambiguous backgrounds, significantly improving foreground–background discriminability. Comprehensive evaluations conducted on the AI-TOD and LEVIR-Ship benchmarks demonstrate that HDC-DETR achieves superior performance in detecting tiny objects, especially under complex remote sensing scenarios. These results validate the effectiveness of the proposed method in simultaneously enhancing fine-grained detail perception and spatial contextual modeling.