Bi-temporal remote sensing change detection remains constrained by two major challenges: pseudo-change interference in complex scenes, which becomes more pronounced when a single model handles urban and mountainous scenes, and the substantial morphological disparity between compact building changes and large-scale irregular landslide changes. To address these challenges, we propose DPVCD-Net, a Difference-Prior-Guided Pseudo-Video Change Detection Network. DPVCD-Net reformulates bi-temporal images as a pseudo-video sequence and introduces a learnable perceptual frame to establish an early difference prior for subsequent feature learning. The network integrates two complementary mechanisms: difference guidance and prior fusion. Specifically, DP-CSA derives channel-wise and spatial attention from the perceptual feature and residually modulates the complete spatiotemporal feature set, strengthening genuine-change responses and suppressing pseudo-changes across different scenes. Meanwhile, PF-STM (Prior-Fused Spatiotemporal Multi-Scale Modeling) consists of two sequential stages: ST-MDCM (Spatiotemporal Multi-scale Dilated Convolution Module) performs within-scale multi-receptive-field spatiotemporal aggregation, while EA-CSCA (Edge-Aware Cross-Scale Cross-Attention module) introduces boundary-aware cross-scale interaction for structural refinement. Their sequential coupling strengthens multi-scale and structural modeling for morphologically disparate change targets. Experimental results on the GVLM-CD, WHU-CD, and LBFD-CD datasets show that DPVCD-Net achieves F1 scores of 0.9101, 0.9275, and 0.9089, respectively, supporting its effectiveness in suppressing scene-dependent pseudo-change interference and performing unified detection of morphologically disparate building and landslide changes within a single shared-parameter model.
ABSTRACT Accurate and up‐to‐date 3D building information is increasingly essential for urban planning, disaster management, and geospatial analytics. At the same time, rapid advances in deep learning (DL) have enabled the automated extraction of building‐related information from remote sensing imagery. The reviewed literature indicates that most studies focus on 2D building footprint extraction, whereas the incorporation of building height information is comparatively less addressed. In this context, this review synthesises DL‐based methods for building height estimation from optical satellite and aerial imagery. The reviewed studies are classified according to their analytical objectives, modelling strategies, and input data sources. We analyse research trends and commonly used datasets, followed by an in‐depth discussion of five major pipelines for building height estimation. The paper concludes with recommended evaluation and minimum reporting standards, key challenges, and practical guidance for future work.
Cross-view geolocation aims to localize a ground-view query image by retrieving the best match from a database of geo-tagged aerial-view and satellite-view images. The perspective gap and urban complexity lead to appearance variations and spatial misalignment between query and reference images, making the task challenging. Although many existing works perform well on single benchmarks, they often overfit dataset characteristics and fail to adapt to unseen regions. To address this problem, we propose SenseGeo, a cross-view localization framework that improves generalization while maintaining accuracy. SenseGeo adopts a two-phase training strategy consisting of Phase I and Phase II. Phase I performs supervised learning with global comparison and local constraints. Phase II, termed label expanded adaptation phase (LEAP), generates pseudo labels for new regions and alternates between global comparison and joint optimization with local constraints. We further introduce two modules within the framework. Slice-guided perspective-aware encoding (SPAE) extracts perspective-aware local features from global representations, and position-prior alignment with local enhancement (PALE) imposes positional consistency using multiple priors to alleviate cross-perspective misalignment. Experiments show that SenseGeo achieves competitive performance within the same region and surpasses existing methods in cross-region evaluation, demonstrating robust generalization for real-world cross-view geolocation.
Object detection in high-resolution aerial imagery has become a central topic in computer vision, yet it remains hindered by three challenges: (i) tiny, densely distributed targets are easily overwhelmed by cluttered backgrounds, leading to missed detections and inaccurate localization; (ii) existing methods inadequately refine shallow features with global semantics, making fine-grained textures and boundary cues vulnerable to background interference; (iii) quadratic-complexity feature interactions cause high computational overhead, limiting their practicality for high-resolution aerial images. To address these problems, we propose MambaSOD, a small-object-oriented detection framework for aerial imaging that integrates deformable detail enhancement while achieving a favorable trade-off between detection accuracy and computational cost. First, we propose SODMamba, an SS2D-based state-space backbone adapted for aerial small object detection, which strengthens global contextual modeling with linear computational complexity. Second, IDWTUP restores high-frequency cues during upsampling to preserve tiny-object details. Third, Enhance-X performs scale-aware deformable feature recalibration to refine shallow representations with contextual semantics. Extensive experiments on multiple aerial datasets demonstrate that MambaSOD achieves state-of-the-art performance at less than half the computational cost of mainstream detectors. On the challenging VisDrone2021 dataset, it improves AP and AP50 by 5.3% and 6.6%, respectively, demonstrating its superiority in small object detection.
Accurate registration of sparse and low-overlap LiDAR point clouds plays an essential role in photogrammetric mapping, 3D scene reconstruction and geospatial data analysis. However, existing learning-based registration methods typically perform keypoint detection and feature description as separate tasks, resulting in inconsistent feature representations and unreliable correspondences in challenging outdoor environments. To address this issue, this paper presents a geometry-aware correspondence learning framework (GACNet) that jointly optimizes keypoint detection and descriptor extraction within a unified multi-task architecture. Specifically, a lightweight geometry-aware module is designed to integrate positional encoding with geometric self-attention, enabling effective modeling of long-range geometric dependencies and enhancing feature discriminability. In addition, a shared feature extraction backbone with task-specific prediction heads simultaneously learns keypoint saliency and local descriptors, promoting consistency between keypoint localization and feature representation. Extensive experiments conducted on the 3DMatch and KITTI Odometry benchmarks demonstrate that the proposed method achieves superior registration performance and robustness compared with state-of-the-art approaches while maintaining high computational efficiency. In particular, GACNet achieved 100% registration recall on the challenging KITTI Odometry dataset. The results indicate that the proposed framework provides an effective solution for large-scale LiDAR point cloud registration and 3D geospatial applications in complex real-world environments.
