Beach slope is a fundamental morphometric parameter governing coastal dynamics, wave dissipation, and shoreline stability, and serves as a key basis for coastal hazard prediction, wave runup and overtopping assessment, and the design of coastal engineering structures. However, due to the limited spatial coverage and high costs of conventional monitoring techniques, no high-resolution, national-scale dataset of beach slopes has previously been available for China until now. In this study, we develop an automated and transferable framework for deriving beach slopes from ICESat-2 ATL03 photon data and produce the first systematic national-scale dataset of China's sandy beach slopes. The workflow integrates (1) data preprocessing, including shoreline-orthogonal correction of ICESat-2 ground tracks, vertical datum harmonization to EGM2008, and selection of high-confidence photons (confidence >= 2); (2) adaptive photon clustering based on an OPTICS (Ordering Points To Identify the Clustering Structure) algorithm coupled with a moving-window filter for accurate beach photon extraction; and (3) slope inversion using low-order Fourier polynomial delineation of the foreshore combined with robust linear regression. The resulting dataset encompasses 688 along-track profiles covering 243 sandy beaches along China's coastline. Validation against high-precision reference data, including RTK-GNSS surveys and DEM-derived slopes, yielded strong agreement (RMSE = 0.012, R2 = 0.89), demonstrating the reliability of the proposed method across diverse geomorphic settings. Spatially, the results reveal a distinct north-south attenuation pattern, with steeper, more variable slopes in temperate northern coasts and gentler, low-gradient beaches in the tropical south. This novel dataset provides the first quantitative characterization of beach-slope spatial patterns at the national scale, offering critical parameters for storm-surge modeling, coastal-erosion assessment, and sustainable coastal management. Furthermore, leveraging open-accessible, global-scale ICESat-2 data, this framework is designed for easy transferability across coastal regions, offering a scalable solution for worldwide beach monitoring and assessment.

Land-air bimodal robots (LABR) provide efficient ground locomotion and flexible aerial mobility, making them suitable for complex navigation tasks. Traditional methods cannot distinguish between traversable obstacles (such as curtains) and rigid obstacles, resulting in unnecessary aerial transitions. To address these challenges, DS-LABRNav is proposed as a navigation framework that integrates a Vision-Language Model (Align-DS-V) into land-air planning. This framework consists of an event-triggered Obstacle-Induced Takeoff Trigger (OITT) and a modality-guided occupancy map update. OITT selectively activates the VLM reasoning only prior to a proposed aerial maneuver while the robot is safely on the ground. The occupancy map update then incorporates semantic reasoning results, enabling the planner to re-optimize the trajectory and prioritize land modalities for improved energy efficiency. Specific experiments demonstrate that DS-LABRNav improves navigation efficiency and path optimality while reducing mode-switching costs. Compared to traditional map-based approaches, the framework achieves energy savings of up to 70% in scenarios with traversable obstacles and unknown slopes, validating its effectiveness in real-world LABR deployments.

Integrated 3D reconstruction from aerial-ground images is essential for generating high-precision urban 3D models, yet severe variations in viewpoint, scale, and rotation make robust feature matching highly challenging. To address these limitations, this study introduces a rotation-robust detector-free matching network coupled with multi-view track refinement for incremental Structure from Motion (ISfM). The proposed workflow features four key modules. First, rotation-aware feature extraction replaces traditional convolutions with an Omnidirectional State Space Block (OSS Block) that selectively scans across eight symmetrical directions to model long-range spatial dependencies and synthesize rotation-invariant feature maps. Second, multi-scale attention transformation utilizes quadtree attention to build a hierarchical token pyramid that isolates high-association token regions and discards irrelevant areas, capturing long-range context with linear computational complexity. Third, bi-directional feature matching executes a symmetric coarse-to-fine matching scheme where coarse alignment computes dual-direction Softmax confidence matrices under mutual nearest neighbor constraints, and fine alignment uses a multi-layer perceptron to regress sub-pixel coordinate offsets. Finally, multi-view track refinement employs an integrated indexing structure to evaluate localized spatial proximity and link disjoint sub-tracks to the highest-confidence anchor point, ensuring stable feature repeatability across the ISfM pipeline. By using real aerial-ground datasets, experimental results demonstrate that the proposed method improves AUC at 5° pose error by 93.9

High-precision and robust autonomous localization is an essential technology for uncrewed systems to achieve reliable navigation. Traditional map-based positioning methods can provide high localization accuracy, but as the size of the environment increases, these methods often require significant computational resources and memory, making them unsuitable for resource-constrained platforms. To address this, we propose a lightweight localization method based on sparse maps, designed to reduce computational overhead while maintaining high precision and robustness. Our approach combines light detection and ranging (LiDAR) and inertial measurements by extracting planar features from local regions for scan matching. This method reduces data processing complexity, ensuring both real-time performance and localization accuracy. Unlike traditional methods that match LiDAR data with the global map for every frame, we perform global map matching only at specific intervals, applying pose corrections through a time threshold mechanism, thereby significantly reducing computational burden. Additionally, to improve the accuracy and stability of state estimation, this article integrates LiDAR data at specific time intervals with the results of map matching, along with planar feature matching results extracted from local regions, to achieve precise pose correction. Combined with a nonlinear geometric observer featuring strong convergence properties, this approach enhances localization accuracy and robustness through multisensor fusion. Experimental results on publicly available large-scale datasets and our self-collected datasets demonstrate that our method outperforms existing map-based localization methods by achieving the lowest state update time and reducing computational costs by over 50%, while maintaining robust performance in complex environments. Data will be publicly available on https://github.com/iron2523/sparse-map_localization

Urban region representation learning has emerged as a fundamental approach for diverse urban analytics tasks, where each neighborhood is encoded as a dense embedding vector for effective downstream applications. However, existing approaches suffer from insufficient multi-modal alignment and inadequate spatial relationship modeling, limiting their representation quality and generalizability. To address these challenges, we propose UrbanMMCL, a novel self-supervised framework that integrates multi-modal multi-view contrastive pre-training with unified fine-tuning for comprehensive urban representation learning. UrbanMMCL employs a dual-stage architecture. First, cross-modal contrastive learning aligns diverse data modalities including remote sensing imagery, street view imagery, location encodings, and Vision-Language Model (VLM)-generated textual descriptions. Second, multi-view adaptive graph contrastive learning captures complex spatial relationships across human mobility, functional similarity, and geographic distance perspectives. The framework then fine-tunes all parameters with the learned representations for effective adaptation to downstream tasks. Comprehensive experiments demonstrate that UrbanMMCL consistently outperforms state-of-the-art methods across pollutant emission prediction, population density estimation, and land use classification with minimal fine-tuning requirements, thereby advancing foundation model development for diverse Geo-AI applications.

Spherical images, with their 360 degrees omnidirectional coverage, have become a crucial data source for 3D reconstruction and visual localization. However, feature matching in large-scale, high-resolution spherical images presents significant challenges, primarily due to high computational costs (O(n2)) and the trade-off between efficiency and accuracy. This study proposes a fast feature matching method that combines graph-order scheduling (GOS) and cascade hashing. First, a spatial adjacency graph is constructed using data from the onboard position and orientation system (POS). This graph is dynamically partitioned into memory-efficient subscenes using breadth-first search (BFS). By employing overlap-aware subscene scheduling, the computational complexity is reduced from O(n2) to O(k & sdot; m2), where k represents the number of subscenes and m denotes the number of images per subscene, with the condition that m << n. Next, reprojection-based rectification is applied to mitigate geometric distortions inherent in spherical images, converting them into a cubic map representation. Feature matching is then performed using a three-level cascade hashing technique. To enhance efficiency while maintaining accuracy, a three-stage filtering strategy is implemented, which includes coarse screening via locality-sensitive blending, refinement through low-dimensional hashing, and final matching based on Hamming distance. Experiments conducted on three datasets demonstrate that the proposed method outperforms traditional algorithms in feature matching, Structure from Motion (SfM)-based 3D reconstruction, and visual localization. The approach achieves a speedup ratio ranging from 4.0 to 30.0 in feature matching compared to SIFTGPU, and the number of reconstructed 3D points in SfM increases by 40%-90%. Overall, this method offers an efficient and reliable solution for feature matching of spherical images.

This article addresses high-precision trajectory tracking for fully actuated hexarotor unmanned aerial vehicles (UAVs) subject to model uncertainty and external disturbances. To exploit the repetitive nature of UAV missions, an adaptive iterative learning control (AILC) scheme is developed, which incrementally enhance tracking accuracy using historical data form previous executions. The controller is designed based on the decoupled translational dynamics of a six-rotor tilted-propeller platform and incorporates adaptive laws that jointly estimate and compensate for reference input errors, aerodynamic drag, residual attitude deviations, and lumped disturbances. Rigorous convergence and boundedness are established via a composite energy function (CEF) analysis in the iteration domain. The approach is validated via MATLAB simulations, PX4-based software-in-the-loop (SITL) in Gazebo and real-flight experiments on a laboratory hexarotor platform. To the best of current knowledge, this work: 1) represents the first application of AILC to a fully actuated hexarotor UAV, showing significant improvements in trajectory-tracking accuracy over benchmark methods; and 2) demonstrates the platform’s potential for attitude-stable logistics tasks such as liquid transport through real-flight tests.

Broadband radiance unfiltering, the spectral restoration of filtered sensor observations, is a fundamental prerequisite for accurate Earth Radiation Budget (ERB) quantification. The newly launched FY-3F Earth Radiation Measurement-II (ERM-II) instrument introduces a dedicated longwave channel, expanding spectral dimensionality beyond its predecessors and necessitating a specialized retrieval framework. This study establishes an operational, scene-dependent unfiltering model for ERM-II that explicitly addresses sensor-specific optical characteristics. A comprehensive spectral library was generated using radiative transfer simulations driven by high-precision global priors. Crucially, to mitigate radiometric distortion caused by sub-pixel scene heterogeneity, we implemented a Point Spread Function (PSF)-weighted convolution scheme. This method integrates collocated high-resolution MERSI-III cloud masks to accurately characterize the effective scene radiance perceived by the ERM-II detector. Based on this framework, optimal unfiltering architectures were determined: a linear fusion of shortwave and total channels for shortwave unfiltering, and distinct single-channel strategies for daytime (linear) and nighttime (nonlinear) longwave retrieval. Validation against Clouds and Earth’s Radiant Energy System (CERES) products demonstrates that the proposed multi-channel fusion effectively compensates for spectral response mismatches. The model achieves a coefficient of determination (R2) exceeding 0.96, with RMSEs constrained below 2.1 W⋅m−2⋅sr−1 and absolute biases within 0.21–1.59 W⋅m−2⋅sr−1. These results represent a substantial quantitative improvement over legacy FY-3C algorithms, reducing systematic biases by up to 96%. This study provides a rigorous physical foundation for the ERM-II mission, preparing it to provide valuable complementary measurements for global climate records in the post-CERES era.

Autonomous object detection in remote sensing requires systems that can discover new categories and assign them usable labels during deployment. Existing Open-World Object Detectors identify unknown objects but leave them unnamed until manual annotation. In contrast, Open-Vocabulary Detectors recognize unseen categories only with provided prompts at test time, lacking autonomous discovery or naming. This work presents HSGDet, a detector that achieves both discovery and semantic assignment at test time without external prompts. This method introduces DHGA that navigates a hierarchical semantic graph to perform scene-conditioned coarse-to-fine classification of detected objects. It leverages spatial co-occurrence patterns from surrounding scene context to produce classification confidence scores. High-scoring regions are identified as known objects, while low-scoring regions are flagged as unknown detections. Unknown regions pass to CR2T, which synthesizes text embeddings by fusing visual features, hierarchical parents, and scene context, enabling prompt-free labeling and vocabulary expansion. This approach enables prompt-free semantic labeling and supports autonomous vocabulary expansion without requiring external models. Results demonstrate that HSGDet outperforms state-of-the-art methods by a large margin of 6.6 points in Known mAP and 9.9 points in Unknown Recall. It also reduces Wilderness Impact by 36\%, enabling scalable and autonomous aerial monitoring.

Indoor positioning acts as a fundamental infrastructure for Internet of Things (IoT) and artificial intelligence systems. However, conventional landmark-based positioning methods suffer from degraded performance in complex indoor scenarios, where complicated path layouts and varied pedestrian behaviors hinder reliable scene recognition. This work presents an integrated framework for indoor scene recognition and pedestrian localization. The core module adopts the Minimum Matching Group of Pedestrian Steps (MMG-PS) to detect scene transitions, and a machine learning model is deployed to fuse real-time sensor data and historical trajectory patterns for enhanced recognition accuracy. To adapt to different indoor environments, two dedicated Access Point (AP) selection strategies are developed: Principal Component Analysis (PCA) is applied to corridor-like linear spaces, while the Eight-Diagram (ED) algorithm is customized for open halls. The proposed system achieves pedestrian positioning and scene transition detection simultaneously. Extensive experiments conducted in an office complex and a shopping mall demonstrate that our method yields superior performance compared with traditional approaches.

The integration of aerial and ground images has been a promising solution in 3D modeling of complex scenes, which is seriously restricted by finding reliable correspondences. The primary contribution of this study is a feature matching algorithm for aerial and ground images, whose core idea is to generate intermediate views to alleviate perspective distortions caused by the extensive viewpoint changes. First, using aerial images only, sparse models are reconstructed through an incremental structure from motion (ISfM) engine due to their large scene coverage. Second, 3D Gaussian splatting (3DGS) is then adopted for scene rendering by taking as inputs sparse points and oriented images. For accurate view rendering, a render viewpoint determination algorithm is designed using the oriented camera poses of aerial images, which is used to generate high-quality intermediate images that can bridge the gap between aerial and ground images. Third, with the aid of intermediate images, reliable feature matching is conducted for match pairs from render-aerial and render-ground images, and final matches can be generated by transmitting correspondences through intermediate views. Using real aerial and ground datasets, the validation of the proposed solution has been verified in terms of feature matching and scene rendering and compared comprehensively with widely used methods. The experimental results demonstrate that the proposed solution can provide reliable feature matches for aerial and ground images with an obvious increase in the number of initial and refined matches, and it can provide enough matches to achieve accurate ISfM reconstruction and complete 3DGS-based scene rendering.

Deformation monitoring of long-span bridges is essential for evaluating their structural health and safety. Traditional methods are labor-intensive and low-frequency. Image-based approaches offer precise deformation results but deteriorate with distance; thus, they cannot meet the requirements of long-span bridges. To fill these gaps, this study proposes a real-time deformation monitoring approach by integrating cameras with inertial sensors (RTDCI). Cameras capture the relative deformation of targets using computer vision. Multiple cameras are serially integrated to extend the measurement range. Inertial sensors compensate for camera motion errors alongside structural deformation. A multi-camera deformation measurement model is developed to process both images and inertial data. Laboratory and field experiments were conducted. Results demonstrate that RTDCI achieves an error of 0.32 mm with a 30 Hz frequency and 34 frames per second (FPS) in a vibrating laboratory environment. Field experiments exhibit millimeter-level accuracy when compared with hydrostatic leveling and ground-based interferometric radar. It offers an alternative deformation monitoring approach for long-span infrastructures and can be extended to complex construction environments.

Robust pavement crack detection in complex scenes remains a significant challenge. This stems not merely from the scarcity of annotated data, but more critically, from the severe lack of pattern diversity within existing datasets. Key variations in morphology, scale, background texture, and imaging conditions are often underrepresented, which fundamentally impedes the generalization capability of recognition models. While generative approaches (e.g., GANs and diffusion models) offer a potential path to augment this diversity synthetically, they commonly suffer from poor background realism, entangled structural-appearance representations, and a lack of precise control over generated defects. This paper presents the Crack Diffusion Generator (DiffCrack), a diffusion-based framework designed for semantic-structural controllability in crack image synthesis. DiffCrack decouples crack geometry and visual appearance through two conditioning inputs: a binary mask to anchor spatial layout, and a Hierarchical Prompt Attention (HPA) module to independently modulate attributes such as width, depth, color, and texture. This design enables the controllable and targeted generation of diverse, photorealistic crack patterns that are often missing from real-world datasets. Extensive experiments on real datasets demonstrate that training with DiffCrack-generated images enhances the F1-score of segmentation models by up to 23% on average under complex scene conditions. This result validates that DiffCrack is a scalable, pattern-aware data augmentation tool. By addressing the critical bottleneck of data diversity, our framework offers a principled pathway to improving the robustness and generalization of infrastructure inspection models.

Optical remote sensing images frequently suffer from extensive cloud coverage, resulting in the loss of valuable raw data and severely hindering their Earth observation capacity. Cloud removal methods employing auxiliary synthetic aperture radar (SAR) data have been explored in recent years instead of using optical cloud-free reference images from the same area on a different date, where uncertainties arise due to surface changes and unavailability of cloud-free images is often encountered. However, the inherent speckle noise in SAR data brings challenges in interpreting texture information from SARs into optical images for cloud removal. In this study, we propose a data-driven noise-aware multisensor cloud removal (NAM-CR) method for multispectral images by fusing SAR data. The network is composed of a self-attention multimodal aggregator (SMMA) and deep-shallow feature cross-fusion (DSFF) modules for better integration of the texture information from SAR data and spectral information from optical images. To further reduce speckle noise effects, the proposed method is equipped with a novel cloud-guided noise optimization loss function. We trained and tested our method on 100,000 patch triplets from the SEN12MS-CR dataset containing global all-season Sentinel optical and SAR images. The experimental results demonstrate that NAM-CR outperforms the other six state-of-the-art cloud removal methods and achieves outstanding performance with the highest PSNR of 34.22 dB and minimal parameter usage. The source code is accessible to the public at http://jiasen.tech/.

The Ancient Shu Roads constitute a significant historical and cultural heritage attesting to the economic and humanistic interactions between China’s ancient central plains and the Ba-Shu regions. To achieve the digital preservation and management of the Ancient Shu Road relics, it is necessary to conduct 3-D reconstruction to obtain their photorealistic models. However, these relics are typically located in dense forests and cliffs, under such occluded environments, conventional visual 3-D reconstruction methods fail to meet the high-fidelity requirements of cultural heritage digitization. The recently emerged 3-D Gaussian splatting (3DGS) approach shows potential for real-scene reconstruction in complex environments, but its original pipeline follows pinhole camera model, making it unsuitable for fisheye images with severe distortion. To address these challenges, we propose a fisheye-adaptive 3DGS framework for modeling realistic scenes: First, a visual-inertial structure-from-motion (SfM) method is employed to recover real-scale sparse geometry initialization; Next, we introduce a hybrid distortion field and an octahedron projection model into 3DGS, which enables joint optimization of lens distortion and scene representation parameters, thereby remarkably suppressing the rendering artifacts caused by fisheye distortion and building high-fidelity 3-D models. We collected a close-range dataset with handheld devices across several representative Ancient Shu Road heritages, upon which we conducted comprehensive evaluation of our method, validating its superiority in 3-D reconstruction tasks under occluded environments, and offering a viable new solution for cultural heritage digital preservation.

Beaches provide essential ecological functions and support socio-economic resilience, yet accurate mapping is hindered by systematic limitations in global Digital Elevation Models (DEMs). A critical challenge remains in the intertidal zone, where frequent tidal inundation creates extensive data voids, disrupting the continuity of coastal topography. To bridge this fundamental data gap, we present NZ-BeachTopo30 which is a national-scale and full-coverage 30 m beach topography dataset for New Zealand constructed by fusing ICESat-2 photon-counting altimetry with Sentinel-2 multispectral time series. The dataset is available at 10.5281/zenodo.17785546 (Wang, 2025). Using DeltaDTM as a high-precision baseline for the stable backshore, we trained an XGBoost model on ICESat-2 control points and Sentinel-2 spectral-geometric features to reconstruct the missing intertidal topography specifically. SHAP analysis was further employed to interpret the physical driving mechanisms of these predictors. Validation against airborne Lidar confirmed that the dataset accurately recovers elevations in previously void zones with an RMSE of 0.94 m. By integrating these predictions with the DeltaDTM baseline, the final national-scale product achieves robust accuracy with an R2 of 0.75 and an RMSE of 1.17 m. This targeted integration significantly expanded valid topographic coverage by 145.8 % from 79.9 to 196.5 km2. It delivers the first spatially continuous and full-coverage beach topography dataset for New Zealand. Given the global availability of ICESat-2 and Sentinel-2, NZ-BeachTopo30 offers a scalable solution for worldwide applications and provides a robust foundation for inundation modeling and coastal management.

Using fiducial markers for visual localization provides a consistent solution for underwater positioning. However, image degradation caused by water turbidity often leads to the failure of traditional, hand-crafted detection methods. To address this challenge, we introduce DeepTurbid, a system for underwater marker detection and pose estimation. Building on existing ArUco marker systems, DeepTurbid defines the mapping of marker control points in challenging underwater environments and employs a high-resolution neural network for keypoint prediction and marker ID decoding. Moreover, by leveraging the underwater imaging model and the optical properties of underwater scenes, we propose an underwater marker image generation scheme and an adaptive heatmap labeling. This approach generates a diverse marker image dataset, spanning multiple water types and degradation levels for network training. We evaluate DeepTurbid in challenging real-world turbid underwater environments, and experimental results demonstrate that our method significantly outperforms existing approaches in terms of marker detection robustness and pose estimation accuracy. The code and dataset are publicly available at https://github.com/fanyi-meng/DeepTurbid.

Feature matching dominates the time costs in structure from motion (SfM). The primary contribution of this study is a GPU data schedule algorithm for efficient feature matching of Unmanned aerial vehicle (UAV) images. The core idea is to divide the whole dataset into blocks based on matrix band reduction (MBR) and achieve efficient feature matching via GPU-accelerated cascade hashing. First, match pairs are selected by using an image retrieval technique, which converts images into global descriptors and searches high-dimension nearest neighbors with graph indexing. Second, compact image blocks are iteratively generated from a MBR-based data schedule strategy, which exploits image connections to generate image blocks and increase the usage of GPU computing power. Third, guided by the generated image blocks, feature matching is executed sequentially within the framework of GPU-accelerated cascade hashing, and initial candidate matches are refined by combining a local geometric constraint and RANSAC-based global verification. For further performance improvement, these two steps are designed to execute in parallel in GPU and CPU. Finally, the performance of the proposed solution is evaluated by using large-scale UAV datasets. The results demonstrate that it increases the efficiency of feature matching with speedup ratios ranging from 77.0 to 100.0 compared with KD-Tree based matching methods due to its high usage of GPU computing power. Besides, it achieves comparable accuracy in both relative and absolute bundle adjustment (BA). The proposed algorithm is an efficient solution for feature matching of large-scale UAV images.

Traffic forecasting is a critical task in intelligent transportation systems, requiring accurate modeling of spatiotemporal dependencies among traffic sensors. Traditional deep-learning methods face two key challenges: (1) decoupled spatial-temporal pipelines that process and fuse spatial and temporal dimensions separately fail to capture their intricate interdependencies; and (2) state-of-the-art (SOTA) models relying on Transformer architectures often struggle to balance computational efficiency with representational capacity. To address these limitations, we propose ST-Camba, a novel decoupled-free spatiotemporal graph fusion state space model that unifies spatial and temporal dimensions within a single framework. ST-Camba is the first to integrate a spatial dimension axis into state space equations, enabling effective coupled spatiotemporal modeling through graph convolutions while inheriting the linear complexity advantage of Mamba series models. Additionally, we design an Adaptive Spatial Structure (ASS) Injector and a Lerp-based Gated Unit (LGU) to facilitate adaptive spatial structure capture and control information flow in spatiotemporal modeling. Extensive experiments on flow and speed prediction tasks across standard datasets demonstrate ST-Camba's superiority. Specifically, on the PEMS07 dataset, our model achieves a 1.8% reduction in MAE compared to other baselines, while reducing computational costs by up to 14.5%. This work underscores the necessity of coupled spatiotemporal modeling and provides a theoretical foundation for scalable solutions in urban traffic systems.