[Objectives] Driven by the rapid advancement of next-generation artificial intelligence, the Geographical Knowledge Graph (GeoKG), as a structured knowledge representation technique that inherently integrates multi-dimensional spatiotemporal geographic features, has emerged as a core driver for advancing geographic cognitive intelligence and promoting digital and intelligent transformation across various industrial fields. Capable of formalizing implicit geographic rules and unifying scattered, fragmented geographic data resources, GeoKG possesses excellent scalability and prominent practical value in interdisciplinary geographic research and engineering service scenarios. Given the scattered and unsystematic status of existing applied research, systematic collation of diversified GeoKG practical applications, in-depth analysis of typical application scenarios, generalization of mainstream technical implementation pathways, and comprehensive clarification of prevailing research deficiencies and practical challenges can effectively fill the gaps in current review studies, providing solid theoretical foundations and practical guidance for the in-depth theoretical innovation and large-scale scalable application of GeoKG. [Analysis] This paper systematically analyzes the core functional attributes and multi-scenario application systems of GeoKG for geographic intelligent services. It first introduces the concept of GeoKG and related research on its applications. Then, it elaborates the unique functional advantages of GeoKG in spatial relationship reasoning, geographic semantic retrieval and complex geographic scene cognition, fully demonstrating its indispensable role in optimizing and enhancing the basic spatial cognition and perception capabilities of mainstream geographic intelligent models. Focusing on typical intelligent service scenarios in surveying, mapping and geographic information disciplines, this paper summarizes representative technical implementation paths of GeoKG in spatiotemporal location service provision, multi-scale GIS spatial analysis, automated intelligent mapping, and intelligent remote sensing interpretation, and verifies its crucial supporting effect on transforming traditional data-centric geographic information services into high-value knowledge-driven intelligent services. Furthermore, this paper explores the practical application performance of GeoKG in public security emergency management, ecological resource and environmental governance, as well as refined urban operation and comprehensive governance, and confirms its outstanding comprehensive advantages in providing reliable and professional knowledge support for industrial intelligent decision-making and high-precision public service delivery. [Prospect] Combining the cumulative progress of existing academic research and the continuously growing practical demands of industrial intelligent construction, this paper further clarifies multiple frontier development directions for the in-depth expansion and iterative upgrading of GeoKG application research. Future exploratory studies should continuously focus on three key research orientations, including dynamic iterative updating and scalable expansion of geographic knowledge systems, deep fusion and unified modeling of multi-source heterogeneous spatiotemporal data, and adaptive collaborative capability enhancement between GeoKG and emerging Large Language Models (LLMs). This work provides valuable academic references and feasible innovative insights for follow-up theoretical exploration, technical optimization and large-scale industrial implementation of geographical knowledge graphs.
[Background] The Beijing-Tianjin-Hebei region is facing severe water scarcity and an increasing imbalance between water supply and demand. Groundwater is an important water source supporting regional socioeconomic development, agricultural production, ecological security, and urban water use. Its long-term variation is closely related to regional water security and sustainable development. Therefore, investigating the future variation of groundwater resources in the Beijing-Tianjin-Hebei region is of great scientific significance and practical value. [Methodology] This study proposes a deep learning model with a geo-spatiotemporal coordination mechanism, namely the GCN-Transformer-BiLSTM model, which integrates the spatial feature extraction capability of the Graph Convolutional Network (GCN), the global temporal dependency modeling capability of the Transformer, and the local temporal dynamic modeling capability of the Bi-directional Long Short-Term Memory (BiLSTM). Based on this model, groundwater storage change simulation and prediction were conducted. Groundwater Storage Anomaly (GWSA) was selected as the indicator representing regional groundwater storage variations. Taking the Beijing-Tianjin-Hebei Plain as the study area, the model was constructed using the GWSA dataset and major influencing factor datasets from 2002 to 2024, and was further applied to multi-scenario prediction of GWSA for the period from 2025 to 2035. [Results] Experiments have shown that, compared to the depletion observed over the past 20 years, future groundwater reserves in the Beijing-Tianjin-Hebei Plain exhibit scenario-dependent and relatively complex spatiotemporal dynamics. Compared with the groundwater depletion pattern over the past two decades, groundwater storage under different scenarios generally shows a gradual alleviation of deficit, but the recovery magnitude varies among scenarios. Specifically, the SSP1 scenario shows relatively stable recovery, the SSP2 scenario presents moderate recovery, while the SSP5 scenario has larger fluctuations and relatively weaker recovery. From the perspective of spatial distribution, groundwater storage changes under different scenarios generally show a pattern of more serious deficits in the southern region and relatively lower deficits in the northern region. Under the high-emission scenario, regional differences in groundwater storage changes become more pronounced. [Conclusions] The deep learning model proposed in this paper, which incorporates a spatio-temporal coordination mechanism, can to some extent account for the spatial autocorrelation and heterogeneity of groundwater reserves, as well as their temporal dynamics. It realizes the simulation and multi-scenario prediction of groundwater storage changes in the study area, and provides a scientific reference for regional groundwater resource regulation, groundwater protection, and sustainable water resource management. In addition, the results demonstrate the effectiveness of integrating spatial and temporal information for groundwater modeling, and highlight the potential of deep learning methods in supporting decision-making under different development scenarios.
[Objective] Aiming at the key issues in hyperspectral remote sensing image classification, such as the insufficient generalization ability of models under small-sample conditions, the uninterpretable decision-making process caused by the "black box" nature of deep learning, and the weak noise immunity and reliance on statistical assumptions of traditional algorithms, this study aims to explore a novel classification method that combines physical mechanism-driven approaches with engineering practicality to break through the limitations of existing methods. [Methods] This paper proposes a land cover classification algorithm for hyperspectral remote sensing images based on Spectral Fingerprint Iterative Matching (SFIMC). First, redundant bands are eliminated through band correlation analysis. Second, a five-level discretization method is adopted to transform spectral responses into integer-coded "digital fingerprints". Then, a labeled "Standard Spectral Fingerprint Template Library" is constructed based on training samples, and bitwise matching is performed using the Hamming distance. Finally, a multi-level progressive threshold mechanism is designed to balance accuracy and coverage through progressive matching from strict to loose, thereby enhancing the identification capability of mixed pixels. [Results] Experiments based on EO-1/Hyperion imagery of the urban-rural ecotone in the study area demonstrate that the proposed method achieves an Overall Accuracy (OA) of 88.82% and a Kappa coefficient of 0.85, with significant advantages in Producer's Accuracy (PA) and User's Accuracy (UA). Compared with commonly used supervised classification algorithms such as Support Vector Machine (SVM: 63.58%, 0.45), Maximum Likelihood Classification (MLC: 70.59%, 0.60), Spectral Angle Mapper (SAM: 47.06%, 0.28), and Spectral Information Divergence (SID: 35.29%, 0.15), the OA is improved by 25.24, 18.23, 41.76, and 53.53 percentage points, respectively, while the Kappa coefficients are increased by 0.40, 0.25, 0.57, and 0.70, respectively. This indicates superior performance over traditional methods. Under the constraint of 30m spatial resolution, the classified patches exhibit better spatial continuity than the comparative methods, and the capability to identify spectral transition zones at class boundaries is enhanced. This effectively alleviates the common "salt-and-pepper" noise and spatial fragmentation issues found in traditional methods, demonstrating excellent adaptability particularly in mixed-pixel intensive areas where woodlands and cultivated lands interlace. Furthermore, comparative experiments under small-sample conditions show that when only 10 training samples per land-cover type are used, the proposed method still achieves an OA of 71.18% and a Kappa coefficient of 0.63. These figures represent improvements of 25.89, 4.12, 17.65, and 34.12 percentage points in OA, and increases of 0.38, 0.08, 0.26, and 0.47 in Kappa coefficients compared to SVM (45.29%, 0.25), MLC (67.06%, 0.55), SAM (53.53%, 0.37), and SID (37.06%, 0.16), respectively. The results validate its applicability and robustness in small-sample scenarios. [Conclusions] This paper constructs a new interpretation path from "analog spectra" to "digital fingerprints". The classification basis of this method is the high similarity of spectral fingerprints of the same type of land cover. Each step of the decision-making process, from band optimization to Hamming distance matching, has clear mathematical definitions and physical meanings, and the classification results can quantify confidence through a multi-level threshold mechanism. In addition, the method requires no complex statistical distribution assumptions, features a simple process and high computational efficiency, improves classification accuracy and robustness to a certain extent, and endows the model with clear physical interpretability. This provides a lightweight, transparent, and engineering-friendly new solution for intelligent hyperspectral remote sensing interpretation, showing good applicability in small-sample and emergency monitoring scenarios.
[Objectives] To address the formidable challenge of effectively identifying and enhancing the extremely weak spectral features of alteration minerals in Gaofen-5 (GF-5) hyperspectral imagery under complicated and strong noise backgrounds, this study proposes a comprehensive joint processing framework. Satellite-borne sensors inevitably suffer from atmospheric attenuation, thermal effects, and push-broom imaging mechanisms, which introduce severe mixed noises that easily submerge the narrow and shallow diagnostic absorption valleys in the shortwave infrared (SWIR) region. Consequently, we develop a framework adopting a physical mechanism of spatial-spectral dual-domain decoupling, with "denoising for fidelity and nonlinear enhancement" as its fundamental core. [Methods] First, an advanced spatial-spectral total variation regularized tensor robust principal component analysis (SSTV-TRPCA) model under computational relaxation is employed. By imposing smoothing constraints exclusively on the spatial dimension while protecting the spectral dimension from over-smoothing, this model meticulously decomposes the hyperspectral image tensor into a low-rank background component and a sparse anomaly component. This critical process effectively removes global mixed noise while intentionally preserving weak mineralization information and residual structural noise within the purified sparse residual space. Moreover, a computational relaxation strategy utilizing an isotropic Gaussian kernel convolution is introduced to approximate the proximal mapping of the total variation, significantly reducing the computational complexity of the non-convex optimization and enabling processing of large-scale datasets. Subsequently, a fractional-order Teager-Kaiser energy operator (F-TKEO) is constructed to exploit the distinct physical morphological differences between spectral and spatial features. By generating generalized binomial weights based on the Grünwald-Letnikov definition, this operator introduces a cross-correlation mechanism between adjacent fractional orders. It nonlinearly amplifies the targeted weak absorption features within the sparse space and adaptively cancels out non-correlated high-frequency pseudo-anomaly interferences, enabling the precise localization of mineral endmembers and the high-fidelity enhancement of weak spectral signatures. [Results] Extensive quantitative simulation experiments validate the superiority of the proposed framework. Under a 15 dB extreme strong noise environment, the SSTV-TRPCA model successfully isolates the background, achieving an impressive reconstruction signal-to-noise ratio (SNR) of 40.1 dB. This performance yields a significant improvement of 10.4 dB over the current best baseline method, Global Low-Rank Total Variation (Global LRTV, 29.7 dB), and substantially outperforms Standard TRPCA (19.3 dB) and one-dimensional Discrete Wavelet Transform (DWT, 25.2 dB). This effectively avoids the pervasive over-smoothing of weak diagnostic features typically caused by aggressive global denoising. Furthermore, the proposed F-TKEO operator demonstrates exceptional nonlinear amplification capabilities. For the kaolinite doublet absorption feature and the weak secondary diagnostic features of various typical minerals, the operator achieves a substantially improved contrast-to-noise ratio (CNR) of 2.73. This represents absolute enhancements of 1.49 and 1.77 compared to the traditional integer-order Teager-Kaiser Energy Operator (Traditional TKEO, CNR=1.24) and standard Fractional-Order Differentiation (FOD, CNR=0.96), respectively, significantly outperforming traditional methods in nonlinear feature amplification. [Conclusions] In the practical application to real-world hyperspectral data acquired from the Beiya gold deposit, the proposed methodology successfully identifies the critical sodium-rich sericite absorption feature located at 2 192.5 nm. This specific feature exhibits a subtle wavelength shift compared to standard spectral positions, indicating the presence of hydrothermal alteration fluid metasomatism. This spatial-spectral dual-domain decoupling framework effectively addresses the critical bottleneck of extracting deeply hidden weak spectral features from low-SNR satellite observation data, offering a novel, reliable technical approach for hyperspectral geological exploration and large-scale mineral mapping.
[Objectives] Urban green spaces constitute a vital component of urban ecosystems. Rapid and accurate acquisition of their spatial distribution is therefore essential for assessing urban ecological quality and guiding sustainable development planning. However, current extraction approaches often depend heavily on large quantities of manually annotated samples and show limited generalization capacity. Due to the scarcity of labeled datasets and the lack of specialized algorithms designed explicitly for complex urban environments, it remains a significant challenge to achieve fine-scale, large-area extraction of such areas without human intervention. [Methods] To overcome these limitations, this paper proposes an unsupervised Open-Vocabulary Semantic Segmentation (OVSS) framework that synergizes vegetation spectral priors with the CLIP (Contrastive Language-Image Pre-training) cross-modal model for the first time, aiming to achieve annotation-free, high-precision, and transferable extraction of urban green spaces. The proposed method combines cross-modal semantic understanding with remote-sensing spectral information to construct a comprehensive framework that bridges semantic reasoning and pixel-level recognition. Specifically, the workflow consists of three primary stages: First, a large language model is utilized to generate natural‑language prompts that both encode vegetation near‑infrared spectral signatures and align with human semantic perceptions of greenery. Second, leveraging the dual‑encoder architecture of the CLIP model, false-color multispectral remote sensing images are aligned with these text prompts at the pixel level to produce a coarse semantic segmentation mask of urban green spaces. Third, building upon the OVSS‑derived coarse mask, the Distribution of the Normalized Difference Vegetation Index (NDVI) is statistically analyzed within the masked regions, and the Youden Index is subsequently applied to determine an optimal adaptive threshold. This threshold is then used to refine the initial results, yielding a high-resolution and detailed green space extraction map. [Results] Wuhan City serves as the study area for empirical evaluation, using Sentinel-2 multispectral satellite imagery. Experimental results demonstrate that the proposed framework achieves a precision of 85.21%, a recall of 95.67%, an F1-score of 90.15%, and an Overall Accuracy (OA) of 88.40%. Compared with traditional threshold-based segmentation results derived from various vegetation indices, the proposed method exhibits a remarkable improvement in overall performance. Furthermore, it is particularly superior in delineating precise boundary details and recognizing small-scale, fragmented green patches when evaluated against mainstream global land cover datasets, demonstrating stronger robustness in diverse and complex urban landscapes. [Conclusions] This study verifies that the proposed unsupervised OVSS mask-driven framework enables accurate and large-scale identification of urban green spaces without any manual annotation. It thus provides a novel and intelligent technical pathway for wide-area urban green space monitoring and offers a valuable reference for related research in ecological assessment and sustainable urban development.
[Background] Discrete Global Grid Systems employ specific recursive subdivision schemes to partition the Earth’s surface into multi-resolution hierarchical structures. Among various grid types, equal-area hexagonal grids possess preferable characteristics because they offer global uniformity and identical cell areas at each resolution level, which makes them highly suitable for organization, management, and analysis of geographic location-based big data. Nevertheless, due to the lack of efficient global rendering acceleration techniques for spherical equal-area hexagonal grids, existing mainstream visualization methods that rely on geometric entity construction and rendering are hard to satisfy the interactive performance requirements of massive multi-resolution grid visualization. [Objective] To address this limitation, this study exploits the parallel computing capability of GPU fragment shaders to batch-render spherical equal-area hexagonal cells. Instead of constructing a large quantity of geometric entities, the proposed method enables efficient visualization of global hexagonal grids at the hundred-million-cell scale. [Methods] First, based on the weak duality relationship between hexagons and rhombuses on polyhedral surfaces, a tile-based partitioning strategy and spherical triangulation scheme are proposed. In this design, hexagonal cells are aggregated into rhombic tiles, which serve as parallel rendering batches. This organization strategy effectively balances CPU-GPU workload distribution and substantially reduces the amount of vertex data transmitted to the graphics pipeline. Second, to replace explicit geometry rendering, a fragment-shader-based texture mapping method obeying equal-area projection is developed. Via Leeuwen equal-area projection, the method establishes an analytical mapping from continuous texture coordinates to discrete cell indices. In this way, texture sampling is adopted instead of geometric construction, allowing hexagonal cells to be rendered fully in parallel within the fragment shader. Third, to address the precision degradation in high level grids caused by single-precision floating-point truncation, a two-stage precision optimization strategy is introduced, thereby extending usable precision and effectively eliminating texture offsets at higher resolution levels. [Results] Experimental results demonstrate that the proposed method is only marginally affected by viewpoint distance and maintains stable rendering performance even when visualizing hundred-million-scale grid cells. Compared with state-of-the-art methods, which achieve a frame rate of approximately 20 fps, the proposed approach increases rendering performance to about 85 fps. In addition, the method does not depend on high-performance graphics hardware and remains stable across a wide range of devices. These results indicate that the proposed method has strong portability and practical applicability. [Conclusions] This paper presents a novel fragment-shader-based texture mapping approach for the efficient visualization of massive spherical equal-area hexagonal grids. The proposed method is geometry-free, computationally efficient, and straightforward to implement. It provides an effective technical foundation for the practical deployment of DGGSs in lightweight web-based platforms and other resource-constrained environments.
[Objectives] Landslides represent one of the most destructive geological hazards globally, particularly in high-density mountainous cities where extreme rainfall events and rapid urbanization intersect. Accurate landslide susceptibility assessment is the cornerstone of disaster risk reduction. However, traditional data-driven models, such as conventional machine learning algorithms, often rely on the assumption of independent samples, thereby generating static susceptibility maps. These approaches inherently fail to capture the complex spatiotemporal autocorrelation of geological environments and cannot adequately respond to the non-stationary nature of future climate change. Consequently, there is an urgent need for a long-term, dynamic evaluation framework. [Methods] To bridge this gap, this study introduces an advanced Bayesian spatiotemporal statistical model based on the Integrated Nested Laplace Approximation (INLA) algorithm. Using the Hong Kong Special Administrative Region as a typical high-density urban study area, we constructed a comprehensive evaluation indicator system incorporating topographic, geological, meteorological, and vegetation factors. The model explicitly incorporates spatial and temporal random effects into a generalized linear mixed model framework. Specifically, an intrinsic Conditional Autoregressive (iCAR) specification was utilized to smooth local spatial heterogeneity, while a first-order autoregressive (AR1) process was employed to extract the evolutionary trends in the temporal dimension. Based on the Enhanced Natural Terrain Landslide Inventory (ENTLI), we precisely reconstructed the spatiotemporal evolutionary history of landslide susceptibility in Hong Kong from 2010 to 2021. Furthermore, by integrating downscaled multi-scenario climate data from the Coupled Model Intercomparison Project Phase 6 (CMIP6) and high-resolution global future population grid datasets, we dynamically projected the regional landslide risk patterns up to the year 2100 under three Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, and SSP5-8.5). [Results] The results demonstrate that the Bayesian spatiotemporal model exhibits excellent predictive performance and generalization capability. Posterior inference confirms that slope steepness and extreme intense rainfall (annual maximum daily precipitation) serve as the primary static and dynamic driving factors governing the spatiotemporal pattern of regional landslides, with regression coefficients of 0.610 and 0.579, respectively. Crucially, by extracting annual temporal intercepts, the model successfully uncovers a significant non-linear fluctuating pattern characterized by “high-low alternation” at the interannual scale. This dynamic temporal component objectively reflects the stage-wise lagged adjustment mechanism of the slope system in response to extreme climate forcing. For future multi-scenario projections, the spatial pattern of extremely high susceptibility areas remains relatively stable under topographic constraints. However, driven by intensified future precipitation, the spatial extent of medium-susceptibility areas will experience significant expansion. Validation of these projections using recently added independent samples shows that the spatial prediction accuracy (AUC) remains above 0.78 across all scenarios, confirming the robustness of the long-term projections. The amplification effect on future geological disaster risks is particularly pronounced under the high-emission scenario (SSP5-8.5), which sustains non-linear decadal fluctuations. Moreover, bivariate spatial coupling analysis of landslide susceptibility and population distribution reveals a severe composite risk. Traditional high-density built-up environments, such as the Mid-Levels of Hong Kong Island and the northern foothills of Kowloon, will continuously face the dual challenge of “high susceptibility and high exposure”. [Conclusions] In conclusion, compared with the limitations of traditional static models in characterizing temporal non-stationarity, the fully integrated spatiotemporal dynamic projection framework developed in this study can integrate climate dynamic evolution and spatial effects to effectively quantify long-term composite risks. Ultimately, this research provides crucial decision-making support and a reliable methodological foundation for high-density cities to formulate climate-resilient spatial planning and implement refined disaster management strategies.
[Objective] Urban waterlogging emergency scenes represent the outcome of simulating complex urban waterlogging emergency systems through digital means, serving as a vital tool for users to perceive and comprehend urban waterlogging risks. However, given the diverse user types involved in urban waterlogging emergency scenarios, most simulations neglect user concerns analysis. This results in generated scenes that struggle to convey waterlogging risks to different user groups effectively. [Methodology] To address these shortcomings, this study achieves user concerns modelling for urban waterlogging emergency scenes by exploring the intricate relationship between user concerns information and scene elements. Specifically, based on existing urban waterlogging and emergency response procedures, we analyse users' social attributes and emergency response requirements. Combining this with the occupational structure of the urban population and the urban waterlogging emergency management system, we categorise user types within urban waterlogging emergency scenes. Then, the LDA, Word2Vec, and PageRank methods were employed to extract concern information matching different types of emergency users from user-generated data sources, such as social media posts and emergency plans. Subsequently, drawing upon geographical scene elements, the constituent elements of the urban waterlogging emergency scene in different phases were proposed. Building upon this foundation, ontology modelling and its fusion methods were utilised to establish the concerns model for different types of emergency users within urban waterlogging emergency scenes. [Results] Drawing upon the 2021 Zhengzhou, Henan rainstorm waterlogging incident, this study established a detailed user classification system encompassing emergency decision-makers, rescuers, and the general public. Through text mining of user-generated content related to urban waterlogging emergencies, information on concerns for different types of emergency users in waterlogging emergency scenes was extracted. Ontology methods were then employed to achieve semantic mapping between emergency users' concerns and the constituent elements of different phases of the scene. The effectiveness of the constructed user concern model was evaluated through a combination of prototype system development and questionnaire surveys. The model demonstrated an accuracy of 0.923, completeness of 0.945, and satisfaction level of 0.879. Calculated using the proposed effectiveness formula, the model achieved an effectiveness of 91.7%, and the scene adapted in this paper has reduced the average response time by 41.0% and increased the task success rate by 21.5% among the general public with the strongest demand for information filtering. [Conclusions] The results demonstrate the model's high reliability. Therefore, integrating concerns information from different types of emergency users regarding scene elements during urban waterlogging emergency scene simulations can effectively enhance the user universality, thereby enhancing the awareness and understanding of waterlogging risks among different types of emergency users.
[Objectives] Within high-density multi-port trade bays, the overlap of ship trajectories and the coexistence of multiple crude oil unloading methods(e.g., near-shore berths within ports, ship-to-ship offshore transfers, and offshore pipeline transportation) create challenges for accurately calculating maritime crude oil imports at specific ports, particularly in identifying the precise moments and locations where unloading activities occur. [Methods] To address this issue in maritime crude oil transportation scenarios, we first used the Geohash geocoding algorithm to generalize the spatial information of AIS ship trajectories and extract ship-stopping characteristics. We then proposed a destination-port text-matching rule and a destination-port bidirectional search method to assign destination port attributes to dispersed ship-stopping areas around the port. Finally, following the principles of obstacle avoidance and proximity integration, we consolidated these stopping areas into crude oil tanker unloading zones, ensuring the sematic reliability of unloading area extraction. The resulting unloading zones effectively capture the diversity of crude oil unloading methods and their spatial relationships with coastal port facilities. Under the premise of unloading identification criteria consistent with real-world transportation scenarios, we compared several modified shipload calculation methods based on AIS data and validated them at major crude oil import ports in Japan. [Results] Among the tested method, the approach based on load condition transitions performed best in estimating maritime crude oil imports. Its results were closest to official statistics, with a coefficient of determination (R2) of 0.95 and the smallest root mean square error (RMSE). Compared with four alternative calculation methods (draft change ratio by ship size, draft change ratio by individual ship, tonnes per centimeter immersion, and ship load adjustment considering ballast water), the RMSE was reduced by 16.41%, 22.90%, 60.06%, and 9.16%, respectively. When compared with common clustering algorithms, such as K-means and DBSCAN, our proposed method improved the reliability of port allocation by 0.12 and 0.21, respectively, when extracting crude oil unloading areas within Ise Bay. It also demonstrated greater stability in import estimation accuracy, with the standard deviation of relative errors reduced by 0.80 and 0.56, respectively. [Conclusions] The proposed technical framework leverages the complementary advantages of spatiotemporal and textual information in AIS big data to achieve precise characterization of crude oil tanker unloading behaviors and annual monitoring of maritime crude oil imports within high-density multi-port trade bays. This approach enables real-time port trade monitoring and visualization of ship movement patterns using continuously updated AIS spatiotemporal big data.
[Objectives] Previous review studies on digitalization for culture heritage have largely summarized existing work from single perspectives-such as data acquisition, data processing, or visualization—they often fail to comprehensively identify current challenges or provide systematic recommendations for future development. [Methods] To address this gap, this paper adopts a multi-dimensional analytical framework, examining research of 3D digitalization for culture heritage research from five perspectives: classification of heritage objects, data acquisition methods, data processing workflows, 3D modeling approaches, and 3D model representation techniques. [Conclusions] Based on a systematic analysis of 10 024 scientific publications on 3D digitalization of cultural heritage published worldwide sinnce 1982 past three decades, the results depict that visualization remains the primary objective of the majority of 3D digitalization projects. However, in recent years, the integration of semantic and attribute information has emerged as an increasingly critical demand within the field. Accordingly, future research is expected to focus on the definition and design of standardized 3D data models and data formats to support semantic modeling, enabling 3D digital models not only to be visualized but also to facilitate computation, querying, selection, and interoperability. In addition, artificial intelligence and large-scale models are anticipated to play an increasingly significant role across all stages of the 3D digitalization workflow.
[Objectives] In the digital preservation of cultural heritage, acquiring 3D models that simultaneously possess high geometric accuracy and visual fidelity remains a significant challenge. Traditional scanning technologies often suffer from inherent limitations in scan data, such as noise, outliers, and uneven point density, while the emerging 3D Gaussian Splatting (3DGS) technique frequently exhibits insufficient geometric accuracy due to a lack of robust geometric constraints. To address these issues, this paper proposes a novel multi-modal fusion 3D Gaussian Splatting method designed to achieve high-fidelity dense reconstruction of curatorial artifacts. [Methods] The methodology begins with a rigorous data alignment process, where a sparse yet geometrically significant scanned point cloud is constructed using a proposed feature-aware sampling strategy. This sparse point cloud serves as a high-quality geometric prior to initialize and guide the optimization process of 3DGS. This fusion mechanism is pivotal; it not only enables the model to learn fine-grained appearance information from high-resolution multi-view images but also effectively constrains and regularizes the geometric structure using the reliable scan-derived prior, thereby mitigating scale drift and geometric distortion. Furthermore, unlike standard 3DGS which yields only a radiance field for rendering, this study innovatively leverages the continuous scene representation learned by the fusion model to explicitly extract geometry. A probabilistic sampling strategy is introduced to densify, detail-enhance, and denoise the input discrete geometry, transforming the neural representation into a superior dense point cloud. [Results] Extensive experimental results on a dataset of representative artifacts demonstrate that the proposed method exhibits significant advantages over baseline methods. Quantitatively, the generated models achieve higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) in rendering, while the final dense point clouds show lower Chamfer Distance and Hausdorff Distance compared to ground truth, indicating sub-millimeter geometric precision. [Conclusions] This research provides an effective solution for the high-precision dense reconstruction of cultural artifacts, validates the application potential of neural rendering models in generating high-quality geometric assets, and offers a new approach for creating digital heritage models that surpass the quality of single-source data, suitable for applications ranging from digital archiving to virtual restoration.
[Objectives] Large-scale cliff carvings are massive in dimension and geometrically complex in surface structure. The 3D reconstruction of large-scale cliff carvings through close-range photogrammetry is therefore characterized by large polygon counts and irregular mesh topology, making it difficult to implement standardized UV unwrapping and texture transfer. Ultimately, it hinders the real-time rendering and interactive visualization of the models in WebGL environments. To this end, this study proposes an integrated method for the 3D reconstruction and lightweight processing of large-scale cliff carvings specifically designed for WebGL display. [Methods] Three processing stages are established. In the data acquisition stage, following a comprehensive comparison of the applicability of close-range photogrammetry and Nap-of-the-Object photogrammetry, a stratified and zonal aerial image acquisition strategy using consumer-grade Unmanned Aerial Vehicles (UAVs) is proposed, targeting the complex mountainous environment characterized by large cliff carvings occluded by dense vegetation. In the reconstruction stage, a photogrammetric pipeline integrating Structure from Motion (SfM) and Multi-View Stereo (MVS) is employed to generate a high-precision model. In the lightweight processing stage, asset production standards from the digital game and real-time rendering industries are introduced, where topological reconstruction is the prerequisite, and standardized UV unwrapping and multi-channel Physically Based Rendering (PBR) texture baking are performed to systematically transfer geometric structure and surface detail from the high-precision model to a low-polygon counterpart, yielding a lightweight model ready for real-time WebGL rendering. [Results] Taking the Qishan Double Buddhas (cliff carvings of the Tang Dynasty with an overall height exceeding 30 meters, located in Meishan, Sichuan, China) as the experimental subject, the study compresses their original high-precision model, which contains approximately 120 million polygons, to a WebGL-compatible low-polygon model of approximately 500 000 polygons, where the surface details from the high-precision model are baked into texture maps and applied to the low-polygon model, ensuring that rich visual fidelity is retained after substantial geometric simplification. The digital micro-exhibition titled “Bashu Buddhism: Sichuan Grottoes in the Digital Era,” designed and implemented with this model, achieved smooth web-based browsing and interaction. The exhibition was also selected by the Ministry of Education of China in the joint exhibition of the “Promoting Fine Traditional Chinese Culture” initiative (university group) and “Inheriting Cultural Heritage” digital micro-exhibition in 2024, providing practical validation of the proposed approach. [Conclusions] The proposed method consolidates the photogrammetric reconstruction pipeline and WebGL-oriented lightweight processing into an integrated workflow, substantially reducing modeling complexity while preserving visual fidelity in the conversion of a high-precision reconstruction model into standardized digital assets that are real-time-renderable and WebGL-compatible. It resolves the tension between reconstruction accuracy and rendering efficiency in the 3D digital reconstruction of large-scale cultural relics, providing a technical reference for the digital preservation and online dissemination of immovable cultural relics.
[Objectives] In the digital preservation and virtual reconstruction of ancient buildings, integrating air-ground heterogeneous point clouds is vital yet highly challenging. Fusion methods relying on artificial targets are labor-intensive, whereas traditional target-free registration algorithms are prone to severe structural ghosting when handling scenes with sharp curvature transitions and large flat walls. Additionally, the highly symmetrical layouts of ancient buildings often trap algorithms in local optima. To overcome these limitations, this paper proposes a robust registration strategy based on hierarchical optimization Generalized Iterative Closest Point (GICP) to improve both the accuracy and robustness of multi-source heterogeneous point cloud fusion. [Methods] The proposed methodology begins with a multi-scale voxel downsampling and Statistical Outlier Removal (SOR) filtering pipeline. This step effectively mitigates the uneven point density and high-frequency noise associated with UAV photogrammetry while balancing computational efficiency and preserving critical geometric details. In the coarse registration phase, Fast Point Feature Histograms (FPFH) descriptors are extracted and coupled with the Sample Consensus Initial Alignment (SAC-IA) algorithm. By encoding multi-dimensional topological geometric signatures, this phase establishes reliable initial alignments and avoids the local extrema caused by bilateral structural symmetries. In the fine registration phase, a hierarchical GICP algorithm with a "wide-to-narrow" dynamically shrinking thresholding strategy is introduced. During the initial iterations, a wide distance threshold utilizes the abundant and geometrically reliable flat walls to drive rapid global alignment. In the later stages, the threshold is strictly tightened to narrow limits, which mathematically suppresses the influence of flat regions and dampens noisy edge mismatches. This shifts the optimization focus to local high-curvature structures, allowing the Gaussian probability model to adaptively fit intricate geometries, such as the swallowtail ridges. [Results] Extensive experiments were conducted using air-ground joint mapping datasets from two representative ancient buildings in Southern Fujian: the Nanyue Ancestral Temple (Case 1) and Weide Hall (Case 2). Quantitative and qualitative evaluations demonstrate the superior performance of the proposed method. In Case 1, the proposed strategy achieved a Root Mean Square Error (RMSE) of 6.3 mm, while reducing the processing time by approximately 95% (from 725.573 s to 31.634 s) compared to the single-scale baseline, and effectively eliminated the "double-eyelid" ghosting artifacts at complex swallowtail ridges. The absolute accuracy was further verified using homonymous feature points, yielding a mean point-to-point error of 7.5 mm. In Case 2, the proposed algorithm reached a sub-millimeter registration accuracy with an RMSE of 0.9 mm, representing a significant accuracy breakthrough over standard GICP and point-to-plane ICP baselines. [Conclusions] This research successfully addresses the core conflicts of macro-micro scale disparities and global structural symmetry in heritage digitization. The proposed hierarchical optimization strategy significantly enhances the quality of air-ground multi-source point cloud integration, providing a high-fidelity and geometrically precise digital asset for ancient building restoration, structural analysis, and Historical Building Information Modeling (HBIM) applications with strong engineering scalability.
[Objectives] Digital protection of traditional Chinese architecture is an important means of preserving architectural cultural heritage. Digital reconstruction forms a key component of this process, and image restoration of traditional Chinese building is central to achieving accurate reconstruction. Most existing image restoration methods rely on manually locating the restoration areas, which is not only time-consuming and labor-intensive but also difficult to ensure the accuracy. In addition, current restoration models lack specificity, and there is an urgent need for a systematic image restoration method tailed to traditional Chinese architecture. [Methods] To address these issues, this study proposes a two-stage automatic image restoration network, DLFT. In the first stage, the DeepLabv3+ algorithm is used to automatically segment foreground occlusion regions, and the extracted results are used as mask inputs in the second stage. In the second stage, Fast Fourier Convolution (FFC) is embedded into a Transformer-based restoration network, and a compound perceptual loss function is designed in the decoding stage to further optimize performance. It should be noted that the restoration tasks in this paper include two aspects: (1) image restoration of building facades and (2) remote sensing image restoration of building roofs. [Results] Comparative experiments conducted on two datasets, building facades and architectural remote sensing images, show that under a mask occlusion ratio of 20%-30%, the proposed method outperforms the second-best model (DWTNet). Specifically, the traditional image quality metrics PSNR and SSIM increased by 0.90% and 0.94% on one dataset, and by 0.34% and 0.23% on the other, respectively. Compared with representative restoration models such as DWTNet and Hourglass, DLFT exhibits superior restoration performance. In ablation experiments, the PSNR increased by 2.2% and 1.8%, respectively, after introducing the fast Fourier convolution; on this basis, it further increased by 0.2% and 0.5% after incorporating the composite perceptual loss function, confirming the crucial role of both modules in enhancing model accuracy. Additionally, for the specific task of traditional architectural style classification, the restored images showed improvements in three metrics, macro-averaged precision, macro-averaged recall, and macro-averaged F1-score. This comprehensive and accurate evaluation of traditional architectural restoration effects further validates the effectiveness of the proposed method in improving the performance of specific tasks. [Conclusions] The proposed DLFT network achieves intelligent and accurate restoration of traditional Chinese architectural images, facilitating deeper understanding and research on the appearance of traditional buildings. It provides higher-quality data sources and scientific data support for village planning, tourism development, cultural inheritance preservation, and three-dimensional reconstruction, offering robust technical support for the digital protection and sustainable development of traditional Chinese architecture.
[Objectives] Historical character is a vital outward expression of a city's cultural heritage and a vessel for collective memory. The primary components of historical character are historical building elements. However, the historic building elements in cities are diverse, complicated and widely distributed. Urban renewal and architectural heritage protection urgently require identifying the distribution of historical building elements and cultural protection areas. However, traditional methods based on manual surveys are both inefficient and subjective. These methods are difficult to capture the exact distribution of historical building elements over a wide area. [Methods] Therefore, this paper proposes a machine learning framework for detecting and classifying traditional historical building elements from street view images, and has successfully extracting and classifying traditional style buildings in the central urban area of Beijing. The main contributions are as follows: (1) combining knowledge of historical architecture with computer vision recognition logic, designing a classification system that includes components and structures; (2) developing a manually annotated dataset with 1 200 images of traditional historical building, which adopts crowdsourced web imagery as the fundamental data source; (3) presenting a Multi-scale Consistency Strategy(MCS) that enhances object detection model capacity across varying scales. [Results] The optimized MCS-Co-DETR achieves an mAP@[0.5:0.95] of 0.554, representing a 3.2% improvement over the original Co-DETR, the model yielded 62 860 detection instances in Beijing's central urban area from street-view imagery. The results reveal that royal architecture are predominantly clustered within the core area inside the Second Ring Road. Independent structures such as steles, pagodas, and sculptures exhibit a dispersed layout, while traditional components of courtyard complexes show strong spatial correlations. High-density areas significantly align with the boundaries of historical and cultural preservation zones. The method also identifies spatial patterns such as the layout of elements along the Beijing Central Axis, demonstrating its ability to accurately capture scattered traditional architectural elements and alleys. In addition, comparative experiments with other mainstream detection models show that the MCS-Co-DETR model proposed in this study has better adaptability to complex urban street view scenes. [Conclusions] This method provides a practical and scalable solution for cities aiming to map and protect large historic areas, demonstrating the potential of street view imagery and deep learning to enrich the recognition and management of historical heritage.Moreover, this study enriches the technical system of urban cultural heritage protection, lays a foundation for subsequent research on intelligent management of historical landscape areas, and has important theoretical significance and practical value for promoting the sustainable development of urban culture and the inheritance of excellent traditional architectural culture.
[Objectives] Digital documentation of architectural heritage is frequently constrained by indoor environments, shadowed regions, and low-light conditions, resulting in insufficient brightness, severe shadow interference, blurred texture details, and color distortion. These degradations significantly affect texture mapping quality, aerial triangulation accuracy, and three-dimensional (3D) reconstruction performance. To address these challenges, this study proposes an effective low-light image enhancement approach specifically designed for architectural heritage documentation and digital surveying applications. [Methods] A Gray-guided Multi-Scale Illumination Attention EnlightenGAN (GI-MSIA-EnlightenGAN) is developed based on the original EnlightenGAN framework. The proposed method introduces a gray-guided illumination encoding branch to explicitly model illumination information and provide reliable global brightness priors. Meanwhile, a multi-scale illumination attention mechanism is incorporated to strengthen the representation of low-light regions and enhance local feature extraction under complex lighting conditions. By integrating gray-guided illumination features with multi-scale attention representations, the network simultaneously improves image brightness while preserving structural boundaries, texture details, and color fidelity, thereby reducing over-enhancement and suppressing artifacts commonly observed in conventional low-light enhancement methods. [Results] Extensive experiments were conducted on both the publicly available LOL dataset and a self-constructed architectural heritage texture dataset collected from Yongding Gate in Beijing. On the LOL dataset, the proposed GI-MSIA-EnlightenGAN achieved a Peak Signal-to-Noise Ratio (PSNR) of 8.928 dB, a Structural Similarity Index Measure (SSIM) of 0.534 6, and a Learned Perceptual Image Patch Similarity (LPIPS) score of 0.391 6, outperforming SCI, URetinex-Net, Zero-DCE, and the original EnlightenGAN in terms of overall enhancement quality. On the self-built architectural heritage dataset, compared with the original EnlightenGAN, the proposed method demonstrated consistent improvements in image entropy, local contrast, and texture preservation, effectively restoring brick joint patterns, shadow-layer structures, and material details while maintaining more natural color consistency. Furthermore, aerial triangulation and 3D reconstruction experiments indicate that the enhanced images improve the detectability of low-light features, achieving 463 successfully calibrated images and a median number of 42 244 keypoints, which contributes to enhanced front-end image usability and feature extraction capability for subsequent photogrammetric processing. [Conclusions] The proposed GI-MSIA-EnlightenGAN effectively balances illumination enhancement, texture preservation, and structural consistency for low-light architectural heritage images. By integrating gray-guided illumination modeling with multi-scale attention mechanisms, the proposed framework exhibits superior robustness in preserving fine architectural details and improving visual quality under challenging lighting conditions. Moreover, it provides more reliable texture information for digital surveying, cultural heritage documentation, and 3D reconstruction workflows, demonstrating significant application potential in the digital preservation of architectural heritage under complex low-light environments.
[Objectives] With the ongoing integration of digital technologies into the cultural and tourism sectors, issues surrounding the protection of cultural heritage and the sustainable development of heritage tourism have become research frontiers, and the interpretation of heritage value has emerged as a central theme. Current research increasingly emphasizes human-centered perspectives, micro-scale analysis, and quantitative methodologies. Visitors’ sensory experiences are fundamental to assessing the quality of street spaces. Given the multifaceted values, intricate material environments, and varied demands for renewal in heritage streets, there is a need for more precise and context-specific approaches to their restoration and transformation. This article explores the interrelationships among perceptions of heritage value, spatial components of heritage sites, and tourists’ behavioral preferences. In doing so, it offers a theoretical basis and methodological framework to support value-oriented and sophisticated renovation of cultural heritage resources. [Methods] This study focuses on the spatial characteristics of streets within cultural heritage sites and proposes a "space-preference-value" research framework. By integrating an improved Stated Preference (SP) questionnaire with virtual reality (VR) technology, the research conducts tourism experiments within virtual geographic environments. Utilizing econometric models such as the discrete choice model and moderated utility model, the study investigates tourists' street selection preferences and perceptions of heritage value based on their spatial awareness. Furthermore, it explores the moderating effect of selection preferences on the perception of heritage value. [Results] (1) Tourists prefer heritage streets with higher green visibility, transparency, and alignment rates, which serve recreational, commercial, and sightseeing purposes, and feature strong imagery of special materials; (2) Respondents most highly perceive the spiritual value of heritage streets, followed by environmental, artistic, economic, and historical values, with scientific value perceived as the weakest; (3) Certain street spatial indicators significantly influence tourists' value perceptions: alignment rate positively affects perceptions of historical and artistic value; transparency rate positively affects perceptions of artistic value; (4) Tourist preferences show three modes of influence concerning the relationship between street spatial indicators and heritage value perceptions: positive regulation, negative regulation, and no regulation. [Conclusions] The study proposes a “space-preference-value” mechanism to establish the potential associations between heritage space and heritage value, thus enriching the theoretical framework of factors influencing heritage value perception and providing a more comprehensive attribution perspective for future research. In terms of planning and design practice, based on the above research results, the study puts forward two value-oriented integrated optimization strategies—cultural-historical tourism and leisure-entertainment tourism—which offer practical guidance for improving the environmental quality of heritage street interfaces and enhancing tourists' experiences and perceptions of heritage value.
[Objectives] The acceleration of urbanization has led to a critical trend of homogenization in traditional village landscapes and their disembedding from local contexts, underscoring an urgent demand for systematic and holistic methodologies to safeguard these landscapes as vital, living carriers of regional cultural heritage. Focusing on the Hetao Plain—a quintessential region characterized by Mongol-Han cultural convergence and Yellow River irrigation agriculture—this study aims to construct a computable, retrievable, and comparable representational system for local knowledge inherent in traditional village landscapes. Analyzing 26 designated traditional villages within this area, the research seeks to develop an analytical framework that can effectively identify, structure, and evaluate place-based wisdom to support scientifically informed conservation planning and decision-making. [Methods] This research is grounded in the theoretical framework of local knowledge and operationalized through the DIKW (Data-Information-Knowledge-Wisdom) hierarchy model. It integrates multi-source datasets, including spatial data (e.g., DEM, remote sensing imagery, village boundaries) and textual data (e.g., local chronicles, survey transcripts), using ArcGIS 10.5 for geospatial processing and analysis. The core of the method involves deconstructing and restructuring local knowledge into a formalized genealogy. This is achieved by establishing a four-level classification system ("Element-Context-Meaning-Practice") and implementing a three-tier systematic coding scheme based on "Element-Relation-Type." Subsequently, spatial clustering and hotspot analysis techniques are applied to uncover typological patterns and identify core landscape elements shaped by the region's diverse natural and human factors. Furthermore, a tailored local knowledge valuation system, constructed based on the genealogical structure and national evaluation standards, is employed to conduct a quantitative assessment of the preservation status of both explicit (tangible) and implicit (intangible) knowledge dimensions across the studied villages. [Results] The results from the quantitative evaluation reveal a significant and pronounced preservation gap. Explicit knowledge, pertaining to tangible landscape elements and spatial configurations, scores relatively high (77.3). In contrast, implicit knowledge—encompassing skills, customs, rituals, values, and socio-ecological practices—shows markedly weaker conservation outcomes (69.9). This disparity indicates a prevalent lack of synergistic conservation between material and socio-cultural dimensions. In response to these findings, the study proposes targeted strategies for the Hetao Plain's agro-pastoral ecotone. These include an "ecological-cultural" coordinated restoration strategy designed to reintegrate functional landscapes with their cultural meanings, and a "digital-living" integrated pathway aimed at leveraging digital tools for archiving and dissemination while fostering the living transmission of intangible heritage within communities. [Conclusions] This study demonstrates that the genealogical structuring and formal encoding of local knowledge can significantly enhance the accuracy of landscape identification and deepen the interpretability of its embedded socio-ecological contexts. It provides a novel analytical framework to diagnose and address the pressing issue of landscape homogenization. By transitioning local knowledge from a narrative, experiential form into a structured, computable, and comparable format, the research offers a transferable methodological toolkit and theoretical reference. This contribution is pivotal for advancing the systematic preservation, dynamic inheritance, and sustainable development of traditional village landscapes, not only in the Hetao Plain but also in other comparable cultural-geographical units facing similar challenges..
[Objectives] Data from Pre-Qin period burials remains fragmented across diverse sources and urgently requires systematic integration. Knowledge graph technology provides effective support for uncovering latent correlations and conducting deep analyses among archaeological burial data. [Methods] Based on archaeological textual data from Pre-Qin burials, this study analyzes multi-element knowledge representation methods of archaeological burials to construct the schema layer of the Pre-Qin archaeological burial knowledge graph. During the construction of the data layer, the entity and relation joint extraction model which is called TPLinker-joint-extraction was employed to extract and integrate knowledge from unstructured Chinese archaeological burial texts. The extracted entities, along with their attributes and relationships, were then stored in Neo4j graph database, achieving the construction of the Pre-Qin period burials knowledge graph. Building on this foundation, macro and micro scale analyses were conducted respectively on the nine major watersheds and the Dadianzi Cemetery to study and analyze in detail Pre-Qin burial characteristics. [Results] The TPLinker model achieved an entity recognition F1 score of 90.57% and a relationship extraction F1 score of 86.64%, demonstrating its capability to handle complex knowledge extraction tasks effectively. This validates the method's suitability for processing burial entities and their relationships within Chinese archaeological burial texts. The research reveals that pit tombs were the predominant burial structure during the Pre-Qin period, with the Yangtze River basin containing the highest concentration of burials by a significant margin. Among all types of burial objects, pottery jars emerged as the most common funerary objects in total across all major watersheds. Through spatial cluster analysis and zonal classification of the Dadianzi Cemetery, we calculated the coefficient of variation and correlation coefficients for burial scale and the value of funerary objects across different periods. Findings indicate that a certain degree of social differentiation had emerged at the Dadianzi Cemetery, and that the social structure of the Lower Xiajiadian Culture was stratified. [Conclusions] This study constructs the Pre-Qin period burials knowledge graph through textual information extraction, and proposes a multi-scale analysis method based on this knowledge graph. It promotes a more comprehensive discussion of the cultural differences of tombs at the watershed level, and the case study of the Dadianzi Cemetery reveals the social stratification of the Lower Xiajiadian Culture. It provides a scientific research paradigm and reference value for the study of ancient social structure based on burial data.
[Problem] Against the backdrop of intensifying urban heat island effects and the increasing frequency of extreme climate events, improving urban ventilation capacity has become a critical strategy for enhancing urban climate resilience and thermal comfort. Urban ventilation corridors (UVCs), as essential pathways facilitating airflow and heat exchange, play a key role in mitigating heat accumulation and pollutant retention. However, existing methods for identifying UVCs commonly face a fundamental trade-off among accuracy, computational efficiency, and spatial scalability. High-fidelity approaches, such as Computational Fluid Dynamics (CFD), provide detailed physical representations of airflow but are computationally expensive and difficult to apply to large-scale urban environments. In contrast, efficient GIS-based methods, including Least Cost Path (LCP) and Circuit Theory Models (CTM), lack sufficient physical consistency with airflow dynamics, particularly under complex and multi-directional wind conditions, resulting in limited capability to capture realistic wind-field response structures. [Methods] To address these challenges, this study proposes a Wind-Oriented Dual-Probability Method(WODP), an anisotropic framework that integrates wind-driven potential and spatial morphological constraints. Inspired by hydrological flow modeling, the method introduces a direction-sensitive Digital Surface Model (DSM) through azimuthal uplift, encoding the influence of terrain and building morphology on airflow resistance. A dual-probability mechanism, combining dominant wind-direction weighting and inverse-slope weighting, is developed to represent both meteorological forcing and underlying surface constraints. A roulette-wheel selection algorithm is further employed to simulate stochastic perturbations and directional variability of airflow, enabling probabilistic sampling of flow paths. By aggregating multiple simulations, a spatial airflow accumulation field is generated. Subsequently, multi-level segmentation based on adaptive thresholding and morphological skeleton extraction are applied to derive a hierarchical ventilation corridor network. [Results] Compared with traditional path-based methods such as CTM and LCP, the proposed method more effectively captures dominant airflow structures while maintaining high computational efficiency, achieving a speed 1-2 orders of magnitude faster than CFD. Experiments conducted in the central urban area of Wuhan demonstrate that the WODP method can stably identify ventilation corridor networks with high coherence and strong spatial continuity in typical urban environments. A single simulation requires only about 15 minutes and exhibits strong responsiveness to varying wind directions. Under multi-directional and multi-resolution conditions, the method shows robust scale adaptability and result stability. In addition, the airflow accumulation-based analytical framework supports hierarchical interpretation of ventilation structures, enabling the simultaneous identification of primary urban-scale corridors, secondary branches, and localized micro-scale pathways, thereby revealing a multi-scale and highly interpretable urban ventilation pattern. [Conclusions] WODP provides an effective compromise between high-cost CFD simulations and physically simplified GIS-based approaches. By incorporating anisotropic wind-field dynamics and probabilistic flow modeling, it achieves both physical interpretability and computational scalability. The proposed method offers a transferable and efficient technical pathway for large-scale, multi-directional urban ventilation corridor identification, hierarchical analysis, and refined urban climate planning. The wind environment simulation tool developed in this study can be obtained from the following website：
[Purpose] Images captured by in-vehicle cameras in low-light conditions often suffer from insufficient brightness, glare interference, and limited labeled data, which severely degrades downstream road element extraction. Existing unsupervised low-light enhancement methods mainly rely on RGB or HSV spaces, but they tend to introduce visual artifacts and ignore the impact of overexposed glare regions on dark areas, leading to information loss. [Method] To address these issues, this paper proposes Road Low-light Enhance Net (RoadLENet), an unsupervised low-light image enhancement network that integrates the vision-language foundation model Contrastive Language-Image Pre-training (CLIP) with multi-color-space feature fusion. RoadLENet can suppress strong light interference, improve the brightness of low-light vehicle-mounted images, and recover detailed road information without requiring paired annotated data, thereby providing higher-quality inputs for downstream road element extraction. Specifically, an overexposed-region weighting mechanism is designed based on the HVI (Horizontal-Vertical-Intensity) color space. This mechanism reduces the learning weight of strong-light regions and effectively weakens their negative influence during enhancement. On this basis, RoadLENet uses efficient frequency-domain attention to extract features from both RGB (Red-Green-Blue) and HVI color spaces. In the RGB branch, an improved illumination curve estimation strategy is introduced to preserve richer color and texture information. By fusing features from the two color spaces, the proposed method reduces artifacts, suppresses noise, and maintains fine structural details. In addition, a CLIP-guided prompt learning loss is incorporated to constrain the semantic consistency of the enhanced results and improve their perceptual naturalness. [Results] Extensive experimental results demonstrate that RoadLENet achieves significant enhancement performance on public datasets, including LOL, GladNet, and LoLI-Street, as well as on a self-constructed vehicle-mounted image dataset. On the self-constructed dataset, the NIQE score of RoadLENet decreases to 2.905 4, which is 0.164 6 lower than that of the second-best method, and CLIPIQA reaches the optimal value of 0.006 6. These results indicate that the proposed method has advantages in perceptual naturalness and semantic consistency. Qualitative visualization further shows that RoadLENet can effectively brighten dark regions, preserve the details of road elements, and suppress the negative influence of overexposed glare regions on enhancement results. Moreover, after enhancement with RoadLENet, the F1 scores of DeepLabV3Plus for lane lines and traffic markings increase by 3.49% and 5.66%, respectively, while the mean intersection over union improves by 6.11%. For U-Net, the corresponding improvements are 3.28%, 4.47%, and 3.63%, respectively. These results demonstrate that the proposed method not only improves the visual quality of low-light vehicle-mounted images, but also effectively enhances the reliability of downstream road element extraction tasks.
