Recent advancements in novel view synthesis, particularly Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have driven substantial progress in 3D reconstruction from street-view visual streams. By optimizing continuous or discrete scene representations through differentiable rendering, these computational paradigms have rendered complex urban environment modeling increasingly tractable. Yet the massive volume of crowdsourced street view videos (CSVVs) on social media platforms remains largely intractable for these methods due to their non-standardized nature, which often causes reconstructions to fail after substantial computational resources are wasted. To address this challenge, we introduce Crowdsourced Radiance Fields (CSRF), a framework designed to harness CSVVs. CSRF enhances reconstruction viability and reduces the processing time wasted on failed initializations through a dual-stage architecture: (1) a lightweight quality-assessment module that rapidly filters severely degraded content, and (2) a progressive joint-optimization strategy that reconstructs scenes using locally overlapping TensoRF-based radiance fields while simultaneously estimating camera poses, guided by monocular depth and optical flow priors. To systematically evaluate the capabilities of 3D reconstruction from CSVVs, we present CS3D100, a challenging benchmark comprising 100 diverse user-uploaded CSVVs (drone, dashcam, mobile, and vlog footage) from social media platforms. Extensive experimental evaluations demonstrate that the proposed CSRF framework yields the most competitive quantitative results on the CS3D100 dataset within the scope of the evaluated zero-prior baselines. This performance advantage is validated across a comprehensive suite of criteria, encompassing sequence processing success rate, perceptual rendering quality, computational resource efficiency, and the newly introduced composite metrics, CUR and REI. Furthermore, the framework maintains comparable reconstruction capabilities when applied to standard reference benchmarks. Our framework achieves an absolute reconstruction success rate of 31
Thematic maps are essential tools for communicating spatial information in academic publications, yet their long-term design evolution across different linguistic publishing contexts remains insufficiently understood. This study presents a large-scale, data-driven analysis of thematic map design in Chinese- and English-language academic cartography from 1990 to 2020. We develop an AI-assisted cartographic data mining framework that integrates vision-language models for understanding heterogeneous academic documents and identifying thematic maps with computer vision techniques for map element detection and image segmentation. Based on this framework, we compile a corpus of 45,732 research articles from 16 representative cartography and geographic information science journals and extract 23,928 high-quality thematic maps for systematic analysis. Map design is characterized along three dimensions: map elements, color design, and layout structure. The results reveal broadly shared cartographic conventions between Chinese- and English-language academic maps, while the degree of convergence varies across design dimensions. English-language maps generally contain slightly richer cartographic elements and greater chromatic diversity, whereas Chinese-language maps show a stronger preference for neutral color schemes. Both groups, however, exhibit highly similar centered and balanced layout structures. Temporal analysis further shows strong convergence in map-element configuration, parallel enrichment of color design with persistent cross-language differences, and comparatively stable layout structures characterized mainly by incremental changes in space allocation and alignment. Overall, academic map design appears to evolve toward shared cartographic conventions while retaining contextual differences associated with publishing traditions.
Place representation provides a foundational data infrastructure for a wide range of urban applications. Although existing models effectively capture the physical dimensions of space and have increasingly incorporated social dimensions, such as urban functions and activity patterns, they rarely explicitly integrate higher-order social semantics of places, such as collective experiences, emotional perceives, or social identities. This study proposes SGAP-Place, a social semantics-augmented representation learning framework for urban places. The framework integrates point-of-interest data, origin–destination flows, street-view imagery, and residential profiles. An evidence-constrained large language model first converts modality-specific information into structured descriptions and synthesizes them into four dimensions of social semantics (e.g., social identity, everyday usage patterns, Cultural/ collective memory, and Sensory experience). These semantic features are then integrated with the original modality-specific features through intra-modal gated fusion and cross-modal interaction. A graph-attention module further captures spatial dependencies among neighboring places, while self-supervised contrastive learning is used to obtain transferable place embeddings. The framework was evaluated in Shenzhen using three downstream tasks. SGAP-Place achieved the strongest overall performance across the three tasks, with improvements of 1.96 R^2 , and 14.41 R^2 . Ablation experiments showed that the generated social semantics provided substantial predictive information and that staged fusion was more effective than direct summation or concatenation of modalities. These findings indicate that social semantics can serve as an explicit and useful information layer in place representation. SGAP-Place provides an approach for incorporating higher-order social meanings into place representation.
This study develops and evaluates an integrated geoinformation framework for modeling urban surface thermal energy deprivation (USTED) and its spatial association with land surface temperature (LST) in Tabriz, Iran, from 2015 to 2024. The framework integrated Landsat- and MODIS-derived LST, municipal building and land use/land cover (LULC) datasets, demographic information, and terrain variables within a GIS based multiple decision analysis workflow. For this goal, after selecting the relevant criteria, the Analytic Network Process (ANP) was integrated with Monte Carlo simulation (MCS) and global sensitivity analysis (GSA) to quantify weight-related uncertainty and improve model robustness. Bivariate spatial association and geovisualization techniques were then used to examine daytime and nighttime USTED–LST relationships, and the resulting susceptibility patterns were evaluated against winter energy-consumption records. Spatiotemporal tracking indicated an eastward shift of high-high clusters and the persistence of energy-stress zones in dense residential areas. Detailed results shows that nighttime high-high clusters expanded within urban centers (p < 0.01), consistent with thermal retention and urban heat-island behavior. On the other hand, daytime high-low clusters were concentrated in compact built-up zones, reflecting relative cooling compared with barren surroundings. Overall, the proposed framework supports uncertainty-aware mapping of urban thermal-energy susceptibility and provides a decision-support basis for targeting energy-efficiency and heat-mitigation interventions in semi-arid cities.
Predicting continuous environmental variables from sparse spatial observations remains challenging because global machine learning models can capture nonlinear relationships but often have limited ability to adapt to local spatial heterogeneity, whereas existing similarity-based methods still rely mainly on weighted-average prediction. This study proposes Pattern Similarity Learning (PSL), which uses pattern similarity as a spatially adaptive indicator to guide local nonlinear prediction. PSL first augments the original environmental covariates with three pattern-derived indicators: geocomplexity, positive outlier strength, and negative outlier strength. These indicators describe local heterogeneity and local anomalousness, allowing each covariate to carry both its original value and its local spatial context. For each target location, PSL then measures covariate-space similarity to identify relevant training samples and fits a similarity-weighted local Random Forest using the pattern-enhanced features. In a soil organic carbon prediction case across the conterminous United States (CONUS), five-fold spatial cross-validation shows that PSL achieved the best overall pooled performance among the six models. Compared with Random Forest, PSL increased R² by 0.017 (7.11
The effectiveness of large language models (LLMs) in transportation mode detection (TMD) remains underexplored, creating a significant research gap in understanding how these models process trajectory data. This study uses TMD as a diagnostic setting to examine which trajectory formats and prompting strategies best support LLM-based inference in human mobility, and further analyzes misclassification patterns and the trajectory formats whose reasoning is most susceptible to hallucinations. Using the Geolife dataset, we evaluate pre-trained and fine-tuned LLMs across 14 trajectory formats, categorized into overview information, coordinate-based, and spatial encoding. Meanwhile, two response strategies are compared: direct answer and Chain-of-Thought (CoT) reasoning. The results show that fine-tuning improves accuracy across all formats. The coordinate-based format with timestamps achieves the highest accuracy of 81.7
Illegal dumping, commonly referred to as fly-tipping in the UK, is a pressing urban management challenge that compromises both municipal budgets and environmental equity. Traditional solutions are largely reactive and unable to identify future fly-tipping events, as they are based solely on historical fly-tipping logs. To address these limitations, we adapt the existing A3T-GCN architecture to sparse fly-tipping prediction and extend it with multisource urban contextual features. The graph component is designed to represent local spatial dependence in fly-tipping risk, providing an operational modelling analogy to the neighbourhood-level spillover implied by Broken Windows Theory. This adapted framework extends crime-count-based prediction by incorporating temporal dynamics and contextual urban information. In addition to historical incident counts, the model incorporates 557 points of interest and land-use features extracted from OpenStreetMap. We simulate the framework using real-world fly-tipping data from Lancaster, UK, that cover the period from 2019 to 2025. The results show that A3T-GCN achieves a 10.4
Solar photovoltaic (PV) systems have great potential to supply renewable energy and support the transition of urban energy structures. While integrating PV technology into public facilities is a promising strategy for expanding PV deployment in urban areas, the PV potential of standardized bus stops, which are numerous and widely distributed across cities, remains insufficiently assessed. To address this gap, this study proposes a framework to quantify the PV potential of bus stop roofs by integrating bus stop point-of-interest (POI) data, street view images, and solar trajectory simulations. The research results reveal two key findings. Firstly, the theoretical annual electricity generation of bus stop photovoltaic (BSPV) systems in Guangzhou, China, can reach 5,531.67 MWh, which is equivalent to powering approximately 84,196 LED streetlights for one year. Secondly, BSPV potential exhibits distinct spatio-temporal variability: it follows a unimodal seasonal pattern (peaking in early summer and dropping to its minimum in March), while significant spatial heterogeneity is observed. Specifically, the central urban districts (e.g., Yuexiu, Liwan, Haizhu, and Tianhe) show relatively low BSPV potential due to severe building shading. In contrast, peripheral districts (including Huadu, northern Baiyun, Huangpu, Panyu, and Nansha) demonstrate notably higher BSPV potential. Additionally, areas along the Pearl River also possess considerable PV potential. The study provides a data-driven solution for urban planners to deploy distributed energy systems via public infrastructure. Furthermore, the findings offer scientific guidance for decision-making related to urban sustainable development and low-carbon transport planning.
Reliable cycling network data is crucial for bicycle research and planning. OpenStreetMap (OSM) is a globally available and open-source network dataset, increasingly used to represent bicycle infrastructure supply, mostly as a means to achieve a further goal (e.g., modelling cycling). However, infrastructure classification varies across studies and common standards are lacking, which hampers transferability and comparison. This study introduces BikeNEAT, an easy-to-use classification tool that utilises OSM data to classify bicycle infrastructure. We analyse OSM tagging patterns in the German context and investigate how bicycle infrastructure categories can be identified from OSM tags. Our contribution presents a standalone classification framework for the German context that derives bicycle infrastructure categories from complex combinations of OSM tags. The framework explicitly addresses heterogeneous OSM tagging practices, incorporates infrastructure directionality, and allows the extracted indicators and categories to be used independently of a predefined assessment workflow. The study presents the publicly available BikeNEAT code and describes how the classification framework can be applied to any OSM dataset. Using a reference dataset from Freiburg im Breisgau for validation, we show that the method performs well, with nearly 95
Conventional expert interpretation remains the principal approach for geomorphological mapping. However, its application over large and heterogeneous areas can be time-consuming and influenced by interpretative variability, particularly in complex fluvial landscapes. This study develops a convolutional neural network (CNN)-based workflow to support geomorphological mapping in Northern Thailand using a multispectral and terrain input stack. The model was trained using an expert-interpreted reference geomorphology map and twelve geospatial layers derived from Landsat 8–9 Operational Land Imager imagery and Shuttle Radar Topography Mission data, including multispectral bands, normalised difference vegetation index, elevation, slope, curvature, topographic position index, and topographic wetness index. A High-Resolution Boundary U-Net (HRBoundaryUNet) model was implemented and compared with selected benchmark segmentation models using mean intersection over union (mIoU) and macro F1-score. To reduce spatial autocorrelation and patch-overlap leakage, model evaluation was performed using a spatially buffered stratified validation scheme, with performance further assessed using per-class intersection over union (IoU) and F1-score. Spatial discrepancy analysis was also applied to distinguish direct agreement, boundary generalisation, and spatial misclassification between the CNN-enhanced map and the reference map. Field investigations from unmanned aerial vehicle surveys and soil coring were used as point-based supporting evidence to corroborate geomorphological interpretation. The results demonstrate that CNN-based segmentation can serve as a reproducible and semi-automated mapping support tool for regional geomorphological mapping. The proposed workflow complements expert-based mapping by providing quantitative spatial evaluation, improving consistency in map generation, and offering a transparent basis for subsequent expert interpretation.
Mangrove ecosystems are central to global climate mitigation and coastal resilience, yet their monitoring remains inherently complex due to tidal inundation, spectral ambiguity and persistent cloud cover across tropical regions. The emergence of cloud-based geospatial platforms, particularly the Google Earth Engine (GEE), has transformed mangrove remote sensing from scene-based analysis to scalable, multi-sensor, time-series frameworks operating at planetary scale. This study presents a systematic review and critical synthesis of 162 peer-reviewed studies published between 2017 and 2025, examining the structural evolution, methodological foundations and ecological implications of GEE-enabled mangrove research. Bibliometric assessment indicates rapid expansion of the field, reflecting a transition from early methodological experimentation to widespread application. Thematic evolution reveals a clear shift from two-dimensional extent mapping to increasingly dynamic and process-oriented analyses, including disturbance monitoring, hydroperiod assessment, biomass estimation and blue carbon quantification. Methodologically, the literature is strongly anchored in ensemble machine learning approaches, particularly Random Forest, supported by multi-sensor integration of optical time-series (Landsat, Sentinel-2) and Synthetic Aperture Radar (SAR) to mitigate atmospheric and tidal constraints. Despite these advances, critical structural limitations persist. Research output and study-area focus remain geographically concentrated in a limited number of countries, whereas several ecologically significant mangrove regions remain underrepresented, indicating a mismatch between research capacity and ecosystem distribution. Furthermore, although classification accuracy has improved substantially, persistent uncertainties arising from tidal variability, spectral similarity, training data bias and limited model transferability constrain ecological interpretation. These findings indicate that GEE has successfully enabled large-scale mangrove monitoring but has not yet fully translated computational advances into ecological understanding or operational decision-making. Future research should prioritize cross-regional model validation, the integration of structural datasets such as spaceborne light detection and ranging (LiDAR), and the development of interpretable, process-based frameworks to better connect remote sensing outputs with ecosystem functioning. Strengthening these directions is essential to ensure that GEE-based monitoring systems effectively contribute to global mangrove conservation, blue carbon accounting and climate-resilient coastal management.
In the current climate change scenario, it is increasingly important to evaluate mountain areas potentially at risk in an efficient and rapid way. Permafrost is one of the main factors determining surface stability in mountain regions. The monitoring network of ground surface thermistors is limited and in several cases the data and metadata are not of easy access. Therefore, modelling is a valid alternative for the prediction of the ground surface temperature (GST). This work presents an update version of the physics-based model PERMACLIM (PERMACLIM 2.0) for the calculation of the mean annual ground surface temperature (MAGST) and for mapping permafrost conditions. PERMACLIM 2.0 presents several updates mainly regarding the realization of the snow depth maps, the calculation of the GST and the classification of the MAGST for detecting permafrost conditions in terms of presence, probability, possibility and absence of permafrost aggradation. The model was tested for the hydrological years 2020 and 2022, on the entire upper Valtellina, in the Northern Central Italian Alps, returning results with a spatial resolution of 10 m. The overall RMSE for MAGST ranges between 0.6 and 0.8 °C. The permafrost conditions were not favourable for the 67.8
The rapid response of police officers plays a crucial role in managing traffic incidents, as delayed responses increase the risk of secondary crashes and prolonged blockages. This delay in crash response varies spatially, posing a critical challenge to effective incident management. However, the spatial variability in police response times and its influence on response time prediction have received limited scholarly attention. Prior to model development, spatial dependence is assessed using global Moran’s I statistic, while spatial heterogeneity is explored through choropleth mapping and geographically weighted regression. Based on this evidence, an Integrated Spatial Deep Neural Network (ISDNN) is developed as a spatially explicit prediction framework to serve as an initial decision-support tool in pre-dispatch operations. The model incorporates radial basis functions and periodic kernels to embed spatial and temporal characteristics, enabling the representation of spatially varying response processes. The proposed model is evaluated using traffic crash records and police operational data from Bartlett, Tennessee, United States. Compared with several non-spatial and spatial baseline models, the ISDNN achieves lower prediction errors (RMSE = 0.43; MAE = 0.33; MAPE = 23.85
Realistic road modeling is a core component of three-dimensional geographic information systems. Traditional methods usually emphasize geometric precision and rarely consider the effects of environmental conditions and natural aging. Thus, they generate visually clean and static road models that fail to dynamically represent the evolution of road states. To overcome this limitation, we propose a novel method for generating realistic 3D road models with multiple styles. The method abstracts a road structure as a top-down directed acyclic graph, decomposing geometry into semantically independent but functionally interrelated nodes. Coupled with a parameter propagation mechanism and a physically based decal library, the method enables the automatic generation of 3D road models that adhere to user-defined styles across temporal, environmental, and functional factors. We evaluated the method through experiments covering nine style categories and two user studies, validating its effectiveness in generating realistic road representations with diverse influencing factors. Compared with common generative AI and procedural methods, the method exhibits superior structural controllability and environmental interactivity. This work advances road modeling by introducing a parametric scheme that explicitly maps the semantic context into realistic representations, providing practical support for applications such as landscape design and traffic simulation.
Causal inference for spatial cross-sectional data is an essential approach to reveal the driving mechanisms of geographical phenomena. As an emerging method for causal inference with spatial cross-sectional data, Geographical Convergent Cross Mapping (GCCM) is easily disturbed by “false neighbors” caused by spatial autocorrelation, and this can lead to misjudgment of causal direction or overestimation of causal strength. In view of this problem, this paper proposes a Geographical Convergent Cross Mapping method accounting for Spatial Autocorrelation (SA-GCCM). The method introduces spatial path roughness to measure the smoothness of attribute value changes between two points, and constructs an iterative neighbor screening mechanism to eliminate the “false neighbors” problem caused by spatial autocorrelation. In addition, the proposed method identifies true causal relationships by examining how the predictive performance changes with the spatial path roughness. To verify the effectiveness of the SA-GCCM method, this study conducted experiments on two typical datasets with clear causal relationships: nighttime-light data and topographic factors, and net primary productivity and climatic factors. The experimental results show that SA-GCCM can effectively address the issues of causal direction misjudgment and causal strength overestimation in the original GCCM method. This study not only improves the robustness of the GCCM method, but also provides methodological support for handling geographical data with spatial autocorrelation.
As significant carriers of China’s historical and cultural heritage, grotto temples and rock carvings (GTRC) warrant systematic investigation into their spatio-temporal distribution and spatially associated factors to support effective conservation and management. This study employs GIS spatial analysis, the optimal parameters-based geographical detector (OPGD), and the Multiscale Geographically Weighted Regression (MGWR) model to analyze the spatial distribution, resource richness, and multidimensional spatial associations with national-level GTRC from natural, economic, and cultural perspectives. The results show that: (1) Spatially, national-level GTRC exhibit an uneven pattern characterized by “dense in the central region, more in the east and less in the west”, with high-density areas concentrated along the Henan–Shanxi border. Resource richness shows significant clustering, forming high-value areas in the Central Plains and Sichuan-Chongqing regions. (2) Temporally, the Sui and Tang dynasties were the peak construction period, and the trajectory of the center of gravity shows a shift in distribution direction from “northwest-southeast” to “southwest-northeast”. (3) The OPGD results indicate that economic development exhibits the strongest explanatory power, followed by population density and intensity of religious belief, with economic-religious interaction showing the strongest explanatory power. (4) The MGWR analysis reveals marked regional heterogeneity: in Henan, distribution is more strongly associated with economic, demographic, and religious factors, whereas in the Sichuan-Chongqing region it is more closely related to religious and cultural forces. These findings clarify the multidimensional factors associated with national-level GTRC patterns and provide theoretical and methodological support for their conservation and sustainable utilization.
Landslide prediction is a core issue in disaster risk management. In rainfall-driven mountainous regions, landslide susceptibility evolves dynamically in response to changing hydro-meteorological conditions, yet most existing approaches remain static or retrospective and lack explicit temporal predictability. This study proposes a spatiotemporal modelling framework that formulates landslide susceptibility as a time-evolving system state driven by precipitation. The framework follows a two-stage workflow, in which dynamic environmental forcing is forecast prior to susceptibility estimation. A U-Net spatiotemporal model is first employed to predict short-term precipitation patterns, which are then integrated with static environmental attributes within an Extreme Gradient Boosting (XGBoost) susceptibility model. Model interpretability is enhanced using SHAP to quantify the relative contributions of dynamic and static predictors. The framework is demonstrated in Nepal using ten years of daily precipitation data to generate temporally explicit landslide susceptibility maps. The results show that the model achieves an accuracy of 90.54
Relief shading is widely used in cartography and the geosciences, but its application in Antarctic coastal ice-free areas remains limited because these landscapes are spatially restricted, low-relief, and often lack sufficient tonal contrast for effective terrain depiction. Using unmanned aerial vehicle (UAV) data acquired during the 2018 Chinese Antarctic Research Expedition, this study produces the first enhanced topographic map of the Broknes Peninsula, Larsemann Hills, supported by a cartography-grade shaded-relief base. We develop and validate a UAV-enabled digital relief-shading workflow tailored to polar ice-free terrain, including (1) UAV structure-from-motion (SfM) processing with DEM quality improvement through dome-error mitigation and ICESat-2-based vertical correction; (2) cartography-oriented DEM generalization using line integral convolution (LIC) with fidelity diagnostics; (3) multi-directional shaded-relief construction through gradient-domain Poisson editing of three normalized hillshades (azimuths 60°, 180°, and 300°); and (4) a reproducible quantitative evaluation framework integrating tonal-distribution analysis, the shadow recovery index (SRI), highlight suppression index (HSI), and texture gain index (TGI). Results show that the proposed workflow effectively suppresses extreme tones, improves shadow legibility, and enhances local texture expression, thereby producing a visually coherent and geomorphically plausible shaded-relief basemap for topographic mapping in Antarctic ice-free areas. Overall, this study provides a transferable workflow for high-precision shaded-relief production in low-relief polar terrain and supports more effective topographic mapping in Antarctic coastal ice-free regions.
Automating spatial analysis remains challenging due to its dependence on predefined workflows, static schemas, and expert-driven configurations, which limit adaptability across diverse datasets and analytical contexts. Recent advances in large language models (LLMs) provide new opportunities to address these challenges by enabling systems to interpret human intent, reason over structured data, and generate executable analytical workflows. This study presents an automated Spatial Query and Analysis (SQA) framework that leverages LLMs to translate natural language queries into validated spatial operations. The framework interprets user prompts and classifies them as either general or spatial queries. General queries are answered directly by a language model, whereas spatial queries are processed through a multi-agent reasoning pipeline that performs semantic interpretation, spatial analysis, peer-review validation, and code execution checks to ensure accuracy and consistency. In addition to query-based interaction, users can upload new spatial datasets that are automatically validated and integrated into the system through dynamic knowledge augmentation. Evaluation using a large metropolitan university campus dataset shows that SQA framework achieves improved spatial reasoning performance compared with single-agent baselines, including configurations with metadata access and error-correction mechanisms. The framework also incorporates a two-step error recovery process that successfully resolves common failure types, including table or schema mismatches, coordinate reference system (CRS) misuse, and column or spatial reasoning errors. SQA framework demonstrates consistent accuracy across single-table, multi-table, and 3D queries, producing integrated outputs in map, chart, and tabular formats.
This study presents a comparative scientometric analysis of 526 publications from the Web of Science Core Collection and China National Knowledge Infrastructure (CNKI) database to decode the global research trajectory on urban green space (UGS) carbon sinks and to delineate China’s distinctive role. Analysis reveals a clear evolution in research focus and output. Early work (before 2011) established foundational concepts. The subsequent period (2011–2015) witnessed significant methodological diversification, driven by advances in remote sensing and quantitative carbon flux assessment. Following the Paris Agreement in 2015, research emphasis shifted markedly towards policy-integrated applications and the exploration of synergies between carbon sequestration and sustainable urban development. A key finding is the rapid ascendancy of Chinese research output post-2020, a surge that has established China as a leading contributor to the field and is closely aligned with the national “dual carbon” goals. The synthesis identifies a critical methodological shortcoming: an “ecological analogy trap,” in which an over-reliance on forest-derived biomass equations systematically overlooks urban-specific anthropogenic factors. The methodological frontier is increasingly defined by the integration of multi-source remote sensing (e.g., LiDAR-optical fusion) and machine learning to overcome urban heterogeneity, though challenges in model transferability persist. The study concludes that while international scholarship increasingly emphasizes the coupling of carbon sink analysis with socio-ecological equity, China’s policy-driven research has prioritized the optimization of technological pathways for carbon accounting. To advance the field, future research must develop urban-specific protocols, leverage advanced computational techniques, and foster interdisciplinary, multi-scale frameworks to provide robust decision-support for achieving urban carbon neutrality.
