Spatial Data and Intelligence

Urban Building Exteriors are increasingly important in urban analytics, driven by advancements in Street View Imagery and its integration with urban research. Multimodal Large Language Models (LLMs) offer powerful tools for urban annotation, enabling deeper insights into urban environments. However, challenges remain in creating accurate and detailed urban building exterior databases, identifying critical indicators for energy efficiency, environmental sustainability, and human-centric design, and systematically organizing these indicators. To address these challenges, we propose BuildingView, a novel approach that integrates high-resolution visual data from Google Street View with spatial information from OpenStreetMap via the Overpass API. This research improves the accuracy of urban building exterior data, identifies key sustainability and design indicators, and develops a framework for their extraction and categorization. Our methodology includes a systematic literature review, building and Street View sampling, and annotation using the ChatGPT-4O API. The resulting database, validated with data from New York City, Amsterdam, and Singapore, provides a comprehensive tool for urban studies, supporting informed decision-making in urban planning, architectural design, and environmental policy. The code for BuildingView is available at https://​github.​com/​Jasper0122/​BuildingView.

Urban fires occur frequently, posing significant threats to public safety in modern cities. Although previous studies have attempted to predict fire risks using regression analysis, traditional regression models are limited in handling complex nonlinear relationships among variables and accurately assessing the weight of each indicator. Additionally, at the level of prefecture-level cities and above in China, a comprehensive fire risk prediction indicator system has not yet been established. This study aims to enhance the accuracy and practicality of urban fire risk prediction. Based on multi-source data from prefecture-level and above cities in China, this research integrates fire risk, social economy, natural meteorology and spatial location to construct a fire risk prediction indicator system from three dimensions: hazard, vulnerability, and exposure. Five machine learning algorithms were employed for modeling and comparison to quantify fire risks and explore their spatiotemporal distribution. The findings reveal that XGBoost performs the best in predicting fire frequency and direct losses, demonstrating strong generalization and reliability. Further spatiotemporal analysis indicates that fire risks fluctuate upward over time, particularly in years with lower humidity, higher sunlight and high wind speed exposure. Spatially, coastal economically developed cities exhibit higher fire-related economic losses, while fire frequency is more concentrated in the central and eastern regions. Conversely, fire risks are relatively lower in the western and northeastern regions.

To improve the efficiency of traffic information prediction, prevent traffic congestion, and reduce navigation conflicts, a mid-to-long-term ship traffic prediction method based on a spatiotemporal diffusion attention network is proposed. First, a method for constructing a maritime network is proposed, addressing the different structural characteristics of waterways and road networks. Next, a spatiotemporal diffusion attention network is built to enhance prediction accuracy. Then, using AIS data from the Huangpu River region as an example, a comparative analysis is conducted with a baseline model under the LibCity traffic prediction framework. Finally, the prediction results are visualized and analyzed. Experimental results show that the proposed model improves the evaluation metrics (MAE, MAPE, RMSE, \( R^2 \)) compared to SOTA, with an improvement range of 1.6%–14.9%. The method surpasses the comparison models, thus validating its feasibility and effectiveness.

Objectives: Pollutants emitted from ships pose significant environmental challenges, particularly to air quality in port-adjacent regions, while greenhouse gas (GHG) emissions exacerbate global warming. Existing emission inventories for Tianjin Port often overlook the Domestic Emission Control Areas (DECA) policy. Additionally, forecasting models relying on multiple data sources face practical constraints. Methods: This study compiled a high spatial-temporal resolution emission inventory of pollutants and GHGs from ships in Tianjin Port in 2018, using AIS data and DECA policies. Four types of time series models—based on Transformer, MLP, TCN, and RNN—were employed to predict emissions. Results: The results showed that \(\text {SO}_\text {X}\) and CO\(_2\) are the primary pollutants and GHGs, with oil tankers, dry bulk carriers, and container ships being the main emission sources. Emissions from the main engine accounted for over 80% in channels, while auxiliary engine and boiler emissions were lower. However, in berths and anchorages, main engine emissions were almost negligible. Anchoring and docking contributed significantly, with emissions from these areas accounting for 94.94% of total emissions. Time series prediction results indicated that SCINet outperformed other models in low-value emission prediction. Conclusions: This study aligned with the DECA policy to develop an emissions inventory for Tianjin Port in 2018, examining the emission patterns of pollutants and GHGs from multiple perspectives. It also achieved emissions forecasting under a single data source condition.

Automatic identification of landslide disasters from remote sensing images based on deep learning models is an effective method, but most of the existing models use multi-temporal images or geological data to improve the accuracy, which suffers from problems of high model complexity and data acquisition difficulty. We proposed a Mv2_SA_DeepLabv3+ model, which is a lightweight deep learning network for detecting landslide disaster occurrence areas from remote sensing images captured in a single temporal image. Mv2_SA_DeepLabv3+ model is based on the Deeplabv3+ network structure. Firstly, it used the lightweight MobileNetV2 network as the backbone network for feature extraction to reduce the complexity of the model. Secondly, it improves the Atrous Spatial Pyramid Pooling (ASPP) module to obtain more features to improve the model accuracy. Finally, the loss function is optimized by combining CrossEntropy Loss and Dice Loss to solve the problem of positive and negative category imbalance of the samples. Furthermore, two publicly available landslide disaster remote sensing image datasets were utilized to validate the proposed model. The experiment results indicated that the Mv2_SA_DeepLabv3+ model shows a good performance, with the Kappa coefficient and the F1 score of 0.83 and 86.84%, which outperforms the original UNet (improved of 0.11 and 8.34%). These results demonstrate the model can obviously overcome the interference of irrelevant information, so as to accurately recognize the landslide area. Besides, the model has a low computational complexity, which facilitates the rapid deployment of the model and provides data support for further disaster emergency rescue.

Existing methods often ignore the correlations between variables or fail to consider their dynamic and delayed nature, which may affect the prediction accuracy. The main goal of this study is to address the challenges in time series prediction by effectively capturing the inter-dependencies between multiple variables. We propose the LCformer model, a new framework based on the Transformer architecture that integrates the lagged correlation information of multiple variables and its own historical information. The model identifies relevant exogenous variables using an exogenous variable filter (EVF), and employs a novel additive attention embedding (AAE) layer as well as cross-attention mechanism to simulate the lagged dependencies between these exogenous variables and the target variable. Experimental evaluation shows that the LCformer model is competitive with state-of-the-art methods, confirming its ability to accurately capture complex relationships in time series data, and learning multivariate lagged correlations, which can significantly improve prediction accuracy.

The ubiquity of mobile phone signaling data (SD) plays a pivotal role in mobility analysis; however, privacy policies and public concerns over personal location exposure severely limit the availability of public SD trajectory datasets. To address privacy concerns in human mobility analysis, synthetic trajectory generation has emerged as a promising solution. Existing trajectory generation methods primarily target GPS data, which exhibits high spatiotemporal continuity, leaving a critical gap in SD-specific trajectory synthesis. To bridge this gap, we propose SignalingTraj, a diffusion-based framework tailored for high-fidelity SD trajectory generation. Leveraging the generative capacity of diffusion models, SignalingTraj addresses the unique challenges of SD data, including low spatial resolution and irregular sampling. Specifically, we integrate spatial attributes and handover correlations of base stations (BSs) into a pre-training strategy using Variational Graph Autoencoders (VGAE), converting BSs into continuous representation vectors. Experiments on real-world SD datasets demonstrate that SignalingTraj synthesizes high-fidelity, privacy-preserving trajectories that closely mirror original data distributions. SignalingTraj outperforms existing methods in data fidelity and utility, positioning it as a robust solution for generating scalable, synthetic SD trajectories to support diverse mobility analysis applications.

The prediction of the Estimated Time of Arrival (ETA) in maritime traffic plays a crucial role in enhancing voyage planning, logistics efficiency, and overall maritime safety. This study introduces a segmented labeling approach leveraging Large Language Models (LLMs) for ETA prediction, highlighting their advanced reasoning capabilities. Building on recent progress in time-series forecasting, we propose a novel framework that utilizes LLMs for ship trajectory prediction and ETA estimation. The framework employs a few-shot in-context learning approach, structured prompt generation, and iterative refinement mechanisms to enhance prediction accuracy and reliability. Comprehensive experiments on a large-scale maritime dataset demonstrate the framework’s strong performance, particularly for short-distance routes, achieving notable improvements in prediction accuracy compared to traditional and deep learning-based methods. These results suggest that LLMs offer a promising direction for advancing time-series forecasting in the maritime sector.

Understanding human behavior is pivotal to the development of embodied artificial intelligence. Hand trajectories, as a critical medium of human interaction, offer a valuable lens through which to explore and interpret human actions. In this work, we propose a self-guided diffusion model for hand trajectory prediction (HTDiff). HTDiff consists of two main stages: The unconditional trajectory reconstruction stage learns from existing hand motion data to generate trajectory samples that conform to human motion patterns. In the conditional trajectory prediction stage, historical trajectories serve as contextual information to adaptively guide the pretrained reconstruction model in predicting future hand trajectories. Furthermore, we employ a Transformer architecture as the decoder within the diffusion model to fully exploit its strengths in sequence modeling, enabling the capture of temporal dependencies and complex motion patterns. Experiments conducted on public datasets reveal that HTDiff surpasses existing baseline methods in hand trajectory prediction, achieving the best performance across four evaluation metrics.

Aiming at the challenges of data integrity and reliability arising from the sparsity of ship trajectory data, this study proposes a ship sparse trajectory repair method combining feature correlation analysis and the interpretability of the SHapley Additive exPlanations (SHAP) model. Firstly, the methodology involves an analysis of sparse points and outliers within ship trajectory data to identify and address data omissions and anomalies. Subsequently, a comprehensive index of feature correlation is employed to select relevant features to trajectory repair, thereby reducing information redundancy and enhancing the precision of the repair process. Finally, utilizing the interpretability of the SHAP model, an interpretable ship trajectory repair model is constructed on a neural network framework, facilitating trajectory recovery and attribution analysis. The experimental outcomes indicate that the proposed method significantly enhances the accuracy and reliability of trajectory repair. By integrating feature correlation analysis with the interpretability of the SHAP model, this study not only refines the accuracy of ship trajectory repair but also provides a new idea for the interpretability of trajectory data repair models.

In recent years, with the profound changes in the global energy supply and demand structure, large oil tankers have become key participants in international energy transportation, playing an important role in ensuring the safety and stability of energy transport. Accurate prediction of tanker routes has become a crucial task to guarantee this. However, existing long-distance route prediction methods rarely consider trajectory consistency during data preprocessing, and recursive predictions suffer from error accumulation, leading to low prediction accuracy. To address these issues, our study proposes a new route prediction framework. The framework introduces an IMO and MMSI matching method during data preprocessing to resolve inconsistencies in historical trajectory data caused by changes in MMSI. Furthermore, to better address the issue of continuous position drift in trajectories, this study proposes an outlier cyclic deletion method. After extracting OD trajectory data based on the buffer zone, this study combines the Transformer model with the Long Short-Term Memory (LSTM) network model, leveraging their strong ability to capture long-term dependencies in time-series data. An Encoder-LSTM architecture-based route prediction model is then constructed, alleviating the decline in prediction accuracy caused by error accumulation. Experimental results show that the proposed framework significantly improves the accuracy and reliability of route prediction.

Location-based social networks (LBSNs) contain a variety of heterogeneous information and are widely used for tasks such as POI recommendation and user relationship inference. However, existing models often overlook the sequential nature of POIs in user check-ins and fail to capture the long-term dependencies of POIs in the check-ins, resulting in suboptimal embedding quality. To address the above issues, this paper proposes a model for learning Sequential Features of check-ins for User Relationship Inference (SFURI), which aims to improve user embeddings. First, we utilize a Bidirectional Long Short-Term Memory network (BiLSTM) to learn the sequential information of POIs in the check-ins, which effectively captures users’ dynamical preferences for POIs. Second, we exploit an attention mechanism to learn the long-term dependencies of check-in times in the check-ins and a feedforward neural network (FNN) to optimize the global features of times in the check-ins, which effectively captures users’ temporal preferences for POIs. The SFURI model enhances user embeddings, resulting in better user relationship inference. Experimental results show that the SFURI model outperforms the baselines across all the three datasets, demonstrating its effectiveness in learning user embeddings.

This study utilizes population, road, building, community and Point of Interest (POI) data to conduct spatial distribution analysis and network analysis of Beijing’s fire stations. Geodetector is employed to analyze the explanatory power of various factors on fire risk. Based on the spatial analysis results, replacing the traditional Euclidean distance coverage radius by network travel time, the Location Set Covering Problem (LSCP) model is used to add new fire stations, and the Maximal Covering Location Problem (MCLP) model is applied to select key fire station locations, thereby optimizing the spatial distribution of Beijing’s fire stations. On the basis of setting \({p}_{1}=100\) for the first MCLP model solution, we set \({p}_{2}=50\) and conducted the second model solution to explore the secondary coverage of the model. The results show that our research can cover the demand points in the study area well, and the distribution among all Rings is more balanced, which has guiding significance for the layout of large fire stations.

With the vigorous development of mobile internet technology and the widespread use of smart devices represented by mobile phones, crowdsourcing mobile trajectory collection has emerged as a significant means of data acquisition due to its efficiency, flexibility, data diversity, and low cost. However, variations in the types of intelligent terminals and user habits have led to inconsistent quality in the crowdsourcing data obtained. Especially in complex environments with signal shielding, the reliability of location precision is low, making it difficult to meet the application needs. Addressing these issues, this paper analyzes and extracts multiple features from crowdsourcing multi-source observation data collected by various models of intelligent terminals (such as Xiaomi and HUAWEI). In open scenarios, it employs multi-source fusion methods based on trajectory segmentation and clustering for data processing. In shaded environments, it utilizes map-matching algorithms to correct errors, thereby enhancing positional accuracy. Field tests and validations have demonstrated that the positioning accuracy of crowdsourcing data has been improved by the algorithm proposed in this paper. This proves the feasibility and effectiveness of this method, which can provide assistance in the analysis and processing of crowdsourcing data.

In recent years, point-of-interest (POI) recommendation has been extensively studied, with existing methods typically modeling user preferences through the integration of multi-factor information (e.g., temporal, spatial, and categorical features) and capturing the periodicity and discontinuity of user check-in sequences. However, these approaches struggle with data sparsity, missing data, and noisy data, leading to suboptimal user representations. To effectively mitigate the above problems, we utilize contrastive self-supervised learning techniques to achieve data augmentation and apply them to the next POI recommendation task. Specifically, we propose DACL (Data Augmentation through Contrastive Self-supervised Learning), a novel framework that unifies next POI recommendation and contrastive self-supervised learning (SSL) via a multi-task strategy. Furthermore, DACL introduces five tailored data augmentation operations to generate high-quality contrastive views, mitigating data limitations while enhancing robustness. Extensive experiments on two real-world datasets (NYC and TKY) demonstrate that DACL significantly outperforms state-of-the-art baselines, achieving \(14.3\%\) and \(4.3\%\) improvements in Recall@10 and NDCG@5, respectively, while maintaining superior robustness against noisy and sparse scenarios.

The driving forces of urbanization research in previous studies are socio-economic, environmental and technological development. We argue that government policies play an important role in shaping urbanization, especially in Socialist countries. In the past two decades, many super cities have emerged in China, and their size and spatial patterns are constantly changing. This paper aims to develop a deductive method to reveal the development patterns of Beijing under the influence of governmental policies and predict the future development of the city. Specifically, we first propose the concept of road network-based urban texture to describe the level of urbanization. Second, we use linear functions and neural networks to learn the spatio-temporal development patterns from Beijing’s survey and mapping data from 2005 to 2015. Finally, we validate the proposed method using the survey and mapping data of 2022. The experimental results confirm that our hypothesis in findings reasonable and the proposed method is able to predict the future development of cities.

Evaluating the landscape ecological risk index (LERI) at the landscape level and identifying its driving mechanism is crucial for ensuring regional ecological stability. Firstly, we developed an LERI evaluation model to comprehensively evaluate its spatial-temporal evolution in the Nanjing Metropolitan Area from 2000 to 2020. Secondly, we applied the geographically weighted random forest model to explore the driving mechanisms of different factors on LERI. Finally, we analyzed the correlations between various driving factors. The results indicated that: (1) the spatial distribution pattern of LERI remained stable, with a general “high in the north, low in the south” pattern. The area of medium-risk zones decreased by 1.25 × 10³ km² (a decline of 5.76%), while the area of high-risk zones increased by 0.39 × 10³ km² (a rise of 8.11%), with the areas of extremely low, low, and extremely high-risk zones changing relatively steadily; (2) the normalized difference vegetation index (NDVI), annual average precipitation (AAP), and net primary productivity (NPP) are the main driving factors of LERI evolution, with population density (PD) playing a secondary role. The effects of annual average temperature (AAT) and nighttime lights (NTL) are relatively small, and the spatial distribution of factor importance varies significantly; (3) there is a consistently strong positive correlation between driving factors such as AAP, NDVI, NPP, and AAT. These findings aim to provide scientific support for regional landscape ecological risk management.