Spatial Data and Intelligence

Graph-based clustering plays an important role in the clustering area. However, in general clustering tasks, the graph structure of data does not exist, so the strategy for constructing the graph is crucial for the performance of the subsequent tasks. In the subsequent comparison task, existing methods fail to consider the class information and will introduce false-negative samples in the random negative sampling, causing poor performance. To this end, we propose a multi-view comparison clustering framework based on clustering guidance and adaptive encoder. First, the graph is constructed adaptively according to the generative perspective of the graphs. The adaptive process is designed to induce the model to exploit the high-level information behind data and utilize the non-Euclidean structure. Then, representations can be optimized by aligning with clustered class information, and simultaneously, the optimized representations can promote clustering, leading to more powerful representations and clustering results. Extensive experiments on five datasets demonstrate that our method achieves new state-of-the-art results on clustering tasks.

In the field of Intelligent Transportation Systems (ITS), the challenge of performance degradation in lightweight object detection models on edge devices is significant. This issue primarily arises from environmental changes and shifts in data distribution. The problem is twofold: the limited computational capacity of edge devices, which hinders timely model updates, and the inherent limitations in the generalization capabilities of lightweight models. While large-scale models may have superior generalization, their deployment at the edge is impractical due to computational constraints. To address this challenge, we propose a cloud-edge collaborative continual adaptation learning framework, specifically designed for the DETR model family, aimed at enhancing the generalization ability of lightweight edge models. This framework uses visual prompts to collect and upload data from the edge, which helps to fine-tune cloud-based models for improved target domain generalization. The refined knowledge is then distilled back into the edge models, enabling continuous adaptation to diverse and dynamic conditions. The effectiveness of this approach has been validated through extensive experiments on two datasets for traffic object detection in dynamic environments. The results indicate that our learning method outperforms existing techniques in continual adaptation and cloud-edge collaboration, highlighting its potential in addressing the challenges posed by dynamic environmental changes in ITS.

AIS provides a huge amount of maritime traffic data containing spatial and temporal information in a limited area. Trajectory clustering based on AIS data is a pre-task in intelligent maritime domain, providing typical movement patterns of vessels for follow-up studies in navigation safety and maritime supervision. This paper presents an AIS trajectory clustering method incorporating discrete Fréchet distance and Douglas-Peucker (DP) algorithm, based on improved density-based spatial clustering of applications with noise (DBSCAN). Experimental results on the dataset of vessels entering and leaving the Taiwan Strait in November 2017 demonstrate the effectiveness of our method.

Nodes clustering is an important approach to partition heterogeneous information networks based on the features and adjacent matrices from different metapaths. Some scholars have adopted contrastive learning methods on the basis of deep clustering, which has achieved promising clustering performance. Despite this, few of them pay attention to redundant information in features, while also not considering both the semantics and structure of the nodes. To fill these gaps, a Structure and Semantic Contrastive Learning for Nodes Clustering in HINs (SSCHC) method is proposed. Specifically, the proposed method explores the high-order neighbor relationship of the node by reconstructing the adjacency matrix containing path and processing the redundant information in the features. In addition, we design a structure and semantic contrastive learning module to obtain more comprehensive information about the nodes. Extensive experiments on several real-world benchmarks demonstrate the effectiveness of the proposed SSCHC method compared with the state-of-the-art baselines.

As a new form of data management, the global discrete grid can describe and exchange geographic information in a standardized way on a global scale, which can be used for efficient storage and application of large-scale global spatial data, and it is a digital multi-resolution geo-reference model, which helps to establish a new data model and is expected to make up for the deficiencies of the existing spatial data in the aspects of organization, processing and application. The representation of vector data based on hexagonal isoproduct projection of global discrete grids fundamentally solves the problems of data redundancy, geometric deformation, and data discontinuity that occur when multi-vector data are represented in grids. In this paper, different gridded methods are proposed for different types of vector data and remote sensing data to achieve efficient gridded processing of multi-source data. For the gridded vector data, a quantifiable accuracy evaluation index system is established to evaluate the accuracy of the gridded vector data in terms of geographic deviation, geometrical features and topological relationships, and for the gridded remote sensing data, a Kyoto evaluation index system is constructed based on the levels of information, image structure, and texture features, which further proves the usability of the hexagonal gridded vector-based and remotely sensed data. The evaluation method is generally applicable to all gridded vector and remote sensing data based on hexagonal grids and can be used to evaluate the usability of hexagonal grid data.

A co-location pattern is a set of spatial features whose instances are frequently located together in geo-space. In real world, different instances have different distributions and different values. However, existing methods for mining pattern ignore these differences. In this paper, we propose a novel method for mining regional high utility co-location pattern by considering both instance distribution and value. First, local regions are obtained based on fuzzy density peak clustering. Then, the regional high utility co-location pattern is defined, and an efficient algorithm for mining the patterns in local regions is presented by pruning unpromising patterns. The experiment results show the patterns are meaningful and the mining algorithm is efficient.

Due to the heterogeneity of spatial data, spatial co-location patterns are not all global prevalent patterns. There are regional prevalent patterns that can only appear in specific local areas. Regional co-location pattern mining (RCPM) is designed to discover co-location patterns like these. The regional co-location patterns can reveal the association relationships among spatial features in the local regions. However, most studies only divide the functional regions through density of instances, ignoring the spatial correlation within, which makes the identification results biased towards a higher number of instances (such as restaurants, convenience stores, etc.), and may not present the functional characteristics of regional differences effectively. In the stage of RCPM, we propose a new algorithm for mining regional co-location patterns. By using the method of representation learning to extract the feature vectors of POI types with the help of the word embedding model, and then the functional areas of the city are divided. This method uses word vector to represent the semantic information of words, so that semantically similar words are close to each other in the representation space, and the division of regions is more reasonable. Compared to the existing algorithms, our method demonstrates a greater potential, as evidenced by experimental results.

From the perspective of optimizing maritime logistics, a key focus in the field of maritime information research has been how to extract behavioral patterns and deep behavioral characteristics of vessels from vast amounts of shipping statistics. Additionally, aligning these characteristics with infrastructure such as berths for effective association and recommendation to vessels is a critical requirement for the evolution of intelligent maritime systems. Traditional methods primarily focus on the behavioral trajectories of vessel navigation, failing to explore the geographical interconnections between vessels and port infrastructure. In light of this, this paper proposes a framework for deep mining of shipping information based on knowledge graph technology. Utilizing AIS data and spatial data of port facilities, it constructs a semantic relationship in the form of triplets between vessels, berths, and waterways, and semantically models vessel behaviors. Effective identification of vessels is achieved based on various semantic information. Simultaneously, based on the berthing semantic relationship between vessels and berths, a reverse semantic knowledge graph of berths is constructed with respect to vessel type, size, and class. This study compares different graph embedding methods, dimensionality reduction techniques, and classification approaches to achieve optimal experimental results. The findings indicate that the vessel type recognition accuracy in the proposed framework reached 83.1%, and the number of Identical Relationships between the recommended and original berths in similar berth recommendations was 3.755. The experiments demonstrate that the framework can provide a technical foundation for deep mining of vessel behavior, vessel type identification, and berth recommendation, as well as a semantic basis for large-scale maritime models.

As the typical representative of marine big data, the Argo plan conducts high-quality and scientific anomaly detection on Argo data, which is an important step in ocean science big data. However, in classical anomaly algorithms, Argo anomaly detection mostly has low accuracy, poor efficiency, and neglects the spatial continuity of Argo data. In the research on anomaly detection of spatial and regional data, graph anomaly detection has achieved excellent results. In the research of graph anomaly detection, depth based classification as a downstream anomaly detection method performs well, but at the same time, there are also problems of hyper sphere collapse and performance flipping. This article focuses on the research work related to the above issues: (1) Based on the study of Argo data and graph data, combined with the three-dimensional spatial characteristics of Argo buoy data, a novel graph data construction method is proposed. (2) Propose to incorporate neural transformation learning into the architecture, improve data learning expression ability, and further improve the shortcomings of graph neural classification, enabling it to adapt to the spatiotemporal and multi-dimensional characteristics of Argo buoy data for outlier detection. This article conducts experiments on five simulation datasets to demonstrate that the improved idea outperforms five state-of-the-art graph anomaly detection algorithms in various indicators, successfully improving the problems of hyper sphere collapse and performance flipping, and enhancing the accuracy and robustness of graph anomaly detection; The effectiveness of graph construction was demonstrated by comparing it with classical anomaly algorithms on real Argo sample data.

Computing shortest distance estimation for road networks is an important component of map service systems. Existing embedded-based shortest path distance estimation methods either have a long training time or the model training time is reduced by sacrificing the estimation accuracy. To address the above problems, this paper proposes a Road Graph Convolutional Networks and Distance2Vector (RGCNdist2vec), which is suitable for road network scenarios, as an embedding method of road network vertices. Used to capture network structure information. In the aspect of sampling model training samples, a three-stage sampling method based on graph logical partition is designed, which can select a small number of high-quality samples for model training. In order to verify the validity of the model and sampling scheme, experiments were carried out on four real road network datasets and compared with existing relevant models. The results show that the proposed model has high estimation accuracy, and the training time of the model is nearly 4 times lower than that of the existing baseline model.

Community search in heterogeneous information network (CSH) based on deep learning methods has received increasing attention. However, almost all the existing methods are semi-supervised learning paradigms, and the learning models based on meta path only consider the end-to-end relationship of meta path, ignoring the intermediate information of meta path. To address these issues, a CSH method based on Self-supervised Graph Neural Network (SGNN) is proposed. The model training is self-supervised by contrastive learning between the network schema view and the meta path view, and the two views capture the local and global information of the meta path from different angles. We then introduce a greedy algorithm called \(k{\text{-}}core\) and \({\mathcal{K}}{\text{-}}sized\) attribute-scores maximization community search (\(k{\mathcal{K}}{\text{ - ASMcs}}\)) to explore target communities. A large number of experiments on real datasets have verified the effectiveness and efficiency of the proposed method.

Studying spatial behavioral conflicts is the main method to understand the conflicts between tourists and residents. However, academic research on spatial conflict mostly involves urban macro-level discussions, and is not deeply involved in micro-scale spatial conflict. And most of them analyze spatial conflict from a qualitative perspective, lacking quantitative research on tourism spatial conflict. In this paper, we identify the spatial conflict areas in the township and quantitatively analyze the structural characteristics of the conflict areas through the method of multi-intelligence body simulation.

The safety monitoring of hydraulic structures is an important measure to ensure the safe construction and operation of water diversion projects. The traditional data analysis and prediction of water conservancy monitoring mostly uses geometric models, and the accuracy of short-term prediction is reasonable, while the accuracy of long-term prediction is greatly reduced. Moreover, the traditional time series analysis method only considers the temporal correlation of the monitoring time series, but does not consider the spatial correlation between the multivariate monitoring time series, and can not make full use of the spatio-temporal correlation information between the multivariate monitoring data. To solve the above problems, this paper proposes a spatio-temporal prediction method, ARIMA-b-DLSSVM, which integrates multiple time series. The model is based on least square support vector machine (LSSVM) for multivariate data fusion, auto-regressive integral moving average (ARIMA) model for trend extraction, bisquare spatial basis to establish spatial correlation of monitoring data, and discounted least square method (DLS) for model optimization. The results show that the accuracy of ARIMA-b-DLSSVM long-term prediction is higher than that of traditional model and single machine learning model. The spatio-temporal fusion of multivariate data can better predict the spatio-temporal sequence changes of diversion tunnels with large fluctuations.

The transformer-based method is a popular choice for medium and long-term traffic prediction. However, it still suffers from some problems. The first is that spatial position embedding has poor interpretability. Additionally, the spatial-temporal correlation learning can struggle to reflect the actual complexity of traffic networks relationships. To address the above problems, we propose a traffic prediction framework for dynamic adaptive spatial-temporal graph transformer (DyAdapTransformer). Our method uses the method of random walk to embed the spatial position. The analyzability between transition probability and spatial position representation enhances the interpretability of the model. When learning spatial-temporal correlation, a method of dynamic adaptive graph attention network is proposed. We compared with our framework with four baselines on three datasets. The results show that DyAdapTransformer has a better predictive performance.

Spatiotemporal prediction is a research topic in urban planning and management. Most existing spatiotemporal prediction models currently face challenges. More specifically, most prediction models are sensitive to missing data, meaning most prediction models are only tested on spatiotemporal data assuming no missing data. Although missing data can be imputed, spatiotemporal prediction models with the capability of handling missing data are needed. In this study, we propose a novel missing-data-tolerant causal graph attention model called CGATM to address the above challenges. To enable the CGATM model to be tested on spatiotemporal data with missing data, we propose a novel missing data handling mechanism that automatically handles missing data according to the probability of data missing patterns. To improve the nonlinear fitting ability of the CGATM model, we propose a novel causal graph attention method that represents geospatial heterogeneity by adjacent nodes with different weights. In addition, we design the CGTAM model as an Imputer-Predictor architecture and define a novel loss function to optimize model parameters. The proposed model was validated on three real-world spatiotemporal datasets (traffic dataset, PM2.5 dataset, and temperature dataset). Experimental results showed that the proposed model has better prediction performance under four missing scenarios, and outperforms eight existing baselines regarding prediction accuracy.

The problem of small sample classification is to identify image categories that have not appeared in the training concentration when marking the scarce sample samples of the training data set. Such tasks are of great significance in the recognition of remote sensing scenarios. It is a problem worth studying in this field. As we all know, training a deep learning model for classification requires a considerable labeling data set, which makes the production of training data sets huge. In this article, we propose a MADB feature extraction model based on Mixed Attention Module as a base model to extract features. Using RccaEMD module as the measurement algorithm to distinguish the classification of remote sensing scenarios. In NWPU-RESISC45 dataset, AID dataset, and UC-Merced dataset, it proves that our method has achieved higher accuracy than the current advanced methods of this field.

With the emergence of a large number of remote sensing data sources, how to effectively use the useful information in multi-source data for better earth observation has become an interesting but challenging problem. In this paper, the deep learning method is used to study the joint classification of hyperspectral imagery (HSI) and light detection and ranging (LiDAR) data. The network proposed in this paper is named convolutional neural network based on multiple attention mechanisms (MatNet). Specifically, a convolutional neural network (CNN) with an attention mechanism is used to extract the deep features of HSI and LiDAR respectively. Then the obtained features are introduced into the dual-branch cross-attention fusion module (DCFM) to fuse the information in HSI and LiDAR data effectively. Finally, the obtained features are introduced into the classification module to obtain the final classification results. Experimental results show that our proposed network can achieve better classification performance than existing methods.

Few-shot learning image classification (FSLIC) is a task that has gained enhanced focus in recent years, the cost of collecting and annotating large number of data samples in some specialised domains is expensive, Few-shot remote scene classification (FRSSC) is of great utility in scenarios where sample is scarce and labelling is extremely costly, the core problem of this task is how to identify new classes with scarce and expensive few-shot samples. However, existing work prefers complicated feature extraction in various ways and the enhancement results are not satisfactory, this paper aims to improve the effectiveness of FSLIC not only through complicated feature extraction but also by exploring alternative approaches. Here are multiple avenues to improve the performance of few-shot classifiers. Training with a scarce data in a few-shot learning (FSL) task often results in a biased feature distribution. In this paper, we propose a method to address this issue by calibrating the support set data feature using sufficient base class data. (Our data distribution calibration method (DC) is on top of feature extractor), requiring no additional parameters. And the feature extraction model is further optimised and the feature extractor of DC-2DEC is optimised with the task of dealing with the spatial context structure of the image i.e. rotation prediction pretext, specifically rotation prediction. We refer to the proposed method as DC-2DEC, and we apply it to few-shot learning classification in RS image (RS image) scene recognition. Through experiments conducted on traditional few-shot datasets and RS image datasets, we validate the algorithm and present corresponding experimental results. These results demonstrate the competitiveness of DC-2DEC, highlighting its efficacy in few-shot learning classification for RS images.

High-definition (HD) map is a fundamental component of autonomous driving systems, as it can provide precise environmental information about driving scenes. Recent work on vectorized map generation could produce merely \(65\%\) local map elements around the ego-vehicle at runtime by one tour with onboard sensors, leaving a puzzle of how to construct a global HD map projected in the world coordinate system under high-quality standards. To address the issue, we present GNMap as an end-to-end generative neural network to automatically construct HD maps with multiple vectorized tiles which are locally produced by autonomous vehicles through several tours. It leverages a multi-layer and attention-based autoencoder as the shared network, of which parameters are learned from two different tasks (i.e., pretraining and finetuning, respectively) to ensure both the completeness of generated maps and the correctness of element categories. Abundant qualitative evaluations are conducted on a real-world dataset and experimental results show that GNMap can surpass the SOTA method by more than \(5\%\) F1 score, reaching the level of industrial usage with a small amount of manual modification. We have already deployed it at Navinfo Co., Ltd., serving as an indispensable software to automatically build HD maps for autonomous driving systems.

The high privacy and accuracy of trajectory data make data integrity and security critical. However, existing algorithms cannot ensure the integrity of temporal attributes and resist common operations in the normal use of data, which limits the use of data. For this reason, this paper proposes a semi-fragile water-marking algorithm for trajectory data that takes into account spatiotemporal features. The proposed algorithm utilizes the minimum area bounded rectangle (MABR) to group the trajectory data. Finally, the fragile watermarks generated from spatial and temporal attributes are embedded into an embedded domain with geometric invariance in stages using the multiple quantization index modulation (MQIM) technique. Experimental results show that the proposed algorithm is extremely robust to geometric attacks. Meanwhile, it can accurately identify whether the temporal and spatial attributes have been tampered with and the type of tampering. The proposed algorithm balances robustness and tampering detection capability, providing a feasible solution for the security protection of trajectory data.

The prevention and crackdown of fraud calls have been paid more and more attention by industrial and academic societies. Most current researches based on machine learning ignore the imbalanced data distribution characteristic between normal and fraudulent call users, and the outputs neglect the probability fluctuation range of the suspected fraudulent calls. To overcome these limitations, we first construct user behavioral feature vector by a random forest method. Secondly, we propose a novel hierarchical sampling method to overcome the class imbalance problem. Thirdly, we propose a novel fraud call recognition method based on HPO-LGBM (the Bayesian hyper parameter optimization based on random forest and Light Gradient Boosting Machine). Finally, we further evaluate the method’s performance with a DRI (dynamic recognition interval) model. Experimental results on public datasets show that the proposed HPO-LGBM holds a 92.90% F1 value, a 91.90% AUC, a 92.92% G-means, and a 92.37% MCC in fraud call recognition. In addition, the proposed HPO-LGBM model can further give the dynamic recognition interval of the output result, behaving more robust than other models (i.e., LR, RF, MLP, GBDT, XGBOOST, LGBM).

The rapid development of information technology has brought about the emergence of urban multi-source big data and the improvement of computing power, and promoted the emergence of new technologies such as artificial intelligence, making the study paradigm of urban modelling face change. Based on the concept of smart city, this paper sorts out the types of urban big data in the information age, and puts forward new opportunities and challenges brought by big data to urban modelling research; Secondly, this paper lists a number of cases for reference on how big data can provide services to the public in the form of network and infrastructure; Finally, the latest progress in the innovation of new technologies such as artificial intelligence, deep learning, data mining, etc. caused by the improvement of computing power is discussed.