A divide-and-conquer method for space–time series prediction

Space–time series can be partitioned into space–time smooth and space–time rough, which represent different scale characteristics. However, most existing methods for space–time series prediction directly address space–time series as a whole and do not consider the interaction between space–time smooth and space–time rough in the process of prediction. This will possibly affect the accuracy of space–time series prediction, because the interaction between these two components (i.e., space–time smooth and space–time rough) may cause one of them as dominant component, thus weakening the behavior of the other. Therefore, a divide-and-conquer method for space–time prediction is proposed in this paper. First, the observational fine-grained data are decomposed into two components: coarse-grained data and the residual terms of fine-grained data. These two components are then modeled, respectively. Finally, the predicted values of the fine-grained data are obtained by integrating the predicted values of the coarse-grained data with the residual terms. The experimental results of two groups of different space–time series demonstrated the effectiveness of the divide-and-conquer method.     

Building Agent‐Based Walking Models by Machine‐Learning on Diverse Databases of Space‐Time Trajectory Samples

We introduce a novel scheme for automatically deriving synthetic walking (locomotion) and movement (steering and avoidance) behavior in simulation from simple trajectory samples. We use a combination of observed and recorded real‐world movement trajectory samples in conjunction with synthetic, agent‐generated, movement as inputs to a machine‐learning scheme. This scheme produces movement behavior for non‐sampled scenarios in simulation, for applications that can differ widely from the original collection settings. It does this by benchmarking a simulated pedestrian's relative behavioral geography, local physical environment, and neighboring agent‐pedestrians; using spatial analysis, spatial data access, classification, and clustering. The scheme then weights, trains, and tunes likely synthetic movement behavior, per‐agent, per‐location, per‐time‐step, and per‐scenario. To prove its usefulness, we demonstrate the task of generating synthetic, non‐sampled, agent‐based pedestrian movement in simulated urban environments, where the scheme proves to be a useful substitute for traditional transition‐driven methods for determining agent behavior. The potential broader applications of the scheme are numerous and include the design and delivery of location‐based services, evaluation of architectures for mobile communications technologies, what‐if experimentation in agent‐based models with hypotheses that are informed or translated from data, and the construction of algorithms for extracting and annotating space‐time paths in massive data‐sets.     

基于支持向量机的GNSS时间序列预测

在深度学习的理论框架下,针对预测全球卫星导航系统(GNSS)时间序列,传统的经验风险最小化预测模型误差大精度低,泛化性能差且对历史数据的经验依赖大的问题.提出一种采用结构风险最小化原则的基于支持向量机(SVM)的时间序列预测模型.通过和多层的BP神经网络预测模型预测效果比较,结果证明SVM预测模型拥有更好的时间序列预测效果.     

Thinking about Space‐Time Connections: Spatiotemporal Scheduling of Individual Activities

This article presents a spatiotemporal model for scheduling applications that is driven by the events and activities individuals plan and manage every day. The framework is presented using an ontological approach where ontologies at different levels of generalization, e.g. domain, application, and task ontologies, are linked together through participation and inheritance relationships. S_Events are entered into a schedule as a new S_Entry, or modifications can be made to existing entries including reschedule, postpone, change location, and delete as schedules vary over time. These schedule updates are formalized through changes to planned start and end times and the planned locations of S_Entries are expressed using SWRL, a semantic web rule language. SWRL is also used for reasoning about schedule changes and the space‐time conflicts that can occur. The sequence of entries in a schedule gives rise to S_trajectories representing the locations that individuals plan to visit in order to carry out their schedule, adding an additional spatial element to the framework. A prototype Geoscheduler application maps S_Entries against a timeline, offering a spatiotemporal visualization of scheduled activities showing the evolution of a schedule over space‐time and affecting spatiotemporal accessibility for individuals.     

The Effect of Data Inaccuracy on Tests of Space‐Time Interaction

Space‐time event data are often subject to deficiencies in: (1) locational accuracy; (2), temporal accuracy; and (3) completeness. This work explores how these failings in the quality of input data may affect the results of global space‐time interaction tests. While previous work has partially investigated the impact of locational inaccuracy on the results of these tests, more work remains. The impacts of temporal inaccuracy and incomplete data reporting on the results of these tests remain completely unexplored. This study examines the influence of these problems individually and collectively, using a series of simulations. Findings demonstrate that even in cases of slight inaccuracy or underreporting, the consequences on results are potentially severe. Although the study is couched in terms of data inaccuracy, its relevance to situations where inaccuracy is replaced with uncertainty is self‐evident.     

A Space‐Time Raster GIS Data Model for Spatiotemporal Analysis of Vegetation Responses to a Freeze Event

Many past space‐time GIS data models viewed the world mainly from a spatial perspective. They attached a time stamp to each state of an entity or the entire area of study. This approach is less efficient for certain spatio‐temporal analyses that focus on how locations change over time, which require researchers to view each location from a temporal perspective. In this article, we present a data model to organize multi‐temporal remote sensing datasets and track their changes at the individual pixel level. This data model can also integrate raster datasets from heterogeneous sources under a unified framework. The proposed data model consists of several object classes under a hierarchical structure. Each object class is associated with specific properties and behaviors to facilitate efficient spatio‐temporal analyses. We apply this data model to a case study of analyzing the impact of the 2007 freeze in Knoxville, Tennessee. The characteristics of different vegetation clusters before, during, and after the 2007 freeze event are compared. Our findings indicate that the majority of the study area is impacted by this freeze event, and different vegetation types show different response patterns to this freeze.     

Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting

Traffic forecasting is a challenging problem due to the complexity of jointly modeling spatio‐temporal dependencies at different scales. Recently, several hybrid deep learning models have been developed to capture such dependencies. These approaches typically utilize convolutional neural networks or graph neural networks (GNNs) to model spatial dependency and leverage recurrent neural networks (RNNs) to learn temporal dependency. However, RNNs are only able to capture sequential information in the time series, while being incapable of modeling their periodicity (e.g., weekly patterns). Moreover, RNNs are difficult to parallelize, making training and prediction less efficient. In this work we propose a novel deep learning architecture called Traffic Transformer to capture the continuity and periodicity of time series and to model spatial dependency. Our work takes inspiration from Google’s Transformer framework for machine translation. We conduct extensive experiments on two real‐world traffic data sets, and the results demonstrate that our model outperforms baseline models by a substantial margin.     

Integrating domain knowledge and graph convolutional neural networks to support river network selection

Deep learning is increasingly being used to improve the intelligence of map generalization. Vector-based map generalization, utilizing deep learning, is an important avenue for research. However, there are three questions: (1) transforming vector data into a deep learning data paradigm; (2) overcoming the limitation of the number of samples; and (3) determining whether existing knowledge can accelerate deep learning. To address these questions, taking river network selection as an example, this study presents a framework integrating hydrological knowledge into graph convolutional neural networks (GCNNs). This framework consists of the following steps: constructing a dual graph of river networks (DG_RN), extracting domain knowledge as node attributes of DG_RN, developing an architecture of GCNNs for the selection, and designing a fine-tuning rule to refine the GCNN results. Experiments show that our framework outperforms existing machine learning and traditional feature sorting methods using different datasets and achieves good morphological consistency after the selection. Furthermore, these results indicate that DG_RN meets the data paradigm of graph deep learning, and the framework integrating existing characteristics (i.e., Strahler coding, the number of tributaries, the distance between proximity rivers, and upstream drainage area) mitigates the dependence of GCNNs on plenty of samples and enhance its performance.     

Constructing a Near Real‐time Space‐time Cube to Depict Urban Ambient Air Pollution Scenario

This study adopts a near real‐time space‐time cube approach to portray a dynamic urban air pollution scenario across space and time. Originating from time geography, space‐time cubes provide an approach to integrate spatial and temporal air pollution information into a 3D space. The base of the cube represents the variation of air pollution in a 2D geographical space while the height represents time. This way, the changes of pollution over time can be described by the different component layers of the cube from the base up. The diurnal ambient ozone (O₃) pollution in Houston, Texas is modeled in this study using the space‐time air pollution cube. Two methods, land use regression (LUR) modeling and spatial interpolation, were applied to build the hourly component layers for the air pollution cube. It was found that the LUR modeling performed better than the spatial interpolation in predicting air pollution level. With the availability of real‐time air pollution data, this approach can be extended to produce real‐time air pollution cube is for more accurate air pollution measurement across space and time, which can provide important support to studies in epidemiology, health geography, and environmental regulation.     

城市扩展极限学习机模型

城市空间结构及其扩展的模拟是城市科学管理与规划的重要前提,本文基于极限学习机提出了顾及不同非城市用地转化为城市用地差异与强度的城市扩展元胞自动机模型（ELM-CA）。模型验证表明：①ELM-CA模型的模拟精度达到70.30%,相比于逻辑回归和神经网络分别提高了2.21%和1.54%,FoM系数分别提高了0.025 9和0.017 9,Kappa系数分别提高了0.024 7和0.016 9,且Moran I指数接近于实际值,说明极限学习机模型较逻辑回归和神经网络能更有效模拟城市扩展的空间形态及其变化;②ELM模型的训练时间仅为神经网络的1/3左右,体现了ELM学习速度的优势;③在小样本情况下,逻辑回归和神经网络都受到明显的影响,而极限学习机还能保持良好的性能,这个特点使其在样本难以获取的情况下具有明显的优势。两个时相的城市扩展模拟与真实数据的比较表明：基于极限学习机的城市扩展元胞自动机模型（ELM-CA）,简化了CA模型的复杂度,并在小样本情况下能有效提高模拟精度,适合于复杂土地利用条件下城市扩展模拟与预测。     

数据驱动的定量遥感研究进展与挑战

定量遥感是从原始遥感观测信息中定量推算或反演出地学参量的理论与方法.传统定量遥感主要基于模型驱动,强调通过数学或物理模型完成推算和反演.随着人工智能技术的发展和普及,数据驱动的方式也逐渐受到广泛关注,其强调的是通过机器学习等方式挖掘遥感观测数据中所包含的信息,完成地学参量的定量反演.在强大计算能力的支持下,数据驱动的方...     

Visualising Crime Clusters in a Space‐time Cube: An Exploratory Data‐analysis Approach Using Space‐time Kernel Density Estimation and Scan Statistics

For an effective interpretation of spatio‐temporal patterns of crime clusters/hotspots, we explore the possibility of three‐dimensional mapping of crime events in a space‐time cube with the aid of space‐time variants of kernel density estimation and scan statistics. Using the crime occurrence dataset of snatch‐and‐run offences in Kyoto City from 2003 to 2004, we confirm that the proposed methodology enables simultaneous visualisation of the geographical extent and duration of crime clusters, by which stable and transient space‐time crime clusters can be intuitively differentiated. Also, the combined use of the two statistical techniques revealed temporal inter‐cluster associations showing that transient clusters alternatively appeared in a pair of hotspot regions, suggesting a new type of “displacement” phenomenon of crime. Highlighting the complementary aspects of the two space‐time statistical approaches, we conclude that combining these approaches in a space‐time cube display is particularly valuable for a spatio‐temporal exploratory data analysis of clusters to extract new knowledge of crime epidemiology from a data set of space‐time crime events.     

利用SVM与ARIMA组合模型进行大坝变形预测

由于大坝位移时间序列数据受各种复杂因素的影响,具有非平稳和非线性等特征,因此,利用传统、单一的时间序列预测模型较难准确地描述大坝位移变形的复杂规律。综合考虑大坝位移时间序列非线性和线性特征,本文提出了一种SVM和ARIMA相结合的时间序列预测模型。将大坝变形的时间序列分为非线性部分和线性部分。针对非线性部分,利用SVM进行滚动预测,并与NAR动态神经网络进行对比,试验表明SVM处理非线性问题具有相对的优势;针对线性部分,通过ARIMA模型对其进行单步滚动预测,综合两项预测结果得到组合模型的预测值。结合大坝实测资料对组合模型进行检验,试验结果表明,SVM-ARIMA组合模型的预测精度高,能更好地描述大坝位移的变化趋势,具有一定的实用价值。     

顾及空间异质性的大范围地面沉降时空预测

大速率、不均匀的地面沉降已经威胁到人类的生产活动,高精度的沉降预测结果对于地质灾害的精准防控具有重要意义。为掌握地面沉降的演化规律,利用现场观测数据或InSAR数据开展了多项预测研究。然而,由于空间异质性的存在,大范围地面沉降的准确预测仍然是一项挑战。在这项研究中,从数据驱动的角度提出了一种顾及空间异质性的大范围地面沉降时空预测方法 STLSTM (Spatio-temporal Long Short-Term Memory)。首先,通过聚类识别地理空间中的均质子区;然后,在每个子区中,一个特别的长短期记忆LSTM (Long ShortTerm Memory)网络被用来捕捉局部位置的非线性特征;最后,利用预训练的网络对未来时刻的地面沉降进行定量预测。在实验部分,哨兵1号影像数据被用来比较STLSTM与其他8种基准方法的性能,利用空间统计指标分析了模型的有效性。结果表明,STLSTM在152 s内达到了最高的预测精度（71.4%）,且能够有效弱化空间异质性对大区域沉降预测任务的影响。总之,这项研究将空间异质性处理策略融合到深度学习模型中,实现了高精度、高时效的大范围地面沉降时空预测。     

A GIS Toolbox for Measuring and Mapping Person‐Based Space‐Time Accessibility

This article presents a toolbox to compute and map person‐based accessibility indicators, based on classical time geography concepts. The intent is to provide GIS and urban planning practitioners with a user‐friendly and easily customizable tool. While it relies on well‐known concepts, the toolbox implements a major innovation in person‐based accessibility assessment by taking into account opening hours when measuring the accessibility of urban facilities. The toolbox can be downloaded from http://bit.ly/1h6yg5Z .     

Prediction of Earth orientation parameters by artificial neural networks

 Earth orientation parameters (EOPs) [polar motion and length of day (LOD), or UT1–UTC] were predicted by artificial neural networks. EOP series from various sources, e.g. the C04 series from the International Earth Rotation Service and the re-analysis optical astrometry series based on the HIPPARCOS frame, served for training the neural network for both short-term and long-term predictions. At first, all effects which can be described by functional models, e.g. effects of the solid Earth tides and the ocean tides or seasonal atmospheric variations of the EOPs, were removed. Only the differences between the modeled and the observed EOPs, i.e. the quasi-periodic and irregular variations, were used for training and prediction. The Stuttgart neural network simulator, which is a very powerful software tool developed at the University of Stuttgart, was applied to construct and to validate different types of neural networks in order to find the optimal topology of the net, the most economical learning algorithm and the best procedure to feed the net with data patterns. The results of the prediction were analyzed and compared with those obtained by other methods. The accuracy of the prediction is equal to or even better than that by other prediction methods. Received: 6 February 2001 / Accepted: 23 October 2001     

Mapping hourly dynamics of urban population using trajectories reconstructed from mobile phone records

Understanding the spatiotemporal dynamics of urban population is crucial for addressing a wide range of urban planning and management issues. Aggregated geospatial big data have been widely used to quantitatively estimate population distribution at fine spatial scales over a given time period. However, it is still a challenge to estimate population density at a fine temporal resolution over a large geographical space, mainly due to the temporal asynchrony of population movement and the challenges to acquiring a complete individual movement record. In this article, we propose a method to estimate hourly population density by examining the time‐series individual trajectories, which were reconstructed from call detail records using BP neural networks. We first used BP neural networks to predict the positions of mobile phone users at an hourly interval and then estimated the hourly population density using log‐linear regression at the cell tower level. The estimated population density is linearly correlated with population census data at the sub‐district level. Trajectory clustering results show five distinct diurnal dynamic patterns of population movement in the study area, revealing spatially explicit characteristics of the diurnal commuting flows, though the driving forces of the flows need further investigation.     

A geospatial case‐based reasoning model for oil–gas reservoir evaluation

Delineating the distribution of oil and natural gas resources is a prerequisite of exploitation. The delineation methods usually include conventional techniques of reservoir evaluation and mathematical models. The conventional reservoir evaluation results generally depend on the experts' knowledge and experience in the field. The mathematical methods mostly require accurate models to be proposed. Considering spatial relationships and characteristics of geological reservoir problems, including nonlinearity, complexity and uncertainty, a novel model called geospatial case‐based reasoning for oil–gas reservoir evaluation was proposed in this article. The key components of the new model, including: (1) the joint representation of spatial relationship and attribute features; (2) the model of spatial relationship and attribute similarity joint reasoning; and (3) the methods of establishing weights for the spatial relationship and attribute features, are completely constructed. A case study of the proposed model for gas reservoir evaluation was carried out. Compared with the backpropagation artificial neural network (BP‐ANN) and the geological empirical evaluation (GEE) methods, the model presented in this article performs 6.38% and 46.81% better than BP‐ANN and GEE, respectively. Furthermore, its execution is simpler, more convenient, and importantly, its utilization hardly requires any professional knowledge of the field.     

A multi‐label cellular automata model for land change simulation

The use of cellular automata (CA) has for some time been considered among the most appropriate approaches for modeling land‐use changes. Each cell in a traditional CA model has a state that evolves according to transition rules, taking into consideration its own and its neighbors’ states and characteristics. Here, we present a multi‐label CA model in which a cell may simultaneously have more than one state. The model uses a multi‐label learning method—a multi‐label support vector machine, Rank‐SVM—to define the transition rules. The model was used with a multi‐label land‐use dataset for Luxembourg, built from vector‐based land‐use data using a method presented here. The proposed multi‐label CA model showed promising performance in terms of its ability to capture and model the details and complexities of changes in land‐use patterns. Applied to historical land use data, the proposed model estimated the land use change with an accuracy of 87.2% exact matching and 98.84% when including cells with a misclassification of a single label, which is comparably better than a classical multi‐class model that achieved 83.6%. The multi‐label cellular automata outperformed a model combining CA and artificial neural networks. All model goodness‐of‐fit comparisons were quantified using various performance metrics for predictive models.     

基于RNN的空气污染时空预报模型研究

针对空气污染物时间序列中包含缺失值以及现有时间序列预报模型缺乏对时序特征状态建模的问题,该文构建了基于缺失值处理算法和RNN(循环神经网络)的时空预报框架。对空气污染物时序数据设计了3种缺失值处理算法(前向递补、均值替代和权重衰减),用缺失标签和缺失时长对缺失值建模,并在此基础上搭建含有全连接层与LSTM层的深度循环神经网络(DRNN)用于时空预报。使用深度全连接神经网络(DFNN)作为DRNN的对照,用京津冀区域的空气质量和气象数据训练模型,并比较不同模型的预测精度。通过实验,比较了3种缺失值处理方法的效果,结果表明,LSTM在空气污染时空序列预测上的表现优于传统的全连接神经网络层,证实了提出的基于深度学习的时空预报框架的有效性。