The urban real estate market is shaped by spatially varying environmental and social determinants, such as the valuation of green spaces, proximity to transport, and distance to central business districts. Among all the spatially varying relationships between prices and housing characteristics, some tend to vary at a global spatial scale, whereas others vary at a local spatial scale. This study applies a random model to specify multi-scale spatial heterogeneity in the rental housing market by utilizing residential rent data in the Tokyo metropolitan area from 2017. The results show that spatially varying determinants impact rental housing prices at the global, moderate, and local scales. Further, we find that the estimation is flexible because the random model determines the spatial scale of each regression coefficient.
Increasing in-town bicycle traffic creates a demand for safe and efficient transportation infrastructure. A significant safety aspect is crossroad layout. Existing solutions such as protected crossroads, roundabouts and standard four-way crossings are investigated in terms of viewing angles between traffic participants. An agent-based simulation helps to generate data, which is further analysed. Special attention is paid to blind spots of vehicles during turns, overall line of sight and human field of view. We can show that especially protected crossroad designs have major advantages. Standard layouts convince in terms of the analysed field of view and possible blind spots. However, they demand extensive shoulder views and head turning especially during right turns. This makes them less safe. Roundabouts show medium results. Exiting this structure always requires a right turn which is, in terms of visibility, the most dangerous action for bicycles. We conclude that protected crossroads can be recommended as the safest approach in comparison to standard and roundabout layouts. Yet, space requirements may restrict in-town realization of this design.
While eXplainable Artificial Intelligence (XAI) has significant potential to glassbox Deep Learning, there are challenges in applying it in the domain of Geospatial Artificial Intelligence (GeoAI). A land use case study highlights these challenges, which include the difficulty of selecting reference data/models, the shortcomings of gradients to serve as explanation, the limited semantics and knowledge scope in the explanation process of GeoAI, and underlying GeoAI processes that are not amenable to XAI. We conclude with possibilities to achieve Geographical XAI (GeoXAI).
In this paper, we present the algorithm aimed to segment the type of geographical space where points are a substantial component. The research problem falls within the mainstream of Automated Unit Design (AUD). The objective function of the solution is a balance between the size of segmented units expressed as an attribute of points datasets and its agreement with constraints provided by the geographic space. An algorithm has three free parameters; two of them allow one to control the objective function: the size of segmented units and allowable deviation from the size. The paper contains a case study where we show how our approach segment the geographic space of the City of Poznań.
The use of machine learning models (ML) in spatial statistics and urban analytics is increasing. However, research studying the generalisability of ML models from a geographical perspective had been sparse, specifically on whether a model trained in one context can be used in another. The aim of this research is to explore the extent to which standard models such as convolutional neural networks being applied on urban images can generalise across different geographies, through two tasks. First, on the classification of street frontages and second, on the prediction of real estate values. In particular, we find in both experiments that the models do not generalise well. More interestingly, there are also differences in terms of generalisability within the first case study which needs further exploration. To summarise, our results suggest that in urban analytics there is a need to systematically test out-of-geography results for this type of geographical image-based models.
Data related to households or addresses needs be published in an aggregated form to obfuscate sensitive information about individuals. Usually, the data is aggregated to the level of existing administrative zones, but these often do not correspond to formal models of privacy or a desired level of anonymity. Therefore, automatic privacy-preserving spatial clustering methods are needed. To address this need, we present algorithms to partition a given set of locations into k-anonymous clusters, meaning that each cluster contains at least k locations. We assume that the locations are given as a set T ⊆ V of terminals in a weighted graph G = (V, E) representing a road network. Our approach is to compute a forest in G, i.e., a set of trees, each of which corresponds to a cluster. We ensure the k-anonymity of the clusters by constraining the trees to span at leastterminals each (plus an arbitrary number of non-terminal nodes called Steiner nodes). By minimizing the total edge weight of the forest, we ensure that the clusters reflect the proximity among the locations. Although the problem is NP-hard, we were able to solve instances of several hundreds of terminals using integer linear programming. Moreover, we present an efficient approximation algorithm and show that it can be used to process large and fine-grained data sets.
For human-centered mobile navigation systems, a computational landmark selection model is critical to automatically include landmarks for communicating routes with users. Although some empirical studies have shown that landmarks selected by familiar and unfamiliar wayfinders, respectively, differ significantly, existing computational models are solely focused on unfamiliar users and ignore selecting landmarks for familiar users, particularly in indoor environments. Meanwhile, it is unclear how the importance of salience metrics employed by machine learning approaches differs from that reported by human participants during landmark selection. In this study, we propose a LambdaMART-based ranking approach to computationally modelling indoor landmark selection. Two models, one for familiar and one for unfamiliar users, respectively, were trained from the human-labelled indoor landmark selection data. The importance of different salience measures in each model was then ranked and compared with human participants’ self-report results of a survey. The evaluation results demonstrate that familiarity does indeed matter in the computational modelling of indoor landmark selection. The ranking differences of salience measures in the trained models show that the salience varies with the familiarity of wayfinders. Moreover, the calculated intraclass correlation coefficients (0.62 for familiar, 0.65 for unfamiliar) illustrate the median consistency between the computational results on feature importance and the self-reported importance results by human participants, confirming the reliability and interpretability of the proposed approach.
This study builds a bridge between the literatures for geographically weighted regression (GWR) and compositional data analysis (CoDA). GWR allows the modeling of spatial heterogeneity in regression models and is increasingly used in various fields. CoDA provides unique and useful tools for compositional data, which are restricted by a constant-sum constraint. Although compositional data are common in many scientific areas, it is not until recently that increasingly sophisticated statistical methods have been deeply investigated. Many types of spatial models based on geostatistics, spatial statistics, and spatial econometrics for compositional data have been proposed. However, there is less attention to both spatial heterogeneity and the constant-sum constraint. In this study, we propose a GWR model for compositional data. This allows us to model spatially varying relationships while considering the constant-sum constraint. We applied this model to analyze household income compositions at the county level in the US. The interpretational usefulness of the results of spatially varying compositional semi-elasticities is empirically performed.
This paper describes the novel development and application of a multi-scale geographically weighted discriminant analysis (MSGWDA). This is applied to a case study of survey data of attitudes to a proposed motorbike / scooter ban in Han Noi, Vietnam. It uses discriminant analysis to examine attitudes to the ban in relation to travel purposes, distances, respondent age and so on. The main part of the paper focuses on describing the novel MSGWDA approach, and the results indicate the varying scales of relationship between the different input variables and the categorical responses variable. The paper also reflects on the pervasive logic of the approaches used to fit multiscale geographically weighted bandwidths (for example in regression). These have historically been based on the iterative back-fitting approaches used in GAMs, but risk missing potentially important variable interactions amongst un-evaluated bandwidths because of the sequence of their application. It is argued that although pragmatic in the 1990s, it may be possible to apply more deterministic approaches with increased memory and readily accessible computing power in order to better navigate such highly dimensional search spaces.
Many classic exploratory data analysis tools in quantitative geography, designed to measure global and local spatial autocorrelation (e.g. Moran’s I statistic), have become standard in modern GIS software. However, there has been little development in amending these tools for visualization and analysis of patterns captured in spatiotemporal data. We design and implement a new open-source Python library, VASA, that simplifies analytical pipelines in assessing spatiotemporal structure of data and enables enhanced visual display of the patterns. Using daily county-level social distancing metrics during 2020 obtained from two different sources (SafeGraph and Cuebiq), we demonstrate the functionality of the developed tool for a swift exploratory spatial data analysis and comparison of trends over larger administrative units.
