Spatiotemporal Hotspot Analysis for Movement Data

Abstract

The cornerstone for cluster analysis was already laid in the 19th century, when John Snow investigated the cause of a Cholera outbreak in Soho, London. The COVID crisis has shown, that today, more than 150 years later, the need for spatial cluster analysis and hotspot detection is as relevant as ever. Rooted in geography, hotspot analysis is often focused merely on the spatial aspect of the problem. But as the characteristics and intensity of a hotspot may vary over time, it is just as important to include the temporal aspect as well. Therefore, the need for spatio-temporal methods to investigate clustering problems arises. Since there have been first approaches to tackle the problem on the plane, the present thesis investigates these methods and their applicability to research problems with data associated to networks. To investigate the existing methodological toolkit, various approaches to spatial, temporal and spatio-temporal statistics and methods are introduced that have primarily been designed for research problems on the plane. To see if they can be validly transferred to a network shaped problem, notions such as point patterns on networks, network spatial autocorrelation and distance concepts on networks are examined. The proposed approach to investigate spatio-temporal problems on networks is a two-stage process. First, the time series for every link is split into time slices of equally sized time intervals. Then, for each of those slices a spatial hotspot analysis is performed. This results in a time series for each link, characterizing every time slice as no point of interest or hotspot / coldspot with significance levels 1%, 5% or 10%. After that, for every link its corresponding time series is subjected to a pattern detection process that includes a trend test to identify diminishing, persistent and intensifying hotspots. In total, nine different patterns are defined that can be detected this way. For the convenience of reducing the workload of investigation, during this thesis a Python tool was developed, that handles the above-described process. Another benefit, that results from the established tool is, that it offers to possibility for visualisation of the results. Finally, the proposed approach and the corresponding Python implementation were tested in form of a case study on taxi movement data in the city of Vienna to investigate a possible shift of hotspots caused by the launch of the new main train station Wien Hauptbahnhof. The data was provided by the data owner "Taxi 31300" and the Austrian Institute of Technology. The results showed that the implemented approach indeed identified points of interested in terms of hotspot detection. On the other hand, the formulated hypothesis of a hotspot shift could only be confirmed to some extent. Unfortunately, a couple of flaws in the Python packages that were used for the implementation were discovered during the evaluation process. These were reported to and acknowledged by the developers. In conclusion, it became apparent during the analysis of the case study, that the spatio-temporal hotspot analysis tool can only assist and support the user by providing a framework for their analysis. There are still certain parameters that have to be decided on by the researcher and therefore are exposed to subjectivity. This may consequently lead to a distortion of the results. The implemented tool for spatio-temporal hotspot analysis represents a first prototype that aims at the analysis of a specific data source. Additional work is necessary to enhance this prototype to a more sophisticated and user- friendly implementation that can be made publicly available and be effectively used for spatio-temporal hotspot analysis on networks.