Introduction and Motivation
Travel demand models are an essential tool in transport planning because they are used to estimate the impacts of changes in transport supply, land-use developments, and related interventions. Most travel demand models are based on the logic of the four-step model, according to which travel demand emerges from four interrelated decisions. In the first step, individuals decide which activity (e.g. work, education, daily shopping) they wish to perform and at what time. Based on this, they choose a destination for that activity. In the third step, they select a transport mode, and finally, they determine a route along the road network or via public transport. [1]
Travel demand models require large amounts of data. These include travel behavior and socio-demographic data, but also information on travel times between origins and destinations and, crucially, on where destinations are located. OpenStreetMap (OSM) has become a popular data source in transport modeling to retrieve street and path networks for routing and travel time calculations. In recent years its role has expanded towards a second and equally important question: where are the destinations that generate trips, and how attractive are they for different activity types? A small greengrocer, for example, should not be assumed to attract the same number of trips as a large supermarket.
Traditionally, such data have been collected manually or purchased from commercial providers. While such data products may offer high quality, their sources and processing steps are non-transparent. Moreover, data for an entire study area can be expensive. OSM has therefore become an attractive alternative. Yet current use of OSM in travel demand modeling remains heterogeneous and is rarely documented. The aim of this study is to show how domain-specific value can be created from OSM data in transport science by presenting a transparent and reproducible data processing pipeline for deriving destination attractiveness data. The workflow has been implemented in Python and R and is available as the open-source tool OSM2Attractiveness[2].
Methodology
The method consists of several consecutive steps and is illustrated in Figure 1. First, OSM elements are assigned to one or more activity types. Activity types categorize similar occasions and contexts of activities, such as work, leisure, etc. For this purpose, a tag list is created for each activity type (e.g. Figure 2), and the OSM data are filtered accordingly.
Second, the area and exact type of the resulting points of interest (POI) are determined which is necessary for the estimation of trip attraction [1]. If a way or relation represents a building, the area can be calculated from its polygon; if the number of floors is available, the floor area can also be derived. If an activity-related area is mapped outside buildings, such as a zoo or playground, the land area may already be a suitable measure. In other cases, especially when activities are usually conducted inside buildings, land area alone is insufficient, and the methodology therefore assigns buildings within the area and derives their usable floor space, for example for hospitals or schools. Specific cases must also be considered when many destinations are located within a larger destination complex, e.g., multiple shops within a shopping mall.
If a location is mapped only as a node, attractiveness may still be estimated by considering the exact type of the POI. Where a space-related measure is required, e.g., for shopping destinations, two fallback strategies can be applied: using the surrounding building area or assigning a typical average size based on the store name or brand.
Third, attractiveness respectively trip attraction is calculated, which is defined as the expected number of trip ends attracted by a destination per day. Simply using floor areas is not sufficient because different facility types generate different amounts of travel. A car dealership generates substantially fewer trips per square meter than a supermarket. The calculation requires trip attraction rates that describe the relationship between POI counts or floor space of an exact type and trip ends. In Germany, one relevant source is VerBau [3]; comparable datasets exist for the United States, Great Britain, Australia, New Zealand, and other countries [4-6]. It should be noted that the calculated attractiveness is not the actual number of daily visitors. The real number of trips to a destination additionally depends on the distance between trip origins and other destinations within the gravity-based destination choice model [7]. In the end the results are aggregated on the level of traffic analysis zones (TAZ).
Findings
The methodology has been applied successfully in several German case studies, including Hamburg, Karlsruhe, and Stuttgart, using the open-source agent-based travel demand model mobiTopp [8]. Across these applications, the results were evaluated by visually inspecting the resulting attractiveness distributions at the level of TAZ. This iterative plausibility assessment proved to be an important part of the workflow, because it allowed systematic refinement of the process. In addition, data quality and completeness were assessed for the intended use. The results indicate that some categories can be used more robustly than others [9].
The resulting attractiveness data were suitable for operational model use and demonstrated that OSM can support not only network modeling but also destination modeling in a domain-specific and reproducible way. Aggregated results for the Hamburg study area are shown in Figure 3 as an example [10].
Discussion and Summary
We developed an approach to calculate the attractiveness of TAZ for a wide range of activity types using OpenStreetMap data. The same methodology can also be applied not only to zones but to individual POIs, which opens opportunities for more disaggregated destination choice modeling.
At the same time, the study also highlights limitations. Reasonable attractiveness measures require a certain level of data quality. Multi-story buildings, for instance, often lack information on the number of floors and the inner distribution of types of use, which limits the quality of floor-space-based estimates. We therefore recommend complementing OSM, where necessary, with additional datasets such as building models or brand-specific reference values. Furthermore, data completeness differs strongly between activity types. Shopping-related information is often relatively stable, whereas private business or specific service functions may be more unevenly covered [10]. Nevertheless, for some activity types we found OSM to be more complete and more up to date than official data sources. We therefore recommend that trust in OSM and official sources be evaluated separately for each activity type and region.
A major contribution is the creation of a transparent and transferable workflow that allows modelers to use OSM more systematically. Because OSM is available worldwide, the approach can be extended not only to additional activities but also to many other regions, including cross-border study areas. In the longer term, the structured nature of OSM data also creates opportunities for stronger automation, faster model updates, multi-temporal analyses, and transport-oriented quality metrics. For the OSM community, this paper demonstrates a concrete and scientifically grounded domain-specific use case.
Acknowledgements
We gratefully acknowledge the OpenStreetMap community and all contributors whose continuous mapping efforts make this work possible.
The map-based figures in this paper use data from OpenStreetMap available under the Open Database Licence (ODbL).