Spatial interaction models for active travel commuting

Ivann Schlosser

ivann.schlosser.19@ucl.ac.uk

Centre for advanced spatial analysis (CASA)

Hello and thank you for having me today at FOSS4G. My name is Ivann and I will be talking about spatial interaction models for active travel commuting, with an open source and reproducible approach. I have been doing this work at the centre for advanced spatial analysis for almost a year and we expect to publish a paper soon, and today I will cover the software related parts of it.

Transport mode choice

The use of a personal vehicle remains a dominant mode of transport even in cities. The urban environment has been extensively developed with vehicle use in mind. But this is changing in light of research indicating the negative impact that cars have on cities and their inhabitants. To name a few, noise and particle pollution, congestion, requirement for large parking areas, collision danger (Hidalgo 2020; Fleury et al. 2021; Saunders et al. 2013).

Active travel (AT) represents a simple yet powerful alternative to private vehicle use. A great amount of car trips on short distances could be made by AT.

Contributing to this shift in priorities involves having accurate representation of walkable and cyclable networks. And a quantitative understanding of AT mobility.

Developing further models for predicting the use of such modes of transport is important for promoting the right change in the transport infrastructure.

The London Plan

Puts active travel in the center of the discussion on redevelopment of local areas and global transport connectivity.

Policy D1 London’s form, character and capacity for growth

Area assessments should cover the elements … 4) Existing and planned transport networks (particularly walking & cycling networks) and public transportation connectivity.

Cities mentioning AT as a priority in their mobility strategy are spread around the world: Chicago, New York, Singapour, Paris and more…

The general context is that we are observing a general shift in priority when it comes to mobility. The research of the last few decades has revealed the negative effects that cars have on cities, such as noise & particle pollution, congestion, use of for large parking space.

To all these problems, active travel represents a simple, cheap, yet very powerful solution. To help effectively implement it, we develop a more quantitative understanding of this type of mobility and the tools to facilitate this.

Transport mode choice

Basically get us from this, or even this, to this and this.

Context

The QUANT model (Batty and Milton 2021) for benchmarking transport or employment change from the point of view of sustainability and benefit for commuting journeys. A country scale SIM. Ultimately provide a user friendly, open source, highly efficient model capable of running country scale scenarios within seconds to the relevant actors.

Methodology

The many open source tools out there aim at solving various aspects, combining them into one workflow that allows to build a full, self-sufficient model from scratch with ease of scalability, review some of the tools on the way and develop when missing appeared relevant.

The 3 pillars

In order to get a more comprehensive understanding of the various methods, data sources, existing software available, we experimented with workflows that would combine a testing process of the various steps involved.

Network

• Data sources
• Profiles based on mode of transport
• Software
• Consistency, quality

Routing

• Origins and Destinations
• Software
• Performance

Spatial interaction

• Scalability, Robustness
• Quality of fit
• Reusability

We have at CASA a model called QUANT that already implements a large scale SIM for other modes of transport (train, underground, bus, car), and this work was dedicated to developing this new mode for the global model. It’s ultimate aim is to provide a web based open source tool allowing to benchmark various transport infrastructure change on a country scale.

With this, we embark on a journey to develop the right methods to get networks, perform routing and model commuting by AT.

But first, let’s take a step back and consider the problem, let us divide it into 3 main part:

The : network, routing, spatial interaction.

Network

• OpenStreetMap
• Ordnance Survey

One being the main reference for open source georeferenced information in the world, the other being the official data set of roads in the UK.

We look at OpenStreetMap and ordnance survey data as primary sources. We can see that one (OS) is developed mainly for cars, while OSM seems richer and more diverse.

Network: cycling and walking

We build network profiles from OSM based on the road segment classifications using the osmnx python package. We observe however that, due to the classification types, the resulting network profiles might not give a truly realistic picture of the network. To counter that, we also keep a general network, removing only motorways since it is illegal to use them. We also use an example of a network where segments are additionally weighted based on the preference of certain modes to use them, provided by the dodgr package in R.

Routing

Centroids

• Geometric
• Network
  • Subset of road nodes inside urbanised areas (exclude parks, water bodies etc…)
• Commute
  • population weighted to workplace zone

The flow data is aggregated by MSOA level, so we need to define the actual locations from and to which we are routing. We use 3 different types of centroids, the most commonly used one being the geometric one. Next comes the network centroid, which is taken from a set of nodes that are inside urbanised areas, so it excludes parks, water bodies etc… The final one is what we called the commute centroids and the actually a pair pair of points, the population weighted centroid on the one hand and the workplace zone centroid, taken from the ons workplace tone definition of the ONS.

Routing

Comparing commonly used packages: tidygraph (Pedersen 2023) and sf_networks (Meer et al. 2022), wrappers of igraph (Csardi and Nepusz 2006), dodgr (Padgham 2019), cppRouting (Larmet 2022).

Benchmark

routing_benchmark <- microbenchmark::microbenchmark(...)

The cppRouting package is used to have a local, self-sufficient workflow, without compromising the performance.

Our next step was to identify the possibility to do routing on a pretty large network, for different set of ODs, requiring a very efficient tool. By benchmarking various famous and common tools, we found one to be much more efficient than the others, the cppRouting package in R and used it for the rest of the work.

Spatial interaction model

This brings us to the final step of our workflow, building the actual spatial interaction model.

Model

Flow data

Using the wu03ew table at MSOA level from the 2011 census, we build flow matrices.

Area of residence	Area of workplace	Bicycle	On foot
E02000001	E02000001	33	1304
E02000001	E02000014	0	0
E02000001	E02000016	0	0

at
1337
0
0

SIM

A doubly constrained spatial interaction model that is calibrated on the distance matrices (in km) and flow, using the foot and bike variables from the flows.

\[ T_{ij}=A_i O_i B_j D_j exp(-\beta d_{ij}) \]

Package

To perform the this step, a set of functions in R and C++ were developed as a package called cppSim and published on github and (hopefully) soon on CRAN.

model <- cppSim::run_model(flows = flow_matrix
                           ,distance = distance_matrix
                           ,beta = beta_best_fit)

Performance

Notably, it takes about 50 milliseconds for one model run on London’s \(983\times 983\) OD matrix.

glm and dependent packages were running out of memory / taking very long. Further development can be done to provide a full set of functionalities around routing, and SIMs as one single R package that would be highly efficient.

We create flow matrices using the commuting data by mode of transport from the 2011 census.

We could not find a tool to do the doubly constrained spatial interaction model for the London area, so we developed it and published it on GitHub as a package called cppSIM, it currently does just that, but does it very efficiently. Hopefully on CRAN soon and I am happy to discuss the various functionalities that can further be developed in a SIM workflow.

Notably it takes about 50 milliseconds to do a run for London.

Other solutions did not complete the run.

Results

• Greater impact for walking, with high quality of fit especially for commute centroids on all networks.

  • Probably due to the better estimate of intra flows

• Less impact on cycling, although regular networks outperform slightly the custom weighted ones.

• When combined, the best of both worlds seems to emerge, with high quality of fit, more homogeneous results across networks and the commute centroids being slightly better with full OSM and OS networks.

Running this model on the various modes/network/centroid combinations discussed, gives us the following picture, when maximising quality of fit with respect to the distance exponent:

We see different maximum values, with high similarity in the results for cycling than walking, highlighting the great sensitivity to the different distances used. Commute centroids appear to significantly increase the model fit for walking and less so for cycling. When combining the two modes into AT, we observe a kind of best of both worlds picture, where the results are more consistent between networks and centroids with a slightly better quality and stability when using commute centroids. A nice results given that they correspond more intuitively to the actual locations a population is trying to join on during commute.

Conclusion

• The emergence good quality open source network data has promoted the development of powerful open tools to manipulate and use them.

• We use these tools with the 2011 UK CENSUS data on foot and bicycle commuting flows to:

  • Compare the different approaches that can be adopted (network profile, routing locations)
  • Bring awareness to the data consistency and it’s geographic spread
  • Develop a reproducible, self-sufficient workflow that is highly efficient even for large scale areas and networks.
  • The final step of our analysis required the development of new tools.

Links

Active-travel modelling: a methodological approach to networks for walking and cycling commuting analysis

Paper on arxive

cppSim - fast and memory efficient doubly constrained SIMs

cppSim package

To conclude, we can say that the growing need for alternative transport models and the availability of high resolution network data can allow for a free and open source approach to such models with few scale limitations, we are currently scaling up this analysis to all of GB and it works pretty well and look forward to developing more tools around this topic and of course providing them open source and free.

Links the paper on arxive and the package on GitHub in the qr codes.

Thank you for your attention !

Thank you for your attention !

References

Batty, Michael, and Richard Milton. 2021. “A New Framework for Very Large-Scale Urban Modelling.” Urban Studies 58 (15): 3071–94. https://doi.org/10.1177/0042098020982252.

Csardi, Gabor, and Tamas Nepusz. 2006. “The Igraph Software Package for Complex Network Research.” InterJournal Complex Systems: 1695. https://igraph.org.

Fleury, Vanessa, Rebecca Himsl, Stéphane Joost, Nicolas Nicastro, Matthieu Bereau, Idris Guessous, and Pierre R. Burkhard. 2021. “Geospatial Analysis of Individual-Based Parkinson’s Disease Data Supports a Link with Air Pollution: A Case-Control Study.” Parkinsonism & Related Disorders 83 (February): 4148. https://doi.org/10.1016/j.parkreldis.2020.12.013.

Hidalgo, César A. 2020. “Trillion Dollar Streets.” Environment and Planning B: Urban Analytics and City Science 47 (7): 1133–35. https://doi.org/10.1177/2399808320949295.

Larmet, Vincent. 2022. cppRouting: Algorithms for Routing and Solving the Traffic Assignment Problem. https://CRAN.R-project.org/package=cppRouting.

Meer, Lucas van der, Lorena Abad, Andrea Gilardi, and Robin Lovelace. 2022. “Sfnetworks: Tidy Geospatial Networks.” https://CRAN.R-project.org/package=sfnetworks.

Padgham, Mark. 2019. “Dodgr: An r Package for Network Flow Aggregation.” Transport Findings, February. https://doi.org/10.32866/6945.

Pedersen, Thomas Lin. 2023. Tidygraph: A Tidy API for Graph Manipulation. https://CRAN.R-project.org/package=tidygraph.

Saunders, Lucinda E., Judith M. Green, Mark P. Petticrew, Rebecca Steinbach, and Helen Roberts. 2013. “What Are the Health Benefits of Active Travel? A Systematic Review of Trials and Cohort Studies.” PLoS ONE 8 (8). https://doi.org/10.1371/journal.pone.0069912.