stplanr: A Package for Transport Planning

Robin Lovelace

University of Leeds
r.lovelace@leeds.ac.uk

Richard Ellison

University of Sydney
richard.ellison@sydney.edu.au

Abstract

Tools for transport planning should be flexible, scalable and transparent. The stplanr package demonstrates and provides a home for such tools, with an emphasis on spatial transport data and non-motorized modes. stplanr facilitates common transport planning tasks including: downloading and cleaning transport datasets; creating geographic ‘desire lines from origin-destination (OD) data; route assignment, via the SpatialLinesNetwork class and interfaces to routing services such as CycleStreets.net; calculation of route segment attributes such as bearing and aggregate flow; and `travel watershed’ analysis. This paper demonstrates this functionality using reproducible examples on real transport datasets. More broadly, the experience shows open source software can form the basis of a reproducible transport planning workflow. stplanr, alongside other packages and open source projects, could provide a more transparent and democratically accountable alternative to the current approach which is heavily reliant on proprietary and technically and financially inaccessible software.

Note

This paper has now been peer reviewed and published by the R Journal. Please see the published version at journal.r-project.org and cite it as Lovelace and Ellison (2018).

The code presented in this paper requires stplanr 0.8.5 or earlier, which can be installed as follows:

remotes::install_github("ropensci/stplanr", ref = "v0.8.5")

Introduction

Transport planning can broadly be defined as the process of designing and evaluating transport interventions (de Dios Ortuzar and Willumsen 2011) usually with the ultimate aim of improving transport systems from economic, social and environmental perspectives. This inevitably involves a degree of subjective judgment and intuition. With the proliferation of new transport datasets — and the increasing availability of hardware and software to make sense of them — there is great potential for the discipline to become more evidence-based and scientific (Balmer, Rieser, and Nagel 2009). Transport planners have always undertaken a wide range computational activities (Boyce and Williams 2015), but with the digital revolution the demands have grown beyond the capabilities of a single, monolithic product. The diversity of tasks , and need for democratic accountability in public decision making, suggests that future-proof transport planning software should be:

This paper sets out to demonstrate that open source software with a command-line interface (CLI) can provide a foundation for transport planning software that meets each of these criteria. R provides a strong basis for progress in this direction because it already contains functionality used in common transport planning workflows. , and greatly improved R’s spatial abilities (Bivand, Pebesma, and G’omez-Rubio 2013), work that is being consolidated and extended in the recent package.

Building on these foundations a number of spatial packages have been developed for applied domains including: disease mapping and modelling, with packages such as and (Kim and Wakefield 2016; Brown and Zhou 2016); spatial ecology, with the adehabitat family of packages (Calenge 2006); and visualisation, with packages such as SpatialEpi, diseasemapping and Brown (2016). However, there has been little prior work to develop R functionality designed specifically for transport planning, with the notable exceptions of TravelR (a package on R-Forge last updated in 2012) and tidytransit (a package for handling General Transit Feed Specification (GTFS) data).

The purpose of stplanr is to provide a toolbox rather than a specific solution for transport planning, with an emphasis on spatial data and active modes. This emphasis is timely given the recent emphasis on sustainability (Banister 2008) and ‘Big Data’ (Zheng et al. 2016) in the wider field of transport planning. A major motivation was the lack of R packages, and open source software in general, for transport applications. This may be surprising given the ubiquity of transport problems;1 R’s proficiency at handling spatial, temporal and travel survey data that describe transport systems; and the growing popularity of R in applied domains (Jalal et al. 2017; Moore and Hutchinson 2017). Another motivation is the growth in open access datasets: the main purpose of early versions of the package was to process open origin-destination data (Lovelace et al. 2017).

R is already used in transport applications, as illustrated by recent research that applies packages from other domains to transport problems. For instance, Efthymiou and Antoniou (2012) use R to analyse the data collected from an online survey focused on car-sharing, bicycle-sharing and electric vehicles. Efthymiou and Antoniou (2012) also used R to collect and analyse transport-related data from Twitter using packages including , and . These packages were used to download, parse and plot the Twitter data using a method that can be repeated and the results reproduced or updated. More general statistical analyses have also been conducted on transport-related datasets using packages including and (Diana 2012; Cerin et al. 2013). Despite the rising use of R for transport research, there has yet been to be a package for transport planning.

The design of the R language, with its emphasis on flexibility, data processing and statistical modelling, suggests it can provide a powerful environment for transport planning research. There are many quantitative methods in transport planning, many of which fit into the classic ‘four stage’ transport model which involves the following steps (de Dios Ortuzar and Willumsen 2011): (1) trip generation to estimate trip freqency from origins; (2) distribution of trips to destinations; (3) modal split of trips between walking, cycling, buses etc.; (4) assignment of trips to the transport route network. To this we would like to add two more stages for the big data age: (0) data processing and exploration; and (5) validation. This sequence is not the only way of transport modelling and some have argued that its dominance has reduced innovation. However it is certainly a common approach and provides a useful schema for classifying the kinds of task that stplanr can tackle:

The automation of such tasks can assist researchers and practitioners to create evidence for decision making. If the data processing and analysis stages are fast and painless, more time can be dedicated to visualisation and decision making. This should allow researchers to focus on problems, rather than on clunky graphical user interfaces (GUIs), and ad-hoc scripts that could be generalised. Furthermore, if the process can be made reproducible and accessible (e.g. via online visualisation packages such as shiny), this could help transport planning move away from reliance on ‘black boxes’ (Waddell 2002) and empower citizens to challenge decisions made by transport planning authorities based on the evidence (Hollander 2016). There are many advantages of using a scriptable, interactive and open source language such as R for transport planning. Such an approach enables: reproducible research; the automation and sharing of code between researchers; reduced barriers to innovation as anyone can create new features for the benefit of all planners; easier interaction with non domain experts (who will lack dedicated software); and integration with other software systems, as illustrated by the use of to generate JavaScript for sharing interactive maps for transport planning, as used in the publicly accessible Propensity to Cycle Tool (Lovelace et al. 2017). Furthermore, R has a strong user community which can support newcomers (stplanr was peer reviewed thanks to the community surrounding ROpenSci). The advantages of using R specifically to develop the functionality described in this paper are that it has excellent geo-statistical capabilities (Pebesma et al. 2015), visualisation packages (e.g. tmap, ggplot2), support for logit models (which are useful for modelling modal shift), and support for the many formats that transport datasets are stored in (e.g. via the haven and rio packages).

Package structure and functionality

The package can be installed and loaded in the usual way (see the package’s README for dependencies and access to development versions):

install.packages("stplanr")
library(stplanr)

As illustrated by the message emitted when stplanr is loaded, it depends on . This means that the spatial data classes commonly used in the package will work with generic R functions such as summary, aggregate and, as illustrated in the figures below, plot .

Core functions and classes

The package’s core functions are structured around 3 common types of spatial transport data:

For ease of use, functions focussed on each data type have been developed with names prefixed with od_, line_ and route_ respectively. A selection of these is presented in Table 1. Additional ‘core functions’ could be developed, such as those prefixed with rn_ (for working with route network data) and g_ functions for geographic operations such as buffer creation on lat/lon projected data (this function is currently named buff_geo). We plan to elicit feedback on such changes before implementing them.

With a tip of the hat to the concept of type stability (e.g. as implemented in ), we also plan to make the core functions of stplanr more type-stable in future releases. Core functions, which begin with the prefixes listed above, could follow ’s lead and return only objects with the same class as that of the input. However there are limitations to this approach: it will break existing functionality and mean that output objects have a larger size than necessary (line_bearing, for example, does not need to duplicate the spatial data contained in its input). Instead, we plan to continue to name functions around the type of input data they take, but are open minded about function input-output data class conventions, especially in the context of the new class system implemented in .

A class system has not been developed for each data type (this option is discussed in the final section). The most common data types used in stplanr are assumed to be data frames and spatial datasets.

Transport datasets are very diverse. There are therefore many other functions which have more ad-hock names. Rather attempt a systematic description of each of stplanr’s functions (which can be gleaned from the online manual) it is more illuminating to see how they work together, as part of a transport planning workflow. As with most workflows, this begins with data access and ends with visualisation.

Accessing and processing transport data

Gaining access to data is often the first stage in transport research. This is often a long and protracted process which is thankfully becoming easier thanks to the ‘open data’ movement and packages such as tigris for making data access from within R easier .

stplanr provides a variety of different functions that facilitate importing common data formats used for transport analysis into R. Although transport analysis generally requires some transport-specific datasets, it also typically relies heavily on common sources of data including census data. This being the case, stplanr also includes functions that may be useful to those not involved in transport research. This includes the read_table_builder function for importing data from the Australian Bureau of Statistics (ABS) and the UK’s Stats19 road traffic casualty dataset. A brief example of the latter is demonstrated below, which begins with downloading the data (warning this downloads ~100 MB of data):

dl_stats19() # download and extract stats19 road traffic casualty data
#> [1] "Data saved at: /tmp/RtmpppF3E2/Accidents0514.csv"
#> [2] "Data saved at: /tmp/RtmpppF3E2/Casualties0514.csv"
#> [3] "Data saved at: /tmp/RtmpppF3E2/Vehicles0514.csv"  

Once the data has been saved in the default directory, determined by tempdir(), it can be read-in and cleaned with the read_stats19_ functions (note these call format_stats19_ functions internally to clean the datasets and add correct labels to the variables):

ac <- read_stats19_ac()
ca <- read_stats19_ca()
ve <- read_stats19_ve()

The resulting datasets (representing accident, casualty and vehicle level data, respectively) can be merged and made geographic, as illustrated below:

library(dplyr)
ca_ac <- inner_join(ca, ac)
ca_cycle <- ca_ac %>%
  filter(Casualty_Severity == "Fatal" & !is.na(Latitude)) %>%
  select(Age = Age_of_Casualty, Mode = Casualty_Type, Longitude, Latitude)
ca_sp <- SpatialPointsDataFrame(coords = ca_cycle[3:4], data = ca_cycle[1:2])

Now that this casualty data has been cleaned, subsetted (to only include serious cycle crashes) and converted into a spatial class system, we can analyse them using geographical datasets of the type commonly used by stplanr. The following code, for example, geographically subsets the dataset to include only crashes that occured within the bounding box of a route network dataset provided by stplanr (from version 0.1.7 and beyond) using the function bb2poly, which converts a spatial dataset into a box, represented as a rectangular SpatialPolygonsDataFrame:

data("route_network") # devtools::install_github("ropensci/splanr")version 0.1.7
proj4string(ca_sp) <- proj4string(route_network)
bb <- bb2poly(route_network)
proj4string(bb) <- proj4string(route_network)
ca_local <- ca_sp[bb, ]

The above code chunk shows the importance of understanding geographical data when working with transport data. It is only by converting the casualty data into a spatial data class, and adding a coordinate reference system (CRS), that transport planners and researchers can link this important dataset back to the route network. We can now perform GIS operations on the results. The next code chunk, for example, finds all the fatalities that took place within 100 m of the route network, using the function buff_geo:

rnet_buff_100 <- geo_buffer(route_network, width = 100)
ca_buff <- ca_local[rnet_buff_100, ]

These can be visualised using base R graphics, extended by , as illustrated in Figure . This provides a good start for analysis but for publication-quality plots and interactive plots, designed for public engagement, we recommend using dedicated visualisation packages that work with spatial data such as .

plot(bb, lty = 4)
plot(rnet_buff_100, col = "grey", add = TRUE)
points(ca_local, pch = 4)
points(ca_buff, cex = 3)

Creating geographic desire lines

Perhaps the most common type of aggregate-level transport information is origin-destination (‘OD’) data. This can be presented either as a matrix or (more commonly) a long table of OD pairs. An example of this type of raw data is provided below (see ?flow to see how this dataset was created).

data("flow", package = "stplanr")
head(flow[c(1:3, 12)])

Although the flow data displayed above describes movement over geographical space, it contains no explicitly geographical information. Instead, the coordinates of the origins and destinations are linked to a separate geographical dataset which also must be loaded to analyse the flows. This is a common problem solved by the function od2line. The geographical data is a set of points representing centroids of the origin and destinations, saved as a SpatialPointsDataFrame. Geographical data in R is best represented as such Spatial* objects, which use the S4 object engine. This explains the close integration of stplanr with R’s spatial packages, especially sp, which defines the S4 spatial object system.

data("cents", package = "stplanr")
as.data.frame(cents[1:3, -c(3, 4)])

We use od2line to combine flow and cents, to join the former to the latter. We will visualise the l object created below in the next section.

l <- od2line(flow = flow, zones = cents)

The data is now in a form that is much easier to analyse. We can plot the data with the command plot(l), which was not possible before. Because the SpatialLinesDataFrame object also contains data per line, it also helps with visualisation of the flows, as illustrated in Figure .

Allocating flows to the transport network

A common problem faced by transport researchers is network allocation: converting the ‘as the crow flies’ lines illustrated in the figure above into routes. These are the complex, winding paths that people and animals make to avoid obstacles such as buildings and to make the journey faster and more efficient (e.g. by following the route network).

This is difficult (and was until recently near impossible using free software) because of the size and complexity of transport networks, the complexity of realistic routing algorithms and need for context-specificity in the routing engine. Inexperienced cyclists, for example, would take a very different route than a heavy goods vehicle. stplanr tackles this issue by using 3rd party APIs to provide route-allocation.

Route allocation is undertaken by functions such as and . These allocate a single OD pair, represented as a text string to be ‘geo-coded’, a pair of of coordinates, or two SpatialPoints objects, representing origins and destinations. This is illustrated below with route_cyclestreet, which uses the CycleStreets.net API, a routing service “by cyclists for cyclists” that offers a range route strategies (primarily ‘fastest’, ‘quietest’ and ‘balanced’) that are based on a detailed analysis of cyclist wayfinding:2

route_bl <- route_cyclestreets(from = "Bradford", to = "Leeds")
route_c1_c2 <- route_cyclestreets(cents[1, ], cents[2, ])

The raw output from routing APIs is usually provided as a JSON or GeoJSON text string. By default, route_cyclestreet saves a number of key variables (including length, time, hilliness and busyness variables generated by CycleStreets.net) from the attribute data provided by the API. If the user wants to save the raw output, the save_raw argument can be used:

route_bl_raw <- route_cyclestreets(from = "Bradford", to = "Leeds", save_raw = TRUE)

Additional arguments taken by the route_ functions depend on the routing function in question. By changing the plan argument of route_cyclestreet to fastest, quietest or balanced, for example, routes favouring speed, quietness or a balance between speed and quietness will be saved, respectively.

To automate the creation of route-allocated lines over many desire lines, the line2route function loops over each line, wrapping any route_ function as an input. The output is a SpatialLinesDataFrame with the same number of dimensions as the input dataset (see the right panel in Figure ).

routes_fast <- line2route(l = l, route_fun = route_cyclestreet)

The result of this ‘batch routing’ exercise is illustrated in Figure . The red lines in the left hand panel are very different from the hypothetical straight ‘desire lines’ often used in transport research, highlighting the importance of this route-allocation functionality.

plot(route_network, lwd = 0)
plot(l, lwd = l$All / 10, add = TRUE)
lines(routes_fast, col = "red")
routes_fast$All <- l$All
rnet <- overline(routes_fast, "All", fun = sum)
rnet$flow <- rnet$All / mean(rnet$All) * 3
plot(rnet, lwd = rnet$flow / mean(rnet$flow))

To estimate the amount of capacity needed at each segment on the transport network, the overline function demonstrated above, is used to divide line geometries into unique segments and aggregate the overlapping values. The results, illustrated in the right-hand panel of Figure , can be used to estimate where there is most need to improve the transport network, for example informing the decision of where to build new bicycle paths.

Limitations with the route_cyclestreet routing API include its specificity, to one mode (cycling) and a single region (the UK and part of Europe). To overcome these limitations, additional routing APIs were added with the functions route_graphhopper, route_transportapi_public and viaroute. These interface to Graphhopper, TransportAPI and the Open Source Routing Machine (OSRM) routing services, respectively. The great advantage of OSRM is that it allows you to run your own routing services on a local server, greatly increasing the rate of route generation.

A short example of finding the route by car and bike between New York and Oaxaca demonstrates how route_graphhopper can collect geographical and other data on routes by various modes, anywhere in the world. The output, shown in Table , shows that the function also saves time, distance and (for bike trips) vertical distance climbed for the trips.

ny2oaxaca1 <- route_graphhopper("New York", "Oaxaca", vehicle = "bike")
ny2oaxaca2 <- route_graphhopper("New York", "Oaxaca", vehicle = "car")
rbind(ny2oaxaca1@data, ny2oaxaca2@data)
time dist change_elev
17522.73 4885663 87388.13
2759.89 4754772 NA

Modelling travel catchment areas

Accessibility to transport services is a particularly important topic when considering public transport or active travel because of the frequent steep reduction in use as distances to access services (or infrastructure) increase. As a result, the planning for transport services and infrastructure frequently focuses on several measures of accessibility including distance, but also travel times and frequencies and weighted by population. The functions in stplanr are intended to provide a method of estimating these accessibility measures as well as calculating the population that can access specific services (i.e., estimating the catchment area).

Catchment areas in particular are a widely used measure of accessibility that attempts to both quantify the likely target group for a particular service, and visualise the geographic area that is covered by the service. For instance, passengers are often said to be willing to walk up to 400 metres to a bus stop, or 800 metres to a railway station . Although these distances may appear relatively arbitrary and have been found to underestimate the true catchment area of bus stops and railway stations they nonetheless represent a good, albeit somewhat conservative, starting point from which catchment areas can be determined.

In many cases, catchment areas are calculated on the basis of straight-line (or “as the crow flies”) distances. This is a simplistic, but relatively appealing approach because it requires little additional data and is straight-forward to understand. stplanr provides functionality that calculates catchment areas using straight-line distances with the calc_catchment function. This function takes a SpatialPolygonsDataFrame that contains the population (or other) data, typically from a census, and a Spatial* layer that contains the geometry of the transport facility. These two layers are overlayed to calculate statistics for the desired catchments including proportioning polygons to account for the proportion located within the catchment area.

To illustrate how catchment areas can be calculated, stplanr contains some sample datasets stored in ESRI Shapefile format (a commonly used format for distributing GIS layers) that can together be used to calculate sample catchment areas. One of these datasets (smallsa1) contains population data for Statistical Area 1 (SA1) zones in Sydney, Australia. The second contains hypothetical cycleways aligned to streets in Sydney. The code below unzips the datasets and reads in the shapefiles.

data_dir <- system.file("extdata", package = "stplanr")
unzip(file.path(data_dir, "smallsa1.zip"))
unzip(file.path(data_dir, "testcycleway.zip"))
sa1income <- as(sf::read_sf("smallsa1.shp"), "Spatial")
testcycleway <- as(sf::read_sf("testcycleway.shp"), "Spatial")
# Remove unzipped files
file.remove(list.files(pattern = "^(smallsa1|testcycleway).*"))

Calculating the catchment area is straightforward and in addition to specifying the required datasets, only a vector containing column names to calculate statistics and a distance is required. Since proportioning the areas assumes projected data, unprojected data are automatically projected to either a common projection (if one is already projected) or a specified projection. It should be emphasised that the choice of projection is important and has an effect on the results meaning setting a local projection is recommended to achieve the most accurate results.

remotes::install_github("ropensci/stplanr")
catch800m <- calc_catchment(
  polygonlayer = sa1income,
  targetlayer = testcycleway,
  calccols = c("Total"),
  distance = 800,
  projection = "austalbers",
  dissolve = TRUE
)

By looking at the data.frame associated with the SpatialPolygonsDataFrame that is returned from the calc_catchment function, the total population within the catchment area can be seen to be nearly 40,000 people. The catchment area can also be plotted as with any other Spatial* object using the plot function using the code below with the result shown in Figure .

plot(sa1income, col = "light grey")
plot(catch800m, col = rgb(1, 0, 0, 0.5), add = TRUE)
plot(testcycleway, col = "green", add = TRUE)

This simplistic catchment area is useful when the straight-line distance is a reasonable approximation of the route taken to walk (or cycle) to a transport facility. However, this is often not the case. The catchment area in Figure initially appears reasonable but the red-shaded catchment area includes an area that requires travelling around a bay to access from the (green-coloured) cycleway. To allow for more realistic catchment areas for most situations, stplanr provides the calc_network_catchment function that uses the same principle as calc_catchment but also takes into account the transport network.

To use calc_network_catchment, a transport network needs to be prepared that can be used in conjunction with the previous datasets. Preparation of the dataset involves using the SpatialLinesNetwork function to create a network from a SpatialLinesDataFrame. This function combines a SpatialLinesDataFrame with a graph network (using the package) to provide basic routing functionality. The network is used to calculate the shortest actual paths within the specific catchment distance. This process involves the following code:

unzip(file.path(data_dir, "sydroads.zip"))
sydroads <- as(sf::read_sf(".", "roads"), "Spatial")
file.remove(list.files(pattern = "^(roads).*"))
sydnetwork <- SpatialLinesNetwork(sydroads)

The network catchment is then calculated using a similar method as with calc_catchment but with a few minor changes. Specifically these are including the SpatialLinesNetwork, and using the maximpedance parameter to define the distance, with distance being the additional distance from the network. In contrast to the distance parameter that is based on the straight-line distance in both the calc_catchment and calc_network_catchment functions, the maximpedance parameter is the maximum value in the units of the network’s weight attribute. In practice this is generally distance in metres but can also be travel times, risk or other measures.

netcatch800m <- calc_network_catchment(
  sln = sydnetwork,
  polygonlayer = sa1income,
  targetlayer = testcycleway,
  calccols = c("Total"),
  maximpedance = 800,
  distance = 100,
  projection = "austalbers"
)

Once calculated, the network catchment area can be used just as the straight-line network catchment. This includes extracting the catchment population of 128,000 and plotting the original catchment area together with the original area with the results shown in Figure :

plot(sa1income, col = "light grey")
plot(catch800m, col = rgb(1, 0, 0, 0.5), add = TRUE)
plot(netcatch800m, col = rgb(0, 0, 1, 0.5), add = TRUE)
plot(testcycleway, col = "green", add = TRUE)

Modelling and visualisation

Modelling mode choice

Route-allocated lines allow estimation of route distance and cirquity (route distance divided by Euclidean distance). These variables can help model the rate of flow between origins and destination, as illustrated in the left-hand panel of Figure . The code below demonstrates how objects generated by stplanr can be used to undertake such analysis, with the line_length function used to find the distance, in meters, of lat/lon data.

l$d_euclidean <- line_length(l)
l$d_rf <- routes_fast@data$length
plot(l$d_euclidean, l$d_rf,
  xlab = "Euclidean distance", ylab = "Route distance")
abline(a = 0, b = 1)
abline(a = 0, b = 1.2, col = "green")
abline(a = 0, b = 1.5, col = "red")

The left hand panel of Figure shows the expected strong correlation between Euclidean (\(d_E\)) and fastest route (\(d_{Rf}\)) distance. However, some OD pairs have a proportionally higher route distance than others, as illustrated by distance from the black line in the above plot: this represents : the ratio of network distance to Euclidean distance :

\[ Q = \frac{d_{Rf}}{d_E} \]

An extension to the concept of cirquity is the ‘quietness diversion factor’ (\(QDF\)) of a desire line , the ratio of the route distance of a quiet route option (\(d_{Rq}\)) to that of the fastest:

\[ QDF = \frac{d_{Rq}}{d_{Rf}} \]

Thanks to the ‘quietest’ route option provided by route_cyclestreet, we can estimate average values for both metrics as follows:

routes_slow <- line2route(l, route_cyclestreet, plan = "quietest")
l$d_rq <- routes_slow$length # quietest route distance
Q <- mean(l$d_rf / l$d_euclidean, na.rm = TRUE)
QDF <- mean(l$d_rq / l$d_rf, na.rm = TRUE)
Q
QDF

The results show that cycle paths are not particularly direct in the study region by international standards . This is hardly surprisingly given the small size of the sample and the short distances covered: \(Q\) tends to decrease at a decaying rate with distance. What is surprising is that \(QDF\) is close to unity, which could imply that the quiet routes are constructed along direct, and therefore sensible routes. We should caution against such assumptions, however: It is a small sample of desire lines and, when time is explored, we find that the ‘quietness diversion factor with respect to time’ (\(QDF_t\)) is slightly larger:

(QDFt <- mean(routes_slow$time / routes_fast$time, na.rm = TRUE))

Models of travel behaviour

There are many ways of estimating flows between origins and destinations, including spatial interaction models, the four-stage transport model and gravity models (‘distance decay’). stplanr aims eventually to facilitate creation of many types of flow model.

At present there are no functions for modelling distance decay, but this is something we would like to add in future versions of stplanr. Distance decay is an especially important concept for sustainable transport planning due to physical limitations on the ability of people to walk and cycle large distances .

We can explore the relationship between distance and the proportion of trips made by walking, using the same object l generated by stplanr.

l$pwalk <- l$On.foot / l$All
plot(l$d_euclidean, l$pwalk,
  cex = l$All / 50,
  xlab = "Euclidean distance (m)", ylab = "Proportion of trips by foot"
)

Based on the right-hand panel in Figure , there is a clear negative relationship between distance of trips and the proportion of those trips made by walking. This is unsurprising: beyond a certain distance (around 1.5km according the the data presented in the figure above) walking is usually seen as too slow and other modes are considered. According to the academic literature, this ‘distance decay’ is non-linear and there have been a number of functions proposed to fit to distance decay curves . From the range of options we test below just two forms. We will compare the ability of linear and log-square-root functions to fit the data contained in l for walking.

lm1 <- lm(pwalk ~ d_euclidean, data = l@data, weights = All)
lm2 <- lm(pwalk ~ d_rf, data = l@data, weights = All)
lm3 <- glm(pwalk ~ d_rf + I(d_rf^0.5),
  data = l@data, weights = All, family = quasipoisson(link = "log")
)

The results of these regression models can be seen using summary(). Surprisingly, Euclidean distance was a better predictor of walking than route distance, but no strong conclusions can be drawn from this finding, with such a small sample of desire lines (n = 42). The results are purely illustrative, of the kind of the possibilities created by using stplanr in conjuction with R’s modelling capabilities (see Figure ).

plot(l$d_euclidean, l$pwalk,
  cex = l$All / 50,
  xlab = "Euclidean distance (m)", ylab = "Proportion of trips by foot"
)
l2 <- data.frame(d_euclidean = 1:5000, d_rf = 1:5000)
lm1p <- predict(lm1, l2)
lm2p <- predict(lm2, l2)
lm3p <- predict(lm3, l2)
lines(l2$d_euclidean, lm1p)
lines(l2$d_euclidean, exp(lm2p), col = "green")
lines(l2$d_euclidean, exp(lm3p), col = "red")

Visualisation

Visualisation is an important aspect of any transport study, as it enables researchers to communicate their findings to other researchers, policy-makers and, ultimately, the public. It may therefore come as a surprise that stplanr contains no functions for visualisation. Instead, users are encouraged to make use of existing spatial visualisation tools in R, such as tmap, leaflet and ggmap .

Furthermore, with the development of online application frameworks such as shiny, it is now easier than ever to make the results of transport analysis and modelling projects available to the public. An example is the online interface of the Propensity to Cycle Tool (PCT). The results of the project, generated using stplanr, are presented at zone, desire line and Route Network levels . There is great potential to expand on the principle of publicly accessible transport planning tools via ‘web apps’, perhaps through new R packages dedicated to visualising transport data.

Future directions of travel

This paper has demonstrated the great potential for R to be used for transport planning. R’s flexibility, powerful GIS capabilities and free accessibility makes it well-suited to the needs of transport planners and researchers, especially those wanting to avoid the high costs of market-leading products. Rather than ‘reinvent the wheel’ (e.g. with a new class system), stplanr builds on existing packages and classes to work with common transport data formats.

It is useful to see stplanr, and R for transport planning in general, as an addition tool in the transport planner’s cabinet. It can be understood as one part of a wider movement that is making transport planning a more open and democratic process. Other developments in this movement include the increasing availability of open data and the rise of open source products for transport modelling, such as SUMO, MATSim and MITSIMLAB . stplanr, with its focus on GIS operations rather than microscopic vehicle-level behaviour, can complement such software and help make better use of new open data sources.

Because transport planning is an inherently spatial activity, stplanr occupies an important niche in the transport planning software landscape, with its focus on spatial transport data. There is great potential for development of stplanr in many directions. Desirable developments include the additional of functions for modelling modal split, for examample with functions to create commonly distance decay curves which are commonly found in active travel research and improving the computational efficiency of existing functions to make the methods more scalable for large databases. Our priority for stplanr however, is to keep the focus on geographic functions for transport planning. There are many opportunities in this direction, including:

Such spatial data processing capabilities would increase the range of transport planning tasks that stplanr can facilitate. For all this planned development activity to be useful, it is vital that new functionality is intuitive. R has a famously steep learning curve. Implementing simple concepts such as consistent naming systems and ensuring ‘type stability’ can greatly improve the usability of the package. For this reason, much future work in stplanr will go into improving documentation and user-friendliness.

Like much open source software stplanr is an open-ended project, a work-in-progress. We have set out clear motivations for developing transport planning capabilities in R and believe that the current version of stplanr (0.1.6) provides a major step in that direction compared with what was available a couple of years ago. But there is much more to do. We therefore welcome input on where the package’s priorities should lie, how it should evolve in the future and how to ensure it is well-developed and sustained.

References

Balmer, Michael, Marcel Rieser, and Kai Nagel. 2009. “MATSim-T: Architecture and Simulation Times.” Multi-Agent Systems for Traffic and Transportation Engineering, 57–78. https://svn.vsp.tu-berlin.de/repos/public-svn/publications/vspwp/2008/08-03/3aug08.pdf.

Banister, David. 2008. “The Sustainable Mobility Paradigm.” Transport Policy 15 (2): 73–80. https://doi.org/10.1016/j.tranpol.2007.10.005.

Bivand, Roger S, Edzer J Pebesma, and Virgilio G’omez-Rubio. 2013. Applied Spatial Data Analysis with R. Vol. 747248717. Springer.

Boyce, David E., and Huw C. W. L. Williams. 2015. Forecasting Urban Travel: Past, Present and Future. Edward Elgar Publishing.

Brown, Patrick E. 2016. “Maps, Coordinate Reference Systems and Visualising Geographic Data with Mapmisc.” The R Journal 8 (1): 64–91.

Brown, Patrick E., and L. Zhou. 2016. “Diseasemapping: Modelling Spatial Variation in Disease Risk for Areal Data.” https://CRAN.R-project.org/package=diseasemapping.

Calenge, C. 2006. “The Package Adehabitat for the R Software: Tool for the Analysis of Space and Habitat Use by Animals.” Ecological Modelling 197: 1035.

Cerin, Ester, Cindy H P Sit, Anthony Barnett, Man Chin Cheung, and Wai Man Chan. 2013. “Walking for Recreation and Perceptions of the Neighborhood Environment in Older Chinese Urban Dwellers.” Journal of Urban Health 90 (1): 56–66. https://doi.org/10.1007/s11524-012-9704-8.

de Dios Ortuzar, Juan, and Luis G. Willumsen. 2011. Modelling Transport. John Wiley & Sons.

Diana, Marco. 2012. “Studying Patterns of Use of Transport Modes Through Data Mining.” Transportation Research Record: Journal of the Transportation Research Board 2308 (December): 1–9. https://doi.org/10.3141/2308-01.

Efthymiou, Dimitrios, and Constantinos Antoniou. 2012. “Use of Social Media for Transport Data Collection.” Procedia - Social and Behavioral Sciences 48 (August 2016): 775–85. https://doi.org/10.1016/j.sbspro.2012.06.1055.

Hollander, Yaron. 2016. Transport Modelling for a Complete Beginner. CTthink!

Jalal, Hawre, Petros Pechlivanoglou, Eline Krijkamp, Fernando Alarid-Escudero, Eva Enns, and M. G. Myriam Hunink. 2017. “An Overview of R in Health Decision Sciences.” Medical Decision Making, January, 0272989X16686559.

Kim, Albert Y., and Jon Wakefield. 2016. “SpatialEpi: Methods and Data for Spatial Epidemiology.” https://CRAN.R-project.org/package=SpatialEpi.

Lovelace, Robin, and Richard Ellison. 2018. “Stplanr: A Package for Transport Planning.” The R Journal 10 (2): 7–23. https://doi.org/10.32614/RJ-2018-053.

Lovelace, Robin, Anna Goodman, Rachel Aldred, Nikolai Berkoff, Ali Abbas, and James Woodcock. 2017. “The Propensity to Cycle Tool: An Open Source Online System for Sustainable Transport Planning.” Journal of Transport and Land Use 10 (1). https://doi.org/10.5198/jtlu.2016.862.

Moore, R. D. (Dan), and David Hutchinson. 2017. “Why Watershed Analysts Should Use R for Data Processing and Analysis.” Confluence: Journal of Watershed Science and Management 1 (1). http://confluence-jwsm.ca/index.php/jwsm/article/view/2.

Pebesma, Edzer, Roger Bivand, Paulo Justiniano Ribeiro, and others. 2015. “Software for Spatial Statistics.” Journal of Statistical Software 63 (1): 1–8.

Waddell, Paul. 2002. “UrbanSim: Modeling Urban Development for Land Use, Transportation, and Environmental Planning.” Journal of the American Planning Association 68 (3): 297–314.

Zheng, Xinhu, Wei Chen, Pu Wang, Dayong Shen, Songhang Chen, Xiao Wang, Qingpeng Zhang, and Liuqing Yang. 2016. “Big Data for Social Transportation.” IEEE Transactions on Intelligent Transportation Systems 17 (3): 620–30. https://ieeexplore.ieee.org/abstract/document/7359138/.


  1. Many people can think of things that could be improved on their local transport networks, especially for walking, cycling and wheel-chairs. But most lack the evidence to communicate the issues, and potential solutions, to others.↩︎

  2. An API key is needed for this function to work. This can be requested (or purchased for large scale routing) from cyclestreets.net/api/apply. See ?route_cyclestreet for details. Thanks to Martin Lucas-Smith and Simon Nuttall for making this possible.↩︎