As the name suggests, origin-destination (OD) data represents movement through geographic space, from an origin (O) to a destination (D). Sometimes also called ‘flow data’, OD datasets contain details of trips between two geographic points or, more commonly, zones (which are often represented by a zone centroid). Most OD datasets refer to start and end locations with ‘ID’ columns containing character strings such as
zone1. These IDs refer to a geographic feature in a separate geographic dataset. Origin and destination locations are sometimes represented as geographic coordinates.
OD datasets typically contain multiple non geographic attributes. These usually include, at a minimum, the number of trips that take place from the origin to the destination over a given time period (e.g. a typical work day). Additional attributes can include breakdown by the mode(s) of transport used for the trips. Usually only a single mode is captured (trips made by a combination of cycle-train-walk modes are often counted only as ‘train’ trips). Additional disaggregations of overall counts may include trip counts at different time periods.
Many OD datasets omit information. If there is only one time period, then this resides in the metadata for the whole data set. There is rarely any information about the path taken between the start and end points. It is typically the job of the analyst to use a routing service (such as OSRM, Google Directions API, CycleStreets.net or OpenRouteService) or an assignment model (such as those contained in proprietary software such as SATURN and Visum) to identify likely routes with reference to shortest path algorithms or generalised cost minimisation algorithms (which account for monetary plus time and quality ‘costs’).
Despite the rather dull name, OD datasets are a vital part of the modern world: they underpin analysis and models that influence current and future transport systems. Historically, these models, and the OD datasets that drove them, were used to plan for car-dominated cities [@boyce_forecasting_2015]. Now that there is growing evidence of the negative impacts car domination, however, there is a strong argument for transport models being repurposed. Origin-destination data can be part of the solution.
From a health perspective transport planning, supported by OD data and analysed primarily using proprietary software and opaque methods, has failed: roads are now the largest cause of death of young people worldwide, killing more than 1 million people each year [@world_health_organization_global_2018]. Even ignoring problems such as air pollution, obesity and climate change, it is clear that current transport systems are unsustainable. There are other reasons why transport data analysis and software are important [@lovelace_stplanr_2018].
The purpose of this vignette is to introduce OD data, an important component of many transport planning models, with examples based on data and functions from the stplanr package. The aim is to enable you to use OD data to inform more sustainable transport plans, for example by identifying ‘desire lines’ along which policies could cause a modal switch away from cars and towards lower energy modes such as walking, cycling, and public transport.
OD data can be accessed from a range of sources (we will see code that downloads many thousands of OD pairs later in this vignette). Some ‘data carpentry’ may be needed before the OD data is ready for analysis. This vignette does not cover cleaning OD data: we assume you know R and come with ‘tidy’ data [@wickham_tidy_2014], in which each row represents travel between an origin and a destination (typically zones represented by zone IDs), and each column represents an attribute such as number of trips or vehichle counts by mode or straight line distance.1
At its most basic, a single trip can be OD data, as illustrated in the very simple example dataset below.
The concept can be illustrated by an example of travel between two cities. As mentioned above, OD data is a data frame that contains at least two columns, one representing an origin and one representing a destination. The code below creates a data frame with one row representing movement between the cities of Leeds and London:
Most OD datasets have attributes associated with movement between the two places. In this case we can represent the number of trips I make from Leeds to London per year with a new column as follows:
Furthermore, if we have geographic representations of the origin and destination datasets, in a separate object called
p (short for points) we can create a geographical representation of the data as follows:
p = sf::st_as_sf( data.frame( name = c("Leeds", "London"), lon = c(-1.5, -0.1), lat = c(53.8, 51.5) ), coords = c("lon", "lat"), crs = 4326 ) p #> Simple feature collection with 2 features and 1 field #> geometry type: POINT #> dimension: XY #> bbox: xmin: -1.5 ymin: 51.5 xmax: -0.1 ymax: 53.8 #> geographic CRS: WGS 84 #> name geometry #> 1 Leeds POINT (-1.5 53.8) #> 2 London POINT (-0.1 51.5)
This can be plotted as follows (result from mapview shown):
The plot above shows that OD data is fundamentally geographic data, although it is not always represented geographically. To convert the non-geographic OD datasets into a geographic object, we can use the
od_to_sf() function from the
od package as follows:
We just converted the data frame into a ‘geographic data frame’. This can be useful for many things, not least visualisation, as shown below:
This next example uses a small dataset that comes pre-loaded with the package, called
od_data_df. We will copy it into an object called
x for conciseness and to highlight the fact that these operations can be generalised to many OD datasets.
Like all data, the object
od, created in the preceding code chunk, comes from a specific context: the 2011 UK Census questions:
od is a data frame containing aggregated answers to these questions (see
?pct::get_od() for details). It is implicitly geographic: the first two columns refer to geographic entities but do not contain coordinates themselves (OD coordinates are covered below). Other columns contain attributes associated with each OD pair, typically counting how many people travel by mode of transport. OD data can be represented in a number of ways, as outlined in the next sections.
The most useful way of representing OD data is the ‘long’ data frame format described above. This is increasingly the format used by official statistical agencies, including the UK’s Office for National Statistics (ONS), who provide origin destination data as a
.csv file. Typically, the first column is the zone code of origin and the second column is the zone code of the destination, as is the case with the object
od. Subsequent columns contain attributes such as
all, meaning trips by all modes, as illustrated below (we will see a matrix representation of this subset of the data in the next section):
geo_code1 refers to the origin,
geo_code2 refers to the destination.
Additional columns can represent addition attributes, such as number of trips by time, mode of travel, type of person, or trip purpose. The
od dataset contains column names representing mode of travel (train, bus, bicycle etc), as can be seen with
names(od[-(1:2)]). These ‘mode’ columns contain integers in the example data, but contain characters, dates and other data types, taking advantage of the flexibility of data frames.
The ‘OD matrix’ representation of OD data represents each attribute column in the long form as a separate matrix. Instead of rows representing OD pairs, rows represent all travel from each origin to all destinations (represented as columns). The stplanr function
od_to_odmatrix() converts between the ‘long’ to the ‘matrix’ form on a per column basis, as illustrated below:
Note that row and column names are now zone codes. The cell in row 1 and column 2 (
od_matrix[1, 2]), for example, reports that there are 94 trips from zone
E02002361 to zone
E02002393. In the case above, no people travel between the majority of the OD pair combinations, as represented by the
NAs. OD matrices are a relatively rudimentary data structure that pre-date R’s
data.frame class. Typically, they only contained integer counts, providing small and simple datasets that could be used in 20th Century transport modelling software running on limited 20th Century hardware.
Although ‘OD matrix’ is still sometimes used informally to refer to any OD datadset, the long OD pair representation is recommended: OD matrices become unwieldy for large OD datasets, which are likely to be sparse, with many empty cells represented by NAs. Furthermore, to represent many attributes in matix format, multiple lists of OD matrices or ‘OD arrays’ must be created. This is demonstrated in the code chunk below, which represents travel between OD pairs by all modes and by bike:
odmatrix_to_od() can converts OD matrices back into the more convenient long form:
A common, and sometimes problematic, feature of OD data is ‘intra-zonal flows’. These are trips that start and end in the same zone. The proportion of travel that is intra-zonal depends largely on the size of the zones used. It is often useful to separate intra-zonal and inter-zonal flows at the outset, as demonstrated below:
Intra-zonal OD pairs represent short trips (up to the size of the zone within which the trips take place) so are sometimes ignored in OD data analyis. However, intra-zonal flows can be valuable, for example in measuring the amount of localised transport activity and as a sign of local economies.
Another subtly with some (symetric, where origins and destinations can be the same points) OD data is that oneway flows can hide the extent of bidirectional flows in plots and other types of analysis. This is illustrated below for a sample of the
Note that in the second dataset there are only 2 rows instead of 3. The function
od_oneway() aggregates oneway lines to produce bidirectional flows. By default, it returns the sum of each numeric column for each bidirectional origin-destination pair.
The previous representations of OD data are all implicitly geographic: their coordinates are not contained in the data, but associated with another object that is geographic, typically a zone or a zone centroid. This is problematic, meaning that multiple objects or files are required to fully represent the same data. Desire line representations overcome this issue. They are geographic lines between origin and destination, with the same attributes as in the ‘long’ representation.
od2line() can convert long form OD data to desire lines. The second argument is a zone or a centroid dataset that contains ‘zone IDs’ that match the IDs in the first and second columns of the OD data, as illustrated below:
The preceding code chunk created a zones object called
z, the coordinates of which were used to convert the object
l, which are geographic desire lines. The desire line object is stored in as a geographic simple features object, which has the same number of rows as does the object
od and one more column:
The new column is the geometry column, which can be plotted as follows:
By default, plotting
l shows the attributes for each line:
Because these lines have a coordinate reference system (CRS) inherited from the zones data, they can also be plotted on an interactive map, as follows:
Note that in some OD datasets there may be IDs that match no zone. We can simulate this situation by setting the third origin ID of
nomatch, a string that is not in the zones ID:
You should clean your OD data and ensure all ids in the first two columns match the ids in the first column of the zone data before running
The minimal example dataset we’ve been using so far is fine for demonstrating the key concepts of OD data. But for more advanced topic, and to get an idea of what is possible with OD data at a city level, it helps to have a larger dataset.
We will use an example dataset representing commuting in London, accessed as follows (note: these code chunks are not evaluated in the vignette because it starts by downloading 2.4 million rows and could take a few minutes to run). First, we can use the
pct package to download official data from the UK (note the addition of the % active column):
library(dplyr) # get nationwide OD data od_all <- pct::get_od() nrow(od_all) # > 2402201 od_all$Active <- (od_all$bicycle + od_all$foot) / od_all$all * 100 centroids_all <- pct::get_centroids_ew() %>% sf::st_transform(4326) nrow(centroids_all) # > 7201 london <- pct::pct_regions %>% filter(region_name == "london") centroids_london <- centroids_all[london, ] od_london <- od_all %>% filter(geo_code1 %in% centroids_london$msoa11cd) %>% filter(geo_code2 %in% centroids_london$msoa11cd) od_london <- od_all[ od_all$geo_code1 %in% centroids_london$msoa11cd & od_all$geo_code2 %in% centroids_london$msoa11cd, ]
Now that we have the input OD data (in
od_london) and zones (population-weighted centroids in
cents_london in this case), can can convert them to desire lines:
Even after filering flows to keep only those with origins and destinations in London, there are still more than 300k flows. That is a lot to plot. So we’ll further subset them, first so they only contain inter-zonal flows (which are actually lines, intra-zonal flows are lines with length 0, which are essentially points) and second to contain only flows containing above a threshold level of flows:
If we do any analysis on this dataset, it’s important to know how representative it is of all flows. A crude way to do this is to calculate the proportion of lines and trips that are covered in the dataset:
This shows that only 8% of the lines contain more than half (55%) of the total number of trips.
Once you have an OD dataset of a size that can be plotted (20,000 desire lines is quick to plot on most computers) a logical next stage is to plot it, e.g. with
You may be disapointed by the result, which is more of a ‘hay stack’ plot than an intuitive illustration of flows across the city. To overcome this issue, you can set the aesthetics to emphasize with important flows, e.g. by line width in
sf’s plotting system:
This is better, but is still not ideal: the code was not intuitive to write, and the result is still not publication quality. Instead, it makes sense to make a dedicated mapping package such tmap, as outlined in the visualisation chapter of the open source book Geocomputation with R [@lovelace_geocomputation_2019]. As shown in the transport chapter of that book, OD flows can be visualised with the following code:
library(tmap) desire_lines_top <- desire_lines_top %>% arrange(Active) tm_shape(london) + tm_borders() + tm_shape(desire_lines_top) + tm_lines( palette = "plasma", breaks = c(0, 5, 10, 20, 40, 100), lwd = "all", scale = 9, title.lwd = "Number of trips", alpha = 0.5, col = "Active", title = "Active travel (%)", legend.lwd.show = FALSE ) + tm_scale_bar() + tm_layout( legend.bg.alpha = 0.5, legend.bg.color = "white" )
The above plot contains much information, providing a visual overview of the transport pattern in the city, telling us that:
Plotting OD data in this way can tell us much about cities, each of which has a different travel pattern. The below figure, for example, shows the same plotting code applied to Bristol, a more polycentric city with a lower average percentage of travel by walking and cycling. See Section 12.4 of Geocomputation with R for details.
It is possible to group OD data by origin and destination to gain information at the zone level. The code and resulting plot below, for example, summarises the number of people departing from each zone by mode:
zones_london <- pct::get_pct_zones("london") %>% select("geo_code") origin_attributes <- desire_lines_top %>% sf::st_drop_geometry() %>% group_by(geo_code1) %>% summarize_if(is.numeric, sum) %>% dplyr::rename(geo_code = geo_code1) # origin_attributes <- zones_origins <- left_join(zones_london, origin_attributes, by = "geo_code") plot(zones_origins, border = NA)
We can observe a number of features, including that:
The pattern is quite different when we calculate the destinations:
destination_attributes <- desire_lines_top %>% sf::st_drop_geometry() %>% group_by(geo_code2) %>% summarize_if(is.numeric, sum) %>% dplyr::rename(geo_code = geo_code2) %>% mutate_at(vars(-matches("geo_|all")), funs(. / all)) %>% left_join(zones_london, ., by = "geo_code") plot(destination_attributes, border = NA)
Despite the importance of origin-destination datasets for transport research, there are surprisingly few guides dedicated to working with them using open source software. The following suggestions are based on my own reading — if you have any other suggestions of good resources for working with OD data, let me know!
stplanr provides many functions for working with OD data. OD data is an important component in transport planning and modelling that can, if used with creativity and skill, could assist with sustainable transport planning and the global transition away from fossil fuels. There are many other things can be done with OD data, some of which could be supported by future versions of this package. To suggest new features of otherwise get in touch, see the
stplanr issue tracker at github.com/ropensci/stplanr.
It may be difficult to convert between ‘number of trip’ and ‘number of vehicle’ counts for modes in which a single vehicle can contain many people, such as cars (a problem that can be overcome when surveys differentiate between car driver and ‘car passenger’ as distinct modes), buses and trams if occupancy levels are unknown. Typically OD data only report single stage trips, but multi-modal trips such as walk-rail-cycle can be represented when such a combination of modes is represented by a new, unique, mode.↩︎