What is a travel time matrix?
A travel time matrix is a table that shows the travel time between all pairs of a set of locations in an urban area. The locations represent typical origins and destinations, such as everyday services and residential homes, or are a complete set of locations covering the entire area wall-to-wall, such as census polygons or a regular grid.
A travel time matrix is a key piece of information in transportation research and planning. It allows us to study how easily people can reach different destinations, to evaluate the impacts of transport and land use policies, to analyse accessibility patterns and to examine how accessibility is influenced by factors such as the quality of public transport systems, street networks, and land use patterns. Travel time matrices can also be used to identify in which areas and for which groups of people a city works best and worst. Altogether, this makes travel time matrices a critical information to help cities become more equitable and more sustainable, and to foster a good quality of life for their residents.
Successful recent research that either used or produced travel time matrices include the work of the Digital Geography Lab at the University of Helsinki (e.g., Tenkanen and Toivonen (2020), Salonen and Toivonen (2013), or Järv et al. (2018)), the Mobility Network at the University of Toronto (e.g., Farber and Fu (2017), Farber et al. (2014)), and the Access to Opportunities Project (AOP) at the Institute for Applied Economic Research - IPEA (e.g., Pereira et al. (2021), Braga et al. (2023), Herszenhut et al. (2022)).
Load a transport network#
a transport network, and
a set of origins and destinations
In the example below, we first create a
TransportNetwork. To do so, we load an
OpenStreetMap extract of the São Paulo city centre as well as a public transport
schedule in GTFS format covering the same area. Again, we use the sample data
sets that can be installed separately
import r5py import r5py.sampledata.sao_paulo transport_network = r5py.TransportNetwork( r5py.sampledata.sao_paulo.osm_pbf, [ r5py.sampledata.sao_paulo.gtfs, ] )
Studies that compare accessibility between different neighbourhoods tend to use a regular grid of points that covers the study area as origins or destinations. Recently, hexagonal grids, such as Uber’s H3 indexing system have gained popularity, as they assure equidistant neighbourhood relationships (all neighbouring grid cells’ centroids are at the same distance; in a grid of squares, the diagonal neighbours are roughly 41% further than the horizontal and vertical ones).
We prepared such a hexagonal grid for São Paulo, and added the counts of
schools within each cell as separate columns.
id column refers to the H3 address of the grid cells.
import geopandas hexagon_grid = geopandas.read_file(r5py.sampledata.sao_paulo.hexgrid_gpkg) hexagon_grid
|0||89a8100c603ffff||1146||1155||0||POLYGON ((336000.312 7392162.721, 336107.079 7...|
|1||89a8100c617ffff||700||463||1||POLYGON ((335685.425 7392005.022, 335792.187 7...|
|2||89a8100c60fffff||377||257||0||POLYGON ((336005.719 7392502.350, 336112.488 7...|
|3||89a8100c607ffff||743||687||0||POLYGON ((335690.825 7392344.644, 335797.590 7...|
|4||89a8100c6abffff||601||29||0||POLYGON ((335375.940 7392186.943, 335482.700 7...|
|318||89a8100cea7ffff||34||2483||0||POLYGON ((330455.500 7397136.529, 330562.237 7...|
|319||89a8100ce23ffff||382||395||1||POLYGON ((331090.629 7397791.927, 331197.377 7...|
|320||89a8100ce37ffff||10||202||0||POLYGON ((330775.717 7397634.124, 330882.461 7...|
|321||89a8100c8dbffff||19||438||0||POLYGON ((330460.813 7397476.324, 330567.552 7...|
|322||89a8100c8cbffff||12||1375||0||POLYGON ((330466.126 7397816.132, 330572.867 7...|
323 rows × 5 columns
We can use
explore() to plot the
hexagonal grid in a map:
R5py expects origins and destinations to be point geometries. For grid cells,
the geometric center point (‘centroid’) is a good approximisation. One can use
geopandas.GeoDataSeries.centroid to quickly derive a centroid (point)
geometry from a polygon. We will create one data frame for origins, and one for
origins = hexagon_grid.copy() origins["geometry"] = origins.geometry.centroid
destinations = hexagon_grid.copy() destinations["geometry"] = destinations.geometry.centroid
Compute a travel time matrix#
With this, we have all input data sets needed for computing a travel time matrix: a transport network, origins, and destinations. We still need to decide which modes of transport should be used, and the departure time in our analysis.
The modes of transport can be passed as a list of different
r5py.TransportModes (or their
str equivalent). Meanwhile, the
departure must be a
datetime.datetime. If you search for public
transport routes, double-check that the departure date and time is covered by
the input GTFS data set.
import datetime travel_time_matrix = r5py.TravelTimeMatrixComputer( transport_network, origins=origins, destinations=destinations, transport_modes=[r5py.TransportMode.TRANSIT], departure=datetime.datetime(2019, 5, 13, 14, 0, 0), ).compute_travel_times()
The output of
is a table in which each row describes the travel time (
travel_time) from an
from_id), to a destination (
104329 rows × 3 columns
A note about transit travel times and travel time windows
With transit travel times, individuals can face (sometimes significantly) different total travel times depending on when they start their journey. For example, a rider taking a bus that comes every 15 minutes on the 15-minute mark will face a travel time that is 14 minutes longer if they arrive at the stop at 09:01 versus if they arrive at 09:15.
To account for that, the R5 engine computes a travel time for every minute in a
specified interval, and reports a median (or, with some customization a
percentile other than the median) travel time over that intervale. For example,
if the specified window is 60 minutes long and the
departure parameter is set
to 09:00, then the reported travel time will be the median of the 60
minute-by-minute travel times between 09:00 and 10:00.
In r5py the default
departure_time_window (the interval over which to sample)
is set to 10 minutes. This is done to allow for a result that is as close as
possible to a “single” travel time measure.
Be careful as choosing very low intervals (or even not-very-low) intervals
can have some adverse effects on the ability of R5 to find a route to a
destination. If you are working with transit schedules that have very low
frequencies (large gaps in time between subsequent vehicles), you may want to
departure_time_window is set significantly higher than these
headway gaps. The reasons for this are quite technical, but you can read more if
you are interested in the discussion on this issue by visiting this GitHub
TravelTimeMatrixComputer creates an
all-to-all matrix in long format. In other words, the results contain one row
for every combination of origins and destinations. Since we have
323 origins and 323 destinations,
the output travel time matrix is 104329 rows long.
Alternatively, and possibly more intuitively, we can display the travel time
matrix table as a matrix in wide format, using
travel_time_matrix.pivot(index="from_id", columns="to_id", values="travel_time")
323 rows × 323 columns
Travel times from anywhere to a particular place#
Once the travel time matrix is computed, we can use the data to analyse and visualise different measures of accessibility. For instance, we can filter the table to show all rows for which the destination is the Praça da Sé, a public square in the centre of the city. By plotting the travel times in a map, we can quickly assess how long it takes for residents from different parts of the city to reach this square by public transport.
For this we first create a copy of the result data frame, filtered to contain
only rows with
to_id referencing the Praça da Sé. Then, we join this table to
the input hexagonal grid, and drop any records that have
NaN values, i.e., for
which there was no result. Finally, as we did above, we use the
explore() to display the values in a
PRAÇA_DA_SÉ = "89a8100c02fffff" travel_times_to_centre = travel_time_matrix[travel_time_matrix["to_id"] == PRAÇA_DA_SÉ].copy() travel_times_to_centre = travel_times_to_centre.set_index("from_id")[["travel_time"]] hexagons_with_travel_time_to_centre = ( hexagon_grid.set_index("id").join(travel_times_to_centre) ) hexagons_with_travel_time_to_centre
|89a8100c603ffff||1146||1155||0||POLYGON ((336000.312 7392162.721, 336107.079 7...||49.0|
|89a8100c617ffff||700||463||1||POLYGON ((335685.425 7392005.022, 335792.187 7...||45.0|
|89a8100c60fffff||377||257||0||POLYGON ((336005.719 7392502.350, 336112.488 7...||49.0|
|89a8100c607ffff||743||687||0||POLYGON ((335690.825 7392344.644, 335797.590 7...||40.0|
|89a8100c6abffff||601||29||0||POLYGON ((335375.940 7392186.943, 335482.700 7...||55.0|
|89a8100cea7ffff||34||2483||0||POLYGON ((330455.500 7397136.529, 330562.237 7...||44.0|
|89a8100ce23ffff||382||395||1||POLYGON ((331090.629 7397791.927, 331197.377 7...||52.0|
|89a8100ce37ffff||10||202||0||POLYGON ((330775.717 7397634.124, 330882.461 7...||45.0|
|89a8100c8dbffff||19||438||0||POLYGON ((330460.813 7397476.324, 330567.552 7...||44.0|
|89a8100c8cbffff||12||1375||0||POLYGON ((330466.126 7397816.132, 330572.867 7...||49.0|
323 rows × 5 columns
hexagons_with_travel_time_to_centre.explore( column="travel_time", cmap="YlOrBr", tiles="CartoDB.Positron", )
You can clearly see how travel times do not increase uniformly, but are shorter along the major transport axes (metro, railways, bus corridors).
Another quick way of getting an understanding of how well different parts of the city are served by public transport is to aggregate the travel times from or to each cell over the entire study region. Of course, this creates edge effects, so in our limited example, grid cells further outside will have worse over-all accessibility values. However, if an entire city region, e.g., covering the entire public transport network, can be captured in one analysis, unwanted artefacts of the analysis have a smaller impact.
To aggregate travel times, we can use the
groupby() method of pandas’ data frames, and
one of the different aggregation functions available for the resulting
pandas.GroupBy objects. For instance, to show the median travel time
from any cell to any other cell in our grid, we group the results using
median_travel_times = travel_time_matrix.groupby("from_id").median("travel_time") median_travel_times
323 rows × 1 columns
Again, we can join these median travel times to the hexagonal grid to display a nice map:
hexagons_with_median_travel_times = ( hexagon_grid.set_index("id").join(median_travel_times) ) hexagons_with_median_travel_times.explore( column="travel_time", cmap="YlOrBr", tiles="CartoDB.Positron", )
Braga, C. K. V., Loureiro, C. F. G., & Pereira, R. H.M. (2023): Evaluating the impact of public transport travel time inaccuracy and variability on socio-spatial inequalities in accessibility. Journal of Transport Geography, 109, 103590. DOI:10.1016/j.jtrangeo.2023.103590
Farber, S. & Fu, L. (2017): Dynamic public transit accessibility using travel time cubes: Comparing the effects of infrastructure (dis)investments over time. Computers, Environment and Urban Systems, 62, 30–40. DOI:10.1016/j.compenvurbsys.2016.10.005
Farber, S., Morang, M. Z., & Widener, M. J. (2014): Temporal variability in transit-based accessibility to supermarkets. Applied Geography, 53, 149–159. DOI:10.1016/j.apgeog.2014.06.012
Herszenhut, D., Pereira, R. H.M., Portugal, L. D. S., & Oliveira, M. H. D. S. (2022): The impact of transit monetary costs on transport inequality. Journal of Transport Geography, 99, 103309. DOI:10.1016/j.jtrangeo.2022.103309
Järv, O., Tenkanen, H., Salonen, M., Ahas, R., & Toivonen, T. (2018): Dynamic cities: Location-based accessibility modelling as a function of time. Applied Geography, 95, 101–110. DOI:10.1016/j.apgeog.2018.04.009
Pereira, R. H. M., Braga, C. K. V., Servo, L. M., Serra, B., Amaral, P., Gouveia, N., & Paez, A. (2021): Geographic access to COVID-19 healthcare in Brazil using a balanced float catchment area approach. Social Science & Medicine, 273. DOI:10.1016/j.socscimed.2021.113773
Salonen, M. & Toivonen, T. (2013): Modelling travel time in urban networks: comparable measures for private car and public transport. Journal of Transport Geography, 31, 143–153. DOI:10.1016/j.jtrangeo.2013.06.011
Tenkanen, H. & Toivonen, T. (2020): Longitudinal spatial dataset on travel times and distances by different travel modes in Helsinki Region. Scientific Data, 7(1), 77. DOI:10.1038/s41597-020-0413-y