dodgr
is an R package for efficient calculation of
many-to-many pairwise distances on dual-weighted directed graphs, for
aggregation of flows throughout networks, and for highly realistic
routing through street networks (time-based routing considering incline,
turn-angles, surface quality, everything).
Note that most dodgr
algorithms implement parallel
computation with the RcppParallel
library, and by default use the maximal number of available cores or
threads. If you do not wish dodgr
to use all available
threads, please reduce the number manually by first specifying a value
via
::setThreadOptions (numThreads = 1L) # or desired number RcppParallel
Four aspects. First, while other packages exist for calculating
distances on directed graphs, notably igraph
, even that
otherwise fabulous package does not (readily) permit analysis of
dual-weighted graphs. Dual-weighted graphs have two sets of
weights for each edge, so routing can be evaluated with one set of
weights, while distances can be calculated with the other. A canonical
example is a street network, where weighted distances are
assigned depending on mode of transport (for example, weighted distances
for pedestrians on multi-lane vehicular roads are longer than equivalent
distances along isolated walking paths), yet the desired output remains
direct, unweighted distances. Accurate calculation of distances on
street networks requires a dual-weighted representation. In
R, dodgr
is currently the only package
that offers this functionality (without excessive data wrangling).
Second, while igraph
and almost all other routing packages are primarily designed for
one-to-one routing, dodgr
is specifically designed for
many-to-many routing, and will generally outperform equivalent packages
in large routing tasks.
Third, dodgr
goes beyond the functionality of comparable
packages through including routines to aggregate flows throughout a
network, through specifying origins, destinations, and flow densities
between each pair of points. Alternatively, flows can be aggregated
according to a network dispersal model from a set of origin points and
associated densities, and a user-specified dispersal model.
Fourth and finally, dodgr
implements highly realistic
and fully-customisable profiles for routing through street networks with
various modes of transport, and using either distance- or time-based
routing. Routing can include such factors as waiting times at traffic
lights, delays for turning across oncoming traffic, access restrictions,
and the effects of elevation on both cyclists and pedestrians. See the
dedicated vignette on street
networks and time-based routing for more detail.
You can install latest stable version of dodgr
from CRAN
with:
install.packages ("dodgr") # current CRAN version
Alternatively, current development versions can be installed using any of the following options:
# install.packages("remotes")
::install_git ("https://git.sr.ht/~mpadge/dodgr")
remotes::install_git ("https://codeberg.org/UrbanAnalyst/dodgr")
remotes::install_bitbucket ("UrbanAnalyst/dodgr")
remotes::install_gitlab ("UrbanAnalyst/dodgr")
remotes::install_github ("UrbanAnalyst/dodgr") remotes
Then load with
library (dodgr)
packageVersion ("dodgr")
#> [1] '0.2.21'
While dodgr
works with any arbitrary networks, it also
includes numerous functions explicitly intended to be applied to
geodesic coordinates, which are identified whenever input data have
columns labelled “longitude” and “latitude”, or similar. Coordinates for
such data must be in the EPSG:4326 (WGS84) coordinate system.
dodgr
treats coordinates as numbers only, and it is up to
the user to ensure appropriate transformation to WGS84 coordinates prior
to submitting data to dodgr
functions.
dodgr
networksTo illustrate functionality, the package includes an example data set
containing the Open Street Map network for Hampi,
India (a primarily pedestrian village in the middle of a large World
Heritage zone). These data are in Simple Features
(sf
) format, as a collection of LINESTRING
objects. dodgr
represents networks as a simple rectangular
graph, with each row representing an edge segment between two points or
vertices. sf
-format objects can be converted to equivalent
dodgr
representations with the
weight_streetnet()
function:
class (hampi)
#> [1] "sf" "data.frame"
dim (hampi)
#> [1] 236 15
<- weight_streetnet (hampi, wt_profile = "foot")
graph class (graph)
#> [1] "data.frame" "dodgr_streetnet"
dim (graph)
#> [1] 6813 15
The sf
-format network contained 236
LINESTRING
objects, with the
weight_streetnet()
function decomposing these into 6,813
distinct edges, indicating that the sf
representation had
around 29 edges or segments in each LINESTRING
object. The
dodgr
network then looks like this:
head (graph)
geom_num | edge_id | from_id | from_lon | from_lat | to_id | to_lon | to_lat | d | d_weighted | highway | way_id | component | time | time_weighted |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | 1 | 339318500 | 76.47491 | 15.34167 | 339318502 | 76.47612 | 15.34173 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 |
1 | 2 | 339318502 | 76.47612 | 15.34173 | 339318500 | 76.47491 | 15.34167 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 |
1 | 3 | 339318502 | 76.47612 | 15.34173 | 2398958028 | 76.47621 | 15.34174 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 |
1 | 4 | 2398958028 | 76.47621 | 15.34174 | 339318502 | 76.47612 | 15.34173 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 |
1 | 5 | 2398958028 | 76.47621 | 15.34174 | 1427116077 | 76.47628 | 15.34179 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 |
1 | 6 | 1427116077 | 76.47628 | 15.34179 | 2398958028 | 76.47621 | 15.34174 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 |
The geom_num
column maps directly onto the sequence of
LINESTRING
objects within the sf
-formatted
data. The highway
column is taken directly from Open Street
Map, and denotes the kind of “highway” represented by each edge. The
component
column is an integer value describing which of
the connected components of the network each edge belongs to (with
1
always being the largest component; 2
the
second largest; and so on).
Note that the d_weighted
values are often greater than
the geometric distances, d
. In the example shown,
service
highways are not ideal for pedestrians, and so
weighted distances are slightly greater than actual distances. Compare
this with:
head (graph [graph$highway == "path", ])
geom_num | edge_id | from_id | from_lon | from_lat | to_id | to_lon | to_lat | d | d_weighted | highway | way_id | component | time | time_weighted |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | 1 | 339318500 | 76.47491 | 15.34167 | 339318502 | 76.47612 | 15.34173 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 |
1 | 2 | 339318502 | 76.47612 | 15.34173 | 339318500 | 76.47491 | 15.34167 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 |
1 | 3 | 339318502 | 76.47612 | 15.34173 | 2398958028 | 76.47621 | 15.34174 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 |
1 | 4 | 2398958028 | 76.47621 | 15.34174 | 339318502 | 76.47612 | 15.34173 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 |
1 | 5 | 2398958028 | 76.47621 | 15.34174 | 1427116077 | 76.47628 | 15.34179 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 |
1 | 6 | 1427116077 | 76.47628 | 15.34179 | 2398958028 | 76.47621 | 15.34174 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 |
A "path"
offers ideal walking conditions, and so
weighted distances are equal to actual distances.
The many-to-many nature of dodgr
means that the function
to calculate distances, dodgr_distances()
or, for street networks, times, dodgr_times()
,
accepts two vectors or matrices of routing points as inputs (describing
origins and destinations), and returns a corresponding matrix of
pairwise distances. If an input graph has columns for both distances and
weighted distances, and/or times and weighted times, the weighted
versions are used to determine the effectively shortest or fastest
routes through a network, while actual distances or times are summed
along the routes to calculate final values. It is of course also
possible to calculate distances along fastest routes, times along
shortest routes, or any combination thereof, as detailed in the package
vignette on street
networks and time-based routing.
Routing points can, for example, be randomly selected from the
vertices of a graph. The vertices can in turn be extracted with the
dodgr_vertices()
function:
<- dodgr_vertices (graph)
v head (v)
id | x | y | component | n | |
---|---|---|---|---|---|
1 | 339318500 | 76.47491 | 15.34167 | 1 | 0 |
2 | 339318502 | 76.47612 | 15.34173 | 1 | 1 |
4 | 2398958028 | 76.47621 | 15.34174 | 1 | 2 |
6 | 1427116077 | 76.47628 | 15.34179 | 1 | 3 |
8 | 7799710916 | 76.47634 | 15.34184 | 1 | 4 |
10 | 339318503 | 76.47641 | 15.34190 | 1 | 5 |
For OSM data extracted with the osmdata
package (or,
equivalently, via the dodgr::dodgr_streetnet()
function),
each object (vertices, ways, and high-level relations between these
objects) is assigned a unique identifying number. These are retained
both in osmdata
and dodgr
, as the
way_id
column in the above graph
, and as the
id
column in the vertices. Random vertices may be generated
in this case through selecting id
values:
<- sample (v$id, size = 20)
from <- sample (v$id, size = 50)
to <- dodgr_dists (graph = graph, from = from, to = to)
d dim (d)
#> [1] 20 50
Alternatively, the points may be specified as matrices of geographic coordinates:
<- min (graph$from_lon) + runif (20) * diff (range (graph$from_lon))
from_x <- min (graph$from_lat) + runif (20) * diff (range (graph$from_lat))
from_y <- min (graph$from_lon) + runif (50) * diff (range (graph$from_lon))
to_x <- min (graph$from_lat) + runif (50) * diff (range (graph$from_lat))
to_y <- dodgr_dists (graph = graph, from = cbind (from_x, from_y), to = cbind (to_x, to_y)) d
In this case, the random points will be mapped on to the nearest points on the street network. This may, of course, map some points onto minor, disconnected components of the graph. This can be controlled either by reducing the graph to it’s largest connected component only:
<- graph [graph$component == 1, ]
graph nrow (graph)
or by explicitly using the match_points_to_verts()
function with the option connected = TRUE
:
<- match_points_to_verts (v, cbind (from_x, from_y), connected = TRUE)
from <- match_points_to_verts (v, cbind (to_x, to_y), connected = TRUE) to
This function returns an index into the result of
dodgr_vertices
, and so points to use for routing must then
be extracted as follows:
<- v$id [from] # or from <- v [from, c ("x", "y")]
from <- v$id [to]
to <- dodgr_dists (graph = graph, from = from, to = to) d
Flow aggregation refers to the procedure of routing along multiple
ways according to specified densities of flow between defined origin and
destination points, and aggregating flows along each edge of the
network. The procedure is functionally similar to the above procedure
for distances, with the addition of a matrix specifying pairwise flow
densities between the input set of origin (from
) and
destination (to
) points. The following example illustrates
use with a random “flow matrix”:
<- array (runif (length (from) * length (to)), dim = c (length (from), length (to)))
flows length (from)
#> [1] 20
length (to)
#> [1] 50
dim (flows)
#> [1] 20 50
<- dodgr_flows_aggregate (graph = graph, from = from, to = to, flows = flows) f
The result is simply the input graph
with an additional
column quantifying the aggregate flows along each edge:
head (f)
geom_num | edge_id | from_id | from_lon | from_lat | to_id | to_lon | to_lat | d | d_weighted | highway | way_id | component | time | time_weighted | flow |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | 1 | 339318500 | 76.47491 | 15.34167 | 339318502 | 76.47612 | 15.34173 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 | 0.5815996 |
1 | 2 | 339318502 | 76.47612 | 15.34173 | 339318500 | 76.47491 | 15.34167 | 130.000241 | 130.000241 | path | 28565950 | 1 | 93.600174 | 93.600174 | 2.3630118 |
1 | 3 | 339318502 | 76.47612 | 15.34173 | 2398958028 | 76.47621 | 15.34174 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 | 0.5815996 |
1 | 4 | 2398958028 | 76.47621 | 15.34174 | 339318502 | 76.47612 | 15.34173 | 8.890622 | 8.890622 | path | 28565950 | 1 | 6.401248 | 6.401248 | 2.3630118 |
1 | 5 | 2398958028 | 76.47621 | 15.34174 | 1427116077 | 76.47628 | 15.34179 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 | 0.5815996 |
1 | 6 | 1427116077 | 76.47628 | 15.34179 | 2398958028 | 76.47621 | 15.34174 | 9.307736 | 9.307736 | path | 28565950 | 1 | 6.701570 | 6.701570 | 2.3630118 |
An additional flow aggregation function can be applied in cases where only densities at origin points are known, and movement throughout a graph is dispersive:
<- dodgr_flows_disperse (graph = graph, from = from, dens = runif (length (from))) f
For more detail, see the main package vignette, and the second vignette on street networks and time-based routing
All contributions to this project are gratefully acknowledged using
the allcontributors
package following the all-contributors specification.
Contributions of any kind are welcome!
mpadge |
karpfen |
Robinlovelace |
agila5 |
JimShady |
DavisVaughan |
layik |
olivroy |
virgesmith |
richardellison |
coatless |
znmeb |
yihui |
MartinLHazelton |