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).
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).
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.
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, and the effects of elevation on both cyclists and pedestrians.
You can install
Then load with
To 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
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
sf-format network contained 191
LINESTRING objects, with the
weight_streetnet() function decomposing these into 5,845 distinct edges, indicating that the
sf representation had around 31 edges or segments in each
LINESTRING object. The
dodgr network then looks like this:
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:
"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() or, for street networks, 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
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
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
Alternatively, the points may be specified as matrices of geographic coordinates:
from_x <- min (graph$from_lon) + runif (20) * diff (range (graph$from_lon)) from_y <- min (graph$from_lat) + runif (20) * diff (range (graph$from_lat)) to_x <- min (graph$from_lon) + runif (50) * diff (range (graph$from_lon)) to_y <- min (graph$from_lat) + runif (50) * diff (range (graph$from_lat)) d <- dodgr_dists (graph = graph, from = cbind (from_x, from_y), to = cbind (to_x, to_y))
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:
or by explicitly using the
match_points_to_graph() function with the option
connected = TRUE:
This function returns an index into the result of
dodgr_vertices, and so points to use for routing must then be extracted as follows:
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”:
The result is simply the input
graph with an additional column quantifying the aggregate flows along each edge:
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: