IEEE Geoscience and Remote Sensing Magazine - March 2021 - 187

detection of new road networks in developing countries
on the OSM map [107].
ASSIGNING ATTRIBUTES: ROAD TYPES
Topological and geometrical features extracted from road
networks can be very useful to predict street types. Figure 3
depicts three types of streets (i.e., secondary, tertiary, and
residential) across the OSM map. It can be observed that
residential street segments are short (a distinctive geometrical feature) and that tertiary roads are connected with
several residential streets (a distinctive topological feature).
A solution to fix incorrect street type tags in OSM is then
presented in [38]. The authors extract topological features
from OSM road network data to train a neural network classifier that predicts whether a street is of the residential or
pedestrian type. This classifier can be useful to find inconsistent street type tags in OSM.
A multigranular graph representation of street networks
is proposed in [39]. This structure combines the primal
(where nodes are road intersections) and dual (where nodes
are fragments of roads) representations of road networks.
This multigranular representation is used in [5] to extract
features and train a random-forest classifier that is able to
sort streets into 21 different categories in OSM. The method
uses a bag-of-words model computed across geometrical
and topological features of the analyzed streets and their
neighbors. The method in [40] employs graphical models
with geometrical and spatial features, such that the parameters of the model are learned by structured support vector
machines (SVMs) [41]. More recently, the authors of [42]
propose a multilayer conditional-random-field model to
perform a hierarchical classification of street types into
coarse- and fine-grained classes. In [43], the authors propose a graph-convolutional-network-based method for estimating driving speeds along road segments. The output
of this method could be used as an additional feature for
models that classify road types.
ASSIGNING ATTRIBUTES: STRUCTURAL
Detecting multilane roads is important for modeling traffic in urban areas. However, the tag " lanes " used to specify
how many traffic lanes a road has is usually empty in OSM.
Therefore, some works [44], [45] have developed methods to detect multilane roads in OSM data by analyzing
the polygons formed by the road network. This is possible
because, frequently, multilane roads are digitized as multiple parallel highways with terminals in street crossings.
In [44], the authors propose to extract geometrical features
(e.g., area, perimeter, and compactness) from polygons obtained from the road network and train an SVM classifier
to predict whether a road has multiple lanes. After that, a
postprocessing step is performed using a region-growing
algorithm to analyze whether roads connected to the predicted multilane thoroughfares also have multiple lanes.
The method proposed in [45] also uses geometrical features
to train a classifier-a random forest, in this case-but the
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

predictions are used to train a second random-forest classifier that uses geometrical and topological features, such as
the percentage of neighboring roads labeled as multilane
thoroughfares by the first classifier. In [46], the authors
use an MRF and data extracted from remote sensing images (e.g., edge information, detected cars, and contextual
information) to correct OSM road centerline locations and
estimate the width of OSM roads.
EXTRACTING ROAD DATA FROM GPS LOCATIONS
Sequences of users' GPS positions (also called tracking data)
can be harnessed to enrich OSM data. This information can
be obtained by the GPS locations of cars and applications
installed on volunteers' mobile devices. In [47], the authors
propose to find errors in OSM data by analyzing patterns
extracted from GPS positions and OSM mapping information. For instance, indoor corridors wrongly labeled as
tunnels in OSM can be detected by verifying whether the
trajectory data come from a pedestrian or a car.
In [48], the authors propose to use GPS positions and
machine learning models for recommending the addition of
new objects to OSM. For instance, GPS positions can help us
predict a missing street in OSM by observing a linear-shaped
agglomeration of points at some location where there is no
street in OSM but where a road network is nearby. By analyzing the spatiotemporal GPS positions, one can also identify,
for example, that a road in OSM is a motorway because of
the high velocity of objects derived from the GPS information data. The authors in [48] extract several features from
GPS information, such as the density of nodes in the trajectory and the speed of movement. Two types of classifiers
are trained: one to predict the geometry and the other to
forecast object attributes, such as motorways, bicycle lanes,
and one- and two-way streets. For geometry classification,
the k-nearest neighbors (KNN) classifier performed better
than other algorithms, such as logistic regression and random forest. It is observed that the KNN model obtains poor
results when detecting polygonal geometries because of
the lack of data along the boundaries of polygonal objects.
For the classification of geographical object attributes, the

FIGURE 3. Three different street types (secondary, tertiary, and
residential, colored in red, blue, and cyan, respectively) over
the OSM map of the city of Wageningen, the Netherlands.
(Source: OSM.)

187

IEEE Geoscience and Remote Sensing Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2021