IEEE Geoscience and Remote Sensing Magazine - March 2021 - 192

classifier, saving considerable user effort and obtaining an
accurate classifier.
NAVIGATION
In several urban areas, building and road network data in
OSM have a quality similar to commercial map services.
Thus, OSM has been used for navigation/routing applications [73]-[75]. In [73], the authors show that routing
services based on OSM data can attain a real-time, shortest-path computation across large areas for web-based
applications as well as for hand-held devices. In [74], the
authors demonstrate that OSM data can be used for pedestrian routing. The paper proposes a solution for efficient
routing in open spaces (e.g., squares, parks, and plazas). A
more recent and extensive analysis of different strategies to
deal with routing that considers open spaces is presented
in [75].
In [76], the authors offer a method for accurate global
vehicle localization. This work harnesses visual odometry
and OSM data, obtaining better localization results than
methods that rely only on visual odometry approaches.
A probabilistic model for autonomous robot navigation
is proposed in [77]. It uses a Markov chain-Monte Carlo
method to combine semantic terrain information extracted
from 3D-lidar data and OSM information.
TRAFFIC ESTIMATION
Traffic prediction is a challenging task that can be very useful for congestion management and vehicle navigation. The
authors of [78] propose a method to predict four classes of
traffic (i.e., good, slow, congested, and extremely congested) in four cities in China. That work uses data obtained
from POIs on Baidu Maps and geographical objects from
OSM (the number of POIs in OSM is limited in China).
This geographical data together with other features, including weather, temperature, and house prices, are used
to train a machine learning model for traffic prediction.
The authors incorporate traffic information from Baidu
Maps as reference data to train an SVM classifier. They observe that traffic congestion data are very unbalanced because most of time the traffic is good, considering all time
intervals in one day (with exception of the rush hours at 9
a.m. and 5 p.m.). Thus, the authors assign higher weights
to the classes with fewer samples. The results show that,
even using class weighting, the accuracy of the model is
high for the good class and poor for the others. Additionally, that method is compared to the Baidu Maps traffic
prediction system, outperforming it during some time
intervals of the day and obtaining reasonably good performance when the model trained in one city is used for
prediction in another metropolis.
In [19], the authors deal with a similar problem: predicting traffic speed using a regression method. They use
a public data set, the University of Illinois at UrbanaChampaign New York City Traffic Estimates [112], for
their experiments. This data set contains hourly average
192

traffic speed measurements on the New York City road
network, obtained from car trips. Several geographical features are gathered from OSM, such as road length, number of roads connected to the analyzed road, number of
neighbor nodes, and roads in the area, as well as temporal features (e.g., the time and whether it is a workday). To
predict the traffic speed in the target areas, the authors propose a cluster-based transfer model for prediction (CTMP)
method, which first clusters the road features of the source
and target areas. Then, the traffic speed of the target area
is computed based on the nearest-neighbor-roads data of
the source area, which contain traffic speed information.
The CTMP shows better results than other baseline methods, such as neural networks and support vector regression.
More recently, in [79], the authors propose a deep learning method that models road network topology to predict
traffic flow in the city of Chengdu, China. They use neural
networks to model the road network topology and residual
learning [80] to replicate spatiotemporal dependencies.
One limitation of that work is that it requires historic traffic
flow reference data for the target location to predict traffic
flows during different time intervals.
FINE-SCALE POPULATION ESTIMATION
Population distribution at the building level is important
for several tasks, such as urban planning and business development. Population estimation at the building-level
scale can be obtained by aerial interpolation. Although this
technique usually requires 3D building models obtained by
lidar data, the authors of [81] propose to use building footprints and POIs from OSM to predict population distribution based on aerial interpolation. The authors of [82] use a
random-forest model to predict population at the grid level
(i.e., the area of study is divided into grid cells) using Baidu
POIs, mobile user density data, and road networks from
OSM. Then, the grid-level estimations are transformed to
building-level approximations.
In [83], the authors propose a CNN-based method to
perform population density estimation at the grid level.
First, the area of study is divided into grid cells of 200 ×
200 m. Then, for each cell, the authors compute several urban features, such as the building area, number of structures, and quantity of POIs. Finally, they use a full CNN
applied across the urban features of the grid cells to obtain
the population estimation of the corresponding grid cells.
The experimental results show that, by training the model
with data from 14 French cities, the replica attains low error
rates in the validation data extracted from the city of Lyon.
DISCUSSION AND FUTURE WORKS
In this section, we discuss potential promising research avenues at the interface of OSM and machine learning.
APPLICATIONS WITH MULTIMODAL DATA
Although the use of several sources of data (multimodal approaches) has proven to be beneficial for solving numerous
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2021

IEEE Geoscience and Remote Sensing Magazine - March 2021

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