IEEE - Aerospace and Electronic Systems - July 2021 - 19

D€astner et al.
Figure 27.
Confusion matrix for classification results with random forest of
the four class model.
Figure 25.
Speed versus altitude for different classes of aircraft types. The
lines are fit functions of second order.
Figure 28.
Results for fivefold cross validation runs (blue: civil, red: military).
X-axis show the five cross validation samples. Y-axis show
the true prediction for both classes. Left: Results for GBT. Right:
GBT including heat map information.
Figure 26.
Speed and altitude histogram for different classes of aircraft. The
aircraft models were bundled to seven different classes (top) and
four classes (bottom).
fact that correlations between speed and altitude seem to
allow a separation of data.
Commercial ADS-B data also deliver the aircraft
model that can be used as label for classification tasks.
Though there exist lots of different aircraft models, they
can be bundled to a few classes, but finding a reliable and
useful labeling is challenging. Figure 25 shows the relation
between speed and altitude for a set of seven different aircraft
classes. The quadratic relationship between speed and
altitude is based on the energy balance of a flying aircraft.
So the potential energy is given by Epot ¼ mgh, whereas the
kinetic energy is given byEkin ¼ 1/2mv2.
Figure 26 shows their histograms for speed and altitude
as the histograms for only four classes. Distributions
that are close together cannot be separated visually, nor
can they be separated from a classifier. So if similar
JULY 2021
classes are combined to one class, e.g., business, different
airliner to jet, differences become more obvious and can
be separated by a classifier easier. The training results
with a random forest for the four class model with speed,
altitude, and altitude rate as feature vector is shown in
Figure 27. To work against imbalanced data, distributions
of all classes were reduced to 25 000 samples, but military
transporters were still underrepresented.
The data provider ADS-B Exchange delivers a military
flag in the context data, which can be used as label
and supports machine learning. There already have
been promising approaches to train classifier for the realtime
detection of military aircraft, which mainly are
based on features consisting of track kinematics and locations
[72].
This task is very challenging as there is only very low
share of military information available in the data, which
also varies from day to day: from near 1% up to 3%. Facing
this high imbalancing problem, the oversampling
approach was used. Although the fivefold cross validation
results [73] show that the optimal oversampling varies
with the test samples, see Figure 28. Adding the information
from the military activity heat map to the feature
gave the result a more reliable stability. However, a
true prediction between 60% and 70% for both classes
was achievable.
IEEE A&E SYSTEMS MAGAZINE
19
IEEE - Aerospace and Electronic Systems - July 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - July 2021
