IEEE - Aerospace and Electronic Systems - July 2021 - 14

Machine Learning Support for Radar-Based Surveillance Systems
Figure 13.
Clustered harbors and anchoring points (red dots and blue shapes)
based on eight-month AIS data at the Riviera in Italy. The yellow
dots are based on the NGA World Port Index [53].
Figure 11.
Principle of clustering of aggregated points.
over the year, it becomes clearly visible when and where
activities have increased or decreased. January displays the
total increase ofactivity since the information from December
2017 is missing.
Smaller time intervals, e.g., days or weeks, no longer
allow a visual assessment, but instead a generation of a
video or automatic analysis supports the evaluation. So pattern
of life can be created, because for every point in time
and every tile, it is known which aircraft has contributed to.
CLUSTERING OF AREAS OF INTEREST
CLUSTERING OF AREAS OF INTEREST
Airports, heliports, ports, harbors, or typical leaving- or
entry zones are examples for areas of interest a trajectory
normally passes through. They are easy to extract using
single-message reports or plots since their characteristics
allow a discrimination from other data, e.g., low speeds or
low altitudes or simply by the fact that trajectories terminate
there.
Extracted plots can then be clustered using algorithms
like DBSCAN or the hierarchical clustering HDBSCAN
[45], [46], [47]. As these types of clustering algorithms
are not parallelizable, the data need to be partitioned into
tiled grids so that small areas can be processed in parallel,
e.g., with the Spark [21], [22].
Finding the best possible tile size is a challenge
because too large areas can contain an amount of data that
exceeds the process memory and leads to a crash [48],
[49]. On the other hand, if the tiles are too small, areas of
interest can be clustered in different tiles and have to be
merged afterwards.
Recognized clusters of points can be shaped with convex
hulls that are ultimately candidates for areas ofinterest
and origin for further investigation [50], [51]. Figure 11
shows the base principle ofthe processing chain.
Figure 12 shows the result of the clustering of military
aircraft reports from ADS-B Exchange in comparison to
airport information from Jane's database [52]. The challenge
for clustering especially for military aircraft lies in
the discrimination of the data, since military aircraft
switch OFF their ADS-B transponder early while landing
or late while starting. So the regions for discriminated
plots become huge and the extracted center is shifted from
the original position. Figure 13 depicts an example for
clustered ports and harbors based on AIS.
RELATIONS GRAPHS BASED ON CLUSTERING
Figure 12.
Example for some extracted military airports in the north of Germany
based on ADS-B Exchange data (red dot). The yellow dots
mark the airports from a Janes database.
14
As result of the clustering, also the contributing objects to
each cluster are known. Using the identifier, e.g., MMSI
or ICAO, relations between objects and harbors/ airports
can be extracted, such as the traveling behavior between
IEEE A&E SYSTEMS MAGAZINE
JULY 2021
IEEE - Aerospace and Electronic Systems - July 2021

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