See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/372249963

Aircraft Identification Using Machine Learning

Article in International Journal of Innovative Research in Science Engineering and Technology · July 2023

CITATIONS READS
0 1,032

5 authors, including:

Mthobisi Khumalo Ronald Chiwariro

Durban University of Technology Jain University
1 PUBLICATION 0 CITATIONS 20 PUBLICATIONS 38 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ronald Chiwariro on 11 July 2023.

The user has requested enhancement of the downloaded file.

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

Aircraft Identification Using Machine Learning

M. Khumalo1, M. Muduva2, E. Tarambiwa3, V. Musanga4 and R. Chiwariro5

1,2,4
Zimbabwe National Defence University
3
University of Zimbabwe
5
Jain University, India

Abstract— Positive aircraft identification plays a crucial unauthorised missions and a security risk. The authority
role in ensuring the security of the airspace, the safety of the must be able to identify suspicious planes before
populace, state resources and military establishments. something bad happens. This study examined aircraft
Aircraft identification aids in air traffic management by identification using machine learning algorithms [1].
positively identifying each aircraft entering monitored
airspace. Automatic target recognition has allowed the
II. BACKGROUND TO THE STUDY
utilization of machine learning algorithms for the
classification of aircraft types. Machine learning as a sub-
field of artificial intelligence is disrupting many fields by Aircraft identification is mainly achieved through
facilitating computers to learn the rom data they are Identification Friend or Foe system (IFF). As shown in
exposed to on their own. This study dives into machine Figure 1, the transponder searches for an interrogation
learning algorithms to try and pick one that can be best used signal before providing broadcaster information [2]. IFF
for classifying aircraft as friend or foe. In this study, the systems mostly employ radar frequencies. Military and
researchers focused on supervised machine learning for the civilian air traffic control interrogation systems identify
classification task. Various classification algorithms were
friendly aircraft, vehicles, and forces and calculate their
implemented in this study to train models and evaluate their
bearing and range from the interrogator. Both military
accuracy. The algorithms were trained using a dataset made
up of motion features extracted from aircraft flight track and commercial aircraft use the IFF system [3].
data. The study showed that the hat classification of aircraft
can be achieved by training the models using the aircraft
motion features.

Keywords – Aircraft identification, machine learning,

supervised learning

I. INTRODUCTION

This work developed a machine-learning aircraft

identification model. In [1], found that recognising
aircraft entering monitored airspace is important in
determining their identity. The responsible authority can Figure 1: Concepts behind Identification Friend or Foe [4]
make quick decisions if the aircraft is spotted sooner. Some modern systems are now integrating machine-
Aviation authorities must record all flight plans. The learning methods to forecast aerial object behaviour.
records will track every aircraft in the monitored airspace. Graphic image processing dominates current research,
The reported flight plans allow the authority to track and however, automatic aircraft recognition is still in the
identify an aircraft displayed in the monitor room. exploratory stage [3]. Air traffic control uses radar signals
An issue emerges when the authority cannot identify a too [1]. Image-based and radar-based approaches identify
radar-detected aircraft. This emanates from two planes by contour (shape). These approaches have
possibilities. First, the aircraft transponder responding to significant drawbacks, including weather and other
ground station interrogations may malfunction. Second, natural variables limiting image quality [5]. When flying
the aircrew wants to turn off the transponder for security at high speeds, it is hard to snap clear images of most civil
concerns. For the second reason, the aircraft must be on

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 713

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

aircraft because their shapes are similar. Aircraft form IV. LITERATURE REVIEW
depends on sensor distance and direction. These obstacles
make contour-based identification harder to apply in A. Overview of aircraft identification
practice. The study used multiple classification methods Manual binoculars identify aircraft by shape and engine
to train a model to learn from aircraft motion sound. [8]. Weather, height, and visibility affect aircraft
characteristics to solve these obstacles. recognition efficiency. Air traffic control and military
IFF system only positively recognises friendly aircraft applications require aircraft type recognition [3].
and forces [2]. In [6], Friendly forces may not properly Automatic aircraft recognition analyses images to
respond to IFF for a variety of reasons. Equipment identify targets. Automatic aircraft recognition relies on
malfunction, and parties in the area not involved in the image extraction of an aircraft's silhouette and contour.
combat, such as civilian airliners, will not be equipped Aircraft identification requires effective feature usage.
with IFF if an IFF interrogation receives no response or The qualities must be independent of the object's position
an invalid response, resulting in the object not being and orientation and include enough data to differentiate
positively identified as a foe. one object from another. However, the aircraft's
Neural networks and information fusion have been the geometric distortion, which might include shift, scale,
main topics of machine learning studies for aircraft and rotation, is often encountered, therefore image
identification. In these experiments, the effectiveness of patterns must be extracted despite this distortion.
supervised and unsupervised neural networks in aircraft
detection systems was compared [7]. The Adaptive B. Factors (parameters) affecting aircraft
Resonance Theory (ART) has been selected for the identification
unsupervised neural network and the backpropagation Many academics and researchers worldwide have studied
network (BPN) for the supervised one. Other studies have aircraft remote identification. In-depth research into
shown the use of two kinds of input i.e. Radar Cross classification problems has yielded several abstract
Section (RCS) or radar signature and average speed for mathematical models that form the theoretical basis for
aircraft identification. classifiers [9]. Multivariate statistics generally seek to
reduce the number of components needed to conclude the
III. PROBLEM STATEMENT data, according to [10]. Fewer variables make problems
easier to understand and solve. Using feature selection, a
Events such as that which occurred in America on the 11th classification rule's number of variables can be lowered
of September 2001 terrorist attack and the Simon Mann without impacting its performance.
2004 incident of aircraft caught trying to smuggle Considering the classification of an aircraft or event.
weapons are significant incidents driving the need for Radar and imaging sensors collect incident data for
positive aircraft identification. Most ground-based classification. A feature extractor reduces data by
surveillance systems rely on aircraft detection and producing features or attributes that indicate distinct
recognition by the use of radar systems and flight plans kinds of potential occurrences from the measuring
obtained from aviation authorities to designate aircraft system's output. One trait may not be enough to classify
entering monitored air space as a friend or a foe. When events. A feature vector represents a measurement's
the radar system detects an object in the air space, the features as dimensions in feature space [11].
object or aircraft identification code is used to compare Due to aircraft motion performance differences, historical
with authorized flights provided. The challenge arises flight tracks can reveal useful features. In [12], identified
when the aircraft cannot be positively identified using the nine motion performance factors that potentially
given identification code against that from the influence aircraft recognition, however, some are difficult
transponder system. In this study, machine learning to compute and require a high-precision acquisition
algorithms were explored and used to successfully sensor. Thus, this study will examine flight data motion
recognize aircraft. Machine learning algorithms are not aspects such as maximum speed, cruising speed,
yet fully experimented with when it comes to the acceleration, and climb rate. For conceptual clarity, the
identification of aircraft as a friend or a foe, hence the equations in the following subsections were adapted from
study sought to experiment on various algorithms to come [12].
up with the best model.

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 714

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

Aircraft Maximum speed: it was assumed that an aircraft C. Aircraft identification techniques
reaches its maximum speed when the engine is at In this section, the researcher looked at different
maximum thrust. At the maximum speed, the tail of the techniques that are used to distinguish aircraft entering
aeroplane cannot be heat balanced and this state cannot protected airspace as friends or foes.
last long. In the dataset, the researchers approximated the
detected instantaneous maximum speed as the maximum i. Aircraft identification using the identification
speed of the aircraft. friend or foe (IFF) system
In [6], identified IFF as an identification tool used in air
defence operations to identify approaching aircraft. IFF
(1) sends encrypted signals to aircraft via electromagnetic or
At which k is the number of aircraft in the detection radio waves. An aircraft is a friend or an enemy
record and i is the ith detection time point. depending on how it responds. All aircraft have an IFF
transponder that responds to IFF interrogation requests
Cruising speed: it is also known as economic speed, when and indicates its purpose. IFF uses two channels to query
the aircraft does not engage the afterburner, the aircraft and receive aircraft responses.
can stay in the air for the longest time in this state. The IFF can distinguish between friendly and hostile aircraft.
researchers approximated the average speed of the probe These aircraft-connected systems use electromagnetic
to the cruising speed of the aircraft. and radio frequency (RF) sensors to detect friendly,
hostile, and neutral airspace forces [3]. It helps the
military identify allies and potential foes. The device
(2) challenges and verifies passwords as an advanced digital
Where n is the number of probe records of aircraft k. challenge-and-response password system.
Two-channel IFFs transmit 1030 MHz interrogating
Maximum acceleration: this feature indicates the signals and receive 1090 MHz signals. Two military and
maximum capacity of the aircraft to enhance speed. The two civilian aircraft methods are available.
magnitude of the acceleration is related to the propeller's
power, shape and cooling capacity of aircraft. Therefore, ii. Aircraft identification using radar cross-section
the ability of acceleration is an important parameter that and speed
reflects the characteristics of different types of aircraft. The radar cross-section (RCS) measures radar detection.
Bigger RCSs make objects easier to find. The detected
target's power density is compared to the transmitting
(3) source's [7]. RCS components determine each aircraft's
radar cross-section.
The maximum rate of climb: The maximum rate of climb In [3] it was found that RCS can identify aircraft at
reflects the ability of the aircraft to overcome its gravity several frequencies. Compared to an airliner, a stealth
and resistance and is one of the characteristics that can aircraft has flat surfaces, absorbent paint, and features
best reflect the kinematic performance of the attack oriented to reflect the signal elsewhere. RCS advances
aircraft. In the flight track information, the maximum radar stealth technology, especially for ballistic missiles
value of altitude difference in the detection interval is and aircraft. Military aircraft RCS data is usually
considered the maximum rate of climb (RoC). classified.
Aircraft speed that is presented at the radar sensor can be
obtained by using the Doppler principle given by the
(4) following equation:
Since most aircraft have a fixed route, longitude, latitude,
altitude, velocity, heading information, and sensor signal
(5)
types together with the above four characteristics will be
Where fd is the Doppler principle,
considered as the inputs of the proposed classification
v is aircraft speed,
model.
λ is wavelength

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 715

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

θ is the angle between the direction of incoming signal ii. Artificial neural networks
propagation and with the direction of antenna movement Artificial neural networks are computing systems
[1]. inspired by the biological neural networks that constitute
animal brains [18]. According to [19] artificial neuron
D. Machine learning techniques for target functions by receiving a signal then processing it and
identification signalling other neurons connected to it. The output of
Machine learning is the study of creating algorithms that each neuron is calculated by some non-linear function of
produce predictions based on data. Identifying, or the sum of its inputs, where the signal at a connection is
learning, a function f: X → Y that translates the input a real number.
domain X (of data) onto the output domain Y, is the goal In [13], stated that early researchers tried to simulate the
of a machine learning task of possible predictions [14]. nervous system's processing unit so that its method might
The functions are selected from different function classes; be replicated in computing systems, as they noted that
this all depends on the type of machine learning algorithm neural networks are effective at performing recognition
that is being used. If a computer program's performance tasks. More specifically, a neural network is a
at tasks in a class of tasks (T) as measured by a generalization of mathematical models of human
performance measure (P) increases with experience (E), cognition based on the assumption that Information
then it is said to have learned from experience (E) [15]. processing occurs at many simple elements called
To show how quantitatively a machine learning algorithm neurons. Each connection link has an associated
performs, the performance measure p is used. When it connection weight that multiplies the signals
comes to classification problems the ms, the accuracy of transmitted;;and each neuron applies an activation
a model is usually chosen as its performance measure, function to its net input that is typically non-linear to
where accuracy is the proportion for which the system determine its output signals as indicated in Figure 2.
correctly produces the output. The main objective of
experimenting with machine learning algorithms is to
give an overview of the most effective algorithms and
their implementations.

i. Computer vision for target identification

Several universities and research institutions have carried
out and continue to do in-depth research on moving
Figure 2: Mathematical model of a neuron [1]
target-tracking technology [16]. Enormous progress is
being achieved in the area of motion target tracking
technology. For example, [12] did thorough research on iii. Information fusion
the inter-frame difference algorithm hence he proposed In [19] states that information fusion is a technique in
an improved algorithm. combining physical or non-physical information form
A deep learning application called object tracking follows from diverse sources to become a single comprehensive
a collection of initial item detections and creates a unique information to be used as a basis for prediction or
identifier for each one before monitoring the identified estimation of a phenomenon. The prediction or estimation
objects as they move between frames in a video. is then used as the basis for performing decisions or
Following a series of initial item detections, an object actions. Figure 3 illustrates information fusion.
tracking application generates a unique identifier for each
one before tracking the identified objects as they move
between frames in a video [16]. Computer vision has
gained so much traction when it comes to its applications
in artificial intelligence systems. Computer vision has
been used in air defence systems for surveillance
purposes [17].

Figure 3: The concept of information fusion [20]

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 716

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

Referring to [1] for acquiring comprehensive information iv. Naïve Bayes classifier
at the decision level, it was observed that we can choose According to [9] Naïve Bayes classifier is a machine
from a variety of techniques, including the Bayes Method, learning model that applies the Bayes theorem, presented
Dempster-Shafer and System, and Boolean operator in the equation below, for probabilistic classification. By
methods (AND, OR). observing the values (input data) of a given set of features
or parameters, represented as B in the equation, the naïve
iv. Support vector machine Bayes classifier can calculate the probability of the input
Support vector machine (SVM) is a discriminative data belonging to a certain class, represented as A.
classification method that originates from the Equations in this sub-section were adapted from [22]
computational learning theory's structural risk
minimization principle. SVM aims to find the best (7)
classification function to differentiate between units of
classes in training data. By building a hyperplane that For the classification of input data to take place, the
maximizes the margin between two datasets and, as a probabilities of it belonging to each of the existing classes
result, creates the greatest distance between datasets, it is must be determined and the one with the highest
possible to determine the most advantageous probability will be the class to which the input data
classification function for a dataset that can be linearly belongs. Therefore, class a with the highest probability
separated [3]. A visualization of this strategy is illustrated must be found as expressed in the given equation below,
in figure 4. where bi is one of the n features or predictors observed.

(8)

Since a naïve Bayes classifier assumes that all variables

are independent, only a small training data is necessary to
estimate ate parameters necessary for classification. In
Figure 4: Concept behind support vector machines [21] the context of fault diagnosis, classes would represent
faults or a set of faults that a system could develop and
The theory behind support vector machines is that the the predictors would represent the symptoms the system
optimum generalization ability is attained by locating the is presenting. Although naïve Bayes classifiers are easy
maximum margin and, thus, the most ideal hyper-plane. and quick to implement, considering the predictors as
As a result, both the training data and future data exhibit independent variables can be seen as a disadvantage of
the best classification performance. Support vector the method, since in most real fault diagnosis cases, the
machines are designed to maximize the function symptoms can be dependent on each other (a high
presented below about w and b to identify maximum vibration can cause an increase in temperature, for
margin hyperplanes: example).

E. Studies on Aircraft Identification

In this section, the researcher looked at other related
(6)
studies done so far about aircraft identification.
Where t represents training point quantity, i stands for
i. Deep learning for aircraft classification from
Lagrangian multipliers and LP exemplifies the
VHF (Very High Frequency) radar signatures
Lagrangian. Vector w and constant b characterize the
The researchers in this study observed that a full‐wave
hyperplane. The data points that sit on the margin of the
simulation of three size classes of aircraft shows that their
best-separating hyperplane are known as the support
bi-static radar cross-sections are statistically comparable,
vector points. All other data points are disregarded, and
albeit perform differently in time while the plane is flying.
the solution is a linear combination of these support
This difference can be exploited to recognize the size of
vector points.
the aircraft concerning these classes. To deal with various

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 717

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

representations of RCS, the researchers in this study subspace approaches. Range profiles produced by the
suggested either a convolutional neural network (CNN) graphical electromagnetic computing method are used to
or a recurrent neural network (RNN). examine the performances of SVM and three
Machine learning algorithms used in this research were: representative subspace algorithms. The SVM classifier
a 2D‐ CNN classifier of the sparse BS‐RCS, a 1D‐CNN outperforms conventional classifiers in terms of
classifier of the BS‐RCS time series and angles, and an robustness to sample variation, according to experimental
RNN taking as input only the BS‐RCS time series. These results. The researchers concluded that their experimental
three approaches deserve specific comments. In their findings demonstrate that SVM is more resilient to
conclusion they observed that their simulations have variation in feature samples than other classifiers [24].
demonstrated in general, that the BS-RCS values of
aircraft of different sizes (large, medium, and small) are iv. A machine learning based 77 GHz radar target
similar, rendering their discrimination on the sole basis of classification for autonomous vehicles
BS‐RCS values difficult [23]. This study focused on millimetre wave (mmW) radar
which is a powerful essential sensor for state-of-the-art
ii. Target classification using the deep and future autonomous vehicles. Besides the traditional
convolutional networks for SAR images intended functionality of mmW radars in target detection
The focus of this research was on the synthetic aperture and measuring its range and speed, in this work
radar automatic target recognition (SAR-ATR) technique, researchers show that by utilizing the knowledge of
which is typically made up of the extraction of a set of targets’ statistical RCS information, over 90%
features that turn the raw input into a representation and classification accuracy can be achieved for distant targets
a trainable classifier. The accuracy of the classification (range over 50m). For advanced radars with beam-
can be greatly impacted by the feature extractor, which is steering capabilities, the classification accuracy can reach
frequently manually constructed using domain more than 99% in the near range. For the data
knowledge. Deep convolutional networks have lately categorization problem in this study, a machine learning
achieved state-of-the-art performance in several technique based on artificial neural networks (ANN) is
computer vision and speech recognition tasks by utilized [9].
autonomously learning hierarchies of features from huge
training sets. Convolution nets were immediately applied F. Conceptual Framework
to SAR-ATR, but this resulted in significant overfitting Figure 5 depicts the conceptual framework of the study.
because there weren't enough training images.
Researchers proposed a new all-convolutional network
that uses only sparsely linked layers rather than fully
connected ones to lower the number of free parameters.
Experimental findings on the Moving and Stationary
Target Acquisition and Recognition (MSTAR)
benchmark data set show that All Convolution Networks
are significantly more accurate than traditional CNN at
classifying target configuration and version variants, with
an average accuracy of 99% on the classification of ten-
class targets [3].

iii. Radar target classification using support vector

machine and subspace methods Figure 1: Conceptual framework of the proposed model
The researchers found that target classification is an
important area for future research in the radar sector. i. Longitude
They found that a good target electromagnetic scattering An aircraft's east-west position on the surface of the Earth
characteristic for real-time target categorization is the or another celestial body can be determined using the
range profile. To achieve difficult target categorization, geographic coordinate longitude. This position is used to
this study suggested a technique that blends SVM and

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 718

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

assess whether or not the aircraft is flying inside its The proposed implementation process is depicted in
intended azimuth. Figure 7.

ii. Latitude
Similar to longitude, latitude is a different way to describe
how far away from or close to the equator an aircraft is.
It is a metric used to judge whether the position of the
aircraft is within a predetermined route.

iii. Altitude
A measurement of altitude or height is the distance
between an aircraft and a reference datum, typically in the
vertical or upward orientation. Authorities require each
aircraft to fly at a particular altitude.

iv. Speed
The aircraft's change in speed is measured or recognized
by sensors. Our parameter is important to this
investigation since abrupt changes will make aircraft
behaviour suspicious. The percentage of speed values
Figure 7: Classification model application procedure
recorded for the aircraft falling into each of the five-speed
quantiles recorded for the sample of 500 aircraft are used
in this study to observe it. C. Model building
The classification algorithms used are the Gaussian Naïve
v. Heading Bayes, K-Nearest Neighbour (KNN), Decision Tree,
The direction that an aircraft's longitudinal axis is Random Forest, Logistic Regression and SVM. The task
pointing is often stated in terms of degrees from North. was a binary classification problem hence the algorithms
An aeroplane maintains a set heading, and sudden stated above could be trained using the provided data. The
deviations from that heading make an aircraft suspicious. researcher used the append function to define and initiate
the algorithms to train all the models at one go, and also
iv. Sensor Signal the metrics to evaluate their performances were called.
A sensor translates the physical activity to be monitored
into an electrical equivalent for electrical signals to be VI. RESULTS
easily transferred and further processed.

V. METHODOLOGY A. Exploring the categorical features

To effectively build the model, it was first necessary to
Figure 6 depicts the research process adopted. identify the variables or elements that influence aircraft
identification. The dataset utilized for this experiment
was made up of samples with motion features of the
aircraft which were used to classify them as either
friendly or unfriendly. The dataset was segregated into
categorical and numerical variables as shown in figure 8.

Figure 8: Exploring categorical variables

Figure 6: Block Diagram showing the research process

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 719

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

It was discovered that the dataset was made up of two of the training model. The figure below shows features
categorical variables which are ‘adshex’ and ‘type’. importance that was obtained from the information gain
These variables indicate that an aircraft can be identified feature selection technique.
using its call sign (adshex) which is its unique identifier
and by type, that is whether Boeing or Airbus.

B. Target feature distribution

The target feature or the output vector is the feature that
tells whether the detected aircraft was friendly or not. The
following figure depicts the distribution of the target
feature classes.

Figure 14: Feature Selection using information gain

The results revealed that steers and speed variables had a

significant impact on the training of the model. The other
features cannot be ruled out as they also contribute to the
learning of the model. The technique proved that it could
not render desired results for effectively training the
model hence the researchers used another selection
Figure 12: Target class distribution technique. The researchers then used the recursive feature
The research revealed that the target feature was made up selection method to come up with an optimum number of
of two classes depicted by 0 and 1 where 0 is the majority features or columns required for effectively and
class representing ‘friendly’ aircraft while 1 is the efficiently training the model. The algorithm showed that
minority class representing ‘foe’ aircraft. The analysis 25 features out of 30 could be used for effectively training
revealed that the researcher was dealing with a the models.
significantly high imbalanced dataset which required
resampling to deal with the imbalance. The majority class
was made up of 19686 observations while the minority
class comprised only 113 observations.
Figure 15: Feature Selection using RFE

D. Results of classification algorithms

In this work, the researchers experimented with various
machine-learning classification algorithms to come up
with the best model for aircraft identification. The
Figure 13: Resampling the dataset algorithms that were implemented or experimented on are;
Naïve Bayes (GNB), K-Nearest Neighbors (KNN),
To correct the imbalance, the above resampling technique Decision Tree Classifier (DTC), Random Forest
was carried out. Samples from the minority class were Classifier (RF), Logistic Regression (LR), Support
replicated to balance the two classes. Vector Machine (SVM) and Artificial Neural Network
(ANN). The results from the trained algorithms are shown
C. Feature selection in figure 16.
In the experiment, the researchers utilized the information
gain technique to discover relevant or important features

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 720

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

ANN obtained an accuracy score of 99.94% also proving

that it could be highly considered for aircraft
identification purposes. Due to the modest size of the
training dataset, practically all of the necessary features
had been learned in the first iteration. Figure 4.19 below
shows the training and validation accuracy. Figure 4.20
show the training and validation loss.
Figure 16: Summary of model performances

The initial model experiment findings are shown above.

Algorithm performance measurements were based on
Accuracy Score Test, Accuracy Score Train, Recall Score
Train, F1 Score Train, Precision Score Test, and Precision
Score Train. F1 Score Test ranked the best model. Scores
are decimal numbers from 0 to 1 with 1 being 100%.
Random Forest classifier performed best in this
experiment with an F1 Score Test of 99.94% and an
Accuracy Score Test of 99.93%. The above models have
similar accuracy values of 95% to 99.9%. Other data Figure 19: Training and Validation accuracy
reveal that the models identified friendly and foe aircraft The training and validation accuracies are quite
differently. An imbalanced dataset algorithm's accuracy comparable with no significant variation between the two
score was not the ideal metric. Since it combines proving that no overfitting was detected. The model
precision and memory, the F1 score was best. obtained a training accuracy of close to 100% and had a
validation accuracy of more than 98.00%.
Instead of the precision score, the accuracy score was
used to rank models because it gives a percentage of
properly categorised samples when all predicted samples
are considered. Negative samples may outnumber
positive ones in an imbalanced data collection. Even if all
positive samples were misclassified, this would have
increased accuracy. The accuracy score was not the
greatest performance metric for this study's dataset.
Precision score revealed more about the imbalanced
Figure 20: Training and Validation loss
circumstance above.
The aforementioned training loss shows that the model
fits the training data well, and also the validation loss
E. Classification using artificial neural network
showed that the model fit the new data well. The training
(ANN)
and validation losses are both minimal. Both the
ANN was also implemented on the data to evaluate its
validation and training loss stabilize at a certain point.
performance in the aircraft classification task. The ReLu
(Rectified Linear Unit) and Sigmoid functions were used
F. Evaluation and comparison of all the models
as activation functions during the training of the Artificial
After training the classification models with hyper-tuned
Neural Network.
parameters, the models were evaluated and compared
using their accuracy scores. Figure 4.21 below shows
results obtained from evaluation metrics.

Figure 17: ANN Accuracy Score

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 721

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

Since this method was used to examine how the model

will generalize to an independent dataset, the model's
mean value of 99.81% in Figure 23 demonstrated that
prediction could be made with a reasonably high degree
of accuracy.

H. Evaluating the best-performing model

Figure 21: Evaluation of Models using their accuracy

scores
The study showed that machine learning relies on
hyperparameter adjustment. The SVM model
outperformed the ANN in all experiments. The accuracy
score determined the best model after Grid search hyper-
parameter tuning. The SVM was chosen because of its F1
score, Recall, and Precision. All accuracy values are Figure 24: Model Evaluation
above 92%. These findings support the findings of [13]. The SVM model classified aircraft detected with 100%
He suggested utilising a machine learning classification precision, recall, and F1 scores for "friendly" and "foe"
model with motion cues from flight track data to aircraft, respectively. The model's confusion matrix
automatically and reliably identify aircraft types. showed 9 erroneous predictions and 0 false negatives
(Figure 24). The model performs well because it makes a
few erroneous predictions. The model classified aircraft
as friends or adversary with 99.9% accuracy.

I. Summary of results
The researchers found that full flight characteristics data
was needed to train aircraft recognition models. Models
were trained using aircraft motion features to classify
aircraft as friendly or foe. The dataset utilised to train the
models showed an uneven data distribution between
friendly and foe classifications. Exploratory data analysis
Figure 22: Comparison of Models showed that most aspects of the data set were important
The figure above is the comparison of the classification because losing one meant losing essential data on aircraft
algorithms used in this study showing that their accuracy motion characteristics at any given time. One of the
scores were quite high and quite comparable. The objectives was to determine the parameters that affected
accuracies did not vary much hence the study had to rely aircraft identification.
on other metrics to choose the best-performing algorithm. In aircraft identification studies, the SVM model
outperformed all other machine-learning classifiers.
G. Cross-validating the best-performing algorithm When the F1 score was used to rank models trained
without hyper-parameters, the Random Forest classifier
outperformed the others. SVM models outperformed
other classification algorithms after hyperparameter
tuning. The study used multiple performance measures to
evaluate algorithms. Table 4.1 shows performance
algorithm findings.
Figure 23: Cross-validating Support Vector Machine
Model Table 1: Model Summary

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 722

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

VII. CONCLUSION

This study found that less complex categorization

approaches can be used to develop aircraft-type
recognition models. Researchers can test aircraft
identification models with reduced computing
complexity. The study positively identifies aircraft
entering monitored airspaces, helping air traffic
The performance metrics above show that SVM could management and civil and military aviation. Scholars are
predict "friendly" and "foe" aircraft with comparable increasingly studying aircraft type recognition, a vital air
parameters. The model's precision, recall, and F1 score traffic management technique. Military aircraft
suggest that it can predict a batch of 100 planes correctly. identification is crucial to preventing fratricide deaths.
This study tested other categorization algorithms. Air traffic control requires aircraft identification and
landing approval. This paper answered three research
J. Comparison with past work questions. By incorporating ideas that arose during topic
Table 2 compares our research to earlier classification selection and research, this paper’s future work could
systems that used machine learning to identify aircraft. expand its reach.
Reviewing five studies. Despite diverse settings, multiple This paper originally used computer vision to detect and
categorization algorithms scored 90%. Hyper-tuned classify aircraft. It was limited to aircraft identification
models averaged 95%. This study's SVM and Logistic due to a lack of resources and unified optical satellite data
Regression scores averaged 97%–98%. In some research, for aircraft classes. This could be a future research area.
the lack of large datasets may have limited machine Second, the SVM method could incorporate more hyper-
learning algorithms' accuracy. parameters to maximise training to expand the research.
Using CPUs with increased core counts to parallel
Table 2 Comparison with past studies process the algorithm's complexity and GPU clusters to
boost model training can be a solution. Other methods
from the literature study that classify aircraft well can be
chosen. To compare aircraft identification models, more
efficient classification methods could be utilised.

REFERENCE

[1] A. Datumaya, W. Sumari, and I. A. Force, “A

Study on Identification Friend , Foe , or Neutral
Methods : The Performance of Supervised and
Unsupervised NEURAL NETWORKS,” no.
July 2009, 2019.
[2] T. Q. S. Q. H. Wu, Z. Zhang, C. Jiao, C. Li,
“Learn to sense: a meta-learning-based sensing
and fusion framework for wireless sensor
networks,” IEEE Internet Things J., vol. 6, no.
4, pp. 1089–1100, 2019.
Researchers used multiple machine-learning methods to [3] B. Zhao, W. Tang, Y. Pan, Y. Han, and W.
identify aircraft. The table above shows their results. Wang, “Aircraft Type Recognition in Remote
Algorithms will score differently for the same assignment. Sensing Images : Bilinear Discriminative
The researcher's results also differed for different Extreme Learning Machine Framework,” pp. 1–
algorithms when trained on the same dataset. Due of their 24, 2021.
ability to classify things, neural networks always [4] “What is IFF (Identification Friend or Foe)? -
performed well. everything RF.” [Online]. Available:

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 723

 © January 2023| IJIRT | Volume 9 Issue 8 | ISSN: 2349-6002

https://www.everythingrf.com/community/what- Bovolo, “IEEE Journal of Selected Topics in

is-iff Applied Earth Observations and Remote
[5] H. Huang, “Aircraft Type Recognition Based on Sensing,” A Nov. Tech. based Deep Learn. a
Target Track.,” J. Phys. Conf. Ser., vol. 1061, Synth. target database Classif. urban areas
no. 1, 2018. PolSAR data, vol. 11, no. 1, pp. 154–170, 2018.
[6] E. Meyers and N. Prince, “Infantry I . F . F . [18] L. Liu, “A Deep Neural Network-Based Target
( Identification Friend or Foe ) System By,” no. Recognition Algorithm for Robot Scenes,”
11, 2016. Scientific Programming, vol. 2022. 2022. doi:
[7] R. Wu, A. Polk, M. Xi, and T. Kendall, “Big 10.1155/2022/4037625.
Data and Deep Learning Models for Automatic [19] M. Bistron, “Artificial Intelligence Applications
Dependent Surveillance Broadcast ( ADS-B ),” in Military Systems and Their Influence on
pp. 111–114, 2018. Sense of Security of Citizens,” 2021.
[8] “Deep learning-based strategies for the detection [20] M. Ahmadi, B. Slocumb, and Y. Ahmadi,
and tracking of drones using several cameras _ “Deinterleaving of Interfering Radars Signals in
IPSJ Transactions on Computer Vision and Identification Friend or Foe Systems Related
Applications _ Full Text.” papers Deinterleaving of Interfering Radars
[9] X. Cai and K. Sarabandi, “A machine learning Signals in Identification Friend or Foe Systems”.
based 77 GHz radar target classification for [21] G. F. Luger, Artificial Intelligence: Structures
autonomous vehicles,” 2019 IEEE Int. Symp. and Strategies for Complex Problem Solving,
Antennas Propag. Usn. Radio Sci. Meet. vol. 5th. 2005. [Online]. Available:
APSURSI 2019 - Proc., pp. 371–372, 2019, doi: http://www.amazon.com/dp/0321545893
10.1109/APUSNCURSINRSM.2019.8888647. [22] A. C. Müller and S. Guido, Introduction to
[10] “Machine Learning Methods for Radar Machine Learning with Python and Scikit-
Detection, Classification_Estimation, and Learn. 2015. [Online]. Available:
Tracking _ Frontiers Research Topic.” http://kukuruku.co/hub/python/introduction-to-
[11] H. B. M. Pourshamsi, M. Garcia, M. Lavalle, “A machine-learning-with-python-andscikit-learn
machine-learning approach to PolInSAR and [23] J. Fix et al., “Deep learning for aircraft
LiDAR data fusion for improved tropical forest classification from VHF radar signatures,” IET
canopy height estimation using NASA AfriSAR Radar, Sonar Navig., vol. 15, no. 7, pp. 697–
Campaign data,” EEE J. Sel. Top. Appl. Earth 707, 2021, doi: 10.1049/rsn2.12067.
Obs. Remote Sens., vol. 11, no. 10, pp. 3453– [24] J. Liu, N. Fang, Y. J. Xie, and B. F. Wang,
3463, 2018. “Radar target classification using support vector
[12] “Machine Learning-Based Multitarget Tracking machine and subspace methods,” vol. 9, no.
of Motion in Sports Video.” September 2014, pp. 632–640, 2015, doi:
[13] H. Huang et al., “Aircraft Type Recognition 10.1049/iet-rsn.2014.0325.
Based on Target Track,” J. Phys. Conf. Ser., vol.
1061, no. 1, 2018, doi: 10.1088/1742-
6596/1061/1/012015.
[14] J. L. R. Bekkerman, M. Bilenko, Scaling Up
Machine Learning. Cambridge: University
Press, 2012.
[15] T. Mitchell, Machine Learning. McGraw Hill,
1997.
[16] and M. D. L. Li, K. Ota, “Deep learning for
smart industry: efficient manufacture inspection
system with fog computing,” IEEE Trans. Ind.
Informatics, vol. 14, no. 10, pp. 1821–1846,
2018.
[17] S. C. S. De, L. Bruzzone, A. Bhattacharya, F.

IJIRT 158013 INTERNATIONAL JOURNAL OF INNOVATIVE RESEARCH IN TECHNOLOGY 724

View publication stats