https://communications.science.ankara.edu.

tr

Commun.Fac.Sci.Univ.Ank.Series A2-A3
Volume 62, Number 2, Pages 134-152 (2020)
DOI: 10.33769/aupse.690478
ISSN 1303-6009 E-ISSN 2618-6462
Received by the Editors: February 17, 2020; Accepted: July 01, 2020

LPI RADAR WAVEFORM CLASSIFICATION USING BINARY SVM AND

MULTI-CLASS SVM BASED ON PRINCIPAL COMPONENTS OF TFI

Almıla BEKTAŞ and Halit ERGEZER

Çankaya University, Mechatronics Engineering Department, Ankara, TURKEY

Abstract. Since cognition has become an important topic in Electronic Warfare

(EW) systems, Electronic Support Measures (ESM) are used to monitor, intercept
and analyze radar signals. Low Probability of Intercept (LPI) radars are preferred
to be able to detect targets without being detected by ESM systems. Because of their
properties as low power, variable frequency, wide bandwidth, LPI Radar
waveforms are difficult to intercept with ESM systems. In addition to intercepting,
the determination of the waveform types used by the LPI Radars is also very
important for applying counter-measures against these radars. In this study, a
solution for the LPI Radar waveform recognition is proposed. The solution is based
on the training of Support Vector Machine (SVM) after applying Principal
Component Analysis (PCA) to the data obtained by Time-Frequency Images (TFI).
TFIs are generated using Choi-Williams Distribution. High energy regions on these
images are cropped automatically and then resized to obtain uniform data set. To
obtain the best result in SVM, the SVM Hyper-Parameters are also optimized.
Results are obtained by using one-against-all and one-against-one methods. Better
classification performance than those given in the literature has been obtained
especially for lower Signal to Noise Ratio (SNR) values. The cross-validated results
obtained are compared with the best results in the literature.

1. Introduction

Electronic Warfare (EW) systems are involved in LPI Radars as they are unlikely to
intercept and it is hard to analyse them in detail while they track the targets [1]. By
using the information provided by ES systems, Electronic Counter Measures (ECM)

Keyword and phrases. Low probability of intercept radar, support vector machine, principal component analysis

[email protected]; [email protected]
0000-0002-5253-7825; 0000-0002-7280-0416
© 2020 Ankara University
Communications Faculty of Sciences University of Ankara Series A2-A3: Physical Sciences and Engineering
 WAVEFORM CLASSIFICATION 135

are used for interfering the radar operation by providing false information and noise,
that is why LPI Radar waveform recognition is very important for EW systems.
Different techniques are used to insert signal data into the classifiers such as using
raw data, filtered data [2], arrays or Time-Frequency Images. Some of the Time-
Frequency Analysis (TFA) techniques used for LPI Radar waveform recognition are
Choi Williams Distribution (CWD) [1], [3] - [7], Wigner-Ville Distribution (WVD)
[1], [4], [7], Radon-WVD [8], short-time Fourier Transform (STFT) [9] and discrete
Fourier Transform(DFT) [10]. By using results obtained from these distributions,
images that show various frequency over time are generated. These images are called
Time-Frequency Images. In literature, the commonly used TFA technique for LPI
Radar waveform recognition is CWD because WVD-TFI contains interfering ghost
terms and these terms degrade the performance of classification. The effect of
different TFA techniques on LPI Radar classification may be the subject of another
study. Similar to the ones in the literature, TFIs used in this paper are generated by
using CWD. Unlike the methods given in the literature, during the generation of TFIs
simple operations that do not require much processing time are performed. While
converting the time-frequency graph of the CWD transform, the appropriate
parameters used to improve image quality and pixel values are normalized to clarify
the distinctive features.
Various techniques have been proposed for classification by using Time-Frequency
Images. As a LPI Radar Waveform Recognition Technique [3], [6] uses the
Convolutional Neural Network (CNN). The computational cost has become a
problem because of the generated TFI sizes. [3] proposes a sampling averaging
technique (SAT) which provides a higher sampling rate with a lower computational
cost. However, because of the CNN complexity, TFIs are re-sized by using nearest-
neighbour interpolation. But by using interpolation, some distinctive features of TFIs
may be lost. Instead of this, taking the principal components of the signals that carry
most of the information is a more accurate way to decrease computational
complexity. It is indicated in [3] that to reduce the number of signal samples,
consecutive signal samples are averaged by applying coherent summation, but it is
not clear how to guarantee coherency in real-world applications and the effect of
coherency on the performance of classification. In this paper, to reduce the
computational cost of TFA, Fractional Fourier Transform is utilized. [8] uses Wigner
Ville Distribution (WVD) to obtain TFI. Before creating WVD-TFIs, a basic
threshold was applied to the amplitude of the signals to fix the raised spikes that the
phase changes caused. As a classification technique, they propose the Fractional
Fourier transform (FRT) to reduce the computational cost. Proposed LPI Radar
waveform recognition technique excludes the four Poly-time signals (T1, T2, T3,
T4). The classification performances of [7] are lower than other techniques, although
136 A. BEKTAŞ, H. ERGEZER

they classified fewer signal types than others. [4] uses a single-shot multi-box
detector (SSD) to detect both Continuous Wave (CW) and Pulse Wave(PW) signals
and a supplementary classifier to classify signals that cannot be classified using SSD.
In [4], it is mentioned that LPI waveform recognition has been performed for both
CW and Pulsed Signals. It is also stated that recognition of CW LPI signals is
challenging due to the lower peak power values of CW signals. However, when we
evaluate the CW and Pulsed signals having the same average power value, CW
signals are clearly distinguishable and traceable in TFI as a continuous curve along
the time axis. This is not the case for Pulsed signals. For Pulsed signals, small zones
of high intensity are observed in certain regions of the TFI. In our study, CW signals
are not considered for classification, but our approach can also be applied to CW
signals. The rationale behind this is that the highest energy region along the time axis
is selected during the automatic crop of the respective zones in the TFI. Therefore,
it becomes clear that the average power of the signal is important, not the peak
power. LPI Radar waveform recognition techniques based on multi-layer perceptron,
radial basis function and probabilistic neural networks are utilized in [7]. To
minimize the information loss, classification is done based on deep sparse capsule
networks in [11] and they used cross ambiguity functions for feature extraction. [5]
proposed an image fusion algorithm and used CNN for feature extraction. Yet, poly-
phase Frank code was excluded in classification. As a result, the similarity between
Frank and other poly-phase codes is not examined. In this case, to compare [5] with
other techniques that include both poly-phase codes and Frank code does not give an
accurate result. Also, clustering [9], decision trees and SVM [5], [12] are commonly
used techniques for classification.
In this paper Linear Frequency Modulation (LFM), Costas, Binary Phase Shift
Keying (BPSK), Frank, P1, P2, P3, P4, T1, T2, T3, T4 signals are used for
classification. Time-frequency images are generated using Choi-Williams
Distribution [13].
All images are cropped automatically to make them include maximum information
and then resized to a constant dimension for all signals to keep data vectors uniform.
SVMs are used to train the dataset. Both binary classification and multi-class
classification methods are used. For multi-class classification, Directed Acrylic
Graph SVM (DAGSVM) method is used which is based on Decision Directed
Acyclic Graph (DDAG) [23].

Support Vector Machine is an effective classification algorithm that is suitable for

LPI Radar classification due to its good generalization performance. A major
limitation of SVM is the high demand for memory during training. To overcome this
limitation, Principal Component Analysis (PCA) is applied to TFIs. Unlike other
 WAVEFORM CLASSIFICATION 137

methods proposed in the literature, principal components have been used instead of
resizing the images that cause entropy increase.

There are a lot of parameters to be turned in SVMs such as choosing the "right"
kernel, regularization penalties, and the slack variable. These parameters must be
optimized to find the best generalization. Firstly, SVMs are trained by using different
Kernel Functions. We notice that different types of LPI signals can be classified
better with different Kernel Functions. Then, it is decided to apply Hyper-Parameter
Optimization to the Kernel Functions as well. Therefore, the optimal Kernel
Function selection is automatically done by Hyper-Parameter Optimization. Then,
the optimal parameter sets are used for each SVM. 5-fold cross-validation results of
both binary classification and multi-class classification methods are presented. For
both of the methods, the same folds are used to be able to compare the performance
of binary classification method and multi-class classification method. All SNR
values of TFIs between -20dB to 10dB are uniformly distributed and fairly separated
between folds.

The paper is organized as follows. Firstly, the mathematical expressions are given in
Section 2. The TFA technique used in the proposed solution is expressed and LPI
Radar Signal TFIs are shown in Section 3. Then, the proposed solution is explained
in detail in Section 4. Section 5 contains the results and comparisons with results in
the literature and then all sections are concluded in 6.

2. Lpi Radar Signals

Mathematical explanations of LPI Radar signals used for the proposed solution are
given below.

A. Frequency Modulation Continuous Wave (FMCW)

Frequency modulation of FMCW signals is created by using increasing and
decreasing frequency signals. Increasing and decreasing frequencies fi and fd are [1]:
∆𝐹 ∆𝐹
𝑓𝑖 = (𝑓0 − 2
)+𝑡 𝑡 (1)
𝑚

∆𝐹 ∆𝐹
𝑓𝑑 = (𝑓0 + 2
)−𝑡 𝑡 (2)
𝑚
138 A. BEKTAŞ, H. ERGEZER

respectively, where f0 is the carrier frequency, t is time, ΔF is modulation bandwidth

and tm is modulation period. The modulation bandwidth of 250, 500,750 and 1000Hz
are used to generate FMCW signals.
B. Binary Phase Shift Keying (BPSK)
Phase modulation of the BPSK signal is provided by Barker Sequences with various
lengths. A Barker Sequence B = [b0, b1, …, bn] contains +1’s and -1’s of length n ≥
2. In this paper, Barker Sequence length of 7, 11 and 13 are used to generate BPSK
signals.
C. Costas
Costas signal contains a set of frequencies that are chosen from the available
frequencies f1, f2, f3, …, fm. Frequency modulation of the carrier signals is provided
by the set [22]. When creating the dataset to train SVMs, 3 different sets of
frequencies with different lengths are used.
D. Polyphase
Polyphase signals' modulation is applied by phase functions given in equations (4)
to (7). In order to approach a stepped or linear frequency modulation, the number of
phase states is varied and the time spent at each phase state is constant. 5 different
phase functions for polyphase signals are given [22]:
Frank
2𝜋
𝛷𝐹𝑅 (𝑖, 𝑗) = 𝑁 (𝑖 − 1)(𝑗 − 1) (3)
𝑝

P1
−𝜋
𝛷𝑃1 (𝑖, 𝑗) = 𝑁𝑝
(𝑁𝑝 − (2𝑗 − 1)) ((𝑗 − 1)𝑁𝑝 + (𝑖 − 1)) (4)

P2
𝜋 𝑁𝑝 −1 𝜋
𝛷𝑃2 (𝑖, 𝑗) = { 2 [( 𝑁𝑝
) − (𝑁 ) (𝑖 − 1)]} [𝑁𝑝 + 1 − 2𝑗] (5)
𝑝

where i, j=1…Np. Np is the number of phase states.

P3
𝜋
𝛷𝑃3 (𝑘) = 𝑁 (𝑘 − 1)2 (6)
𝑝
 WAVEFORM CLASSIFICATION 139

P4
𝜋
𝛷𝑃4 (𝑘) = ( (𝑘 − 1)2 − 𝜋(𝑘 − 1)), (7)
𝑁𝑝

where k=1, ..., Np and Np is the number of phase states.

E. Polytime
As it is stated earlier in the Polyphase section, the time spent at each phase state is
constant. In Poly-time signals the number of phase states is user-defined and the time
spent at the phase states is varied to approach a stepped or linear frequency
modulation. Phase functions of Poly-time codes are [1]:
T1
2𝜋 𝑗𝑁𝑝
𝛷𝑇1 (𝑡) = 𝑚𝑜𝑑 {𝑁 ⌊(𝑚𝑡 − 𝑗𝑇) 𝑇
⌋ , 2𝜋}
𝑝
(8)
T2
2𝜋 2𝑗−𝑚+1 𝑁𝑝
𝛷𝑇2 (𝑡) = 𝑚𝑜𝑑 {𝑁 ⌊(𝑚𝑡 − 𝑗𝑇) ( 𝑇
) 2 ⌋ , 2𝜋}
𝑝
(9)
T3
2𝜋 𝑁𝑝 ∆𝐹𝑡 2
𝛷𝑇3 (𝑡) = 𝑚𝑜𝑑 {𝑁 ⌊ 2𝑡𝑚
⌋ , 2𝜋}
𝑝
(10)
T4
2𝜋 𝑁𝑝 ∆𝐹𝑡 2 𝑁𝑝 ∆𝐹𝑡
𝛷𝑇4 (𝑡) = 𝑚𝑜𝑑 {𝑁 ⌊ 2𝑡𝑚
− 2
⌋ , 2𝜋},
𝑝
(11)
where j = 0, 1, 2, …, m-1 and m is the number of frequency segments, 𝑁𝑝 is the
number of phase states and T is overall code period.
140 A. BEKTAŞ, H. ERGEZER

3. Time-frequency Analysis Technique: choi-williams

distribution

Choi Williams Distribution (CWD) [13], [14] has been used to create time-frequency
images of the LPI Radar signals. It is a time-frequency analysis technique that is
used to extract data from the signal. CWD is included in Cohen’s generalized class
as a time-frequency distribution. It uses an exponential kernel function while
Wigner-Ville Distribution (WVD) uses a kernel as one. The exponential kernel
function makes CWD different than other distributions. CWD is expressed in
discrete form as:
1 2 2
𝐶𝑊𝐷𝑥 (𝑡, 𝜔) = 2 ∑∞ 𝜏=−∞ 𝑒
−𝑗2𝜔𝜏 ∑∞
𝜏=−∞ 𝑒 −𝜎(𝜇−𝑡) ⁄(4𝜏 ) 𝑥(𝜇 + 𝜏)𝑥 ∗ (𝜇 − 𝜏) (12)
2
√4𝜋𝑛 ⁄𝜎

where t is the time index, ω is the angular frequency, x(µ) is the time signal and x*(µ)
is its complex conjugate, τ is the time delay and σ is scaling factor [15].

Figure 1. 12 LPI Radar signals TFIs that are created by using CWD. All signals created
with a SNR value of 0dB.

CWD images of each LPI Radar signal and No-Modulation signal have created at 16
different SNR values, which varies -20dB to 10dB with a step of 2dBs. All of the
signals are created 20 different times with 20 different random noises to make the
effect of noise more realistic. It is mentioned in Section 2 that some of the signals
 WAVEFORM CLASSIFICATION 141

are created with different parameters. In Figure 1, 12 different LPI Radar signals at
0dB are shown.

4. PROPOSED LPI RADAR WAVEFORM CLASSIFICATION

TECHNIQUE
The block diagram of the proposed technique is given in Figure 2. To recognize 12
different LPI waveforms given in Section 2 and the No-Modulation signal, signals
are sampled at a sampling rate of 20 kHz. Then, Choi Williams Distribution is
applied to the sampled data. The output CWD-TFI is automatically cropped
according to the high energy regions on the image. Cropped images are resized to a
fixed size in order to provide uniformity when creating the dataset. This process has
been repeated for every image which has different waveforms, parameters and SNR
values. A total of N (=33600) images have been created in RGB format, each with a
dimension of 64x64 pixels. The resulting RGB images are transformed into feature
vectors of size M (=64x64x3 = 12288). Since we have used RGB format instead of
transforming TFIs into greyscale images, distinctive features of TFIs are preserved.
All feature vectors are collected into a NxM matrix to create the dataset which is the
input of PCA. With the application of PCA, a new NxKe dataset matrix is created
using the features that contain the highest information. Ke is the number of features
with the highest information in the feature vector. By this way, the data load is
reduced. These principal components are used to train SVM. For one-against-all
SVMs, the signal type to be trained is labelled as +1 and all other signal types are
labelled as -1. For example, to train SVM for BPSK, the BPSK images are labelled
as +1, other classes corresponding to 11 LPI signal types and No-Modulation signal
are labelled as -1. While training the classifier, the parameters including kernel types
and kernel parameters are optimized to obtain the best scenario. For each signal type,
5 SVMs are trained by 5-fold cross-validation. A total of 65 SVMs are trained for
one-against-all method and a total of 390 SVMs are trained for one-against-one
method. At the end of this process, we have PCA centres and trained classifier.
SVM is an algorithm that is generally used to classify the data with a boundary. The
total distance between the closest feature vectors of both and the boundary is called
margin. Assume (xi, yi) is a feature vector where xi is the feature and yi is the class
label for i = 1, 2, …, N. 𝑥𝑖 ∈ 𝐼𝑅 𝑝 and 𝑦𝑖 ∈ {−1, 1}.
A hyperplane is defined by
{𝑥: 𝑓(𝑥) = 𝑥 𝑇 𝛽 + 𝛽0 = 0} (13)
where β is a unit vector. Classification rule is:
142 A. BEKTAŞ, H. ERGEZER

𝐺(𝑥) = 𝑠𝑖𝑔𝑛[𝑥 𝑇 𝛽 + 𝛽0 ]. (14)

The correct classification satisfies 𝑦𝑖 𝑓(𝑥𝑖 ) > 0 ∀𝑖 .

FIGURE 2. The flow diagram of the proposed LPI Radar Waveform classification
technique.

Then, a hyperplane that creates the biggest margin m between points of both classes
can be written as [16]:
𝑦𝑖 (𝑥𝑖𝑇 𝛽 + 𝛽0 ) ≥ 𝑚, 𝑖 = 1, … , 𝑁 (15)
In classification, the feature vectors of classes can be linearly-separable or not. For
a non-linearly separable case, it is possible to create more flexible boundaries that
minimize the amount of misclassified vectors. By designing a hyperplane classifier
into a new k-dimensional space using the vectors which are the closest to the decision
boundary, the non-linear case can also be classified. Mapping vectors into a high
dimensional space provided by kernel operation.
It is stated that the computational complexity of SVM is independent of the
dimensionality of the kernel space [17]. Thus, SVM can be used for a number of
high dimensional models which make it a highly applicable algorithm. Since CWD
creates a 3-dimensional TFI, the independence of computational complexity from
the dimensions has become advantageous for the proposed technique.
Some of the commonly used Kernel functions (Kf) are described as:
 WAVEFORM CLASSIFICATION 143

dth-Degree polynomial: 𝐾𝑓 (𝑥, 𝑥 ′ ) = (1 + 〈𝑥, 𝑥′〉)𝑑 ,

(16)
2
Radial basis: 𝐾𝑓 (𝑥, 𝑥 ′ ) = 𝑒𝑥𝑝 (−𝛾||𝑥 − 𝑥 ′ || ),
(17)
Hyperbolic tangent [16].: 𝐾𝑓 (𝑥, 𝑥 ′ ) = 𝑡𝑎𝑛ℎ(𝑘1 〈𝑥, 𝑥′〉 + 𝑘2 )
(18)
There is no practical method to choose the best kernel function yet. It is also a
challenging problem. For this purpose, No-Modulation signal and 12 LPI Radar
signals are trained by SVM using 3 different kernel functions. Linear, 2 nd order
polynomial and Radial Basis Function are used as Kernel Function for classification.
Then, different Kernel Functions are used for different signal types by using Hyper-
Parameter Optimization.
During the training process, Hyper-Parameter optimization is applied. There may be
multiple boundaries created by SVM that classifies the data. The most important
parameters that characterize the boundary are penalty factor and Kernel
parameters. By using these parameters, dozens of boundaries can be created that
give correct classification. Finding the boundary that provides the most effectivity
with minimum error is the objective of the SVM algorithm. Therefore, it becomes
necessary to optimize the penalty factor and kernel parameters. These parameters
are called Hyper-Parameters of SVM.
Bayesian Optimization [18] method is used for Hyper-Parameter Optimization.
Bayesian Optimization is done by using a probabilistic model and an acquisition
function. The probabilistic model is based on the observations that were made
previously. The probabilistic model gives primary knowledge to find the locations
of potential Hyper-Parameters. Acquisition function is used by defining an
exploration ratio after each iteration and program evaluates whether the next point is
over-exploiting or not. Exploration ratio provides the balance between exploring new
points vs. concentrating near points. By exploring different points according to the
given acquisition function, it results in a model.
As acquisition function, expected improvement (EI) plus is used.
𝑓𝑚𝑖𝑛 −𝜇(𝑥) 𝑓𝑚𝑖𝑛 −𝜇(𝛾)
𝐸[𝐼(𝜆)] = (𝑓𝑚𝑖𝑛 − 𝜇(𝛾))𝛷 ( 𝜎
)+ 𝜎𝜙 ( 𝜎
) (19)

where γ is a model of predictions which follows the normal distribution, Φ(*) is

standard normal distribution function and ϕ(*) is standard normal density function
and fmin is the best-observed value so far [19].
144 A. BEKTAŞ, H. ERGEZER

The blue points in Figure 3 show the explored Hyper-Parameter values with respect
to estimated objective function values. As it can be understood from the model, the
parameters with the highest estimated objective function values are the optimal
Hyper-Parameters.
As a conclusion, the best scenario for SVM is created by optimizing parameters and
using different Kernel functions.

FIGURE 3. The exploration of objective function vs. SVM parameters during Hyper-
Parameter Optimization

However, SVM is a classifier that requires high computational cost for the training
process. In order to reduce the computational cost, principal components of the data
matrix are obtained by computing the eigenvalues and eigenvectors. Then, Ke largest
eigenvalues are chosen to form a transformation matrix A that contains eigenvectors.
Each 1xM (=1x12288) dimensional feature vector x transformed into a new 1xKe
dimensional vector y that carries most of the information [20]. It can be calculated
by
y = ATx (20)

The correlation matrix Sy is determined by using (32)

Sy = E [yyT] = E[ATxxTA] = ATSxA (21)
 WAVEFORM CLASSIFICATION 145

where S is the diagonal matrix of eigenvalues. E [*] indicates the expected value of
*.
By using the mean square error approximation, the Ke value can be specified for
given conditions.
𝐾 −1
𝑥 = ∑𝑀−1 ′
𝑖=0 𝑦(𝑖)𝑎𝑖 and 𝑥 = ∑𝑖=0 𝑦(𝑖)𝑎𝑖 .
𝑒

Representing x by using x’ gives a mean square error of

2
𝐸[‖𝒙 − 𝒙′‖2 ] = 𝐸 [‖∑𝑀−1
𝑖=𝐾𝑒 𝑦(𝑖)𝑎𝑖 ‖ ] (22)

By eigenvector definition, the mean square error is determined using

∑𝑀−1 𝑇 𝑇 𝑀−1
𝑖=𝐾𝑒 𝑎𝑖 𝐸[𝑥𝑥 ] 𝑎𝑖 = ∑𝑖=𝐾𝑒 𝜆𝑖 (23)

As a result, to specify a Ke for a given mean square error can be determined by (23),
[17].

FIGURE 4. Percentage of variance vs. number of components to keep graph for PCA.

In the proposed technique, the mean square error of 5% is used where the percentage
of variance becomes smaller than 0.01%. The result of the change of percentage of
variance value is shown in Figure 4. A trial and error process is applied to set a mean
square error value. In fact, the mean squared error value can be increased with a very
little amount of information loss. Also, it results in a considerable reduction in the
data load. However, trying to reduce the data load with different mean squared error
values which are greater than 5% result in a performance loss in classification. As
can be seen in Figure 4, from a certain point, the change in the percentage of variance
becomes much smaller than before, meaning that from this point, features after Kth
feature will make minor differences in classification.
146 A. BEKTAŞ, H. ERGEZER

By using PCA, around 550 of 12288 features are kept for one-against-all
classification, around 130 of 12288 for one-against-one classification to represent
95% of each feature vector. The computational cost problem of SVM is reduced
depending on PCA. The dataset with principal components makes the SVM an
appropriate classification algorithm for LPI Radar waveform classification.
Since there are C=13 classes, it is needed to train C*(C-1)/2 = 78 binary SVMs.
These SVMs are used to construct DDAG.
5. RESULTS
In the proposed technique, one-against-one and one-against-all methods are applied.
For the one-against-one method, DDAG is used to find the best class. DDAG
structure is given in Figure 5. The Confusion Matrix of the one-against-one method
is shown in Table 6. For the one-against-all method, the maximum score-values of
13 SVMs are used to find the best class. The Confusion Matrix of the one-against-
all method is shown in Table 5.

FIGURE 5. The Decision Directed Acyclic Graph for finding the best class out of 13 classes.
(Not all of the nodes has been shown for the sake of clarity).
 WAVEFORM CLASSIFICATION 147

In addition, there are 4 important parameters that qualify the classification

performance. Accuracy, recall, precision and F-measure. Accuracy, i.e.
classification rate can be calculated by [21]:
𝑇𝑃+𝑇𝑁
𝑂𝑣𝑒𝑟𝑎𝑙𝑙 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (24)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

Where TP is the correctly classified feature vectors with label +1, FP is the
misclassified feature vectors with label +1 and TN is the correctly classified feature
vectors with label -1, FN is the misclassified feature vectors with label -1.
The overall accuracy of the one-against-all method is 98.73%. The poly-phase
signals especially P1 and P4 are misclassified in between. The results of binary
classification and multi-class classification are similar; the overall accuracy of the
one-against-one method is 98.61%.
The recall is the ratio of the total number of correctly classified vectors of the class
with label +1 and all vectors of the class with label +1. The recall value closer to 1
shows that the class with label +1 is correctly classified.
For a given feature vector labelled as +1, the precision gives the percentage that the
given vector is actually labelled as +1. Precision is given by [21]:
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃+𝐹𝑃 (25)

F-Measure represents both the precision and recall by taking their harmonic mean
to save the balance in between. F-Measure is given by [21]:
2×𝑅𝑒𝑐𝑎𝑙𝑙×𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
𝐹_𝑀𝑒𝑎𝑠𝑢𝑟𝑒 = 𝑅𝑒𝑐𝑎𝑙𝑙+𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛
(26)

Classification performance parameters of one-against-all and one-against-one

methods are given in Table 1 and Table 2, respectively.
Table 1. Accuracy, Recall, Precision, F-Measure values for one-against-all method
Accuracy 0.98730
Recall 0.98989
Precision 0.99237
F-Measure 0.99113

Table 2. Accuracy, Recall, Precision, F-Measure values for one-against-one

method
Accuracy 0.98610
148 A. BEKTAŞ, H. ERGEZER

Recall 0.98996
Precision 0.99174
F-Measure 0.99085

The recall, precision and F-measure values cannot be compared to existing methods
because there are no such values presented in these articles.

Table 3. Classification accuracy value comparison with other proposed techniques

Proposed [3] [4] [5] [8]
Technique
BPSK 98.96% 99.00% 66.17% 98.00% NC**
COSTAS 100% 99.00% 77.67% 99.00% NC**
LFM 100% 94.00% 68.58% 99.00% 95.00%
FRANK 100% 89.00% NA 85.00%
65.83%
P1 87.35% 86.00% 86.00% 85.00%
P2 98.8% 100% 68.08% 100% 95.00%
P3 96.69% 91.00% 64.5% 98.00% 85.00%
P4 84.55% 85.00% *64.08% 91.00% 85.00%
T1 100% 97.00% 66.00% 84.00% NC**
T2 100% 94.00% 94.50% 92.00% NC**
T3 98.36% 95.00% 68.83% 91.00% NC**
T4 96.72% 94.00% NC** 99.00% NC**
BFSK*** NC** NC** NC** NC** 95.00%
No Modulation 99.12% NC** NC** NC** 95.00%
*In [4], P3 signal is taken as two different signal objects and one of this part and P4 are considered as
one class. The other part of the P3 signal is taken as class P3. ** NC: Not Considered. *** BFSK signal
stands for Binary Frequency Shift Keying and used only in [8].

Results of the proposed solution and [3], [4], [5] and [8] are given in Table 3. Since
the precision and F-Measure values are not given in any of these, the reliability of
these accuracy values can be arguable. In [3] the Confusion Matrix is created by
using the results that have SNR value of -6dB. But Table 4 and Table 5 show the
classification performance of all the SNR values examined including -18 and -20dB.
Lower SNR values have substantial effects on classification performance. All of the
accuracy values except the Costas Signal are higher in the proposed solution with
the effect of lower SNRs. Therefore, the performance of the proposed solution can
be said to be better. There is an abrupt change of classification performance for SNR
values smaller than -12dB. In our technique, by the application of PCA, the
confusing elements of the images are also eliminated, so the classification
performance at these SNR values are also higher.
 WAVEFORM CLASSIFICATION 149

[4] proposes a different technique to classify Frank, P1, P3 and P4 codes. The
Confusion Matrix is created by using lower SNR values so the accuracy values are
lower than other techniques. But the point where the abrupt change occurs that is
mentioned earlier is -14dB in [4]. For lower SNR values, the results are similar to
the results obtained from the solution that we propose. However, the precision and
F-Measure values are not mentioned in [4] either, so the overall results cannot be
compared properly.
LPI Radar waveform recognition technique that is proposed in [5] gives better results
for P3 and P4 signals. For poly-time signals, especially for T1, T2 and T3, there are
major differences in the accuracy. The results that we obtained are much higher than
the results in [5]. Since the Frank signal is excluded in [5], the similarity with the
other poly-phase codes are not examined in between.
[8] used higher SNR values and fewer signal types for classification. However, their
classification results are lower. For P4 signal, the classification performance is very
close to the result that we obtained. But the resolution of the result of [8] is lower
than ours and 84,55% can also be rounded to 85,00%. So they can be assumed to be
the same. Simulations have been performed on the computer having 3.7 GHz
processor and 16 GB memory. The required time to train one SVM in the one-
against-all method is 1838 seconds on average, including PCA and Hyper-Parameter
Optimization. The required time to train one SVM in the one-against-one method is
328 seconds on average, including PCA and Hyper-Parameter Optimization.

6. CONCLUSIONS
In this paper, an automatic LPI Radar classification method has been performed
using SVM and Principal Component Analysis. SVMs are trained using methods of
one-against-all and one-against-one with DDAG. Up to the PCA step, all distinctive
features have been preserved.
Table 4. Confusion Matrix of the one-against-all method.
SVM RESULTS
BPSK COSTAS FMCW FRANK NOMOD P1 P2 P3 P4 T1 T2 T3 T4
BPSK 0.9896 0.0000 0.0000 0.0000 0.0083 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0021 0.0000
ACTUAL SIGNAL TYPE

COSTAS 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
FMCW 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
FRANK 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
NOMOD* 0.0088 0.0000 0.0000 0.0000 0.9912 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
P1 0.0000 0.0000 0.0000 0.0000 0.0000 0.8735 0.0158 0.0065 0.1011 0.0000 0.0031 0.0000 0.0000
P2 0.0000 0.0000 0.0000 0.0000 0.0000 0.0031 0.9880 0.0029 0.0060 0.0000 0.0000 0.0000 0.0000
P3 0.0000 0.0000 0.0000 0.0000 0.0000 0.0069 0.0232 0.9669 0.0029 0.0000 0.0000 0.0000 0.0000
150 A. BEKTAŞ, H. ERGEZER

P4 0.0000 0.0000 0.0000 0.0000 0.0000 0.1113 0.0191 0.0172 0.8455 0.0069 0.0000 0.0000 0.0000
T1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000
T2 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000
T3 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0035 0.0000 0.0000 0.9836 0.0129
T4 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0328 0.9672
*NOMOD abbreviation is used instead of No-Modulation signal for the sake of simplicity.

Contrary to previous studies, operations such as averaging and converting to grey

level are avoided. The Hyper-Parameters of SVMs has been optimized including
kernel type by using Bayesian Optimization. The performances of methods have
been obtained through simulations using different SNR levels. The overall accuracy
level of 98.60% has been obtained from one-against-one and 98.73% has been
obtained from the one-against-all method. The measure of classification
performance is evaluated by calculating F-measures. It is observed that the
classification performances are satisfactory compared to existing methods, except
for -18 dB and -20 dB SNR values.
For low SNR values, the accuracy levels are also low for existing methods. The
computational complexity of the method is enough for real-world ESM applications
since the number of feature vectors is reduced using SVD. It can be concluded that
the confused waveform cannot be considered as a big problem since the main
objective of the LPI classification is to assign the jamming method against the LPI
Radar, and similar jamming methods can be applied to confused LPI waveforms.
Table 5. Confusion Matrix of the one-against-one method.
SVM RESULTS

BPSK COSTAS FMCW FRANK NOMOD P1 P2 P3 P4 T1 T2 T3 T4

BPSK 0.9969 0.0000 0.0000 0.0000 0.0083 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0021 0.0000

COSTAS 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

FMCW 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

FRANK 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000
ACTUAL SIGNAL TYPE

NOMOD 0.0055 0.0000 0.0000 0.0000 0.9945 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

P1 0.0000 0.0000 0.0000 0.0000 0.0000 0.8438 0.0250 0.0063 0.1250 0.0000 0.0000 0.0000 0.0000

P2 0.0000 0.0000 0.0000 0.0000 0.0000 0.0063 0.9625 0.0281 0.0031 0.0000 0.0000 0.0000 0.0000

P3 0.0000 0.0000 0.0000 0.0000 0.0000 0.0063 0.0469 0.9438 0.0031 0.0000 0.0000 0.0000 0.0000

P4 0.0000 0.0000 0.0000 0.0000 0.0000 0.1063 0.0219 0.0188 0.8531 0.0000 0.0000 0.0000 0.0000

T1 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 1.0000 0.0000 0.0000 0.0000

T2 0.0000 0.0000 0.0000 0.0031 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.9969 0.0000 0.0000

T3 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.9688 0.0313

T4 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0250 0.9750
 WAVEFORM CLASSIFICATION 151

REFERENCES
[1] Pace, P.E., Detecting and Classifying Low Probability of Intercept Radar. 2 nd Ed.,
Artech House, Norwood, MA, USA, 2009.
[2] Tao, R., Li, B., Sun, H., Research Progress of the Algebraic and Geometric Signal
Processing, Defence Technology, 9(1) (2013), 40-47.
[3] Kong, S.H., Kim, M., Hoang, L.M., Kim, E., Automatic LPI Radar Waveform
Recognition Using CNN, IEEE Access, 6 (2018), 4207-4219.
[4] Hoang, L.M., Kim, M., Kong, S.H., Automatic Recognition of General LPI Radar
Waveform Using SSD and Supplementary Classifier, IEEE Transactions on Signal
Processing, 67(13) 2019, 3516-3530.
[5] Gao, L., Zhang, X., Gao, J., You, S., Fusion Image Based Radar Signal Feature
Extraction and Modulation Recognition, Access IEEE, 7 (2019), 13135-13148.
[6] Huang, Z., Ma, Z., Huang, G., Radar Waveform Recognition Based on Multiple
Autocorrelation Images, Access IEEE, 7 (2019), 98653-98668.
[7] Gulum, T., Autonomous Non-Linear Classification of LPI Radar Signal Modulations,
https://calhoun.nps.edu/handle/10945/3302; 2007 [accessed 24 September 2019].
[8] Kishore, T.R., Rao, K.D., Automatic intrapulse modulation classification of advanced
LPI radar waveforms, IEEE Trans. Aerosp. Electron. Syst., 53(2) (2017), 901-914.
[9] Tong, X., Modelling and realization of real time electronic countermeasure simulation
system based on SystemVue, Defence Technology, 2019.
[10] Deng, B., Luan, J., Cui, S., Analysis of parameter estimation using the sampling-type
algorithm of discrete fractional Fourier transform, Defence Technology, 10(4) (2014),
321-327.
[11] Liu, M., Liao, G., Yang, Z., Song, H., Gong, F., Electromagnetic Signal Classification
Based on Deep Sparse Capsule Networks, Access IEEE, 7 (2019), 83974-83983.
[12] Zeng, X., Wang, S., Bark-wavelet Analysis and Hilbert–Huang Transform for
Underwater Target Recognition, Defence Technology, 9(2) (2013), 115-120.
[13] Choi, H.I., Williams, W.J., Improved time-frequency representation of multicomponent
signals using exponential kernels, IEEE Trans. Acoust., Speech, Signal Process., 37(6)
(1989), 862-871.
[14] Hollinger, K.B., Code optimization for the Choi-Williams distribution for ELINT
applications, https://calhoun.nps.edu/handle/10945/4422; 2009 [accessed 24 September
2019].
[15] Liu, Y., Xiao, P., Wu, H., Xiao, W., LPI radar signal detection based on radial
integration of Choi-Williams time-frequency image, Journal of Systems Engineering
and Electronics, (2015), 973-981.
[16] Hastie, T., Tibshirani, R., Friedman, J., The Elements of Statistical Learning, 2nd Ed.,
Springer New York Inc, New York, NY, USA, 2001.
[17] Theodoridis, S., Koutroumbas, K., Pattern Recognition, 4th Ed., Academic Press. Inc.,
Orlando, FL, USA, 2009.
[18] Snoek, J., Larochelle, H., Adams, R.P., Practical Bayesian optimization of machine
learning algorithms, Advances in Neural Information Processing Systems, 25 (2012),
2960-2968.
152 A. BEKTAŞ, H. ERGEZER

[19] Feurer, M., Hutter, F., Hyperparameter Optimization. In: Hutter F, Kotthoff L,
Vanschoren, J., editors, Automated Machine Learning, Cham: Springer International
Publishing, (2019), 3-33.
[20] Duda, R.O., Hart, P.E., Stork, D.G., Pattern Classification, 2nd Ed., John Wiley & Sons,
Inc. New York, 2001.
[21] Powers, D.M.W., Evaluation: From Precision, Recall and F-Measure to Roc,
Informedness, Markedness & Correlation, Journal of Machine Learning Technologies,
2(1) (2011), 37-63.
[22] Lima, A.F., Analysis of low probability of intercept (LPI) radar signals using
cyclostationary processing, https://calhoun.nps.edu/handle/10945/4944, 2002
[accessed 24 December 2019].
[23] Platt, J. C., Cristianini, N., Shawe-Taylor, J., Large margin DAGs for multiclass
classification, Advances in Neural Information Processing Systems, MIT Press,12
(2000), 547–553.