Machine Learning Algorithms For Identification of Low Signal-To-Noise Ratio Radar Signals

Machine learning algorithms for identification of
low signal-to-noise ratio radar signals
by
Anne Young, B.Sc.
A thesis submitted to the

Faculty of Graduate and Postdoctoral Affairs
in partial fulfillment of the requirements for the degree of
Master of Applied Science in Electrical Engineering
Ottawa-Carleton Institute for Electrical and Computer Engineering

Department of Systems and Computer Engineering
Carleton University
Ottawa, Ontario
August, 2021
© 2021
Anne Young
Abstract
Areas such as spectrum monitoring require identification of known and unknown

radar transmitters to identify known and rogue users. Often such identification needs
to done under conditions where the signal-to-noise ratio is low. This thesis proposes
an approach to determine the unknown radar chirp parameters of a linear frequency
modulated (LFM) radar waveform, assuming that the unknown parameters come
from a given set of known chirp parameters. A concatenated output of matched filters
corresponding to the known set of chirp parameters is presented to four well-known
machine learning architectures, namely decision tree (DT), random forest (RnF),
naı̈ve Bayes (NB) and support vector machine (SVM). Realistic radar parameters for
airborne, marine and weather radars were used in the simulations. The robustness
of the classifiers to parameter mismatch and truncation of the radar pulse were also
studied. DT outperformed the other classifiers except for the truncated pulse case
(where NB and SVM performed better). RnF did not perform acceptably.
ii
Acknowledgments
First, I would like to extend my heartfelt gratitude to my supervisors, Prof. Sreeraman

Rajan and Dr. Bhashyam Balaji, for their unparalleled guidance, mentorship, and
scientific advice. They have inspired me to learn and grow as a researcher and as a
person. Secondly, I would like to thank my research group colleague, David Luong,
for his invaluable technical aid, his feedback on my work, and going above and beyond
to help me finish. I could not have completed this thesis without Prof. Rajan, Dr.
Balaji and David Luong in these strange times. I would also like to acknowledge Prof.
Mae Seto and Dr. Chuck Livingstone for their support, belief in me, and the impetus
to undertake this journey.
Finally, I would like to thank my family and friends for their love, support and
encouragement. I have truly been blessed to have my parents, Helena and Dick
Wagenaar, there for me throughout my life. My sons, Luke and Adam, mean the
world to me.
iii
Table of Contents
Abstract ii
Acknowledgments iii
Table of Contents iv
List of Tables vi
List of Figures vii
Nomenclature xi
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Scope of the Research . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contributions of the thesis . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Literature review 7
2.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Radar Signal Processing and Machine Learning 16

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Fundamentals of Radars and Radar Intercept Receivers . . . . . . . . 16
3.3 Linear frequency modulation (LFM) modulation . . . . . . . . . . . . 18
3.4 Wigner-Ville Distribution . . . . . . . . . . . . . . . . . . . . . . . . 19
3.5 Machine Learning algorithms and performance metrics . . . . . . . . 20
3.5.1 Support Vector Machine (SVM) Classifier . . . . . . . . . . . 21
3.5.2 Naive Bayes (NB) Classifier . . . . . . . . . . . . . . . . . . . 23
iv
3.5.3 Decision Tree (DT) Classifier . . . . . . . . . . . . . . . . . . 24
3.5.4 Random Forest (RnF) Classifier . . . . . . . . . . . . . . . . . 25
3.6 Machine Learning Metrics . . . . . . . . . . . . . . . . . . . . . . . . 25
4 Some Common Radars 29

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Weather radars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.3 Marine radars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.4 Airborne radars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5 Results 43
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Inadequacy of the Wigner-Ville Transform . . . . . . . . . . . . . . . 43
5.3 Proposed approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.4 Classifiers Evaluated in This Thesis . . . . . . . . . . . . . . . . . . . 47
5.4.1 Details of Simulations . . . . . . . . . . . . . . . . . . . . . . 49
5.5 Results of Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.5.1 Problem 1: LFM signals with pre-selected chirp rates . . . . . 50
5.5.2 Problem 2: Airborne radar realistic chirp results . . . . . . . . 51
5.5.3 Problem 3: Weather radar realistic chirp results . . . . . . . . 58
5.5.4 Problem 4: Marine radar realistic chirp results . . . . . . . . . 62
5.5.5 Problem 5: All realistic radar chirp rates results . . . . . . . . 69
5.5.6 Problem 6: Robustness of the classifiers against windowing im-
perfections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5.7 Problem 7: Robustness of the classifier against perturbation in
chirp parameters. . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6 Conclusions and Way Forward 80

6.1 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
List of References 83
v
List of Tables
2 Nexrad WSR Weather Radar Parameters . . . . . . . . . . . . . . . . 32

3 Marine radar frequency band allocations [1]. . . . . . . . . . . . . . . 33
4 Typical characteristics of IMO category radio navigation radars. . . . 33
5 Typical characteristics of river class radio navigation radars. . . . . . 34
6 Typical characteristics of pleasure-craft class radio navigation radars. 35
7 Furono FR7111 Marine Radar Parameters . . . . . . . . . . . . . . . 37
8 Sample marine radar chirp rates. . . . . . . . . . . . . . . . . . . . . 38
9 Air radar pulse transmission schemes by function. . . . . . . . . . . . 38
10 ASR-9 airborne radar parameters. . . . . . . . . . . . . . . . . . . . . 40
11 Sample airborne radar chirp rates. . . . . . . . . . . . . . . . . . . . . 41
12 Summary of realistic chirp rates for airborne, weather and marine radars. 42
13 Radar types and signal parameters for a set of radars that are unknown
to the bank of matched filters. . . . . . . . . . . . . . . . . . . . . . . 51
vi
List of Figures
1 Support Vector Machine Hyperplane [1]. . . . . . . . . . . . . . . . . 22

2 WSR 88D Weather Radar scanning pattern . . . . . . . . . . . . . . 30
3 Wigner-Ville transform of an LFM signal. . . . . . . . . . . . . . . . 44
4 Wigner-Ville transform of a noisy LFM signal at a SNR of 0 dB. . . . 44
5 Wigner-Ville transform of an LFM signal at a SNR of −10 dB. . . . . 45
6 Wigner-Ville transform of a noisy LFM signal at a SNR of −20 dB. . 45
7 Example of a feature vector obtained by concatenating all the matched
filter outputs of an ideal chirp. . . . . . . . . . . . . . . . . . . . . . . 47
8 Concatenated match filter approach applied to LFM signal at -10 dB.
No clear peak is evident making it difficult to accurately classify the
chirp rate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
9 Confusion matrix for a NB classifier at −20 dB SNR (Problem 1). . . 52
10 Confusion matrix for a SVM classifier for −20 dB SNR (Problem 1). . 52
11 Confusion matrix for a DT classifier of realistic airborne chirp rates at
SNRs from 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10
dB (bottom left), -15 dB (bottom mid) and −20 dB (bottom right).
(Problem 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
12 Confusion matrix for a NB classifier of realistic airborne radar chirp
rates at SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10
(Problem 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
13 Confusion matrix for a RnF classifier of realistic airborne radar chirp
(Problem 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
vii
14 Confusion matrix for a SVM classifier of realistic airborne chirp rates at
SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
left), -15 dB (bottom mid) and −20 dB (bottom right). (Problem 2) . 56
15 Accuracy versus SNR level - Airborne radar chirp rates. (Problem 2) 57
16 F1 score versus SNR level - Airborne radar chirp rates. (Problem 2) . 57
17 Confusion matrix for a DT classifier of realistic weather radar chirp
rates at SNRs 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10
(Problem 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
18 Confusion matrix for a NB classifier of realistic weather radar chirp
(Problem 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
19 Confusion matrix for a RnF classifier of realistic weather chirp rates at
20 Confusion matrix for a SVM classifier of realistic airborne radar chirp
(Problem 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
21 Accuracy versus SNR level - Weather radar chirp rates. (Problem 3) . 62
22 F1 score versus SNR level - Weather radar chirp rates. (Problem 3) . 63
23 Confusion matrix for a DT classifier of realistic marine radar chirp
(Problem 4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
24 Confusion matrix for a NB classifier of realistic marine radar chirp
(Problem 4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
25 Confusion matrix for a RnF classifier of realistic marine radar chirp
(Problem 4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
viii
26 Confusion matrix for a SVM classifier of realistic marine radar chirp
rates at SNR 10 dB (top left), 0 dB (top mid) , -5 dB (top right), -10
(Problem 4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
27 Accuracy versus SNR level - Marine radar chirp rates. (Problem 4) . 68
28 F1 score versus SNR level - Marine radar chirp rates. (Problem 4) . . 68
29 Confusion matrix for a DT classifier of all realistic radar chirp rates at
30 Confusion matrix for a NB classifier of all realistic radar chirp rates at
left), -15 dB (bottom mid) and −20 dB (bottom right). (Problem 5) 71
31 Confusion matrix for a SVM classifier of all realistic radar chirp rates at
left), -15 dB (bottom mid) and −20 dB (bottom right). (Problem 5) 71
32 Accuracy versus SNR level - All radar chirp rates. (Problem 5) . . . . 72
33 F1 score versus SNR level - All radar chirp rates. (Problem 5) . . . . 72
34 Truncated chirp of an LFM signal. (Problem 6) . . . . . . . . . . . . 73
35 Confusion matrix for a NB classifier of realistic truncated chirp rates at
36 Confusion matrix for a SVM classifier of realistic chirp rates at SNR
10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
37 Accuracy versus SNR level - All radar truncated chirp rates. (Problem
6) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
38 F1 score versus SNR level - All radar truncated chirp rates. (Problem 6) 76
39 Confusion matrix for classifying signals from unknown radars at a SNR
of 10 dB using a DT classifier. (Problem 7) . . . . . . . . . . . . . . . 78
of 0 dB using a DT classifier. (Problem 7) . . . . . . . . . . . . . . . 78
of −5 dB using a DT classifier. (Problem 7) . . . . . . . . . . . . . . 78
ix
x
Nomenclature
Acc Accuracy
Ae Effective Area
AESA Advanced Electronically Scanned Array
ANN Artificial Neural Network
ARSR Air Route Surveillance Radars
ASL Above Sea Level
ASR Airport Surveillance Radars
AUC Area Under the Curve
Az Azimuth
BPSK Barker Phase Shift Keying
BW Bandwidth
CCQP Convex Constrained Quadratic Programming
CFAR Constant False Alarm Rate
CM Confusion Matrix
cm Centimetre
CNN Convolutional Neural Network
CW Continuous Wave
dB Decibel
dBz Decibel (zero weighted)
xi
xii
DNN Deep Neural Network
DT Decision Tree
EL Elevation
ER Error Rate
fd Frequency
FFT Fast Fourier Transform
FM Frequency Modulation
FMCW Frequency Modulated Continuous Wave
FN False Negative
FP False Positive
FRFT Fractional Fourier Transform
FSK Frequency Shift Keying
Gt,r Antenna Gain, Receive and Transmit
GHz Giga Hertz
H Horizontal
HT Hough Transform
Hz Hertz
IPM Intra-pulse Modulation
k Boltzmann’s Constant
Km Kilometre
kW Kilo Watt
IF Intermediate Frequency
IMO International Maritime Organisation
IMOP Intra-modulation on Pulse

xiii
ITU International Telecommunications Union
LFM Linear Frequency Modulation
LFMCW Linear Frequency Modulation Continuous Wave
LPI Low Probability of Intercept
LS Least Squares
LSTM Long Short-term Memory
MF Matched Filter
MHz Mega Hertz
ML Machine Learning
MLE Maximum Likelihood Estimation
ms Millisecond
MSE Mean Squared Error
MW Mega Watt
N Negative
NN Neural Networks
No Receiver Noise
NB Naı̈ve Bayes
NLFM Non-linear Frequency Modulation
nm Nautical Mile
NN Neural Network
NR No Radar
P Positive
PA Pulse Amplitude
PAR Precision Approach Radar

xiv
PC Pulse Carrier
PD Pulse Distribution
Pd Power Density
Pr Received Power
PRC Precision Recall Curve
Prec Precision
PRF Pulse Repetition Frequency
PRI Pulse Repetition Interval
PRR Pulse Repetition Rate
Pt Transmitted Power
PW Pulse Width
QPSK Quadrature Phase Shift Keying
R Range
RCS Radar Cross Section
RF Radio Frequency
RnF Random Forrest
RMSE Root Mean Squared Error
rpm Revolutions per Minute
SNR Signal to Noise ratio
STFT Short-Time Frequency Transform
SVM Support Vector Machine
TF Time Frequency
To Standard Environmental Temperature
TN True Negative
xv
TP True Positive
V Vertical
W Watt
WVD Wigner-Ville Distribution
WH Wigner-Hough
WVH Wigner-Ville Hough

Chapter 1
Introduction
1.1 Motivation
In radio frequency (RF) spectrum monitoring applications, it is of importance to
determine whether radar transmitters in a given location are operating legitimately.
These environments are abundant with different classes of radar waveforms (e.g. wave-
forms corresponding to marine, airborne, weather, and automotive radars). Harbours
and airports are instances of particularly congested electromagnetic spectrum envi-
ronments [2]. Unregulated radars that do not have licenses to operate in the frequency
band for the appropriate geographical area could interfere with legally regulated radar
emitters in these environments, potentially impacting the safe and secure operation of
harbours or airports and creating serious economic damage. Pirates or illegal fishing
vessels, for example, may use illicit radar transmitters, and suppressing their activities
is a clear public safety goal. Therefore, the ability to determine the presence/absence
of a radar waveform and (if present) determine its parameters and classify its type,
is particularly important for spectrum monitoring authorities in environments such
as airports and harbours. Unfortunately, this problem is made difficult by the fact
that unlicensed radars tend to be low probability of intercept (LPI) radars: they
are designed to be difficult to detect. They may, for instance, transmit only during
short periods of time. For these reasons, a technique to classify the radar signals in a
certain area would be of great utility to spectrum monitoring authorities, especially
in low-Signal-to-Noise Ratio (SNR) scenarios such as at long radar ranges.
A waveform commonly employed in commercial radars is the linear frequency
modulated (LFM) waveform. It is therefore of particular interest to estimate the
parameters of LFM waveforms. This thesis proposes an approach to determine the
1
CHAPTER 1. INTRODUCTION 2
unknown radar chirp parameters of a LFM radar waveform, assuming that the un-
known parameters come from a given set of known chirp parameters. Unlike many
competing techniques, the approach studied here works at low signal-to-noise ratios
(SNRs). The proposed identification of the chirp parameters in this thesis is done on
a single-pulse basis without the need to process multiple pulses. This approach en-
ables the rapid identification of emitters when intercepted passively by an intercepting
radar receiver.
Often, one may not have the luxury of collecting a lot of pulses to enhance detec-
tion performance, particularly in the low-SNR scenario. Under these circumstances,
signal processing and decision making has to be conducted with very a limited amount
of data. This problem has not drawn wide attention in the literature, which has hith-
erto typically focused on older radars (such as mechanically scanned array radars) for
which it was a safe assumption that the same waveform was used for several pulses.
Currently, most receivers operate using a threshold detector, and if detection is de-
clared then the signal parameters are estimated. These detection techniques often use
multiple pulses and hence require a high probability of intercept for the individual
pulses [3]. The signal parameter estimation step typically relies on time-frequency
analysis techniques such as the Wigner-Ville transform [4] or the Hough transform,
which fail at low SNRs (as shown later in this thesis). It may be argued that one
could boost the SNR by integrating multiple pulses, but this leads to many practical
difficulties. In any case, the integration process may require more time than may be
available to classify an unknown emitter.
Modern radars are typically advanced electronically scanned array (AESA) radars.
AESA radars can be programmed in such a way that each pulse can have different
waveforms corresponding to different radar modes, subject only to hardware con-
straints. In short, they offer the flexibility of selecting waveforms on a pulse-by-pulse
basis. It is more challenging to classify radar signals when the waveform parameters
vary pulse by pulse. Hence, the need for recognition of waveform parameters on a
pulse-by-pulse basis. The research in this thesis is intended to explore the possibilities
available for the identification of emitters based on noisy single pulse data.
Traditionally, most radar signal processing at the receiver side begins with
matched filtering. This motivates the use of matched filter outputs for classification
purposes, a use which is surprisingly rare in the literature. In this thesis, supervised
machine learning algorithms are used to identify pulsed LFM radars using the out-
puts of a bank of matched filters. Because the approach in this thesis is totally novel
and has never been studied before, the analysis represents only a preliminary appli-
cation of the combination of matched filtering and machine learning. In this work,
only raw matched filter outputs are considered as inputs to the classifiers; there was
no use of feature selection or extraction. In addition, we have not used any of the
state-of-the-art machine learning techniques based on neural networks that have had
tremendous success in solving many real world problems. However, the investigations
carried out in this thesis gives a lower bound on the possible performance benefits of
the proposed approach.
1.2 Scope of the Research

The problem solved in this thesis is as follows: given a LFM radar pulse whose
defining parameters (carrier frequency and chirp rate) are taken from a known set
of parameters, and which may be contaminated with additive white Gaussian noise,
determine the parameters corresponding to the given LFM pulse.
It is assumed that the intercept receiver is tuned properly to receive the radar
signal and that the radar emitters and the intercept receivers are both stationary.
Doppler effects are not considered as they will not impact the performance as only
single pulse LFM signals are considered in this work. It is further assumed that a given
recording contains at most one radar pulse; the problem of distinguishing multiple
LFM pulses within a single recording is left for future work. Likewise, the only type
of interference is additive white Gaussian noise; any other type of interference (e.g.
channel impairments or clutter) is left for future work.
In order to solve this problem, this thesis adopts a technique based on matched fil-
tering (MF). The motivation for MF-based algorithms is that MF optimizes the SNR.
The match filter output is then fed into a machine learning algorithm to classify the
LFM signal. Thus, in this work, no further transformations for extraction of features
are considered. A bank of matched filters corresponding to the chirp parameters from
the known set of parameters is itself considered as the feature vector and is used as
input to the classifiers.
The proposed machine learning methodology will be tested using single LFM
pulses with a set of predetermined chirp rates. The failure to identify the chirp
rate based on the traditional Wigner-Ville transform is demonstrated in this work,

thus motivating the need for the proposed method. Two sets of LFM parameters
are used: a small and simple set of unknown radar types to demonstrate proof of
concept, then a larger set of parameters drawn from real-life weather radars, naval
radars and airborne radars. For the set of realistic LFM parameters, two scenarios
are considered. In the first scenario, the type of radar is assumed to be known and
a machine learning algorithm is used to identify the chirp rate (the “mode”) of the
radar. Effectively, in this scenario, only subsets of the parameter set are considered.
This is equivalent to identifying the sub-types of radars in an environment where it is
known that only one kind of radar would be present but the variety within the radar
is unknown. In the second scenario, both the type and the mode of the radar are
assumed to be unknown and a machine learning algorithm is used to identify both
the radar type and the mode simultaneously. The effect of not having the whole pulse
for classification will also be investigated. Testing of the classifier is done with an
incomplete pulse. Further, to test the robustness of the classification system, testing
is carried out with radar LFM pulses whose chirp parameters are slightly deviant
from the known set.
Four well-known traditional machine learning algorithms are considered in this
work since classification is done with limited training data and computing power.
Emerging learning architectures such as deep learning architectures and their variants
are not considered. They will be considered in future research. The research in
this thesis work is based entirely on simulations. (Due to the COVID-19 pandemic,
collection of real data was not possible.)
1.3 Contributions of the thesis

The following are the contributions of the thesis:
• This is the first technical work that addresses the problem of classifying chirp
rates and radar types using a single pulse waveform under poor SNR conditions
using a bank of matched filter outputs.
• Unlike other papers in the literature, realistic parameters of the radars are used
in this research, i.e., real world radar type waveforms were simulated in this
study for both training and testing.
• Robustness study of the system were also carried out using variants of real world
radar waveforms.
The following peer reviewed publications have resulted from this research work.
• Conference Paper
– A. Young, D. Luong, B. Balaji, and S. Rajan, “Machine Learning Approach

to Chirp rate Estimation of LFM Radars,” Integrated Communications,
Navigation, and Surveillance (ICNS), September 2020.
• Report
– A.Young, C. Livingstone, “Monitoring and Characterisation Receiver Sig-

nal Simulator – Volume 1 Conceptual Design,” Defence Research and De-
velopment Canada, Ottawa Research Centre, March 2019.
• Invited Oral Presentation
– A. Young, “Wireless Communications and Sensing for Space-based Appli-

cations – The Problem Space,” IEEE Global Signal and Information Pro-
cessing (Global SIP) Workshop on Wireless Communications and Sensing
for Space-based Applications, November 2019.
1.4 Thesis Outline

The following is the organization of the thesis:
• Chapter 1 introduces the problem of detecting LFM radar signal parameters

for low Signal to Noise Ratio (SNR) conditions and how machine learning algo-
rithms are viable techniques to perform that parametric characterisation. The
problem is outlined and thesis objectives and framework are then discussed.
• Chapter 2 provides a literature review relevant to this thesis.
• Chapter 3 provides fundamental signal processing and machine learning algo-

rithms needed for this thesis.
• Chapter 4 details the three types of radars and their modes considered in this
research work.
• Chapter 5 provides details about the simulations and the results. Also it dis-
cusses the results and compares them against current experimental findings.
• Chapter 6 highlights the key findings of this thesis, the contributions made to
this area of research and identifies future areas of investigation.
Chapter 2
Literature review
In this chapter, we review previous work on LFM parameter estimation. This is an

important problem because synthetic aperture radars, commonly found on airborne
and spaceborne platforms, typically use LFM radar signals. The detection and para-
metric estimation of the LFM signal, specifically chirp rates, has been a long-standing
problem in the radar domain.
Many current approaches are based on the sampling theorem and fractional broad-
ening [5]. In prior work, this problem has been investigated using various signal
processing methods for LFM parametric estimation, such as the Maximum likeli-
hood estimation (MLE), Dechirp parameter estimation, the Wigner-Ville distribution
(WVD) time-frequency analysis and the Fractional Fourier transform (FrFT).
Song et al. [6] estimated LFM parameters using a discrete chirp Fourier transform
for SNR values of -9 to 6 dB. The Cramér-Rao Bound (CRB) and root mean squared
error (RMSE) are closely correlated at SNR values higher than 0 dB. Zheng et al. [7]
employed a technique based on parameterised centroid frequency chirp rate distri-
bution (PCFCRD) for noisy multi-component LFM signals. Djurovic et al. [8] used
time-frequency (TF) peak position to estimate the instantaneous frequency (IF) for
non-parametric chirp rate estimation without assuming polynomial phase and applied
this approach to high SNR values of 18, 21 and 24 dB.
Zeng et al. [9] proposed a feature-based recognition algorithm for three kinds of
IMOP signals to detect the phase discontinuities of a BPSK signal and the chirp rate
of an LFM signal. Their approach estimated the chirp rate accurately at 90% for SNR
greater than 5 dB and identified errors greater than 10% in phase for SNR values ¡7
dB. The lack of a priori information was identified as a challenge in radar electronic
warfare (EW) recognition.
7
CHAPTER 2. LITERATURE REVIEW 8
Wu et al. [10] applied the STFT to analyse six radar features and used them
to train support vector machine (SVM) k-nearest neighbour classifiers to extract
feature vectors from radar signals, classify activities, and detect falls. A Markov
model classifier was then applied for human act recognition. The Doppler shift was
used to characterise motions with each having a different frequency. It was able to
classify types of falls into categories but had misclassification errors.
The Wigner-Ville Distribution (WVD) provides good estimates if the signal-to-
noise ratio (SNR) is high and has high resolution in the time-frequency plane [11].
Its main disadvantages include the inability to handle multiple incoming LFM signals
arriving at a radar (congested spectrum), its ineffectiveness when receiving low-SNR
LFM signals, and the difficulty of interpreting the received signal when there are
multiple components to that signal. The WVD uses a time-frequency decomposition
of radar signals to extract key parametric information. As the noise level increases,
more signal interference terms appear and the spectrum becomes blurred [12]. The
WVD requires enough time and spectrum over a pulse stream to accurately char-
acterise signals. As SNR levels decrease below 0 dB, this capability becomes more
strained.
Sirianunpiboon [13] detected weak LFM signals using parametric estimation with
the TF techniques of WVD and WH in gaussian noise. The techniques had issues
with low SNR values due to the dominance of cross-terms at these values. An antenna
array of receivers was used to conduct the parametric estimation to find chirp rate
and noted the deficiency of computational complexity at that time. WVH was also
considered as a technique.
Ahmad et al. [14] used the WVD to sort LPI airborne FM radar signals by fre-
quency and phase. The probability of correctly sorting non-LPI radars (PCS) was
100 percent at -12 dB but needed higher SNR values to correctly sort LPI LFM radar
signals. Zhang et al. [15] investigated the parametric estimation of LFM signals based
on a theoretical approach. They noted that the WVD cross terms decreased the abil-
ity to estimate chirp rates and that WVD has trouble when multiple LFM signals
exist.
Gulum et al. [16] extracted FMCW radar signal parameters using the WHT-
RADON transform. This TF approach shows that it can extract radar modulation
parameters down to -9 dB but not on a pulse-to-pulse basis. Also, in [17], a pseudo-
WVD was used to extract signal parameters of FMCW radar signals where the Hough
transform (HT) identified the signal parameters to -6 dB. These simulated parameters
were to be compared to real radar parameters at a later date.
Guner et al. [18] used the WH transform to detect LPI LFMCW signals at SNR
levels from -3 to -14 dB. Eight methods were evaluated for the probability to detect
the chirp rate and estimate if the signal were present. They achieved a PCRE of 0.96
for their approach at -8dB. Geroleo et al. [19] detected and estimated multi-pulse
LFMCW radar signals and their chirp rates using the WVH transform to search for
the sawtooth LFM pattern in the TF domain. The author noted that current TF
techniques that are premised on detection of a single pulse are sub-optimal and its
method is computationally demanding.
Niczyporuk [20] applied the RADON transform to WVD results to detect chirps
in the SNR range of 0 to -6 dB without a priori information. Results showed that the
receiver did not have information on signal duration or time delay and was unable to
correctly detect the signal at SNR values below 0 dB.
Bouochikhi et al. [21] applied the Teager-Huang-Hough transform to analyse
multi-component LFM signals as it has the advantage that cross-spectral signal com-
ponents are free and has good TF and FD resolution. They noted that the maxi-
mum likelihood estimation (MILE) method is computationally complex and that TF
methods are effective for detection and estimation (WVH, WVD, MF), but the TF
spectrogram method suffers from fixed time and frequency resolutions due to fixed
window length and has cross-terms which limits its application to LFM parametric
analysis. WVD is optimal for maximum energy concentration around IF frequency
but has cross terms. The paper tests LFM signals in noise-free and noisy environ-
ments in a SNR range of 5 to 50 dB. The FrFT can be used to estimate LFM signal
parameters but the computational efficiency is dependent upon the search range and
evaluation of the optimum FrFT order.
Sirianunpiboon et al. [22] used a group invariance approach to detect weak LFM
signals and noted that the FrFT for low SNR values had a window length constrained
by chirp rate and was SNR-limited. It was noted that a windowed chirp rate approach
only detected small chirp rates at low SNR and that discrete ambiguity needs high
SNR. Cheng et al. [23] used the MCR-FrFT to detect multilinear chirp signals for
underwater acoustic sensor networks.
Serbes [24] used the FrFT and a Fibonacci search algorithm to estimate LFM
chirp rate for a small range of chirp rates. The MSE was very low for SNR above 7
dB. In Serbes [25], the authors identified the accuracy of an algorithm that achieves
CRB down to -7 dB for chirp rates 0.1 to 0.4. They did not use ML algorithms,
but their method achieved higher performance than competing algorithms with less
computational cost. They noted that a key problem was estimating chirp rates in
noisy environments.
Aldimashki and Serbes [26] estimated the chirp rates and frequency of airborne
radars using the FrFT for the multi-component case and noisy signals at 5 dB for the
single component case. Mean square error versus SNR was plotted for SNR values
from -10 to 10 dB. They concluded that their method incurred minimal computational
cost.
Qiu et al. [27] proposed a BPSK/LFM hybrid model for signal parameter esti-
mation of non-cooperative radar receivers in a congested environment. Their FrFT
and linear TF model estimated chirp rates at SNR greater than 0 dB. Ma et al. [28]
used amplitude characteristics of LFM in FrFT domain to estimate chirp parame-
ters. They were able to detect the chirp rate at SNR values of -6 to 4 dB but LFM
parameter estimation was best at SNRs greater than 2 dB.
Chen et al. [29] proposed a STFT transformation algorithm to do parametric esti-
mation of LFM signals under low SNR and simulated it in MATLAB. They indicated
that their simulation approach could theoretically be applied to various SNR values
but this would take a significant amount of time and computational power. They
considered chirp rates of 5 to 40 dB. Machine learning estimation was not considered
due to computational and storage requirements; although it could give better accu-
racy. Additionally, this approach cannot conduct a true pulse-to-pulse parametric
estimation. Fourer et al. [30] proposed a chirp rate and IF estimation technique with
STFT for FM signals above 0 dB.
The detection and parametric estimation of LFM signals in particular has been an
issue in the radar domain, but approaches based on matched filtering have not been
well explored. A matched filter is the optimum linear filter for detecting a known
signal in the presence of Gaussian noise. Its performance is non-optimum for non-
Gaussian noise and it requires a priori information on radar emitters [31]. Salem
noted that the matched filter is one of the most effective techniques to maximise the
SNR of chirp radar signals but has the drawback where its sinc response sidelobe
levels degrade performance [32].
Other work has focused on applying machine learning approaches to LFM para-
metric characterisation for TF analysis. Several techniques have been proposed to
conduct parametric estimation of LFM signals using machine learning algorithms.
Convolutional neural networks (CNNs) have been used to classify waveform types [33]
and signal parameters [34]. These machine learning algorithms have been used with
various types of preprocessing. For example, [10] used a short-time Fourier transform
(STFT) to recognize micro-Doppler features in radar signals, then applied Markov
and support vector machine (SVM) classifiers. The Wigner-Ville transform has also
been used [35].
Ruiz et al. [36] used non-parametric Bayesian inference to classify images as states.
It did not require a priori information and found similar results compared to a su-
pervised SVM mode. The kappa metric was used to assess the performance accuracy
of the classifier compared to SVM to properly classify 10 images. Confusion matrices
were not used and the author only discussed classifier running time as the factor to
properly classify images.
Lunden et al. [37] used feature vector recognition to recognise signal waveform
features, such as leading and trailing edge of pulse, bandwidth, and calculated them
using a TF Choi-Williams distribution for LFM signals in pulse compression radars.
Bayesian classifiers were also considered. They were able to classify received wave-
forms as one of 8 types using their hierarchical approach.
Fischell and Schmidt [38] used a SVM classifier to estimate target aspect angle
from bistatic acoustic scattering. The classifier was trained on simulated targets and
appeared to be effective in classifying actual data. Seok [39] used Hidden-Markov,
NN and SVM classifiers to classify active sonar targets and the full sonar LFM signal.
Feature extraction was also implemented. The article notes that combining the results
of multiple classifier output reduces the generalisation error.
El-Khamy and Elsayed [40] classified multi-user chirp LFM signals using wavelets.
Neural networks, support vector machines and maximum likelihood estimators were
used to classify the signals. The probability of a correct decision was only consid-
ered for SNR values ranging from 0 to 50 dB. The Neural Network (NN) classifier
was identified as the best classifier and indicated that signals could be classified in
Gaussian white noise.
Boashash and Ouelha [41] used the WVD to assess the spread of signals in the
TF domain and implemented machine learning classifiers to improve non-stationary
signal classification thorough feature extraction. Three classifiers were used for feature
extraction: SVM, Random Forest (RF), and Artificial NN (ANN). The metrics of
sensitivity, specificity and total accuracy were used as metrics and the authors found
that SVM generalizes well and has good accuracy. They propose that for a multi-class
problem, accuracy should be used to evaluate feature selection.
Su et al. [42] did parameter estimation for chirp signals using a complex deep
CNN that uses CV DNN processor for feature extraction. The chirp rate was varied
from 0.04 to 4 in steps of 0.04 with no overlapping values in training and testing data
to avoid overfitting. SNR values were superimposed on the data ranging from -20 dB
to 100 dB in steps of 10 dB. Simulation results indicated that the approach was more
accurate than WVH transform and that there were benefits to using deep learning
systems for signal parameter estimation.
Lee et al. [43] investigated jamming prediction for radar signals using a combina-
tion of ML classifiers (SVM, NN, and RF) to perform feature extraction and classify
the signals. They obtained an accuracy of 98.46% for NN with extracted features,
but 99.36% for LSTM without feature extraction. However, LSTM took longer to
train than NN.
Gao et al. [44] performed fusion image radar signal feature extraction and modu-
lation recognition using NN to extract features and SVM to classify the image. WVD
was not viable due to cross terms. The confusion matrix (CM) showed an accuracy
of 94.3% at 2 dB for RSR feature extraction for the image. The authors noted that
90% recognition at greater than 0 dB is very good performance.
Wang et al. [45] proposed a hybrid ML approach to extract features of seven
modulation types, including LFM signals. STFT was used in the first layer to conduct
feature extraction while NB and SVM classifiers were used to classify the modulation
waveforms. The authors only looked at SNRs from 10 dB to 30 dB and achieved
accuracy rates of 94.42% at 10 dB for LFM using this approach. With noise, it was
93.42% for SVM and less than 80% accuracy at 0 dB.
Noone [46] used SVM to classify radar signals into various categories. Two signals
were recorded at distances of 3 nmi and 6 nmi for three pulse distribution (PD)
modes, and the mean pulse characteristic for each category was calculated at each
distance. A SVM was used to classify and train on a series of pulses (4340 pulses for
three PD modes). Their algorithm was susceptible to SNR changes, could not handle
noisy environments, and needed other pulse characteristics to categorise the complete
pulse stream.
Kwon et al. [47] used deep neural networks to recognise micro-Doppler radar
signals of movements using FFT results as feature vectors. MLP and DNN were used
for multi-class classifiers. The MLP had 89.8% accuracy for the test set while the
DNN had 97.2% accuracy. The results were presented using confusion matrices. The
authors stated that the DNN can be applied as an effective approach without complex
feature extraction.
Hong et al. [48] classified radar signals using CNN. They wrote that a library-
based method cannot properly classify radar signals due to the complexity of new
radar waveforms, but a ML could. The CNN classifier received radar pulse radar
parameters of RF, PRI, pulse amplitude (PA), PW and estimated a PW for each signal
pulse of the total signal. The proposed method assumed an ideal noise environment
for the pulse stream. The probability of success is 90% after 30 iterations and 93%
after 200.
Carretero et al. [49] trained an ANN on CFAR window attributes and found that
this approach outperforms traditional CFAR methods. They performed Monte-Carlo
simulations at SNR values from -10 to 30 dB.
Ren et al. [50] considered LFM, FSK, BPSK, CW and QPSK waveforms as mod-
ern radars have more sophisticated waveforms and inter-pulse information may not
be enough to separate received pulses according to their source emitters. They used
signal feature extraction using wavelets to simplify classifier and computational com-
plexity by generating smaller-dimensional feature vectors. They simulated 150 sam-
ples at SNRs of 5, 10, 15, and 20 dB for a total of 3000 signals. They defined 34
features such as instantaneous frequency and bandwidth. However, they considered
only one chirp rate. Their best accuracy rate was 95.26% at 5 dB.
Kurdzo et al. [51] used a NN with four hidden layers to optimise a set of NLFM
waveforms and increase their resolution via interpolation. Bandwidth and pulse
length were used as inputs to the NN classifier. The need to determine waveform
parameters on a pulse to pulse basis was noted for future work.
Mughai and Kim [52] conducted signal classification and jamming detection of
wideband radios using a Naı̈ve Bayes (NB) classifier to classify the signal at different
jamming levels. Compressive sensing recovered the wide-band spectrum; they then
trained a NB classifier in feature extraction of spectral features.
Xu et al. [53] applied sparse Bayesian techniques for radar target recognition to
binary and multi-tasks classification problems. They used a LV-SVM to perform

binary classification of each task. Three airborne radars were used with the same
frequency and bandwidth for 524 test samples and 500 training samples, a total of
4721 samples. It was noted that the more samples there are, the better the classifier
recognises the radar and that the performance was better than that of kernel SVM.
Selim et al. [33] used a CNN ML algorithm to detect radar signals in a spectrum
sharing situation, in the presence of signal interference (i.e., congested spectrum) to
infer spectrum occupancy and achieved 99.6% accuracy in detecting RF signals. They
looked at LFM Barker BPSK phase modulation (PM) and pulse carrier (PC) for their
intra-pulse modulation (IPM). They characterised radar pulses and assumed that all
pulses were the same width, transmitting pulses every 1 ms. A trained CNN classifier
was tested at low SNR values of 2 to 20 dB and achieved 99.6% accuracy.
Kong et al. [34] classified radar signals with a CNN ML classifier to model the
real part of the signal waveforms in the time domain. They generated eight types of
waveforms for eight radar types, converted radar signals into the time domain using
STFT, and then extracted the time-frequency distribution in frequency domain. A
CNN extracted radar signal characteristics. The authors show that CNN learns more
about features after STFT. Their approach approximates the radar signal with a
data model for the algorithm to recognise, then performs feature extraction. There
were difficulties in correctly identifying unknown radar signals. According to the
authors, they selected CNNs as they have strong nonlinear fitting and self-learning
abilities. Learning rules are simple and easier to implement. However, CNNs have
more difficulty recognising radar signals.
Wang et al. [54] used support vector clustering with time-frequency analysis and
the RADON transform of multi-component LFM signals to detect signal components
using the WVD, ambiguity function and wavelets. Two experiments were conducted
to detect chirps at SNRs above -7 dB. The assertions that chirp rates could be clas-
sified above -8dB were not substantiated with performance metric values or SNR
results in the paper.
Finally, Caroni and Souryal [55] evaluated the performance of matched filters
to detect radars in the 3.5 GHz band and selected a ML approach to filter out in-
band interference. A priori knowledge of PRI and pulse duration was assumed.
They implemented a bank of MFs over the entire baseband. Results showed that
the magnitude detector performed well at an SNR of 5 dB in Gaussian noise as its
performance exceeded commercial devices.
2.1 Summary
The extensive literature review indicates that much work has been performed on
LFM parameter estimation and classification using conventional techniques such as
the Wigner-Ville transform and other time-frequency methods. To a lesser extent,
matched filtering has been studied as well. On the other hand, a good amount
of work has been devoted to applying machine learning algorithms to radar signal
classification as well. However, very little work has been done on combining the
two approaches. Thus, the approach used in this thesis, namely the use of a bank
of matched filters combined with machine learning algorithms to classify a low-SNR
radar signal is novel. This approach is so novel, in fact, that only a preliminary
investigation is possible in this thesis.
Chapter 3
Radar Signal Processing and Machine

Learning
3.1 Introduction
This chapter provides a quick overview of relevant aspects of radar signal processing
and machine learning algorithms needed for understanding this thesis.
The standard radar waveform that is used in many practical systems, linear fre-
quency modulation (LFM), will be defined, as well as the theory of our matched filter
approach used in our model and the Wigner-Ville distribution (WVD) that we will
use a basis of comparison.
Finally, four basic machine learning algorithms will be explained, along with key
performance metrics that will be used to evaluate the model.
3.2 Fundamentals of Radars and Radar Intercept

Receivers
A radar is an electromagnetic system that detects and locates objects by transmitting
a waveform. A radar consists of a radio frequency (RF) front end, i.e., the receiving
and transmitting antennae, and the system processing components to infer the pres-
ence/absence of a target. The radars are typically installed on static (i.e., weather
radars) or mobile platforms (i.e., airborne, automotive, marine, or space radars). The
target needs to be in the antenna main beam in order to be detected (highest SNR
value). Radars use techniques, such as pulse compression, to achieve higher range
16
CHAPTER 3. RADAR SIGNAL PROCESSING AND MACHINE LEARNING 17
resolution without requiring shorter pulses and maintain the energy of longer pulses.
Continuous waveforms are also used to take advantage of the Doppler frequency, fd ,
shift to separate received signal from echoes due to environmental clutter.
An intercept receiver is a radar receiver which is typically wideband and is used
to detect emissions of transmitting radars.
The radar intercept receiver equation is used to infer signal-to-noise ratio of re-
ceived signals from transmitting radars [56]. The power density, Pd of a radar trans-
mitter at a distance, R, from the radar intercept receiver is given by equation1:
Pr
Pd = (1)
4πR2
Directional antennas are employed by radars to direct the radiated energy in a given
direction. Their antenna gain, Gt , is a measure of the radiated power in a direction
compared to that of an isotropic antenna. The power density received by the intercept
receiver with an antenna with gain, Gr , is thus given by
Pt G r
Pd = (2)
4πR2
The receiving antenna with effective area, Ae will capture part of this echoed power
and the received power will be given by equation 3:
Pt G σAe
Pd = (3)
4πR2 4πR2
A radar’s ability to detect a target’s signal is constrained by the energy of the

ambient noise that occupies the same electromagnetic spectrum as the radar and that
interferes with signal detection. Noise sources vary from inherent receiver sources to
other radars. This noise power is a composite value of the thermal noise and other
noise factors. Suffice to say, that for the purposes of this thesis, we assume that the
noise power is given by
N0 = kT0 BFn , (4)
where N0 is the noise from the receiver, T0 is the standard environmental temperature
of 290 degrees kelvin, and B is the radar intercept receiver intermediate frequency
(IF) bandwidth, and k is Boltzmann’s constant.
Therefore, the signal-to-noise ratio (SNR) is given by
Pt G 1
SNR = 2 2
Ae . (5)
(4πR ) kTo BFn
The signal that arrives at a radar receiver from a transmitting radar is comprised
of the radar signal and various sources of noise, such as clutter, other interference
sources and jamming. In this thesis, the effects due to clutter and other interferences
are not considered.
3.3 Linear frequency modulation (LFM) modula-

tion
Conventional pulsed radars transmit narrowband bandpass signals. The equation for
an individual transmitted radar pulse is written as:
xa (t) = a(t) sin(2πfc t + θ(t)) (6)
where a(t) is the constant amplitude pulse envelope, fc is the radar carrier frequency,
and θ(t) represents either a constant phase or phase modulation of the pulse, The
received radar signal amplitude is directly related to the received power in the radar
intercept range equation, Pr .
Most radars use either linear frequency modulation (LFM) or Non-linear Fre-
quency Modulation (NLFM). This thesis will focus on the case of radars that transmit
LFM waveforms. In general, LFM signals can be written in the following form:

β
xLFM (t) = exp jφ0 + 2πj αt + t2 , (7)
2
where φ0 is the initial phase, α is the starting frequency, and β is the chirp rate, i.e.,
β ∈ {β1 , . . . , βm } (8)
where m is the number of possible chirp rates of the LFM signal. In the context of
transmitting radars, these chirp rates correspond to the radar operating modes. The
goal of this thesis is to determine the chirp rate for an arbitrary LFM signal where
the chirp rate is unknown, but is drawn from a set of known chirp rates. Without
loss of generality, we can set φ0 = 0 for the purpose of this thesis. The parameter α
is the carrier frequency of the transmitting radar.
Assuming that we have some a priori knowledge the emitter, say due to regulatory
requirements, we can set α = 0. In practice, at the intercept receiver side, this would
be equivalent to demodulating the signal to baseband.
At the intercept receiver side, we assume that all the signals involved have been
sampled respecting the Nyquist rate to yield discrete-time signals. Thus, the ideal
LFM signal xLFM (t) would be sampled at a constant sampling frequency fs , yielding
the discrete-time signal
" 2 #
β n
xβ [n] = exp 2πj · , n = 0, . . . , N − 1 (9)
2 fs
where N is the length of the sampled signal. In general, a received LFM signal may
be corrupted by noise, yielding
x[n] = xβ [n] + w[n]. (10)
For the purposes of this thesis, we assume that the noise, w[n], is a complex-valued
Gaussian white noise .It is further assumed that any incoming signal contains at most
one LFM signal.
The problem we will consider may be stated as follows: given a signal of the
form xβ [n] + w[n], determine β drawn from a predetermined set of m chirp rates
{β1 , β2 , . . . , βm }.
3.4 Wigner-Ville Distribution

The Wigner-Ville Distribution (WVD) is one of the key techniques in radar signal
processing as it represents a signal in both the time and frequency domains. It can be
used to determine the presence of a specific LFM signal with its modulation rate in
high SNR scenarios [3]. The WVD uses the autocorrelation function to calculate the
power spectrum of a signal and preserves the highest energy of a linearly modulated
signal in the time-frequency domain.
The WVD of a continuous input signal, x(t), is:

Z ∞
τ τ
Wx (t, ω) = x(t + )x∗ (t − )e−jωt dτ (11)
−∞ 2 2
where t is the time variable, ω = 2πf is the angular frequency, and * represents the
complex conjugate of the input signal, x(t).
Z ∞
τ τ
Wx (t, f ) = x(t + )x∗ (t − )e−j2πf t dτ (12)
−∞ 2 2
The WVD transformed to the Fourier domain is:

Z ∞
1 ω0 ω0
Wx (ω, t) = X(ω + )X ∗ (ω − )e−jω0 t dω0 (13)
2π −∞ 2 2
It can be noted from these two equations that:
Wx (t, ω) = Wx (ω, t) (14)
For a discrete signal, x(t) sampled at sampling frequency f s with N samples, the
equation is:
N
X m m
Wx (n, k) = x(n + )x∗ (n − )e−j2πkm/N (15)
m=−N
2 2
In this thesis, we will compare the performance of the proposed machine learning
technique with that of the WVD and show that WVD does not work well in low SNR
conditions.
3.5 Machine Learning algorithms and perfor-

mance metrics
Machine Learning attempts to obtain models from the data by learning the under-
lying relationships.Machine learning algorithms may also be viewed as the use of
statistical methods, and algorithms to make classifications or predictions, uncovering
key relationships that group certain data together. These relationships subsequently
lead to decision making.
There is a wide variety of machine learning algorithms ranging in flexibility and
complexity that can be used to classify radar chirp rates. Each algorithm has its
own assumptions and, based on the underlying assumptions and methodology used
to capture the hidden information, the complexity of the resulting classifier will differ.
In this thesis, four well known machine learning algorithms are considered to classify
the chirp rates. They are Decision Tree (DT), Naive Bayes (NB), Support Vector
Machine (SVM) and Random Forest (RnF). These four machine learning algorithms
or classifiers are described in the following subsections.
3.5.1 Support Vector Machine (SVM) Classifier

Support vector machines (SVM) are binary classifiers. In a binary classifier, one can
choose several hyperplanes to separate the classes. A SVM classifier would produce
the best hyperplane that maximizes distance between the data points in the two
classes. This line is referred to as a maximum-margin hyperplane and is used to
separate data points into two classes [1]. The further away from the hyperplane
that a data point is, the larger that margin between data point and the hyperplane,
the higher the probability that a correct prediction has been made in classification.
The points closest to the separating hyperplane are known as support vectors. The
classifier’s goal is to maximise the distance from the hyperplane to the support vectors.
This maximization leads to a good classification with good confidence even for
unseen data points. In other words, SVM’s strength is in its ability to classify data
points that are outside the classifier training set [57]. This is termed as generalization
in machine learning. A multi-class classification problem can be reduced to a set of
binary classification problems (using a one-vs.-one or a one-vs.-all strategy) as long
as there are no dependencies between any two clases.
An SVM’s learning phase consists of finding a set of parameters by solving a
Convex Constrained Quadratic Programming (CCQP) problem. A set of hyper-
parameters is tuned to find the SVM that provides the most optimal performance
in classifying the set of data. This is known as the model selection phase where the
chosen model is characterised by the lowest estimated generalisation error [58].
The equation for the hyper-plane is
g(x) = wT x + w0 , (16)
where w0 is the weight vector and w0 offset. The distance from a data point to the
Figure 1: Support Vector Machine Hyperplane [1].
hyperplane is defined in Equation (17) and is known as the margin
|wT x + w0 |
b= . (17)
||w||
In order to determine where the optimal placement of the hyperplane is for the
data set comprised of linearly separable values, the SVM classifier allocates data
points into separate feature domains split by the hyperplane where:
u = wT x + w0 < 0 if x ∈ C = −1, (18)
and
u = wT x + w0 > 0 if x ∈ C = 1. (19)
In the above equation C represents the class.

If a point is either quite far from the hyper plane on the positive or negative side,
then u will be a large positive number. Thus, for the binary problem, we can assign
points to either Class 1 or Class −1. The points with the smallest margins form the
hyper-plane and will be referred to as the support vectors. Once the points with the
smallest margin are found, SVM then maximises the margin so that u = 1 and we
1
have the maximum of ||w|| . For an N class problem, there will be N hyper-planes and
regions defined. The complexity of the SVM classifier is proportional to the number
of SVM and not the feature space itself.
The procedure implemented in Wolfram Mathematica is as follows:
• Transform the data to the format of the SVM package.
• Scale the data to a linear range of [-1, +1] or [0,1].
• Apply the relevant kernel to the data.
• Determine the best values for C and γ through cross-validation.
• Test the data.
• Run the trained classifier on the data sets.
3.5.2 Naive Bayes (NB) Classifier

Bayes theorem determines the conditional probability of an event A when another
event B has occurred, i.e.,
P (A)
P (A|B) = P (B|A) × , (20)
P (B)
where P (B|A) represents the probability of event B assuming that event A has oc-
curred, P (A) is the probability of event A, and P (B) is the probability of event
B.
Let X = {x1 , . . . , xm } be the set of of m possible features and y be the associated
class variable. Then, using Bayes’ theorem
P (y, X) P (x1 |y)P (x2 |y) · · · P (xn |y)P (y)

P (y|X) = = . (21)
P (X) P (x1 )P (x2 ) · · · P (xn )
Since the denominator is a known quantity, the class y with maximum probability is
given by
n
Y
ỹ = argmaxy P (y) P (xi |y). (22)
i=1
The probability distributions above could be Bernoulli, Multinomial or Gaussian. In

the binary classification case, the distribution is Bernoulli.
Wolfram Mathematica Naive Bayes variant implements Baye’s theorem where
P (class|x) ∝ P (class)P (x|class) and assumes that each feature, x, is independent
given a class. The class probabilities are P (class|x) ∝ P (class) ni=1 P (xi |class).
Q
P (xi |class) is the probability distribution of the feature xi given a class and P(class)
is the prior probability of the class. Both distributions are estimated from the training
data. Distributions are modeled using a piecewise-constant function. A regularisation
parameter, ”SmoothingParameter”, can be defined to smooth out the function [59].
3.5.3 Decision Tree (DT) Classifier

Decision Trees are a non-parametric supervised learning method used for classifica-
tion. The decision tree algorithm seeks to learn simple decision rules from training
data and then uses those rules to predict the class of the test data.
Let the training data be denoted by T ∈ RN ×D , where N is the number of samples
and D is the number of features. Also, let the label matrix be denoted by L ∈ RN ×K ,
where K is the number of labels/classes. A decision tree recursively partitions the
space of the training data such that the samples having the same class labels will be
grouped together.
Consider the data at node m that is represented by T . For each candidate split
θ = (j, tm ), where tm is the threshold at node m for feature j. The data is then
partitioned into Tl and Tr subsets:
Tl (θ) = (T, y)|Tj ≤ tm , (23)

Tr (θ) = X\Tl (θ)
A function called impurity function F (T ) is introduced. Two common choices for the
impurity functions are
K
X
F (Tm ) = 1 − p2i , (24)
i=1
K
X
F (Tm ) = − pi log2 (pi ), (25)
i=1
where
1 X
pi = I(yi = k) (26)
Nm t ∈R
i m
where I is the indicator function. Equation (24) is called the gini index while equation
(25) refers to entropy. The parameters θ = (j, tm ) are selected such that at node m
the impurity at the left and right nodes are minimized. This is accomplished via
|Tl | |Tr |
Q(T, j, tm ) = F (T ) − F (Tl ) − F (Tr ). (27)
|T | |T |
This is carried out recursively such that the maximum allowable depth of the tree is
reached and the number of samples in a node is either 1 or less than the minimum
number of samples split.
3.5.4 Random Forest (RnF) Classifier

Random forest (RnF) is an ensemble learning method for classification and regression
that operates by constructing a multitude of decision trees, otherwise known as an
ensemble [60]. The forest prediction is obtained by taking the most common class or
the mean-value tree predictions [61].
In order to have a better classification result (reduced variance) different tech-
niques such as bagging and boosting are employed. To reduce the variance, a random
forest bags number of decision trees in order to arrive at a classification result. Sub-
samples are drawn with replacement so that the sub-sample size is the same as the
original input sample size. In addition, sub-samples have a random subset of features.
Although, the bias of the forest may increase, averaging over the ensemble will reduce
the variance leading to a better model.
3.6 Machine Learning Metrics

This section provides a brief summary of the metrics that are commonly used to
evaluate machine learning algorithms.
A confusion matrix [62] is a M × M matrix used for evaluating the performance
of a classification model where M is the number of target classes.
The entries in the confusion matrix for a binary classifier are defined in terms of
the following quantities:
• The positive (P) and negative (N) labels represent the predicted outcomes,
• The true (T) and false (F) labels represent the actual outcomes,
• TP represents the number of times that an actual outcome is true but the
predicted outcome is true,
• FP represents the number of times that an actual outcome is false but the
predicted outcome is true,
• FN represents the number of times that an actual outcome is false but the
predicted outcome is false, and
• TN represents the number of times that an actual outcome is true but the
predicted outcome is false.
The confusion matrix for a binary classifier is defined as:

 
TP FN 
 . (28)
FP TN
It shows the actual values of True Positive (TP), True Negative (TN), False Positive
(FP) and False Negative (FN) rates obtained using the machine learning classifier.
FP errors are known as Type I errors while FN errors are defined as Type II errors
in statistical detection theory. For a M −ary classification problem, the confusion
matrix has M × M elements, where each element can be designated as nij where i
identifies the row and j identifies the column. Each element indicates the number of
times that data from class i that has been classified as belonging to class j.
The total number of cases, N , in the M class matrix is:
M X
X M
N= nij (29)
i=1 j=1
Confusion matrices can also present the percentage values instead of the number of
cases. In this thesis, the number of cases will be represented in the confusion matrices.
It should be noted that an asymmetric confusion matrix can reveal a biased classifier.
The performance of the confusion matrix is evaluated from the categorisation of
the results for the True Positive (TP) , True Negative (TN), False Positive (FP) and
False Negative (FN) rates. From [63], we can define the following performance metrics
to evaluate the performance of the M = 2 confusion matrix.
The misclassification rate is the proportion of negatives cases that were incorrectly
classified as: positive and is defined as
FP + FN
Misclass = . (30)
TN + FP + FN + TN
The specificity rate is defined as the proportion of negatives cases that were classified
correctly and is calculated as:
TN
Spec = . (31)
TN + FP
The accuracy (acc) is defined as follows:
TP + TN
acc = (32)
TP + FP + FN + TN
The error rate is another common parameter, ER, used to assess the performance of
a classifier and is ER = 1 − acc.
The accuracy determined using Equation (32) may not be an viable performance
measure when the number of negative cases is much greater than the number of
positive cases [64]. It should be noted that the accuracy (acc) and the error rate, ER,
depend upon the number of classes and on the total number of data, N .
The F1-score is a harmonic mean of Precision and Recall, and so it gives a com-
bined idea about these two metrics and is defined as follows [65]:
prec ∗ recall
F1 = 2 ∗ (33)
prec + recall
The F1-score is maximum when Precision is equal to Recall. Based on the data, there
will be performance metrics that will better reflect the classifier performance due to
the worst case scenario and the error type that needs to be avoided. These values
can then be inserted into Equation (33) to give us the F1 values for each class.
The above definitions can be extended for M −ary classification. An M −ary
classification problem can be considered as a series of M binary classification problems
so that the above definitions can be utilized.
In this thesis, the confusion matrix has been selected as the means to report
results in our M −class classification problem as it is possible to observe the relations
between the classifier outputs and the true outputs. The accuracy, recall, precision,
and F1 values will be calculated for the confusion matrices of each classifier to assess
their performance.
Chapter 4
Some Common Radars
4.1 Introduction
Three radar types namely, the weather, the air-surveillance and the marine radars are
considered in this work. These radars are typically found in harbours and airports.
Each of the radars can operate in several modes. In this chapter, we discuss the three
radar types and their realistic values for the radar parameters in the different modes
of operation.
4.2 Weather radars

Ground-based weather radars are widely distributed and come in a variety of forms
to support differing observational requirements. Deployed weather radars fall into
two broad categories: (i) high power, fixed, long-range radars in the S band and
C-band and (ii) medium to low-power short-range radars in the C-band and X-band
(primarily X band). The high power long-range radars are tower-mounted large sys-
tems that radiate between 250 kW and 1.1 MW short (0.8 to 2.0 µseconds) pulses
at horizontal and/or vertical polarization at pulse rates between 300 Hz and 500 Hz.
For weather radar applications, cross-polarized radar returns are generally insignif-
icant and horizontal and vertical linear polarizations are sometimes simultaneously
transmitted. Some systems use circularly polarized transmissions. The medium to
low-power short-range radars are often transportable systems that radiate between
100 W and 25 kW short pulses (0.2 to 1.2 µseconds) at both polarizations that radiate
at pulse rates between 1 kHz and 2.5 kHz.
Traditionally, weather radars use circular parabolic dish antennas. Long-range
29
CHAPTER 4. SOME COMMON RADARS 30
radars are mechanically scanned in azimuth between 3 and 6 RPM and are stepped in
elevation over a period of several (approximately 5) minutes. The antenna elevation
angle is fixed or varies slowly for each azimuth scan. A demanding elevation [66]
scanning strategy for a WSR-88D radar is shown in Figure 2. The WSR-88D uses
pulse repetition rates between 320 Hz and 1300 Hz and, for long-range (small elevation
angles) alternates between 4.57 µseconds pulses at 320 Hz and 1.57 µsecond pulses
at 1280 Hz on alternate azimuth scans.
Figure 2: WSR 88D Weather Radar scanning pattern
Short to medium range weather radars are typically portable systems that supply
local weather data within 80 km (often less than 60 km) range of the radar site. Tra-
ditional short-range weather radars also use parabolic dish antennas with mechanical
azimuth and elevation scans. Typical azimuth scan rates vary from 3 RPM to 6 RPM
and many of the operational modes and data acquisition strategies used for larger
systems are also used for the smaller systems. Because of the shorter range require-
ment, antenna beam widths up to 3◦ are useful. Some of the most compact systems
are specialized for rain observations. Many of the short range radars in common use
magnetron transmitters and radiate peak powers between 10 kW and 50 kW. As with
the large long-range radars, transmitted signal and beam scanning requirements are
driven by meteorological observation requirements [66], [67].
More modern portable weather radars use pulse compression technology to re-
duce the required peak radiated power and use solid state transmitters [68]. Typical
transmitted powers are 100 to 500 W but one at 3.5 kW peak power operating in
C-band is also available [69]. Some recent systems use single polarization flat plate
array antennas with mechanical steering, other systems come with single array, and
with mechanical azimuth scans and electronic elevation beam steering. Few weather
radars also use multiple electronically steered phased array antennas panels [70].
In this thesis, we use the NEXRAD WSR 88 S-band long range system [71] as
the example for our weather radar. The system uses a rotating, circular, center-
fed parabolic dish antenna, operates at frequencies in the vicinity of 3.0 GHz and
alternately transmits long and short CW pulses. Typical operating parameters [71]
are given in Table 2. Additional information can be found in [66].
4.3 Marine radars

Marine navigation radaras operate in the S, C and X bands. The International
Telecommunications Union (ITU) recognizes three navigation radar classes: the In-
ternational Maritime Organization (IMO) and fishing boat class, the river class and
the pleasure-craft class. Their frequency ranges are detailed in Table 3.
Following the discussion in [1], the radar properties for each marine radar class
are given in Tables 4, 5 ,and 6. These tables provide the typical features of each radar
class.
All of the radars shown in Tables 4, 5 ,and 6 use uncoded incoherent (magnetron)
pulse transmissions. More recent radars new marine radar systems [72] show a transi-
tion to coherent, internally coded pulse transmissions combined with a reduced-power
range-compression technology.
Under IMO rules, all vessels greater than 200 tons and all passenger vessels require
an X-band navigation radar while vessels greater than 3000 tons require an additional
S-band radar (or, with permission, a second independent X-band radar). Marine
navigation radar antennas are mounted on or above a vessel’s bridge roof (and often
on a ship’s mast) to maximize usable radar range and to minimize human exposure to
microwave radiation. For small pleasure craft, the antennas are typically higher than
2.5 m above sea level (ASL). For large vessels heights above 10 m ASL are common.
Common marine navigation radar antennas radiate signals into a fan beam that
Radar Parameter Feature Value

Range resolution (m) 235
Detection Z10 (dBz) with RCS at 460 km -23 to -25 dB
Angular Coverage
Azimuth Full circle or sector
Elevation Operational limits: −1◦ to 20◦
Antenna
Type S-band, center-fed, parabolic dish
Reflector Aperture 8.54 m diameter; circular
Beam width (one-way, 3dB) 0.96◦ at 2.7 GHz
0.88◦ at 3.0 GHz
Gain 45.8 dB at 2.85 GHz (midland)
Polarisation Linear horizontal
First side-lobe level -29 dB
Steer ability 360◦ azimuth;−1◦ to 45◦ elevation
Rotation Rate 30◦ /s (azimuth and elevation)
Mechanical Limits −1◦ to 60◦
Angular Acceleration 15◦ s−2 (azimuth and elevation)
Pointing Accuracy ±0.2◦
Transmitter
Frequency Range 2.7–3.0 GHz
Peak Power output 500 kW into antenna
Pulse width 1.57 microseconds (short pulse)
4.5 microseconds (long pulse)
PRF
Long Pulse 322–422 Hz ±1.7%
Short Pulse 322–1282 Hz ±1.7%
Table 2: Nexrad WSR Weather Radar Parameters

Band Lower frequency (MHz) Upper frequency (MHz)

S-band 2700 2900
C-band 5250 5725
X-band 9300 9500
Table 3: Marine radar frequency band allocations [1].
Property 2.9 GHz– 5.47 GHz- 8.85 GHz– 9.2 GHz–

3 GHz 5.65 GHz 9 GHz 9.5 GHz
Min. Max. Min. Max. Min. Max.
Antenna
3-dB Beamwidth (deg)
Horizontal 4.0 1.0 2.6 1.0 2.3 0.75
Vertical 30.0 24.0 25.0 18.0 26.0 20.0
Sidelobe attenuation (dB)
Within±10◦ 28 23 29 23 31 23
Outside±10◦ 32 31 35 31 40 30
Gain (dB 28 26 31 28 32 27
Rotation rate (rpm) 60 20 60 14 60 20
Transmitter
Peak power (kW) 75 30 70 50 50 5
Frequency (GHz) 3.08 3.02 5.595 5.485 9.445±0.03 9.375±0.03
Pulse length (micros) 1.2 0.05 1.5 0.07 1.2 0.03
Receiver
(IF) (MHz) 60 45 60 45 60 45
IF bandwidth (MHz)
Short pulse (µs) 28 6 28 6 28 6
Medium/Long pulse(µs) 6 2.5 6 2.5 6 2.5
Noise figure (dB) 8.5 3 8.5 3 8.5 3.5
Table 4: Typical characteristics of IMO category radio navigation radars.


Antenna (Transmitter and Receiver)
Horizontal Beam width 0.95◦
Vertical Beam width 26◦
Side lobe attenuation >25 dB within ±10◦ of mainlobe
>32 dB outside ±10◦ of mainlobe
Gain (dB) 30
Antenna rotation speed 30 rpm
Transmitter
Peak Output Power 5 kW
Frequency -9410 MHz ± 30 MHz
Pulse length (microseconds) 0.05, 0.18, 0.5
Pulse repetition rate(PRR)(Hz) 1000 to 3000
Receiver
Intermediate frequency (IF) 50 MHz
IF bandwidth 15-25 MHz
Noise figure 6 dB
Table 5: Typical characteristics of river class radio navigation radars.


Antenna (Tx and Rx)
Horizontal Beam width 0.95◦
Side lobe attenuation >25 dB within ± 10◦ of mainlobe
>32 dB outside ± 10◦ of mainlobe
Gain (dB) 30
Transmitter
Peak Output Power 5 kW
Pulse length (microseconds) 0.05, 0.18, 0.5
Pulse repetition rate(PRR)(Hz) 1000 to 3000
Receiver
IF bandwidth 15-25 MHz
Noise figure 6 dB
Table 6: Typical characteristics of pleasure-craft class radio navigation radars.

is large in elevation (typically 18◦ to 26◦ ) and narrow in azimuth (typically 0.75◦ to
4◦ ). Azimuth beam widths in the vicinity of 1◦ are common. Array-fed horn antennas
are the most common antenna design and dual polarized (H and V linear polariza-
tions) are becoming common for marine hazard detection, sea surface monitoring and
meteorological observations.
In this thesis, we use the Furuno FR7111 X-band marine radar [73] as a baseline
to construct a marine radar interference source model. The baseline system uses
a rotating, two-dimensional slotted waveguide array, operates at frequencies in the
vicinity of 9.4 GHz and alternately transmits short and long pulses. Typical operating
parameters [73] are displayed in Table 7. Sample marine radar chirp rates are noted
in Table 8.
4.4 Airborne radars

Aircraft search radars are a primary air traffic control tools and can be grouped in
three different classes: air route surveillance radars (ARSRs), airport surveillance
radars (ASRs) and precision approach radars (PARs). The major properties of three
ARSR and ASR systems in current use are shown in Table 9 [74].
ARSRs are long-range, three-dimensional surveillance radars designed to detect
and track transiting aircraft within 400 km of the radar site. The most recent ARSR
systems deployed around the continental US is the ARSR-4 pulse-compression radar
class. ARSR-4 is an L-band radar operating in the 1.2 to 1.4 GHz frequency range.
Radio spectrum reallocation proposals to free more bandwidth for commercial com-
munications would reduce the upper bound of the frequency range to 1390 GHz. Since
the ARSR-4 system uses frequency hopping within its operation band, 4 of the 22
frequency band pairs would be lost.
ASRs are intermediate range radars found at all major airports. These are S-
band instruments operating between 2.7 GHz and 2.9 GHz with peak radiated power
near 1.1 MW for the older, short-pulse ASR-9 class radars still in wide use in North
America and 25 kW for the current pulse-compression ASR-11 class radars. ASRs
are designed to detect and track aircraft with radar cross section greater than 1 metre
squared at altitudes up to 19 km and high-altitude radar ranges up to 105 km.

Radiator Slotted waveguide array
Polarisation Linear horizontal
Radiator length (XN13A) 180 cm
Horizontal Beam width (XN13A) 1.2◦
Side lobe attenuation(XN13A) Within ±10◦ of main-lobe: < −24dB
Outside ±10◦ of main-lobe: < −30 dB
Transceiver module
Modulation Continuous wave
Peak Output Power 10kW
Pulse length (τ )and At 0.125 nm–0.75 nm, τ =0.12µs,
Pulse Repetition rate(PRR) PRR = 2100 Hz
At 1–2 nm, τ = 0.3 µs
PRR = 1200 Hz
At 3–24 nm, τ = 0.8 micros, PRR = 600 Hz
IF bandwidth 25 MHz (Tx τ 0.08 µs and 0.3 µs)
4 MHz (0.8 µs)
Table 7: Furono FR7111 Marine Radar Parameters

Centre Pulse PRF (MHz) BW (MHz)

frequency width (µs)
(MHZ)
Radar 1
9400 0.12 2100 25
Radar 2
9400 0.3 1200 25
Radar 3
9400 0.8 600 25
Table 8: Sample marine radar chirp rates.
Function Sensitivity Range/Alt Az/El Waveform Update

rate
Terminal 1 m2 60 nm/20,000 1.4◦ /5◦ >18 5s
aircraft ft pulses;PRI
surveil- 1ms
lance
En Route 2.2 m2 250 1.4◦ /2◦ >10 12s
aircraft nm/60,000 ft pulses;PRI
surveil- 3ms
lance
Terminal -20 to 5 600 1◦ /1◦ >50 60s
area dBz nm/15,000 ft pulses;PRI
weather 1ms
National -20 to 5 250 1◦ /1◦ >50 300s
scale dBz nm/60,000 ft pulses;PRI
weather 1ms
Table 9: Air radar pulse transmission schemes by function.

Recent designs of ARSR and ASR civilian and military radar use compressed
pulse excitation to increase peak pulse power by increasing the radiated pulse length.
Recent military air surveillance radars use solid-state active array technology with
two-dimensional electronic scan (often combined with a mechanical azimuthal scan) to
maximize radar flexibility and maximize mean time between failures. Many systems
use non-linear FM coding [75] and [76] in the expanded pulse to minimize range
compression side lobes.
Many ASR and ARSR systems use a frequency diversity strategy of transmitted
pulse generation in which radar pulses are transmitted simultaneously at two frequen-
cies (for the ASR4 these are 83 MHz apart and are stepped to create several transmit
pulse frequency pairs) to minimize the effects of scattering nulls on radar target de-
tection. The transmitted frequency plan is programmable and proprietary for each
radar type. To maximize radar blind speed for target motion measurements, pulse
transmissions are sequentially generated with different PRIs selected by proprietary
protocols to yield a maximum-range mean PRF.
Precision approach radars (PARs) operate in the 9.1 to 9.2 GHz Band. PARs
are designed to provide aircraft glide slope data out to a maximum range of 40km.
Depending on the radar design, some PARs use two fan beam antennas, one for el-
evation (narrow beam width 0.75◦ ) and one for azimuth (narrow beam width 1.45◦ )
and others use electronically scanned pencil beams in one or both dimensions. PARs
provide an azimuth sector scan between 20◦ and 30◦ and an elevation sector scan
between 8◦ and 15◦ good For applications where multiple glide-path azimuths are
needed, the antenna beams are mechanically pointed in azimuth to select the glide-
path to be monitored. Real-pulse radars transmit peak power between 180 kW and
300 kW with pulse bandwidths between 4 MHz and 0.5 MHz. Compressed-pulse
PARs can achieve effective performance for peak pulse power less than 20 kW. Radar
pulse repetition rates very between 3400 Hz and 3800 Hz. Most PARs use two simul-
taneously transmitted frequencies and frequency hopping to minimize the effects of
target response nulls.
Air surveillance radars operate at S band frequencies as shown in Table 6. The
long-range systems used to support airport surveillance operate in the L-band between
1.2 and 1.4 GHz, while the airport surveillance S-band radars typically operate from
2.7 to 2.9 GHZ to support local sites.
In this thesis, we will use the ASR-9 airport surveillance S-band system as a
Parameter Value
Radiator Parabolic dish antenna
offset-feed with a main Tx/Rx feed horn
passive Rx only feed horn
Polarisation Selectable circular or linear
Antenna rotation speed 12.5 rpm
Azimuth Beam width 1.3◦
Elevation Beam width 4.8◦
Gain 33.5 dB (Main beam LOW) to 32.5 dB (Main beam HIGH)
Side lobe attenuation 24 dB (Azimuth); 6 dB (Elevation)
Frequency 2700–2900 Hz
Peak Output Power 1.32 MW
PRF 1200 Hz
Pulse width 1.05 microseconds
Table 10: ASR-9 airborne radar parameters.
baseline airport surveillance radar interference source. The baseline system uses a
rotating, circular, offset-fed parabolic dish antenna, operates at frequencies in the
vicinity of 2.7 to 2.9 GHz and alternately transmits long and short CW pulses. Typical
operating parameters are displayed in Table 10.
Typical airborne radar chirp rates are defined in Table 11. The magnitude of the
received signal is determined via the radar range equation.
4.5 Summary
This chapter provided an overview of the various radars considered in this thesis. The
three radars, namely the weather, the marine and the airborne radars can operate
in several modes. The realistic chirp rates for weather, marine, and airborne radars
have been defined based on the parameters contained in Sections 4.2, 4.3, and 4.4.
Table 12 provides a summary. These typical chirp rates are considered as the known
Parameter Value
Radar 1
Centre Frequency (MHz) 9750
Transmit bandwidth (MHz) 50
Transmit pulse length (microseconds) 50
Transmit PRF (Hz) 2000
Rx Complex Sampling Rate (MHz) 50
Radar 2
Centre Frequency (MHz) 9750
Transmit bandwidth (MHz) 1500
Transmit pulse length (microseconds) 50
Transmit PRF (Hz) 1000
Rx Complex Sampling Rate (MHz) 200
Table 11: Sample airborne radar chirp rates.
set of chirp rates in this study. As seen from Table 12, some of the radars have the
same centre frequency but may differ in their bandwidths. In a low SNR condition, it
may be rather impossible to estimate the bandwidth. Resorting to machine learning
techniques to first determine the type of radar and possibly its mode, appears as a
logical approach. The following chapter will provide further insight into the simula-
tions done to explore the idea of using machine learning with matched filter outputs
to classify these radars.
Radar Type Center fre- Sampling fre- Bandwidth Pulse width (µs) Chirp rate (k)
quency (MHz) quency (MHz) (B)MHz MHz/(µs)
Airborne Mode 9750 50 50 50 1
1
Airborne Mode 9750 200 1500 50 30
2
Airborne Mode 9750 50 600 42 14
3
Airborne Mode 9750 50 2.05 21 0.10
4
Weather Mode 1 9500 50 600 1.57 382
Weather Mode 2 9500 50 600 4.57 131
CHAPTER 4. SOME COMMON RADARS
Weather Mode 3 9500 50 600 0.6 1000

Weather Mode 4 9750 50 600 1.6 375
Weather Mode 5 9750 50 600 2.0 300
Marine Mode 1 9410 50 20 0.05 400
Marine Mode 2 9410 50 20 0.18 111
Marine Mode 3 9410 50 20 0.5 40
Marine Mode 4 9410 50 25 0.05 500
Marine Mode 5 9410 50 25 0.18 139
Marine Mode 6 9410 50 25 0.5 50
Table 12: Summary of realistic chirp rates for airborne, weather and marine radars.
42
Chapter 5
Results
5.1 Introduction
In this chapter, the inadequacies of the Wigner-Ville transform and the peak detection
using a bank of matched filters are presented as the motivation for this thesis work.
A bank of M matched filters whose individual filter impulse response correspond to
clean ideal LFM radar waveform for the chirp rate contained in the known set is
constructed and used in this work. Different simulation scenarios are discussed and
the results are analyzed.
5.2 Inadequacy of the Wigner-Ville Transform

The Wigner-Ville transform as described in Chapter 3 was applied on a LFM pulse
with β = 1. Figure 3 shows the Wigner-Ville Transform of a LFM signal in the
absence of noise. The Wigner-Ville transform for the SNR of 0 dB is shown in Figure
4. The line is still somewhat visible. When the SNR is-10 dB and below, the line is
no longer visible as seen in Figures 5 and 6. This motivates consideration of other
approaches to detecting LFM signals particularly in low-SNR conditions.
5.3 Proposed approach

Normally as a first stage in the detection of signals in a radar receiver, matched
filtering is carried out. However, in the radar intercept receiver, we do not have the
43
CHAPTER 5. RESULTS 44
Figure 3: Wigner-Ville transform of an LFM signal.
500
450
600
400
350 400
300
200
250
200
0
150
100 -200
50
0 -400
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Figure 4: Wigner-Ville transform of a noisy LFM signal at a SNR of 0 dB.

500 2500
450 2000
400 1500
350 1000
300 500
250 0
-500
200
-1000
150
-1500
100
-2000
50
-2500
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Figure 5: Wigner-Ville transform of an LFM signal at a SNR of −10 dB.

10 4
500
2.5
450
2
400
1.5
Frequency (arbitrary units)
350
1
300 0.5
250 0
200 -0.5
-1
150
-1.5
100
-2
50
-2.5
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Time (arbitrary units)
Figure 6: Wigner-Ville transform of a noisy LFM signal at a SNR of −20 dB.

information about the transmitted waveform (chirp rate in the case of LFM signals).
Therefore, the traditional radar approach will not be successful in determining the
chirp rate.
The approach we will explore for determining β is to perform matched filtering
of an incoming signal with an unknown chirp rate, which is denoted simply as x[n],
against all the ideal signals xβ1 [n], . . . , xβm [n]. The outputs of these matched filters
are concatenated to form the input which is then fed into a machine learning classifier.
The classifier outputs a prediction of the chirp rate of the signal.
In the case of additive white noise as assumed above, the matched filter hβ [n]
corresponding to a LFM signal xβ [n] is:
x∗β [−n]
hβ [n] = PN −1 (34)
k=0 xβ [k]x∗β [k]
where the superscript ∗ indicates complex conjugation and the denominator is simply
a normalization factor. The output of the matched filter is simply the convolution
x[n] ∗ hβ [n]. This amounts, in fact, to calculating the cross-correlation between the
signals x[n] and xβ [n], so we expect a large output when the incoming signal is
similar to the ideal LFM signal xβ [n] and a small output when the two signals are
very different. In the following, we will assume that the input x[n] is normalized just
like the filter hβ [n] is. In effect, we calculate the normalized cross-correlation instead
of the cross-covariance.
Because we know the set that β is drawn from, we can perform matched filtering of
the incoming signal against all of the signals xβ1 [n], . . . , xβm [n]. Heuristically speaking,
we would expect that the matched filter output with the highest peak will tell us
which of the chirp rates best describes the incoming signal. However, at low-SNR,
this simple peak detection approach may not be robust. In order to develop a more
robust algorithm, we propose a machine learning approach to this problem using the
concatenated matched filter outputs.
The machine learning portion of our procedure involves concatenating all of the
matched filter outputs x[n] ∗ hβ1 [n], . . . , x[n] ∗ hβm [n] into one long vector. This vector
is then considered as a feature vector to be fed into a classifier. An example of such
a feature vector, assuming no noise, is shown in Fig. 7. By training a classifier on
a large number of matched filter outputs for various chirp rates and signal-to-noise
0.100
Matched filter output
0.010
0.001
10-4
0 20 000 40 000 60 000 80 000 100 000 120 000
Feature vector index
Figure 7: Example of a feature vector obtained by concatenating all the matched

filter outputs of an ideal chirp.
ratios (SNRs), we expect to obtain a classifier which will accurately predict the chirp
rate of the incoming signal.
Four machine learning algorithms, namely the Decision Tree, the Naive Bayes,
the Random Forest, and the Support Vector Machine (SVM) are used to classify the
chirp parameter.
One might ask why not simply use the peak of the match filter output to determine
the chirp rate. Figure 8 shows the example of a case where no peak is evident. We
shall see later that the signal is correctly classified correctly using the proposed ML
approach.
5.4 Classifiers Evaluated in This Thesis

There are seven distinct classification problems for which ML classifiers were trained:
• Problem 1: This may be used as demonstration of the proof-of-concept through

a toy problem. The intent here is to classify LFM signals whose center frequen-
cies are all 0 Hz and whose chirp rates are taken from the set β = 0, 1, 2, 3, 4,
Figure 8: Concatenated match filter approach applied to LFM signal at -10 dB. No
clear peak is evident making it difficult to accurately classify the chirp rate.
and 5 Hz/s (6-class problem).
• Problem 2: This problem considers realistic chirp rates of only airborne radars:
The intent here is to classify LFM signals with realistic parameters correspond-
ing to airborne radars as listed in Table 12 (4 modes + no chirp = 5-class
problem).
• Problem 3: This problem considers realistic chirp rates of only weather radars.
The intent here is to classify LFM signals with parameters corresponding to
weather radars as listed in Table 12 (5 modes + no chirp = 6-class problem).
• Problem 4: This problem considers realistic chirp rates of only marine radars.
The intent here is to classify LFM signals with parameters corresponding to
marine radars as listed in Table 12 (6 modes + no chirp = 7-class problem).
• Problem 5: The intent here is to consider realistic chirp rates of all radar types
considered in this research. The intent here is to classify LFM signals with
parameters corresponding to all the radars listed in Table 12 (15 modes + no
chirp = 16-class problem).
• Problem 6: The intent here is to study the robustness of the classifiers against
windowing imperfections. This can occur if the intercept receiver does not turn
on immediately when the radar transmitter’s pulse is received. As a conse-
quence, some part of the pulse is lost. The classifiers trained in Problems 1 and
5 are used for this robustness study.
• Problem 7: The intent here is to study the robustness of the classifier against
perturbation in chirp parameters. The classifier trained for Problem 5 is used
in this study. In practice, this situation arises due to hardware imperfections.
5.4.1 Details of Simulations

For each of Problems 1–5 listed above, four machine learning algorithms were used:
decision tree (DT), naive Bayes (NB), random forest (RF), and support vector ma-
chine (SVM). There are two exceptions: for Problem 1, as it was a proof-of-concept
study, we tried only NB and SVM classifiers. All classifiers were trained using Wolfram
Mathematica’s Classify command, which is a high-level command that automati-
cally optimizes each classifier without human input (e.g., by choosing hyperparame-
ters and selecting appropriate kernels). No classifiers were trained for problems 6 and
7; they used the ones trained for Problem 5.
For Problem 1, LFM signals were simulated with a sampling frequency of fs =
2000 Hz; the number of samples in each sampled signal was N = 10 000. Each signal
was contaminated with additive white Gaussian noise, resulting in simulated signals
with SNRs of 20, 10, 0, −10, −20, −30, and −40 dB. The training set consisted of 200
noisy LFM signals for each β and each SNR, for a total of 8400 signals. The trained
classifiers were then tested on a test set containing 50 LFM signals for each β and
SNR (for a total of 2100 signals), and confusion matrices were generated separately
for each SNR.
For Problems 2 to Problem 5, LFM signals were simulated with a sampling fre-
quency of fs = 1200 MHz, though the number of samples in each sampled signal
remained at N = 10 000. Each signal was contaminated with additive white Gaus-
sian noise, resulting in simulated signals with SNRs of 10, 0, −5, −10, −15, and
−20 dB. The training set consisted of 50 noisy LFM signals for each set of param-
eters and each SNR. Thus, for Problem 2, the training set contained 1500 samples,
while the training sets for Problem 3 to Problem 5 contained 1800, 2100, and 4000
samples respectively. The test set contained 10 LFM signals for each β and SNR,
so the size of the test sets for Problems 2–5 were 250, 300, 350, and 800 samples
respectively. In all cases, confusion matrices were generated separately for each SNR.
In Problem 6, the classifiers for Problem 1 and Problem 5 were subjected to a
further test to understand the effect of imperfect windowing. In this test, the last
20% of each chirp were dropped, and a corresponding number of zeros were added
to the beginning of each chirp in order to pad them back to the original length. An
example of this procedure is shown in 36. This was performed for each of the training
sets described above for Problems 1 and 5. Note that the noise was added after the
chirp was “shifted”, so the first 20% of each noisy chirp corresponds to pure noise.
Finally, in Problem 7 the classifiers for Problem 5 (which were trained on all
the radars in Table 12) were subjected to yet another test in order to characterize
the robustness of the approach considered in this thesis. This time, the classifiers
were presented with simulated LFM signals whose parameters were not drawn from
Table 12. A list of five such “unknown” radars is given in Table 13. For each entry in
this list, the type of radar is given, as well as the radar in Table 12 whose parameters
come closest to each unknown radar. The classifiers cannot possibly classify these
radars correctly, as the radar parameters do not fall under any of the classes which
the classifiers are designed to classify. It is hoped, however, that the classifier returns
the radar in Table 12 which comes closest to a given unknown radar, or at least returns
a radar whose type is the same as the unknown radar’s type. If this is successful, then
the same classifiers used to solve Problem 5 could also be used to classify the type
(but not the mode) of radars whose LFM parameters are unknown to the classifiers.
For this test, the same sampling frequency, signal length, and SNRs that were listed
above for Problems 2–5 were used, with 10 samples for each unknown radar and SNR
(300 samples total).
5.5 Results of Simulations

5.5.1 Problem 1: LFM signals with pre-selected chirp rates
Naive Bayes and SVM classifiers classified the test set perfectly for all signals with
an SNR of −10 dB or above. Therefore, classification above −10 dB is not included
in the thesis. Performance begins to degrade at −20 dB, though at this level of
Label Radar Type Center freq (MHz) β MHz/(µs) Similar to Radar

Radar A Marine 9410 42 Marine Mode 3
Radar B Weather 9500 373 Weather Mode 4
Radar C Air 9750 15 Air Mode 3
Radar D Marine 9410 128 Marine Mode 5
Radar E Air 9750 31 Air Mode 2
Table 13: Radar types and signal parameters for a set of radars that are unknown
to the bank of matched filters.
noise the performance is still quite good, as can be seen in Figs. 9 and 10. How-
ever, performance at −30 dB and below was poor. In summary, good classification
performance was achievable down to an SNR of −20 dB even with the simple, un-
sophisticated approach taken here. This is significantly better than what is possible
with the Wigner-Ville transform.
5.5.2 Problem 2: Airborne radar realistic chirp results

The confusion matrix for a decision tree classifier of realistic airborne radar chirp
rates is shown in Figure 11. DT makes classification errors at SNR =-15 dB with
misclassification of mode 2 as mode 4. At -20 dB, the only modes that have no issues
are no radar and mode 4. The chirp rate that is the hardest to detect at this SNR
level is mode 3.
The confusion matrix result for a naive Bayes classifier of realistic chirp rates for
airborne radars are shown in Figure 12 for the following SNR values 10 dB, 0 dB,
-5 dB, -10 dB, -15 dB and -20 dB. From the confusion matrix, it is clear that this
classifier has performance degradation with decreasing SNR. At -20 dB, this classifier
fails.
The confusion matrix result for a Random Forest classifier of realistic chirp rates
for airborne radars are shown in Figure 13 for the following SNR values 10 dB, 0 dB,
-5 dB, -10 dB, -15 dB and -20 dB. RnF performs well till 0 dB and the performance
deteriorates very rapidly as the SNR worsens.
The confusion matrix result for a Support Vector Machine (SVM) classifier of
5
0 50 0 0 0 0 0 50
1 0 47 3 0 0 0 50
2 0 0 43 7 0 0 50
actual class
3 0 0 1 49 0 0 50
4 0 0 0 0 50 0 50
5 0 0 0 0 5 45 50
50
47
47
56
55
45
predicted class
Figure 9: Confusion matrix for a NB classifier at −20 dB SNR (Problem 1).

0
0 50 0 0 0 0 0 50
1 0 50 0 0 0 0 50
2 2 0 47 0 0 1 50
actual class
3 1 0 2 45 1 1 50
4 0 0 0 3 42 5 50
5 0 0 1 0 1 48 50
53
50
50
48
44
55
predicted class
Figure 10: Confusion matrix for a SVM classifier for −20 dB SNR (Problem 1).
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
Air Mode 1 10 0 0 0 0 10 Air Mode 1 10 0 0 0 0 10 Air Mode 1 10 0 0 0 0 10

actual class
actual class
actual class
 , , ,
No radar 0 0 0 0 10 10 No radar 0 0 0 0 10 10 No radar 0 0 0 0 10 10

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
predicted class predicted class predicted class
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
10
10
10
10
10
11
10
11
15
10
9
Figure 11: Confusion matrix for a DT classifier of realistic airborne chirp rates at
SNRs from 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
left), -15 dB (bottom mid) and −20 dB (bottom right). (Problem 2)
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, , ,

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
12
10
10
12
10
10
24
16
10
8
0
Figure 12: Confusion matrix for a NB classifier of realistic airborne radar chirp rates
at SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
realistic chirp rates for airborne radars are shown in Figure 14 for the following SNR
values 10 dB, 0 dB, -5 dB, -10 dB, -15 dB and -20 dB. The confusion matrix shows
that the SVM classifier can accurately classify all airborne modes till SNR of -10 dB.
Below -10 dB, the degradation is not as steep as RF. Where as no classifier performs
well at -20 dB.
DT has the best classification rate even at -20 dB as shown in Figure. 15. DT has
the best performance with respect to F1 score seen in Figure 16. The performance
of SVM is better than NB. The classification performance of NB is worse than DT.
The classification performance of RnF is the worst of all the classifiers considered for
this type of radar. This also follows for F1 score.
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
 , , ,

10
10
10
10
10
11
10
10
10
14
11
10
9
9
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
10
12
10
12
11
12
12
12
12
8
Figure 13: Confusion matrix for a RnF classifier of realistic airborne radar chirp
rates at SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
 , , ,

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
10
10
10
10
12
10
32
10
9
Figure 14: Confusion matrix for a SVM classifier of realistic airborne chirp rates at
SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom left),
-15 dB (bottom mid) and −20 dB (bottom right). (Problem 2)
Figure 15: Accuracy versus SNR level - Airborne radar chirp rates. (Problem 2)
Figure 16: F1 score versus SNR level - Airborne radar chirp rates. (Problem 2)
Weather Mode 2
Weather Mode 1
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar
No radar 10 0 0 0 0 0 10 No radar 10 0 0 0 0 0 10 No radar 10 0 0 0 0 0 10
Weather Mode 1 0 10 0 0 0 0 10 Weather Mode 1 0 10 0 0 0 0 10 Weather Mode 1 0 10 0 0 0 0 10

 , , ,
actual class
actual class
actual class

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar

, ,
actual class
actual class
actual class

10
10
11
11
10
10
10
10
11
10
10
13
14
8
9
Figure 17: Confusion matrix for a DT classifier of realistic weather radar chirp rates
at SNRs 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
5.5.3 Problem 3: Weather radar realistic chirp results

The confusion matrix for a DT classifier of realistic weather radar chirp rates in shown
in Figure 17 for the following SNR values 10 dB, 0 dB, -5 dB, -10 dB, -15 dB and -20
dB. DT’s classification accuracy started dropping from -5 dB.
The confusion matrix result for a NB classifier of realistic chirp rates for weather
radars are shown in Figure 18 for the following SNR values 10 dB, 0 dB, -5 dB, -10
dB, -15 dB and -20 dB. NB classification accuracy also started dropping from -5 dB
SNR.
The confusion matrix result for a RnF classifier of realistic chirp rates for weather
dB, -15 dB and -20 dB. RnF classification accuracy also started dropping from -10
dB SNR.
Weather Mode 2
Weather Mode 1
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar

 , , ,
actual class
actual class
actual class

10
10
10
10
10
10
10
10
10
10
10
10
10
10
12
10
9
9
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar

, , 
actual class
actual class
actual class

10
13
13
11
10
10
10
17
10
10
27
3
4
Figure 18: Confusion matrix for a NB classifier of realistic weather radar chirp rates
Weather Mode 2
Weather Mode 1
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar

 , , ,
actual class
actual class
actual class

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
No radar
No radar
No radar

, ,
actual class
actual class
actual class

10
10
11
10
10
10
12
16
10
12
10
15
9
7
Figure 19: Confusion matrix for a RnF classifier of realistic weather chirp rates at
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
 , , ,

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
10
10
10
10
12
10
32
10
9
Figure 20: Confusion matrix for a SVM classifier of realistic airborne radar chirp
rates at SNR 10 dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom
realistic chirp rates for weather radars are shown in Figure 20 for the following SNR
values 10 dB, 0 dB, -5 dB, -10 dB, -15 dB and -20 dB. SVM classification accuracy
also started dropping from -10 dB SNR.
DT and SVM have the best classification rate for weather radars as shown in
Figure.21. RnF is better than NB over all the SNRs. The F1 scores follow the same
trend as shown in 22.
Figure 21: Accuracy versus SNR level - Weather radar chirp rates. (Problem 3)
5.5.4 Problem 4: Marine radar realistic chirp results

The confusion matrix for a DT classifier of realistic marine radar chirp rates in shown
The confusion matrix result for a NB classifier of realistic chirp rates for marine
dB, -15 dB and -20 dB. NB classification accuracy also started dropping from -5 dB
SNR.
Figure 22: F1 score versus SNR level - Weather radar chirp rates. (Problem 3)
The confusion matrix result for a RnF classifier of realistic chirp rates for marine
dB, -15 dB and -20 dB. RnF classification accuracy also started dropping from -10
dB SNR.
realistic chirp rates for marine radars are shown in Figure 26 for the following SNR
values 10 dB, 0 dB, -5 dB, -10 dB, -15 dB and -20 dB. SVM classification accuracy
started dropping from -15 dB SNR. Note that “Radar 10” to “Radar 15” in Figure
26 is the same as “Marine Mode 1” to “Marine Mode 5”.
DT has the best classification rate for marine radars as shown in Figure 27. RnF
is better than NB over all the SNRs. The F1 scores follow the same trend as shown
in 28.
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
Marine Mode 1 10 0 0 0 0 0 0 10 Marine Mode 1 10 0 0 0 0 0 0 10 Marine Mode 1 10 0 0 0 0 0 0 10
Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 ,

actual class
actual class
actual class
No radar 0 0 0 0 0 0 10 10 No radar 0 0 0 0 0 0 10 10 No radar 0 0 0 0 0 0 10 10

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 3 0 6 1 0 0 0 10

actual class
actual class
actual class

11
10
10
10
10
10
10
11
10
10
11
10
14
17
10
9
9
Figure 23: Confusion matrix for a DT classifier of realistic marine radar chirp rates
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
 Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 ,

actual class
actual class
actual class

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
11
10
10
10
9
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
Marine Mode 3 1 0 7 1 1 0 0 10 , Marine Mode 3 7 0 1 1 1 0 0 10 , Marine Mode 3 5 1 0 1 3 0 0 10 

actual class
actual class
actual class

15
18
10
28
12
14
10
20
10
24
10
5
0
Figure 24: Confusion matrix for a NB classifier of realistic marine radar chirp rates
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
 Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 , Marine Mode 3 0 0 10 0 0 0 0 10 ,

actual class
actual class
actual class

10
10
10
10
10
10
10
10
13
10
10
10
10
10
10
10
10
10
10
10
7
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
No radar
No radar
No radar
Marine Mode 3 0 2 5 0 2 1 0 10 , Marine Mode 3 0 0 7 0 2 1 0 10 , Marine Mode 3 0 3 1 1 4 1 0 10

actual class
actual class
actual class

10
14
10
10
11
10
11
16
10
10
21
18
5
9
Figure 25: Confusion matrix for a RnF classifier of realistic marine radar chirp rates
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
No radar
No radar
No radar
Radar 10 0 10 0 0 0 0 0 10 Radar 10 0 10 0 0 0 0 0 10 Radar 10 0 10 0 0 0 0 0 10

actual class
actual class
actual class
 , , ,

10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
No radar
No radar
No radar

actual class
actual class
actual class
, ,

10
10
10
10
10
10
10
52
13
63
0
0
Figure 26: Confusion matrix for a SVM classifier of realistic marine radar chirp rates
at SNR 10 dB (top left), 0 dB (top mid) , -5 dB (top right), -10 dB (bottom
Figure 27: Accuracy versus SNR level - Marine radar chirp rates. (Problem 4)
Figure 28: F1 score versus SNR level - Marine radar chirp rates. (Problem 4)
5.5.5 Problem 5: All realistic radar chirp rates results

The confusion matrix for a DT classifier of realistic chirp rates for all radars in shown
The confusion matrix result for a NB classifier of realistic chirp rates for all radars
are shown in Figure 30 for the following SNR values 10 dB, 0 dB, -5 dB, -10 dB, -15
dB and -20 dB. NB classification accuracy also started dropping from -10 dB SNR.
The confusion matrix result for a SVM classifier realistic chirp rates for all radars
are shown in Figure 31 for the following SNR values 10 dB, 0 dB, -5 dB, -10 dB, -15
dB and -20 dB. SVM classification accuracy also started dropping from -10 dB SNR.
DT, NB, and SVM have almost identical performance in terms of classification
rate for all radars till SNR of -5 dB as shown in Figure 32. The deterioration of the
DT is less than NB and SVM. The F1 scores follow the same trend as shown in Figure
33. The performance of three ML classifiers was analysed for all realistic radar chirp
rates. The RnF ML classifier was not evaluated as its performance was significantly
worse than the other three classifiers at higher SNR values.
The marine radars have the same sampling frequency and center frequency with
only changes in the pulsewidth. Secondly, there are three pairs of pulse widths where
the only difference in each pairing is a change in the bandwidth. These pairs are
marine mode 1 and marine mode 4, marine mode 2 and marine mode 5 and marine
mode 3, and marine mode 6. Since there are only differences between these pairs,
the classification results are poorer generally and more so in the more of noise. It
is clear that the marine radar chirp rates are the most difficult type to classify by
all the algorithms. Future work will investigate if more advanced machine learning
algorithms with feature selection can improve upon the results in this thesis.
5.5.6 Problem 6: Robustness of the classifiers against win-

dowing imperfections
As noted previously, the intercept receiver may not be operating when the transmitted
pulse is received at the receiver. When the receiver is turned on, it might miss a
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
Air Mode 1 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 Air Mode 1 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 Air Mode 1 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10
Marine Mode 1 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 10 Marine Mode 1 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 10 Marine Mode 1 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 10
 Marine Mode 2 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 10 , Marine Mode 2 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 10 , Marine Mode 2 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 10 ,
actual class
actual class
actual class
No radar 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 10 No radar 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 10 No radar 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 10
Weather Mode 1 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 10 Weather Mode 1 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 10 Weather Mode 1 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar

Marine Mode 2 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 10 , Marine Mode 2 0 0 0 0 0 1 0 0 0 0 0 0 9 0 0 0 10 , Marine Mode 2 0 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 10 
actual class
actual class
Marine Mode 3 0 0 0 0 10 0 0 0 10 Marine Mode 3 0 0 0 0 10 0 0 0 10

actual class
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Marine Mode 3 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 10
145
157
10
0
0
0
0
0
1
0
0
0
0
0
4
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3
0
0
10
20
10
10
10
10
10
10
10
25
10
0
2
7
8
8
Figure 29: Confusion matrix for a DT classifier of all realistic radar chirp rates at
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
11
10
10
9
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class
10
10
10
10
12
10
10
10
11
12
10
10
11
16
35
10
25
17
12
11
15
37
10
44
15
10
9
9
8
2
3
4
4
4
1
0
1
0
1
3
2
9
Figure 30: Confusion matrix for a NB classifier of all realistic radar chirp rates at
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
No radar
No radar
No radar
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9
Radar 1 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 Radar 1 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 Radar 1 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10
Radar 14 10 Radar 14 10 Radar 14 10
actual class
actual class
actual class
0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0
Radar 15 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 10 , Radar 15 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 10 , Radar 15 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 10 ,
10
10
11
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
11
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
11
9

Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
Radar 10
Radar 11
Radar 12
Radar 13
Radar 14
Radar 15
No radar
No radar
No radar
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9
Radar 1
Radar 2
Radar 3
Radar 4
Radar 5
Radar 6
Radar 7
Radar 8
Radar 9

Radar 14 10 Radar 14 10 Radar 14 10
actual class
actual class
actual class
0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 0 0 1 0 0 2 0 5 0 1 0 1 0 0 0 0 0 0 0 1 0 3 2 0 1 0 1 0 0 0 0 2 0
Radar 15 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 0 10 , Radar 15 0 0 0 1 0 0 0 8 1 0 0 0 0 0 0 0 10 , Radar 15 0 0 0 0 1 1 0 5 0 0 0 0 0 1 2 0 10 
10
10
10
10
10
10
10
10
10
10
11
10
10
10
10
10
11
11
12
10
12
10
11
10
10
10
11
14
11
19
13
20
10
13
10
14
13
9
5
8
3
4
5
Figure 31: Confusion matrix for a SVM classifier of all realistic radar chirp rates at
Figure 32: Accuracy versus SNR level - All radar chirp rates. (Problem 5)
Figure 33: F1 score versus SNR level - All radar chirp rates. (Problem 5)
portion of the pulse as illustrated in Figure 34.
1.0
0.5
1 2 3 4 5
-0.5
-1.0
Figure 34: Truncated chirp of an LFM signal. (Problem 6)
For the proof-of-concept classifier, the chirp signal was truncated to determine
if the classifier could detect and correctly classify the LFM signal when the pulse
was detected at an unknown point in the pulse stream, as shown in Figure 34. For
testing the robustness, 20% of the pulse was removed. The confusion matrix results
of applying the NB, and SVM classifiers to a truncated chirp signal for all realistic
radar modes are shown in Figures 35 and 36. The results for the DT classifier were
inconclusive due to the simplicity of the classifier. Note that the classifiers were
trained on the full pulse but tested with this truncated pulse.
All the four ML classifiers were applied but results only two best classifiers, NB
and SVM are shown in this work. Both DT and RnF performed poorly at all SNRs.
Plots of accuracy and F1 scores at various SNR levels (10 dB, 0 dB, -5 dB, -10
dB, -15 dB and -20 dB) are shown in Figure 37and Figure 38.
From the results, SVM slighlty outperforms NB at 0 dB but NB outperformed
SVM at all other SNR levels. These results are significant as the algorithm indicates
that we can classify 20% truncated pulse up to 0 dB with an accuracy and F1 score
of 0.8. The results for the confusion matrices suggest that the SVM and Naive Bayes
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class
10
10
10
10
17
10
12
10
10
10
10
10
10
10
10
10
10
10
11
18
12
29
10
10
13
10
10
10
10
10
14
60
10
14
3
8
0
3
4
0
8
2
0
0
9
0
3
0
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 1
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class

110
133
152
15
10
13
10
0
1
1
3
3
0
0
0
1
0
3
0
0
0
0
0
1
5
0
0
0
0
9
0
2
0
0
0
0
0
1
0
0
0
0
0
0
7
0
0
0
Figure 35: Confusion matrix for a NB classifier of realistic truncated chirp rates at
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class
10
10
10
10
10
10
10
10
10
10
10
30
10
10
10
10
10
10
10
10
10
10
10
10
48
10
10
18
10
10
10
96
0
0
0
2
0
1
0
0
0
5
0
0
0
0
0
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 1
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 2
Air Mode 3
Air Mode 4
Air Mode 2
Air Mode 3
Air Mode 4
No radar
No radar
No radar
actual class
actual class
actual class

150
160
160
10
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Figure 36: Confusion matrix for a SVM classifier of realistic chirp rates at SNR 10
dB (top left), 0 dB (top mid), -5 dB (top right), -10 dB (bottom left), -15 dB
(bottom mid) and −20 dB (bottom right). (Problem 6)
Figure 37: Accuracy versus SNR level - All radar truncated chirp rates. (Problem
6)
Figure 38: F1 score versus SNR level - All radar truncated chirp rates. (Problem
6)
classifiers are tolerant to truncated pulses in some degree, as long as the SNR value
is not too low.
5.5.7 Problem 7: Robustness of the classifier against pertur-

bation in chirp parameters.
Since DT was the best performing classifier for Problem 5, the DT machine learning
classifier was selected to determine if it could correctly determine the radar type
based on the realistic chirp rates of Table 13. These radars are not nominally close
to the realistic radar parameters. A tag was assigned to each unknown radar type
to indicate which radar it should be similar to (radars with realistic parameters) in
order to evaluate classifier performance.
The trained classifier from Problem 5 was tested using the simulated single LFM
pulse using the parameters given in Table 13. Classification results are shown in
Figures 39, 40, 41, 42, 43, and 44. Since the parameters are not close to the realistic
parameters, the classifiers are not expected to perform perfectly even at high SNR.
Interestingly, these simple classifiers are still able to provide reasonably good classifi-
cation rates up to about -10 dB SNR. Although the classifier does not always return
the specific radar which is “most similar” to the one listed in the table, it selected
another radar within the same class in most cases.
5.6 Summary
In this chapter, seven different classification problems were investigated. Examples
were drawn from realistic parameters in airborne, marine and weather radars. Ro-
bustness against windowed received signal as well as parameter mismatch was also
investigated. The general conclusion is that with the proposed concatenated match
filtered output as the inputs to four well known machine learning algorithms give a
reasonably good classification performance up to an SNR of -10 dB in almost all the
cases studied in this thesis.
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 0 0 0 0 0 0 0 0 0 0 0 10 0 0 0 10
Radar B Weather 0 0 0 8 0 0 0 0 2 0 0 0 0 0 0 0 10
actual
Radar C Air 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 10
Radar D Marine 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 10
Radar E Air 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 10
0 0 10 18 0 0 0 0 2 0 0 0 10 0 10 0
predicted
Figure 39: Confusion matrix for classifying signals from unknown radars at a SNR
of 10 dB using a DT classifier. (Problem 7)
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 2 10
Radar B Weather 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 10
actual
Radar C Air 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 10
Radar D Marine 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 10
Radar E Air 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 10
0 0 10 10 8 0 0 0 10 0 0 0 0 0 10 2
predicted
of 0 dB using a DT classifier. (Problem 7)
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 10
Radar B Weather 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 10
actual
Radar C Air 0 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 10
Radar D Marine 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 0 10
Radar E Air 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 10
0 0 10 10 0 0 0 0 10 0 0 0 0 0 10 10
predicted
of −5 dB using a DT classifier. (Problem 7)
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10 10
Radar B Weather 0 0 0 0 0 0 0 0 10 0 0 0 0 0 0 0 10
actual
Radar C Air 0 0 0 9 0 0 0 0 0 0 0 0 0 1 0 0 10
Radar D Marine 0 1 0 0 0 0 0 0 0 2 0 0 0 0 7 0 10
Radar E Air 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 10
0 1 10 9 0 0 0 0 10 2 0 0 0 1 7 10
predicted
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 1 0 0 1 0 0 0 0 0 1 0 4 1 1 1 10
Radar B Weather 0 0 0 0 0 0 0 0 7 0 0 0 2 1 0 0 10
actual
Radar C Air 0 0 0 9 0 0 0 0 0 0 0 0 1 0 0 0 10
Radar D Marine 0 1 0 0 0 0 1 0 0 1 1 0 1 0 3 2 10
Radar E Air 0 0 10 0 0 0 0 0 0 0 0 0 0 0 0 0 10
0 2 10 9 1 0 1 0 7 1 2 0 8 2 4 3
predicted
Weather Mode 1
Weather Mode 2
Weather Mode 3
Weather Mode 4
Weather Mode 5
Marine Mode 1
Marine Mode 2
Marine Mode 3
Marine Mode 4
Marine Mode 5
Marine Mode 6
Air Mode 1
Air Mode 2
Air Mode 3
Air Mode 4
No radar
Radar A Marine 0 1 0 0 1 0 0 0 1 1 0 0 3 1 1 1 10
Radar B Weather 0 0 0 1 0 1 0 1 2 1 0 0 3 1 0 0 10
actual
Radar C Air 0 0 0 0 0 0 1 3 0 1 1 1 2 0 1 0 10
Radar D Marine 0 2 1 0 1 0 1 0 1 1 0 0 0 0 2 1 10
Radar E Air 0 1 1 1 0 0 0 0 0 2 0 2 0 0 2 1 10
0 4 2 2 2 1 2 4 4 6 1 3 8 2 6 3
predicted
Chapter 6
Conclusions and Way Forward
This thesis has proposed an approach to identify the unknown radar chirp rate, β,
from a known set of chirp rates, of an incoming LFM radar signals at SNR values as
low as -20 dB. Supervised machine learning algorithms have been selected to identify
pulsed LFM radars using the outputs of matched filtering. The approach has been
validated with simulations. The results presented in this thesis represent only a
preliminary application of the combination of matched filtering and machine learning.
No feature extraction is attempted in this thesis. A concatenated matched filter
outputs from the matched filters corresponding to the known set of chirp rates were
presented to the classifiers used in this thesis.
The approach was tested on simulated LFM signals with pre-selected chirp rates.
The method has the drawback that it requires the set of possible LFM parameters
to be known in advance. However, this is not an unreasonable assumption in the
context of spectrum monitoring. This approach can be used to determine the presence
or absence of known radar emitters, or to show the operational or non-operational
state of these emitters potentially even in a low-SNR scenario. Low-SNR scenarios
can occur if the receiver is situated along the side-lobes of the transmitting radar
antenna.
Four types of machine learning classifiers namely, the decision tree, the naive
Bayes, the random forest and the support vector machines are used in this thesis.
Three types of radars namely airborne, marine and weather radars are considered
in this research work along with their modes. Airborne radar has 4 modes, weather
radar has 5 modes while marine radar has 6 modes. Classifiers for each type of radar is
designed and the modes of the radars are identified with single LFM pulses received
from the same type of radar. Thus in the case of classifiers for airborne, weather
80
CHAPTER 6. CONCLUSIONS AND WAY FORWARD 81
and marine radars, the classifier output has 5, 6 and 7 classes. The extra class in
each of these radars corrected to noise class (same as “no radar” class). Thus, the
input features to each of the above classifiers are outputs of a bank of 5, 6 and 7
matched filters. Another 16-way classification is also considered where the classifiers
take the output of a bank of 16 matched filters is also trained to classify the radar
type along with its mode. This three 16-way classifiers based on DT, NB and SVM
are considered. In order to study the robustness of these 16-way classifiers, a set of 5
radars whose chirp parameter is close but not exactly matching any of the parameters
in the known set of chirp parameters is considered. As the modes do not match, this
set of 5 radars is considered similar to certain radars and their types. Test data
consisting of single pulses from these 5 radars is presented to the classifier and the
output of the classifier is analyzed. As one cannot expect the classifier to correctly
identify the mode, if the classifier declares the class corresponding to the radar which
the inputs are similar to, then the classifier is considered to be robust. Furthermore,
the 16-way classifier is checked for its classification robustness to imperfect pulse.
Instead of having a full pulse, only 80% of the pulse is presented and the ability of
the classifier to still output a correct class is verified.
The proof-of-concept toy problem indicated that simple machine learning classi-
fiers trained using concatenated matched filter outputs were able to classify noisy
single pule LFM radar signals.
For classification systems trained with airborne radar single LFM pulses, DT
performed the best, and provided an acceptable performance even at -20 dB when
the classifier was presented with noisy single LFM pulses from airbone radars. DT
and SVM gave nearly identical performance for weather radars, and very good per-
formance until -15 dB For systems trained with marine radar, DT gave the best
performance, with very good performance down to about -10 dB.
In a 16-class problem, DT outperformed other classifiers when there a perfect
match in the parameters and full single pulse LFM signals were assumed at the
receiver end.
Surprisingly DT did not perform well for the truncated chirp. NB and SVM
performed well until an SNR of 0 dB, when the truncation was at 20%. Since the
study was done under noisy conditions, any more truncation would lead to serious
under performance and that was not investigated. The simulation results suggest
that the NB and SVM classifiers are more tolerant to imperfect pulses than DT and
CHAPTER 6. CONCLUSIONS AND WAY FORWARD 82
RnF.
6.1 Limitations
This investigations carried out in the thesis have the following limitations:
• The radar waveform is assumed to be LFM.
• The chirp rate of the LFM signal is a member of a known set of chirp rates.
• The signal is assumed to be adequately sampled.
• The received signal is only assumed to be corrupted by additive noise.
• Only simple machine learning algorithms were considered. For instance, neural
networks were not considered.
• Experiments for real data collection could not be carried out due to the pan-
demic situation. Therefore, this study is purely done based on simulations and
may not reflect accurately the real world performance.
• Only one emitter was considered to be present in the simulations.
• It is assumed in this study that the radar pulse has been correctly de-interleaved.
6.2 Future work

Some possible areas to consider in future research are as follows:
• Assessment of performance with real radar LFM signals.
• Superposition of multiple LFM signals which normally occur in a RF congested

environment.
• Nonlinear FM signals such as those defined in [76], [77].
• Optimization of the SVM parameters by hyperparameter tuning.
• Classification using shallow and deep neural networks.
• Assessment of performance with sophisticated radar waveforms such as ultraw-

ide band noise radars.
List of References
[1] ITU, “Technical characteristics of maritime radio navigation radars,” Tech. Rep.
ITU-R.M.1313, International Maritime Organization, 2018.
[2] Radar Spectrum Engineering and Management: Technical and Regulatory Issues,
2015.
[3] P. E. Pace, Detecting and Classifying Low Probability of Intercept Radars. Artech
House, 2004.
[4] L. Stankovic, M. Dakovic, and T. Thayparan, Time-Frequency Signal Analysis
with Applications. Artech House, 2002.
[5] X. Liu, T. Li, X. Fan, and Z. Chen, “Nyquist zone index and chirp rate estimation
of LFM signal intercepted by Nyquist folding receiver based on random sample
consensus and Fractional Fourier Transform,” Sensors, vol. 19, no. 6, pp. 1477–
1491, 2019.
[6] J. Song, X. Xu, and G. Li, “An accurate parameter estimator for LFM signals
based on zoom modified discrete chirp Fourier transform,” IAENG International
Journal of Computer Science, vol. 46, no. 3, pp. 78–85, 2019.
[7] J. Zheng, H. Liu, and Q. H. Liu, “Parameterized centroid frequency-chirp rate
distribution for LFM signal analysis and mechanisms of constant delay intro-
duction,” IEEE Transactions on Signal Processing, vol. 65, pp. 6435–6447, Dec.
2017.
[8] D. Djurovic, C. Ioana, L. Stankovic, and P. Wang, “Adaptive algorithm for chirp-
rate estimation,” EURASIP Journal on Advances in Signal Processing, 2009.
[9] D. Zeng, H. Xiong, K. Long, and B. Tang, “A fast recognition algorithm for
three kinds of intra-pulse modulation signals,” in 2009 IET International Radar
Conference, pp. 1–4, Apr. 2009.
[10] M. Wu, X. Dai, Y. D. Zhang, B. Davidson, M. G. Amin, and J. Zhang, “Fall
detection based on sequential modeling of radar signal time-frequency features,”
in IEEE International Conference on Healthcare Informatics, pp. 169–174, Sept.
2013.
[11] S. Haykin and T. Bhattacharya, “Wigner-Ville distribution: An important func-
tional block for radar target detection in clutter,” in Proceedings of 1994 28th
Asilomar Conference on Signals, Systems and Computers, vol. 1, pp. 68–72, 1994.
83
84
[12] P. K. Kumar and K. M. M. Prabhu, “Classification of radar returns using Wigner-

Ville distribution,” in 1996 IEEE International Conference on Acoustics, Speech,
and Signal Processing Conference Proceedings, vol. 6, pp. 3105–3108, May 1996.
[13] S. Sirianunpiboon, “Extended array method for detection and direction finding
of a weak linear frequency modulated signal,” in 2011 IEEE International Con-
ference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2560–2563,
May 2011.
[14] A. Ahmad, A. Muhammad Lawan, M. Ajiya, and Z. Yusuf, “Sorting of low prob-
ability of intercept radar signals based on frequency modulation constancy using
Wigner-Ville distribution,” in 2019 IEEE Asia-Pacific Conference on Applied
Electromagnetics (APACE), pp. 1–5, Nov. 2019.
[15] L. Zhang, X. He, Y. Li, M. Xing, and Z. Bao, “Parameters estimation of LFM
signals based on STTFD,” in ICSP 2008 Proceedings, 2008.
[16] T. O. Gulum, A. Y. Erdogan, T. Yildirim, and P. E. Pace, “A parameter extrac-
tion technique for FMCW radar signals using Wigner-Hough-Radon transform,”
in 2012 IEEE Radar Conference, pp. 0847–0852, May 2012.
[17] T. O. Gulum, A. Y. Erdogan, T. Yildirim, and L. Durak Ata, “Parameter ex-
traction of FMCW modulated radar signals using Wigner-Hough transform,”
in 2011 IEEE 12th International Symposium on Computational Intelligence and
Informatics (CINTI), pp. 465–468, Nov. 2011.
[18] K. K. Guner, B. Erkmen, T. O. Gulum, A. Y. Erdogan, T. Yıldırım, and L. Durak
Ata, “Implementation aspects of Wigner-Hough transform based detectors for
LFMCW signals,” in 2016 39th International Conference on Telecommunications
and Signal Processing (TSP), pp. 441–444, June 2016.
[19] F. G. Geroleo, M. Brandt-Pearce, and C. L. Brown, “Detection and estimation of
multi-pulse LFMCW radar signals,” in 2010 IEEE Radar Conference, pp. 1009–
1013, May 2010.
[20] W. Niczyporuk, “LFM radar signal detection in the joint time-frequency do-
main,” in Signal Processing Algorithms, Architectures, Arrangements, and Ap-
plications SPA 2007, pp. 33–37, Sept. 2007.
[21] A. Bouchikhi, A. Boudraa, J. Cexus, and T. Chonavel, “Analysis of multicompo-
nent LFM signals by Teager-Huang-Hough transform,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 50, pp. 1222–1233, April 2014.
[22] S. Sirianunpiboon, S. D. Howard, and S. D. Elton, “A group invariance approach
to a very weak LFM signal detection,” in 2018 IEEE Statistical Signal Processing
Workshop (SSP), pp. 698–701, June 2018.
[23] E. Cheng, S. Chen, and F. Yuan, “Design and detection of multilinear chirp
signals for underwater acoustics sensor networks,” International Journal of Dis-
tributed Sensor Networks, vol. 11, Nov. 2015.
85
[24] A. Serbes, “A method for estimating chirp rate of a LFM signal,” Balkan Journal
of Electrical and Computer Engineering, vol. 6, pp. 9–12, Feb. 2018.
[25] A. Serbes and O. Aldimashki, “A fast and accurate chirp rate estimation al-
gorithm based on FrFT,” in 2017 25th European Signal processing Conference,
2017.
[26] O. Aldimashki and A. Serbes, “Performance of chirp parameter estimation in
the fractional fourier domains and an algorithm for fast chirp-rate estimation,”
IEEE Transactions on Aerospace and Electronic Systems, vol. 56, pp. 3865–3700,
Oct. 2020.
[27] Z. Qiu, J. Zhu, and F. Li, “Multiple BPSK/LFM hybrid modulated signals pa-
rameter estimation and analysis intercepted by non-cooperative radar receiver,”
in 2018 IEEE International Conference on Computational Electromagnetics (IC-
CEM), pp. 1–4, Mar. 2018.
[28] Y. Ma, R. Wang, and J. Du, “Amplitude characteristics of linear frequency
modulation signal in FRFT domain,” in 2013 IEEE International Conference
on Signal and Image Processing Applications, pp. 431–434, Oct. 2013.
[29] Q. Chen, Y. Li, and M. Zhu, “Fast algorithm for parameter estimation of LFM
signals under low SNR,” in AIP Conference Proceedings, vol. 1839, 2017.
[30] D. Fourer, F. Auger, K. Czarnecki, S. Meignen, and P. Flandrin, “Chirp rate
and instantaneous frequency estimation:application to recursive vertical syn-
chrosqueezing,” IEEE Signal Processing Letters, vol. 24, pp. 1724–1728, Nov.
2017.
[31] G. Turin, “An introduction to matched filters,” IRE Transactions on Information
Theory, vol. 6, pp. 311–329, June 1960.
[32] S. Salem, “Enhancement of matched filter response for chirp radar signals using
signal recovery method,” International Journal of Innovative Technology and
Exploring Engineering, vol. 9, no. 2, pp. 561–565, 2019.
[33] A. Selim, F. Paisana, J. Arokkiam, Y. Zhang, and L. Doyle, “Spectrum moni-
toring for radar bands using deep convolutional neural networks,” in 2017 IEEE
Global Communications Conference, pp. 1–6, Dec. 2017.
[34] M. Kong, J. Zhang, W. Liu, and G. Zhang, “Radar emitter identification based
on deep convolutional neural network,” in International Conference on Control,
Automation and Information Sciences, pp. 309–314, Oct. 2018.
[35] L. Wang, H. Ma, Q. Li, and Z. Li, “Multicomponent linear FM signal detection
based on support vector clustering,” in 5th International Conference on Infor-
mation Communications Signal Processing, pp. 743–747, 2005.
[36] P. Ruiz, J. Mateos, G. Camps-Valls, R. Molina, and A. K. Katsaggelos, “Bayesian
active remote sensing image classification,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 52, pp. 2186–2196, Apr. 2014.
86
[37] J. Lunden, L. Terho, and V. Koivunen, “Waveform recognition in pulse compres-

sion radar systems,” in 2005 IEEE Workshop on Machine Learning for Signal
Processing, pp. 271–276, 2005.
[38] E. Fischell and H. Schmidt, “Supervised machine learning for estimation of tar-
get aspect angle from bistatic acoustic scattering,” IEEE Journal of Oceanic
Engineering, pp. 759–769, Oct. 2017.
[39] J. Seok, “Active sonar target classification using classifier ensembles,” Interna-
tional Journal of Engineering Research and Technology, vol. 11, no. 12, pp. 2125–
2133, 2018.
[40] S. El-Khamy and H. Elsayed, “Classification of multi-user chirp modulation sig-
nals using wavelet higher-order statistics features and artificial intelligence tech-
niques,” International Journal of Communications, Network and System Sci-
ences, vol. 5, pp. 520–533, Sept. 2012.
[41] B. Boashash and S. Ouelha, “Automatic signal abnormalilty detection using
time-frequency features and machine learning: A newborn EEG seizure study,”
Knowledge-based Systems, vol. 106, pp. 38–50, Aug. 2016.
[42] H. Su, Q. Bao, and Z. Chen, “Parameter estimation processor for chirp sig-
nals based on a complex-valued deep neural network,” IEEE Access, vol. 7,
pp. 176278–176290, 2019.
[43] G. Lee, J. Jo, and C. Park, “Jamming prediction for radar signals using machine
learning methods,” Security and Communications Networks, vol. 2020, 2020.
[44] L. Gao, X. Zhang, J. Gao, and S. You, “Fusion image-based radar signal feature
extraction and modulation recognition,” IEEE Access, vol. 7, pp. 13135–13148,
2019.
[45] F. Wang, S. Huang, H. Wang, and C. Yang, “Automatic modulation classifi-
cation exploiting hybrid machine learning network,” Mathematical Problems in
Engineering, vol. 2018, p. 14, 2018.
[46] G. Noone, “Radar pulse classification using support vector machine,” in Work-
shop on the Application of radio Science National Committee for Radio Science,
2000.
[47] J. Kwon, S. Lee, and N. Kwak, “Radar application of deep neural networks
for recognizing micro-Doppler radar signals by human walking and background
noise,” in 2018 International Symposium on Antennas and Propagation (ISAP),
pp. 1–2, 2018.
[48] S. Hong, Y. Yi, J. Jo, and B. Seo, “Classification of radar signals with convolu-
tional neural networks,” in 2018 Tenth International Conference on Ubiquitous
and Future Networks (ICUFN), pp. 894–896, 2018.
[49] M. V. i. Carretero, R. I. A. Harmanny, and R. P. Trommel, “Smart-CFAR, a
machine learning approach to floating level detection in radar,” in 2019 16th
European Radar Conference (EuRAD), pp. 161–164, 2019.
87
[50] M. Ren, J. Cai, Y. Zhu, and M. He, “Radar emitter signal classification based
on mutual information and fuzzy support vector machines,” in ICSP2008 Pro-
ceedings, pp. 1641–1646, 2008.
[51] J. M. Kurdzo, J. Y. N. Cho, B. L. Cheong, and R. D. Palmer, “A neural network
approach for waveform generation and selection with multi-mission radar,” in
2019 IEEE Radar Conference (RadarConf ), pp. 1–6, 2019.
[52] O. Mughal and S. Kim, “Signal classification and jamming detection in wide-
band radios using Naive Bayes classifier,” IEEE Communications Letters, vol. 22,
pp. 1398–1401, July 2018.
[53] D. Xu, L. Du, H. Liu, and D. Luo, “Sparse bayesian multitask learning for radar
target recognition,” in 2016 CIE International Conference on Radar (RADAR),
pp. 1–4, 2016.
[54] L. Wang, H. Ma, Q. Li, and Z. Li, “Multicomponent LFM signal detecetion based
on support vector clustering,” in 5th International Conference on Information
Systems and Signal Processing, pp. 743–747, 2005.
[55] R. Caroni and M. Souryal, “Detection of incumbent radar in the 3.5 GHz CBRS
band,” in 2018 IEEE Global Conference on Signal and Information Processing
(GlobalSIP), pp. 241–245, 2018.
[56] M. Skolnik, Introduction to Radar Systems. McGraw-Hill, 2 ed., 1980.
[57] M. Tipping, “The relevance vector machine,” in Advances in Neural Information
Processing Systems, pp. 652–658, 2000.
[58] D. Anguita, A. Ghio, N. Greco, L. Oneto, and S. Ridella, “Model selection for
support vector machines:advantages and disadvantages of the machine learn-
ing theory,” in The 2010 International Joint Conference on Neural Networks
(IJCNN), pp. 1–8, 2010.
[59] S. Archana and K. Elangovan, “Survey of classification techniques in data min-
ing,” International Journal of Computer Science and Mobile Applications, vol. 1,
pp. 268–278, Nov. 2014.
[60] S. L. Crawford, “Extensions to the CART algorithm,” International Journal of
Man-Machine Studies, vol. 31, no. 2, pp. 197–217, 1989.
[61] T. Buskirk, “Surveying the forests and sampling the trees: An overview of clas-
sification and regression trees and random forests with applications in survey
research,” Survey Practice, vol. 11, no. 1, 2018.
[62] R. Kohavi and F. Provost, “Guest editors’ introduction: On applied research in
machine learning,” Machine Learning, vol. 30, pp. 127–132, Feb. 1998.
[63] A. Ragan, “Taking the confusion out of confusion matrices,”
www.towardsdatascience.com, Oct. 2018. Accessed April 2020.
88
[64] M. Kubat, R. C. Holte, and S. Matwin, “Machine learning for the detection of
oil spills in satellite radar images,” in Machine Learning, vol. 30, pp. 195–215,
1998.
[65] F. Giannini, K. Lupinetti, and M. Monti, “Identification of similar and com-
plementary subparts in B-Rep mechanical models,” Journal of Computers and
Information Science, vol. 17, Dec. 2017.
[66] D. S. Zrnic and R. J. Doviak, “System requirements for phased array weather
radar,” tech. rep., National Severe Storms Laboratory, 2018.
[67] M. Gabella, R. Notarpietro, S. Bertoldo, A. Prato, C. Lucianaz, O. Rorato,
M. Allegretti, and G. Perona, Doppler Radar Observations: Weather Radar,
Wind Profiler, Ionospheric Radar, and Other Advanced Applications, ch. “A net-
work of portable, low-cost X-band radars, Doppler radar Observations, network
of portable, low-cost radars”. IntechOpen, 2012.
[68] “Operators manual WR-2100, https://fccid.jp/ad89zwwr2100/user-
manual/user-manual 2470823, (accessed june 29, 2018).”
[69] M. Wada, N. Anraku, H. Yamauchi, and A. Adachi, “Operational usage of solid-
state weather radar,” in WMO technical conference on meteorological and envi-
ronmental instruments and methods of observation (TECO), 2012.
[70] J. L. Salazar, E. A. Knapp, and D. J. McLuaghlib, “Antenna design trade-offs for
a dense, distributed radar network for weather sensing,” in Proc. 33 International
Conference on Radar Meteorology, 2007.
[71] N. A. Press, “Weather radar technology beyond NEXRAD,” tech. rep., Commit-
tee on Weather Radar Technology Beyond NEXRAD, 1982.
[72] A. Nelander and Z. Tóth-Pál, “Modular system design for a new S-band marine
radar,” in International Radar Conference, pp. 1–5, 2009.
[73] Furono Operator’s Manual – Model 7111, 1994.
[74] M. E. Weber, “FAA surveillance radar data as a complement to the WSR-88D
network,” tech. rep., https://www.cv.nrao.edu, 2005.
[75] C. Lesnik, “Nonlinear frequency modulated signal design,” Acta Physica
Polonica A, vol. 116, no. 3, pp. 351–354, 2009.
[76] A. W. Doerry, “Generating nonlinear FM chirp waveforms for radar,” tech. rep.,
Sandia Report SAND2006-5856, 2006.
[77] K. J. Abratkiewicz and P. Samczytiski, “A block method using the chirp rate es-
timation for NLFM radar pulse reconstruction,” Sensors, vol. 19, no. 22, p. 5015,
2019.

Machine Learning Algorithms For Identification of Low Signal-To-Noise Ratio Radar Signals

Uploaded by

Copyright:

Available Formats

Machine Learning Algorithms For Identification of Low Signal-To-Noise Ratio Radar Signals

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning Algorithms For Identification of Low Signal-To-Noise Ratio Radar Signals

Uploaded by

Copyright:

Available Formats

Machine learning algorithms for identification of

low signal-to-noise ratio radar signals

Anne Young, B.Sc.

A thesis submitted to the

Master of Applied Science in Electrical Engineering

Ottawa-Carleton Institute for Electrical and Computer Engineering

Areas such as spectrum monitoring require identification of known and unknown

First, I would like to extend my heartfelt gratitude to my supervisors, Prof. Sreeraman

List of Figures vii

3 Radar Signal Processing and Machine Learning 16

4 Some Common Radars 29

6 Conclusions and Way Forward 80

2 Nexrad WSR Weather Radar Parameters . . . . . . . . . . . . . . . . 32

1 Support Vector Machine Hyperplane [1]. . . . . . . . . . . . . . . . . 22

AESA Advanced Electronically Scanned Array

ANN Artificial Neural Network

ARSR Air Route Surveillance Radars

ASL Above Sea Level

ASR Airport Surveillance Radars

AUC Area Under the Curve

BPSK Barker Phase Shift Keying

CCQP Convex Constrained Quadratic Programming

CFAR Constant False Alarm Rate

CNN Convolutional Neural Network

dBz Decibel (zero weighted)

DNN Deep Neural Network

FFT Fast Fourier Transform

FMCW Frequency Modulated Continuous Wave

FRFT Fractional Fourier Transform

FSK Frequency Shift Keying

Gt,r Antenna Gain, Receive and Transmit

GHz Giga Hertz

IPM Intra-pulse Modulation

IMO International Maritime Organisation

IMOP Intra-modulation on Pulse

ITU International Telecommunications Union

LFM Linear Frequency Modulation

LFMCW Linear Frequency Modulation Continuous Wave

LPI Low Probability of Intercept

LSTM Long Short-term Memory

MHz Mega Hertz

MLE Maximum Likelihood Estimation

MSE Mean Squared Error

NLFM Non-linear Frequency Modulation

PAR Precision Approach Radar

PRC Precision Recall Curve

PRF Pulse Repetition Frequency

PRI Pulse Repetition Interval

PRR Pulse Repetition Rate

QPSK Quadrature Phase Shift Keying

RCS Radar Cross Section

RnF Random Forrest

RMSE Root Mean Squared Error

rpm Revolutions per Minute

SNR Signal to Noise ratio

STFT Short-Time Frequency Transform

SVM Support Vector Machine

To Standard Environmental Temperature

WVD Wigner-Ville Distribution