Page 167 - ITU Journal Future and evolving technologies – Volume 2 (2021), Issue 2

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 2

Naive Bayes classifier K-nearest Neighbor classifier
Ballistic missiles 200 Ballistic missiles 43 94 5 58
Birds 200 Birds 200
Cruise missiles 200 Cruise missiles 200
Fighter jets 200 Fighter jets 200
True Class Fixed wing UAVs 200 200 200 True Class Fixed wing UAVs 189 91 200 5 54 11
HGVs
HGVs 50
Helicopters
Helicopters
Large fixed wing planes 200 Large fixed wing planes 200
Multi-rotor UAVs 200 Multi-rotor UAVs 7 63 130
Rockets & artillery 200 Rockets & artillery 13 25 41 121
Small fixed wing planes 200 Small fixed wing planes 200
Fighter jets
Helicopters
Birds
Fixed wing UAVs
HGVs
Fighter jets
Fixed wing UAVs
Helicopters
Rockets & artillery
Small fixed wing planes
Multi-rotor UAVs
Small fixed wing planes
Multi-rotor UAVs
HGVs
Rockets & artillery
Birds
Ballistic missiles e missiles Large fixed wing planes Ballistic missiles e missiles Large fixed wing planes
Cruis
Cruis
Predicted Class Predicted Class
(a) (b)
Linear Discriminant Analysis classifier Random Forest classifier
Ballistic missiles 200 Ballistic missiles 200
Birds 200 Birds 200
Cruise missiles 200 Cruise missiles 200
Fighter jets
Fighter jets 200 Fixed wing UAVs 200 178 22
True Class Fixed wing UAVs 19 181 200 200 True Class Helicopters 200 200
HGVs
HGVs
Helicopters
Large fixed wing planes 200 Large fixed wing planes 200
Multi-rotor UAVs 200 Multi-rotor UAVs 200
Rockets & artillery 200 Rockets & artillery 200
Small fixed wing planes 88 112 Small fixed wing planes 7 193
Large fixed wing planes
e missiles
Rockets & artillery
Small fixed wing planes
Multi-rotor UAVs
Helicopters
Rockets & artillery
e missiles
HGVs
Large fixed wing planes
Fighter jets
Fighter jets
Small fixed wing planes
Ballistic missiles Birds Fixed wing UAVs HGVs Multi-rotor UAVs Ballistic missiles Birds Fixed wing UAVs Helicopters
Cruis
Cruis
Predicted Class Predicted Class
(c) (d)
Fig. 15 – Confusion matrix for classi ication of targets at 200 different instances using (a) Naive Bayes, (b) K‑nearest neighbor, (c) Linear Discriminant
Analysis classi iers, and (d) Random forest classi iers. The classi ications are obtained through automated hyperparameter optimization.
that is signi icantly less than the SNR threshold for target are used in the simulations [34]. Hyperparameter opti‑
detection. Consequently, the target will be detected. mization in classi ication is also performed using Matlab.
The laser beam in Fig. 13 is used to form meshes at dif‑ For this speci ic target (Tomahawk misile), NB, LDA, and
ferent steering positions shown in Fig. 14. In the simula‑ DT models were able to correctly classify based on its fea‑
tions, we used 7 steering positions, i.e., = −250 ∶ 25 ∶ tures provided in Table 1 as a cruise missile, whereas KNN
250, and each steering position had three 1D arrays i.e. failed. For a better understanding of the used classi ica‑
= 3. Each 1D array had 21 RX elements from the two tion models, the confusion matrices are also provided for
airborne UAVs. The number of laser intersection posi‑ all four classi iers in Fig. 15. To derive the confusion ma‑
tions in each mesh were 21×21. A target highlighted with trices, 200 new samples are created using the parameters
red dots is shown in Fig. 14. The target lays over three given in Table 1. The model is optimized using automated
meshes at each steering position. The estimated features optimized hyperparameter values to minimize the classi‑
of the target from Section 4 were recorded for the target. ication error.
The features of the target are given in Table 1 under the TheresultsgiveninFig.15showthattheNBclassi ierper‑
name Given target. The features of the given target are forms the best among all. There are no misclassi ications
similar to a BGM‑109 Tomahawk missile [33]. with an NB classi ier. The NB classi ier considers the dif‑
In our data set, we used 200 samples per class generated ferent features given in Table 1 as independent that helps
from the Gaussian distribution parameters given in Ta‑ in the best classi ication. The KNN performs poorly com‑
ble 1. A target can be classi ied with the help of training pared to the other classi iers. The classi ication in KNN
data of different aerial targets, and using NB, LDA, KNN, is based on nearest distance and values of many of the
and RF classi iers. The NB, LDA, KNN, and RF classi iers features of the targets e.g., length and velocity shown in
from Statistics and Machine Learning Toolbox of Matlab Fig. 8 are overlapping, therefore, KNN misclassi ies differ‑
© International Telecommunication Union, 2021 153