IEEE - Aerospace and Electronic Systems - July 2021 - 57

Vakil et al.
hard voting, as for the given inputs only one of the methods
performed above 0.80 in accuracy.
Besides implementing voting for decision-level fusion,
Figure 8.
Overview ofcomparison research architecture.
Table 13.
Classification Accuracy of Traditional Probabilistic
Classifiers and Learning Methods
Method and Input Data
Accuracy
Logistic Regression (EO and
Nadir)
Logistic Regression (EO,
Nadir, and North)
Logistic Regression (EO,
Nadir, North, and West)
0.81( 0.13)
0.75 ( 0.17)
0.73 ( 0.17)
Naı¨ve Bayes (EO and Nadir) 0.64 ( 0.04)
Naı¨ve Bayes (EO, Nadir, and
North)
0.64 ( 0.08)
Naı¨ve Bayes (EO, Nadir,
North, and West)
Random Forest (EO, Nadir,
North, and West)
Random Forest (EO, Nadir,
and North)
Random Forest (EO, Nadir,
North, and West)
Gaussian Naı¨ve Bayes (EO
and Nadir)
Gaussian Naı¨ve Bayes (EO,
Nadir, and North)
Gaussian Naı¨ve Bayes (EO,
Nadir, North, and West)
0.63 ( 0.07)
0.64 ( 0.21)
0.69 ( 0.16)
0.67 ( 0.17)
0.73 ( 0.14)
0.71 ( 0.15)
0.70 ( 0.14)
is completed the results are used for decision-level fusion.
For the purposes of comparison research, hard and soft
voting was chosen, using the classification output of logistical
regression, Naı¨ve Bayes, random forest, and GNB to
implement decision-level fusion.
As seen in Table 14, the results for hard and soft voting
showed little difference in terms oftotal accuracy. Given the
input methods being weighed against each other, and the
accuracies they had individually, the result of the decisionlevel
fusion is dependent on the accuracies of the methods
implemented. Soft voting performed marginally better than
JULY 2021
Table 15.
SVM Fusion for EO and RF Data
Scenario
1
2
3
Accuracy
Macro AVG 0.90
Weighted AVG 0.90
IEEE A&E SYSTEMS MAGAZINE
0.76
0.93
1.00
Table 14.
Decision-Level Fusion Comparison
Method and Input Data
Hard Voting (LR, RF, NB, GNB)
(EO and Nadir)
Hard Voting (LR, RF, NB, GNB)
(EO, Nadir, and North)
Hard Voting (LR, RF, NB, GNB)
(EO, Nadir, North, and West)
Soft Voting (LR, RF, NB, GNB)
(EO and Nadir)
Soft Voting (LR, RF, NB, GNB)
(EO, Nadir, and North)
Soft Voting (LR, RF, NB, GNB)
(EO, Nadir, North, and West)
SVM learning was used in order to test the effectiveness
of decision-level fusion. For the SVM decision-level
fusion, the prediction values of the independently trained
standalone EO and standalone RF NN were fed as a
concatenated array of values. As seen above in Table 15,
the late decision fusion implemented via SVM with the
standalone EO and RF NN classification weights could
achieve an accuracy of 88%.
Out of the results of using feature- and decision-level
fusion with the RF histograms and EO frames, the highest
accuracy of all the methods tested was FERNN. Out of the
decision-level fusion methods, late-fusion NN (LFNN)
was the closest in terms of accuracy, with an F1 score of
90.7%. From the results, it can be concluded that for this
dataset of the RF and EO features, a NN benefits more
Accuracy
0.73 (þ/- 0.14)
0.71 (þ/- 0.15)
0.70 (þ/- 0.12)
0.74 (0.13)
0.72 ( 0.11)
0.71 ( 0.13)
Precision Recall F1-Score
0.84 0.81
0.96 0.94
0.73 0.85
0.88
0.88 0.88
0.88 0.88
57
IEEE - Aerospace and Electronic Systems - July 2021

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