Abstract:
In recent years, there has been significant developments in artificial intelligence (AI), with machine
learning (ML) implementations achieving impressive performance in numerous fields. The defence capability
of countries can greatly benefit from the use of ML systems for Joint Intelligence, Surveillance,
and Reconnaissance (JISR). Currently, there are deficiencies in the time required to analyse large Synthetic
Aperture Radar (SAR) scenes in order to gather sufficient intelligence to make tactical decisions.
ML systems can assist through Automatic Target Recognition (ATR) using SAR measurements to
identify potential targets. However, the advancements in ML systems have resulted in non-transparent
models that lack interpretability by the human users of the system and, therefore, disqualifying the use
of these algorithms in applications that affect human lives and costly property.
Current Deep Machine Learning (DML) implementations applied to ATR are still non-transparent and
suffer from over-confident predictions. This study addresses these limitations of DML by investigating
the performance of a Bayesian Convolutional Neural Network (BCNN) when applied with the task
of ATR using SAR images. In addition, the BCNN is used to perform target detection using data
provided by the Council for Scientific and Industrial Research (CSIR). To improve interpretability, a
method is proposed that utilises the epistemic uncertainty of the BCNN detector to visualise high- or
low-confidence regions in each of the SAR scenes.
The results of this research showed that the performance of the BCNN in the task of ATR using SAR
images is comparable to current DML methods from literature. The BCNN achieves a classification
accuracy of 93.1 % which is marginally lower than the performance of a similar Convolutional Neural
Network of 96.8 %. The BCNN outperformed the CNN when the networks were given out-ofdistribution
data. The CNN outputs showed over-confident predictions while the BCNN was able to
indicate its lack of confidence by using the epistemic uncertainty in combination with the predictive
variance in its output.
Using the dataset from the CSIR, uncertainty heat maps were generated that illustrated regions of highand
low-confidence. The regions with the highest uncertainty were located near large collections of
trees and areas near shadows. The high-uncertainty incorrect predictions were fed back into the BCNN,
and results showed a reduction in overall uncertainty and detection performance.