In this paper, a feedback training approach for efficiently dealing with distribution shift in synthetic aperture radar target detection using a Bayesian convolutional neural network is proposed. After training the network on in-distribution data, it is tested on out-of-distribution data. Samples that are classified incorrectly with high certainty are fed back for a second round of training. This results in the reduction of false positives in the out-of-distribution dataset. False positive target detections challenge human attention, sensor resource management, and mission engagement. In these types of applications, a reduction in false positives thus often takes precedence over target detection and classification performance. The classifier is used to discriminate the targets from the clutter and to classify the target type in a single step as opposed to the traditional approach of having a sequential chain of functions for target detection and localisation before the machine learning algorithm. Another aspect of automated synthetic aperture radar detection and recognition problems addressed here is the fact that human users of the output of traditional classification systems are presented with decisions made by “black box” algorithms. Consequently, the decisions are not explainable, even to an expert in the sensor domain. This paper makes use of the concept of explainable artificial intelligence via uncertainty heat maps that are overlaid onto synthetic aperture radar imagery to furnish the user with additional information about classification decisions. These uncertainty heat maps facilitate trust in the machine learning algorithm and are derived from the uncertainty estimates of the classifications from the Bayesian convolutional neural network. These uncertainty overlays further enhance the users’ ability to interpret the reasons why certain decisions were made by the algorithm. Further, it is demonstrated that feeding back the high-certainty, incorrectly classified out-of-distribution data results in an average improvement in detection performance and a reduction in uncertainty for all synthetic aperture radar images processed. Compared to the baseline method, an improvement in recall of 11.8%, and a reduction in the false positive rate of 7.08% were demonstrated using the Feedback-assisted Bayesian Convolutional Neural Network or FaBCNN.

This paper describes the development and testing of low cost Inverse Synthetic Aperture Radar (ISAR) turn table system with a machine learning back-end. The ISAR sensor is based on audio components which mimic the functioning of a radar system but at a much lower cost. The system is also compact enough to fit on a single desk for classroom demonstrations and experiments. The system can record range lines as the turn table revolves and form ISAR images. These images can then be used to train machine learning algorithms to demonstrate the accuracy of such algorithms. The system thus allows for classroom demonstrations of the sensor to classifier chain in a way that is immediately accessible to students.

There have been numerous advancements in machine learning technologies in recent years, which has led to the application of machine learning algorithms to automatic target recognition. Two key challenges for these methods are the lack of sufficient training datasets and non-transparent deep models. In this paper, experiments are conducted that investigate the application of target detection using a model trained on the MSTAR to detect targets in another dataset, as well as the investigation of uncertainty estimates in Bayesian convolutional neural networks and how these outputs can improve confidence in the model’s predictions. The model can correctly detect targets in the test scene’s, as well as targets not seen from the MSTAR dataset. The output of the Bayesian convolutional neural network is used to create uncertainty heat maps. The epistemic uncertainty is the uncertainty created by the model and aleatoric is created by the data. These heat maps are overlaid on SAR images, thereby aiding in explainability by highlighting regions in the SAR images that exhibit high uncertainty from a classification point of view. Hence, uncertainty estimates from the Bayesian model give insight into the confidence of its predictions and show promise to improve trust between users and the model.