Improved Explainability through Uncertainty Estimation in Automatic Target Recognition of SAR Images

Abstract

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.