SAR Image Authentic Assessment with Bayesian Deep Learning and Counterfactual Explanations

Abstract

Synthetic Aperture Radar (SAR) targets vary significantly with observation angles, which leads to the deep neural network being over-confident and unreliable under the condition of limited measurements. This issue can be addressed by means of data augmentation such as simulation or exemplar synthesis for unknown observations, however, the authenticity of simulated/synthesized data is critical. In this study, we propose a Bayesian convolutional neural network (BayesCNN) for SAR target recognition to attain better generalization and assess predictive uncertainty. Based on the measured uncertainty of BayesCNN’s prediction, we further propose to obtain the counterfactual explanation of an unrealistic SAR target generated by the unreliable simulation method. The experiments preliminarily demonstrates the proposed BayesCNN outperforms the counterpart frequentist neural network in terms of accuracy and confidence calibration when observation angles are constrained. In addition, the counterfactual explanation can reveal the non-authenticity of the augmented SAR target, which inspires us to filter the high-quality data, as well as to understand and improve the fidelity of generated SAR image in the future study.

This work was supported by the National Natural Science Foundation of China under Grant 62101459, and the China Postdoctoral Science Foundation under Grant BX2021248.