Uncertainty Estimation for CNN-based SAR Target Classification

Since target classification is an significant technique of synthetic aperture radar (SAR) images, many deep convolutional neural networks (CNNs) have been applied to SAR target classification and achieved a high accuracy recently. However, these CNN-based methods pay no attention to predictive uncertainty, which is closely related to safety and reliability. In practice, it is important to know how confident the model is in the predictive classification result. Therefore, in this study, predictive uncertainty in CNN is to be estimated and calibrated. Due to the fact that the output probabilities of softmax layer cannot indicate how confident the model is, a classical CNN combined with a distance measurement scheme (CNN-DM) is utilized to obtain classification prediction and confidence scores of predictions. The proposed CNN-DM is able to reduce the predictive uncertainty by minimizing the distance between the feature vector of sample and the correct class centroid during training. Experiment is conducted under the condition of insufficient SAR samples, and the results show the proposed method has great performance on both accuracy and confidence.