Boosting SAR ATR with simulated data: Unsupervised domain adaptation in time, frequency, scattering domains

Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) often faces insufficient labeled data, limiting Deep Convolutional Neural Networks (DCNNs) deployment. Simulated SAR data, which is easy to collect and has consistent acquisition principle with measured data, is an important knowledge source for SAR ATR. However, current cross-domain methods, which adhere to general domain adaptation paradigm, lose discriminative features and underuse unlabeled measured data. To overcome these issues, this paper introduces TFSNet, which integrates feature alignment in the Time, Frequency, and Scattering domains, along with class alignment based on multi-domain decision consistency. It selects optimal frequency-domain component to capture distinctive details and integrates electromagnetic scattering physical information to supplement discrimination upon the time domain, enabling comprehensive SAR data understanding. Furthermore, TFSNet reduces the intra-class distance and increases the inter-class distance based on the pseudo-labels generated through the decision consistency across the three domains to learn the rational decision boundaries. Experimental results demonstrate that TFSNet achieves state-of-the-art performance, and ablation studies validate the effectiveness of components.