Unsupervised Domain Adaptation SAR Target Recognition Based on Evidential Deep Learning

Synthetic aperture radar (SAR) target recognition based on deep learning has yielded superior results. However, these methods typically necessitate large-scale labeled SAR data, which is challenging to acquire. Therefore, investigating the knowledge transfer from easily generated simulated SAR data to unlabeled real SAR data is of significant importance. The distribution difference between simulated SAR data and real SAR data can impact recognition performance. To align the distribution of simulated SAR data with that of real SAR data, this paper proposes an unsupervised domain adaptation (UDA) SAR target recognition algorithm based on evidential deep learning (EDL). The algorithm employs an adversarial-based method for domain-level alignment between the simulated and real domains, while also filtering pseudo-labels to achieve finer class-level alignment. To obtain reliable pseudo-labels, the algorithm utilizes EDL with a second-order distribution model based on the Dirichlet distribution, enabling the model to represent decision uncertainty, which integrates model uncertainty and data uncertainty. Experimental results indicate that the algorithm can achieve excellent recognition performance.