Limited Data-Driven Multi-Task Deep Learning Approach for Target Classification in SAR Imagery

Synthetic Aperture Radar (SAR) is an advanced radar, which has an extensive application in many fields. However, the annotation of SAR images requires a lot of professional knowledge, resulting in SAR target classification facing the problem of limited labelled data. Consequently, traditional deep learning methods often fail to achieve high accuracy due to their reliance on substantial training data. Therefore, this paper proposed a limited data-driven multi-task learning (MTL-Net) method for SAR target classification to solve this problem. The method of multi-task learning can share knowledge features between multiple tasks. It can also enhance the model's generalization ability and reduce its reliance on training data. Specifically, the proposed method encompasses two primary tasks, the main task utilizes a complex value network for target classification, and the auxiliary task reconstructs the sub-aperture images to help the complex value network extract more separability features from SAR images. Based on the Moving and Stationary Target Acquisition and Recognition (MSTAR) measured data, the experimental results illustrate that MTL-Net achieved a classification accuracy of 99.59% on the MSTAR dataset. In addition, MTL-Net still achieved good results even with training data of 40% and 60%, effectively solving the problem of limited data.