Class hierarchy-driven regularization model for SAR target classification

The classification of synthetic aperture radar (SAR) images is crucial for assigning accurate class labels to targets of interest. Recently, numerous deep learning-based methods have been proposed for this purpose. Despite their success, these existing methods often neglect critical domain knowledge. Specifically, they fail to account for the inherent hierarchical relationships among target categories. For instance, BTR60 and BTR70 are both armored personnel carriers, while T62 and T72 are tanks, and all these examples fall under the broader category of land targets. These hierarchies encapsulate meaningful intra-class similarities and inter-class distinctions. However, current methods typically treat categories as independent entities, overlooking the potential benefits of leveraging this hierarchical structure. In fact, incorporating such hierarchies can serve as a knowledge-driven regularization mechanism. Therefore, this paper constrains the model to preserve feature similarities among related categories, thereby reducing overfitting and enhancing the discrimination of structurally similar classes. The proposed class hierarchy-driven regularization model for SAR target classification integrates expert-defined category hierarchies into its architecture. By combining empirical domain knowledge with deep feature extraction, this framework aims to improve the accuracy and interpretability of SAR target classification through knowledge-guided regularization. The effectiveness of the proposed hierarchy-driven regularization model was evaluated using two publicly available datasets (i.e., MSTAR and OpenSARShip). The experimental results validate the effectiveness of our proposed framework in enhancing model performance, highlighting the value of incorporating hierarchical relationships for improved SAR image target classification methods.