Decision-Boundary-Aware Adversarial Distillation for Remote Sensing Scene Classification

Abstract

Knowledge distillation (KD) has become a useful solution for deploying lightweight models in remote sensing scene classification. However, existing KD methods mainly focus on transferring static class-level knowledge while overlooking the decision transition behavior between confusing categories, which limits the student’s ability to learn fine-grained decision boundary. To solve the problem, we propose the Decision-Boundary-Aware Adversarial Distillation (DBA-AD) framework. Specifically, a transition-sensitive localization strategy is first introduced to identify the critical region that drives category transition. Based on the obtained transition mask, masked adversarial interventions are performed to generate adversarial samples near decision boundary. Furthermore, a confidence-aware dynamic distillation mechanism is designed to adaptively adjust the distillation strength according to the teacher’s prediction uncertainty. A series of experiments conducted on AID and UCM datasets have validated the effectiveness of the framework. DBA-AD achieves overall accuracies of 95.51% and 98.70%, respectively, outperforming the teacher model and several representative remote sensing scene classification methods. The results indicate that DBA-AD effectively transfers the knowledge about decision boundary and decision transition behavior, leading to improved classification performance without increasing the inference complexity of the model.