Dynamic Priority Replay and Boundary-Aware Distillation for Remote Sensing Scene Domain Incremental Classification

Abstract

The continuous growth of remote sensing imagery introduces substantial distribution discrepancies between previously observed and newly acquired domains, making remote sensing scene classification models susceptible to catastrophic forgetting during domain-incremental learning. To address this challenge, this article proposes a dynamic priority replay and boundary-aware distillation (DPR-BAD) framework. Specifically, a dynamic priority metric is designed by jointly exploiting cross-domain semantic shifts and prediction entropy to identify highly informative historical samples for replay, thereby improving knowledge retention under cross-domain feature distribution shifts. Furthermore, a BAD strategy is introduced to adaptively adjust the distillation strength according to each sample’s contribution to decision boundary stability, enabling stronger preservation of boundary-critical knowledge while alleviating forgetting of previously learned domains. Extensive experiments on four publicly available remote sensing scene classification datasets demonstrate that DPR-BAD consistently outperforms representative continual learning methods, achieving an average accuracy of 93.02% while effectively mitigating catastrophic forgetting without compromising adaptation to new domains.