Domain Adaptive Semantic Segmentation of Remote Sensing Images via Self-Training-Based Dual-Level Data Augmentation

Semantic segmentation models experience a significant performance degradation due to domain shifts between the source and target domains. This issue is particularly prevalent in remote sensing imagery, where a semantic segmentation model trained on images from one satellite is tested on images from another. Previous research has often overlooked the role of data augmentation in enhancing a model's adaptability to target domains. In contrast, this article proposes a novel self-training framework that incorporates data augmentation at both the input and feature levels, yielding excellent results. Specifically, we introduce a regularized online self-training framework that effectively addresses the challenges of overconfidence and class imbalance inherent in self-training. Based on this framework, we implement two robust data augmentation strategies at the input and feature levels to facilitate the learning of cross-domain invariant knowledge. At the input level, we employ a large-scale domain mixing strategy, termed multidomain mixing, to enhance the model's generalization capability. At the feature level, we introduce masked feature augmentation, a masking-based perturbation technique applied to the semantic features of the target domain. This approach enhances the consistency of teacher-student network predictions in the target domain feature space, thereby improving the robustness of the model's recognition of target domain features. The integration of the proposed self-training framework with dual-level data augmentation culminates in our innovative self-training-based dual-level data augmentation (STDA) method. Extensive experimental results on the ISPRS semantic segmentation benchmark demonstrate that STDA outperforms existing state-of-the-art methods, showcasing its effectiveness.