Learning From Vision Foundation Models for Cross-Domain Remote Sensing Image Segmentation

Cross-domain image segmentation plays a crucial role in the field of remote sensing. Current approaches often rely on a mean-teacher model that is integrated from student models to guide the training of the student model itself. However, the feature space of the mean-teacher model exhibits significant domain discrepancy and considerable class overlap, which results in suboptimal performance. Motivated by the idea of learning from stronger teachers, we introduce a robust domain adaptation method called LFMDA. This novel approach is the first to explicitly enhance cross-domain semantic segmentation performance by leveraging vision foundation models (VFMs) within remote sensing applications. Specifically, we propose a prototypical contrastive knowledge distillation loss (PCD) that enables the student model to produce domain-invariant yet category-discriminative features by distilling knowledge from a domain-generalized VFM teacher. Additionally, we introduce a local region homogenization strategy (LRH) to generate high-quality and high-quantity pseudo-labels by incorporating a Segment Anything Model (SAM). Extensive empirical evaluations demonstrate that our method outperforms existing approaches, setting a new state-of-the-art (SOTA) method in domain-adaptive remote sensing image segmentation. The code is available at https://github.com/StuLiu/LFMDA