Additive Content and Style Disentanglement for Domain Generalized Semantic Segmentation of Optical Remote Sensing Images

Supervised learning methods assume training and test data are independently and identically distributed, commonly used in land cover semantic segmentation. However, this assumption doesn't hold in real-world scenarios, as test data stems from unseen domains, leading to domain shifts that degrade model performance. Domain Generalization (DG) techniques address this by focusing on learning domain-invariant features. This paper investigates DG for land cover semantic segmentation in remote sensing optical images by progressively separating content and style features, emphasizing domain-invariant ones. Unlike state-of-the-art methods, it explores extracting only domain-invariant features. The approach is validated on the FLAIR dataset, achieving superior performance compared to existing methods.